nfs_bio.c revision 48225
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 * $Id: nfs_bio.c,v 1.73 1999/06/16 23:27:46 mckusick Exp $ 38 */ 39 40 41#include <sys/param.h> 42#include <sys/systm.h> 43#include <sys/resourcevar.h> 44#include <sys/signalvar.h> 45#include <sys/proc.h> 46#include <sys/buf.h> 47#include <sys/vnode.h> 48#include <sys/mount.h> 49#include <sys/kernel.h> 50 51#include <vm/vm.h> 52#include <vm/vm_extern.h> 53#include <vm/vm_prot.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 <nfs/rpcv2.h> 60#include <nfs/nfsproto.h> 61#include <nfs/nfs.h> 62#include <nfs/nfsmount.h> 63#include <nfs/nqnfs.h> 64#include <nfs/nfsnode.h> 65 66static struct buf *nfs_getcacheblk __P((struct vnode *vp, daddr_t bn, int size, 67 struct proc *p)); 68 69extern int nfs_numasync; 70extern int nfs_pbuf_freecnt; 71extern struct nfsstats nfsstats; 72 73/* 74 * Vnode op for VM getpages. 75 */ 76int 77nfs_getpages(ap) 78 struct vop_getpages_args /* { 79 struct vnode *a_vp; 80 vm_page_t *a_m; 81 int a_count; 82 int a_reqpage; 83 vm_ooffset_t a_offset; 84 } */ *ap; 85{ 86 int i, error, nextoff, size, toff, count, npages; 87 struct uio uio; 88 struct iovec iov; 89 vm_offset_t kva; 90 struct buf *bp; 91 struct vnode *vp; 92 struct proc *p; 93 struct ucred *cred; 94 struct nfsmount *nmp; 95 vm_page_t *pages; 96 97 vp = ap->a_vp; 98 p = curproc; /* XXX */ 99 cred = curproc->p_ucred; /* XXX */ 100 nmp = VFSTONFS(vp->v_mount); 101 pages = ap->a_m; 102 count = ap->a_count; 103 104 if (vp->v_object == NULL) { 105 printf("nfs_getpages: called with non-merged cache vnode??\n"); 106 return VM_PAGER_ERROR; 107 } 108 109 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 110 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 111 (void)nfs_fsinfo(nmp, vp, cred, p); 112 113 npages = btoc(count); 114 115 /* 116 * If the requested page is partially valid, just return it and 117 * allow the pager to zero-out the blanks. Partially valid pages 118 * can only occur at the file EOF. 119 */ 120 121 { 122 vm_page_t m = pages[ap->a_reqpage]; 123 124 if (m->valid != 0) { 125 /* handled by vm_fault now */ 126 /* vm_page_zero_invalid(m, TRUE); */ 127 for (i = 0; i < npages; ++i) { 128 if (i != ap->a_reqpage) 129 vnode_pager_freepage(pages[i]); 130 } 131 return(0); 132 } 133 } 134 135 /* 136 * We use only the kva address for the buffer, but this is extremely 137 * convienient and fast. 138 */ 139 bp = getpbuf(&nfs_pbuf_freecnt); 140 141 kva = (vm_offset_t) bp->b_data; 142 pmap_qenter(kva, pages, npages); 143 144 iov.iov_base = (caddr_t) kva; 145 iov.iov_len = count; 146 uio.uio_iov = &iov; 147 uio.uio_iovcnt = 1; 148 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 149 uio.uio_resid = count; 150 uio.uio_segflg = UIO_SYSSPACE; 151 uio.uio_rw = UIO_READ; 152 uio.uio_procp = p; 153 154 error = nfs_readrpc(vp, &uio, cred); 155 pmap_qremove(kva, npages); 156 157 relpbuf(bp, &nfs_pbuf_freecnt); 158 159 if (error && (uio.uio_resid == count)) { 160 printf("nfs_getpages: error %d\n", error); 161 for (i = 0; i < npages; ++i) { 162 if (i != ap->a_reqpage) 163 vnode_pager_freepage(pages[i]); 164 } 165 return VM_PAGER_ERROR; 166 } 167 168 /* 169 * Calculate the number of bytes read and validate only that number 170 * of bytes. Note that due to pending writes, size may be 0. This 171 * does not mean that the remaining data is invalid! 172 */ 173 174 size = count - uio.uio_resid; 175 176 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 177 vm_page_t m; 178 nextoff = toff + PAGE_SIZE; 179 m = pages[i]; 180 181 m->flags &= ~PG_ZERO; 182 183 if (nextoff <= size) { 184 /* 185 * Read operation filled an entire page 186 */ 187 m->valid = VM_PAGE_BITS_ALL; 188 m->dirty = 0; 189 } else if (size > toff) { 190 /* 191 * Read operation filled a partial page. 192 */ 193 m->valid = 0; 194 vm_page_set_validclean(m, 0, size - toff); 195 /* handled by vm_fault now */ 196 /* vm_page_zero_invalid(m, TRUE); */ 197 } 198 199 if (i != ap->a_reqpage) { 200 /* 201 * Whether or not to leave the page activated is up in 202 * the air, but we should put the page on a page queue 203 * somewhere (it already is in the object). Result: 204 * It appears that emperical results show that 205 * deactivating pages is best. 206 */ 207 208 /* 209 * Just in case someone was asking for this page we 210 * now tell them that it is ok to use. 211 */ 212 if (!error) { 213 if (m->flags & PG_WANTED) 214 vm_page_activate(m); 215 else 216 vm_page_deactivate(m); 217 vm_page_wakeup(m); 218 } else { 219 vnode_pager_freepage(m); 220 } 221 } 222 } 223 return 0; 224} 225 226/* 227 * Vnode op for VM putpages. 228 */ 229int 230nfs_putpages(ap) 231 struct vop_putpages_args /* { 232 struct vnode *a_vp; 233 vm_page_t *a_m; 234 int a_count; 235 int a_sync; 236 int *a_rtvals; 237 vm_ooffset_t a_offset; 238 } */ *ap; 239{ 240 struct uio uio; 241 struct iovec iov; 242 vm_offset_t kva; 243 struct buf *bp; 244 int iomode, must_commit, i, error, npages, count; 245 off_t offset; 246 int *rtvals; 247 struct vnode *vp; 248 struct proc *p; 249 struct ucred *cred; 250 struct nfsmount *nmp; 251 struct nfsnode *np; 252 vm_page_t *pages; 253 254 vp = ap->a_vp; 255 np = VTONFS(vp); 256 p = curproc; /* XXX */ 257 cred = curproc->p_ucred; /* XXX */ 258 nmp = VFSTONFS(vp->v_mount); 259 pages = ap->a_m; 260 count = ap->a_count; 261 rtvals = ap->a_rtvals; 262 npages = btoc(count); 263 offset = IDX_TO_OFF(pages[0]->pindex); 264 265 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 266 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 267 (void)nfs_fsinfo(nmp, vp, cred, p); 268 269 for (i = 0; i < npages; i++) { 270 rtvals[i] = VM_PAGER_AGAIN; 271 } 272 273 /* 274 * When putting pages, do not extend file past EOF. 275 */ 276 277 if (offset + count > np->n_size) { 278 count = np->n_size - offset; 279 if (count < 0) 280 count = 0; 281 } 282 283 /* 284 * We use only the kva address for the buffer, but this is extremely 285 * convienient and fast. 286 */ 287 bp = getpbuf(&nfs_pbuf_freecnt); 288 289 kva = (vm_offset_t) bp->b_data; 290 pmap_qenter(kva, pages, npages); 291 292 iov.iov_base = (caddr_t) kva; 293 iov.iov_len = count; 294 uio.uio_iov = &iov; 295 uio.uio_iovcnt = 1; 296 uio.uio_offset = offset; 297 uio.uio_resid = count; 298 uio.uio_segflg = UIO_SYSSPACE; 299 uio.uio_rw = UIO_WRITE; 300 uio.uio_procp = p; 301 302 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 303 iomode = NFSV3WRITE_UNSTABLE; 304 else 305 iomode = NFSV3WRITE_FILESYNC; 306 307 error = nfs_writerpc(vp, &uio, cred, &iomode, &must_commit); 308 309 pmap_qremove(kva, npages); 310 relpbuf(bp, &nfs_pbuf_freecnt); 311 312 if (!error) { 313 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 314 for (i = 0; i < nwritten; i++) { 315 rtvals[i] = VM_PAGER_OK; 316 pages[i]->dirty = 0; 317 } 318 if (must_commit) 319 nfs_clearcommit(vp->v_mount); 320 } 321 return rtvals[0]; 322} 323 324/* 325 * Vnode op for read using bio 326 */ 327int 328nfs_bioread(vp, uio, ioflag, cred) 329 register struct vnode *vp; 330 register struct uio *uio; 331 int ioflag; 332 struct ucred *cred; 333{ 334 register struct nfsnode *np = VTONFS(vp); 335 register int biosize, i; 336 struct buf *bp = 0, *rabp; 337 struct vattr vattr; 338 struct proc *p; 339 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 340 daddr_t lbn, rabn; 341 int bcount; 342 int nra, error = 0, n = 0, on = 0; 343 344#ifdef DIAGNOSTIC 345 if (uio->uio_rw != UIO_READ) 346 panic("nfs_read mode"); 347#endif 348 if (uio->uio_resid == 0) 349 return (0); 350 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 351 return (EINVAL); 352 p = uio->uio_procp; 353 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 354 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 355 (void)nfs_fsinfo(nmp, vp, cred, p); 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 /* 361 * For nfs, cache consistency can only be maintained approximately. 362 * Although RFC1094 does not specify the criteria, the following is 363 * believed to be compatible with the reference port. 364 * For nqnfs, full cache consistency is maintained within the loop. 365 * For nfs: 366 * If the file's modify time on the server has changed since the 367 * last read rpc or you have written to the file, 368 * you may have lost data cache consistency with the 369 * server, so flush all of the file's data out of the cache. 370 * Then force a getattr rpc to ensure that you have up to date 371 * attributes. 372 * NB: This implies that cache data can be read when up to 373 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 374 * attributes this could be forced by setting n_attrstamp to 0 before 375 * the VOP_GETATTR() call. 376 */ 377 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) { 378 if (np->n_flag & NMODIFIED) { 379 if (vp->v_type != VREG) { 380 if (vp->v_type != VDIR) 381 panic("nfs: bioread, not dir"); 382 nfs_invaldir(vp); 383 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 384 if (error) 385 return (error); 386 } 387 np->n_attrstamp = 0; 388 error = VOP_GETATTR(vp, &vattr, cred, p); 389 if (error) 390 return (error); 391 np->n_mtime = vattr.va_mtime.tv_sec; 392 } else { 393 error = VOP_GETATTR(vp, &vattr, cred, p); 394 if (error) 395 return (error); 396 if (np->n_mtime != vattr.va_mtime.tv_sec) { 397 if (vp->v_type == VDIR) 398 nfs_invaldir(vp); 399 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 400 if (error) 401 return (error); 402 np->n_mtime = vattr.va_mtime.tv_sec; 403 } 404 } 405 } 406 do { 407 408 /* 409 * Get a valid lease. If cached data is stale, flush it. 410 */ 411 if (nmp->nm_flag & NFSMNT_NQNFS) { 412 if (NQNFS_CKINVALID(vp, np, ND_READ)) { 413 do { 414 error = nqnfs_getlease(vp, ND_READ, cred, p); 415 } while (error == NQNFS_EXPIRED); 416 if (error) 417 return (error); 418 if (np->n_lrev != np->n_brev || 419 (np->n_flag & NQNFSNONCACHE) || 420 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) { 421 if (vp->v_type == VDIR) 422 nfs_invaldir(vp); 423 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 424 if (error) 425 return (error); 426 np->n_brev = np->n_lrev; 427 } 428 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) { 429 nfs_invaldir(vp); 430 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 431 if (error) 432 return (error); 433 } 434 } 435 if (np->n_flag & NQNFSNONCACHE) { 436 switch (vp->v_type) { 437 case VREG: 438 return (nfs_readrpc(vp, uio, cred)); 439 case VLNK: 440 return (nfs_readlinkrpc(vp, uio, cred)); 441 case VDIR: 442 break; 443 default: 444 printf(" NQNFSNONCACHE: type %x unexpected\n", 445 vp->v_type); 446 }; 447 } 448 switch (vp->v_type) { 449 case VREG: 450 nfsstats.biocache_reads++; 451 lbn = uio->uio_offset / biosize; 452 on = uio->uio_offset & (biosize - 1); 453 454 /* 455 * Start the read ahead(s), as required. 456 */ 457 if (nfs_numasync > 0 && nmp->nm_readahead > 0) { 458 for (nra = 0; nra < nmp->nm_readahead && 459 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 460 rabn = lbn + 1 + nra; 461 if (!incore(vp, rabn)) { 462 rabp = nfs_getcacheblk(vp, rabn, biosize, p); 463 if (!rabp) 464 return (EINTR); 465 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 466 rabp->b_flags |= (B_READ | B_ASYNC); 467 vfs_busy_pages(rabp, 0); 468 if (nfs_asyncio(rabp, cred, p)) { 469 rabp->b_flags |= B_INVAL|B_ERROR; 470 vfs_unbusy_pages(rabp); 471 brelse(rabp); 472 } 473 } else 474 brelse(rabp); 475 } 476 } 477 } 478 479 /* 480 * Obtain the buffer cache block. Figure out the buffer size 481 * when we are at EOF. nfs_getcacheblk() will also force 482 * uncached delayed-writes to be flushed to the server. 483 * 484 * Note that bcount is *not* DEV_BSIZE aligned. 485 */ 486 487 bcount = biosize; 488 if ((off_t)lbn * biosize >= np->n_size) { 489 bcount = 0; 490 } else if ((off_t)(lbn + 1) * biosize > np->n_size) { 491 bcount = np->n_size - (off_t)lbn * biosize; 492 } 493 494 bp = nfs_getcacheblk(vp, lbn, bcount, p); 495 if (!bp) 496 return (EINTR); 497 498 /* 499 * If B_CACHE is not set, we must issue the read. If this 500 * fails, we return an error. 501 */ 502 503 if ((bp->b_flags & B_CACHE) == 0) { 504 bp->b_flags |= B_READ; 505 vfs_busy_pages(bp, 0); 506 error = nfs_doio(bp, cred, p); 507 if (error) { 508 brelse(bp); 509 return (error); 510 } 511 } 512 513 /* 514 * on is the offset into the current bp. Figure out how many 515 * bytes we can copy out of the bp. Note that bcount is 516 * NOT DEV_BSIZE aligned. 517 * 518 * Then figure out how many bytes we can copy into the uio. 519 */ 520 521 n = 0; 522 if (on < bcount) 523 n = min((unsigned)(bcount - on), uio->uio_resid); 524 525 vp->v_lastr = lbn; 526 break; 527 case VLNK: 528 nfsstats.biocache_readlinks++; 529 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p); 530 if (!bp) 531 return (EINTR); 532 if ((bp->b_flags & B_CACHE) == 0) { 533 bp->b_flags |= B_READ; 534 vfs_busy_pages(bp, 0); 535 error = nfs_doio(bp, cred, p); 536 if (error) { 537 bp->b_flags |= B_ERROR; 538 brelse(bp); 539 return (error); 540 } 541 } 542 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 543 on = 0; 544 break; 545 case VDIR: 546 nfsstats.biocache_readdirs++; 547 if (np->n_direofoffset 548 && uio->uio_offset >= np->n_direofoffset) { 549 return (0); 550 } 551 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 552 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 553 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, p); 554 if (!bp) 555 return (EINTR); 556 if ((bp->b_flags & B_CACHE) == 0) { 557 bp->b_flags |= B_READ; 558 vfs_busy_pages(bp, 0); 559 error = nfs_doio(bp, cred, p); 560 if (error) { 561 brelse(bp); 562 } 563 while (error == NFSERR_BAD_COOKIE) { 564 printf("got bad cookie vp %p bp %p\n", vp, bp); 565 nfs_invaldir(vp); 566 error = nfs_vinvalbuf(vp, 0, cred, p, 1); 567 /* 568 * Yuck! The directory has been modified on the 569 * server. The only way to get the block is by 570 * reading from the beginning to get all the 571 * offset cookies. 572 * 573 * Leave the last bp intact unless there is an error. 574 * Loop back up to the while if the error is another 575 * NFSERR_BAD_COOKIE (double yuch!). 576 */ 577 for (i = 0; i <= lbn && !error; i++) { 578 if (np->n_direofoffset 579 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 580 return (0); 581 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, p); 582 if (!bp) 583 return (EINTR); 584 if ((bp->b_flags & B_CACHE) == 0) { 585 bp->b_flags |= B_READ; 586 vfs_busy_pages(bp, 0); 587 error = nfs_doio(bp, cred, p); 588 /* 589 * no error + B_INVAL == directory EOF, 590 * use the block. 591 */ 592 if (error == 0 && (bp->b_flags & B_INVAL)) 593 break; 594 } 595 /* 596 * An error will throw away the block and the 597 * for loop will break out. If no error and this 598 * is not the block we want, we throw away the 599 * block and go for the next one via the for loop. 600 */ 601 if (error || i < lbn) 602 brelse(bp); 603 } 604 } 605 /* 606 * The above while is repeated if we hit another cookie 607 * error. If we hit an error and it wasn't a cookie error, 608 * we give up. 609 */ 610 if (error) 611 return (error); 612 } 613 614 /* 615 * If not eof and read aheads are enabled, start one. 616 * (You need the current block first, so that you have the 617 * directory offset cookie of the next block.) 618 */ 619 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 620 (bp->b_flags & B_INVAL) == 0 && 621 (np->n_direofoffset == 0 || 622 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 623 !(np->n_flag & NQNFSNONCACHE) && 624 !incore(vp, lbn + 1)) { 625 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, p); 626 if (rabp) { 627 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 628 rabp->b_flags |= (B_READ | B_ASYNC); 629 vfs_busy_pages(rabp, 0); 630 if (nfs_asyncio(rabp, cred, p)) { 631 rabp->b_flags |= B_INVAL|B_ERROR; 632 vfs_unbusy_pages(rabp); 633 brelse(rabp); 634 } 635 } else { 636 brelse(rabp); 637 } 638 } 639 } 640 /* 641 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 642 * chopped for the EOF condition, we cannot tell how large 643 * NFS directories are going to be until we hit EOF. So 644 * an NFS directory buffer is *not* chopped to its EOF. Now, 645 * it just so happens that b_resid will effectively chop it 646 * to EOF. *BUT* this information is lost if the buffer goes 647 * away and is reconstituted into a B_CACHE state ( due to 648 * being VMIO ) later. So we keep track of the directory eof 649 * in np->n_direofoffset and chop it off as an extra step 650 * right here. 651 */ 652 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 653 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 654 n = np->n_direofoffset - uio->uio_offset; 655 break; 656 default: 657 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 658 break; 659 }; 660 661 if (n > 0) { 662 error = uiomove(bp->b_data + on, (int)n, uio); 663 } 664 switch (vp->v_type) { 665 case VREG: 666 break; 667 case VLNK: 668 n = 0; 669 break; 670 case VDIR: 671 /* 672 * Invalidate buffer if caching is disabled, forcing a 673 * re-read from the remote later. 674 */ 675 if (np->n_flag & NQNFSNONCACHE) 676 bp->b_flags |= B_INVAL; 677 break; 678 default: 679 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 680 } 681 brelse(bp); 682 } while (error == 0 && uio->uio_resid > 0 && n > 0); 683 return (error); 684} 685 686/* 687 * Vnode op for write using bio 688 */ 689int 690nfs_write(ap) 691 struct vop_write_args /* { 692 struct vnode *a_vp; 693 struct uio *a_uio; 694 int a_ioflag; 695 struct ucred *a_cred; 696 } */ *ap; 697{ 698 int biosize; 699 struct uio *uio = ap->a_uio; 700 struct proc *p = uio->uio_procp; 701 struct vnode *vp = ap->a_vp; 702 struct nfsnode *np = VTONFS(vp); 703 struct ucred *cred = ap->a_cred; 704 int ioflag = ap->a_ioflag; 705 struct buf *bp; 706 struct vattr vattr; 707 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 708 daddr_t lbn; 709 int bcount; 710 int n, on, error = 0, iomode, must_commit; 711 712#ifdef DIAGNOSTIC 713 if (uio->uio_rw != UIO_WRITE) 714 panic("nfs_write mode"); 715 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc) 716 panic("nfs_write proc"); 717#endif 718 if (vp->v_type != VREG) 719 return (EIO); 720 if (np->n_flag & NWRITEERR) { 721 np->n_flag &= ~NWRITEERR; 722 return (np->n_error); 723 } 724 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 725 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 726 (void)nfs_fsinfo(nmp, vp, cred, p); 727 if (ioflag & (IO_APPEND | IO_SYNC)) { 728 if (np->n_flag & NMODIFIED) { 729 np->n_attrstamp = 0; 730 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 731 if (error) 732 return (error); 733 } 734 if (ioflag & IO_APPEND) { 735 np->n_attrstamp = 0; 736 error = VOP_GETATTR(vp, &vattr, cred, p); 737 if (error) 738 return (error); 739 uio->uio_offset = np->n_size; 740 } 741 } 742 if (uio->uio_offset < 0) 743 return (EINVAL); 744 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 745 return (EFBIG); 746 if (uio->uio_resid == 0) 747 return (0); 748 /* 749 * Maybe this should be above the vnode op call, but so long as 750 * file servers have no limits, i don't think it matters 751 */ 752 if (p && uio->uio_offset + uio->uio_resid > 753 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 754 psignal(p, SIGXFSZ); 755 return (EFBIG); 756 } 757 758 biosize = vp->v_mount->mnt_stat.f_iosize; 759 760 do { 761 /* 762 * Check for a valid write lease. 763 */ 764 if ((nmp->nm_flag & NFSMNT_NQNFS) && 765 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 766 do { 767 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 768 } while (error == NQNFS_EXPIRED); 769 if (error) 770 return (error); 771 if (np->n_lrev != np->n_brev || 772 (np->n_flag & NQNFSNONCACHE)) { 773 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 774 if (error) 775 return (error); 776 np->n_brev = np->n_lrev; 777 } 778 } 779 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 780 iomode = NFSV3WRITE_FILESYNC; 781 error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit); 782 if (must_commit) 783 nfs_clearcommit(vp->v_mount); 784 return (error); 785 } 786 nfsstats.biocache_writes++; 787 lbn = uio->uio_offset / biosize; 788 on = uio->uio_offset & (biosize-1); 789 n = min((unsigned)(biosize - on), uio->uio_resid); 790again: 791 /* 792 * Handle direct append and file extension cases, calculate 793 * unaligned buffer size. 794 */ 795 796 if (uio->uio_offset == np->n_size && n) { 797 /* 798 * special append case. Obtain buffer prior to 799 * resizing it to maintain B_CACHE. 800 */ 801 long save; 802 803 bcount = on; 804 bp = nfs_getcacheblk(vp, lbn, bcount, p); 805 save = bp->b_flags & B_CACHE; 806 807 np->n_size = uio->uio_offset + n; 808 np->n_flag |= NMODIFIED; 809 vnode_pager_setsize(vp, np->n_size); 810 811 bcount += n; 812 allocbuf(bp, bcount); 813 bp->b_flags |= save; 814 } else { 815 if (uio->uio_offset + n > np->n_size) { 816 np->n_size = uio->uio_offset + n; 817 np->n_flag |= NMODIFIED; 818 vnode_pager_setsize(vp, np->n_size); 819 } 820 bcount = biosize; 821 if ((off_t)(lbn + 1) * biosize > np->n_size) 822 bcount = np->n_size - (off_t)lbn * biosize; 823 bp = nfs_getcacheblk(vp, lbn, bcount, p); 824 } 825 826 /* 827 * Issue a READ if B_CACHE is not set. In special-append 828 * mode, B_CACHE is based on the buffer prior to the write 829 * op and is typically set, avoiding the read. If a read 830 * is required in special append mode, the server will 831 * probably send us a short-read since we extended the file 832 * on our end, resulting in b_resid == 0 and, thusly, 833 * B_CACHE getting set. 834 * 835 * We can also avoid issuing the read if the write covers 836 * the entire buffer. We have to make sure the buffer state 837 * is reasonable in this case since we will not be initiating 838 * I/O. See the comments in kern/vfs_bio.c's getblk() for 839 * more information. 840 * 841 * B_CACHE may also be set due to the buffer being cached 842 * normally. 843 */ 844 845 if (on == 0 && n == bcount) { 846 bp->b_flags |= B_CACHE; 847 bp->b_flags &= ~(B_ERROR | B_INVAL); 848 } 849 850 if ((bp->b_flags & B_CACHE) == 0) { 851 bp->b_flags |= B_READ; 852 vfs_busy_pages(bp, 0); 853 error = nfs_doio(bp, cred, p); 854 if (error) { 855 brelse(bp); 856 return (error); 857 } 858 } 859 if (!bp) 860 return (EINTR); 861 if (bp->b_wcred == NOCRED) { 862 crhold(cred); 863 bp->b_wcred = cred; 864 } 865 np->n_flag |= NMODIFIED; 866 867 /* 868 * If dirtyend exceeds file size, chop it down. If this 869 * creates a reverse-indexed or degenerate situation with 870 * dirtyoff/end, 0 them. 871 */ 872 873 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 874 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 875 if (bp->b_dirtyoff >= bp->b_dirtyend) 876 bp->b_dirtyoff = bp->b_dirtyend = 0; 877 878 /* 879 * If the new write will leave a contiguous dirty 880 * area, just update the b_dirtyoff and b_dirtyend, 881 * otherwise force a write rpc of the old dirty area. 882 * 883 * While it is possible to merge discontiguous writes due to 884 * our having a B_CACHE buffer ( and thus valid read data 885 * for the hole), we don't because it could lead to 886 * significant cache coherency problems with multiple clients, 887 * especially if locking is implemented later on. 888 * 889 * as an optimization we could theoretically maintain 890 * a linked list of discontinuous areas, but we would still 891 * have to commit them separately so there isn't much 892 * advantage to it except perhaps a bit of asynchronization. 893 */ 894 895 if (bp->b_dirtyend > 0 && 896 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 897 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) 898 return (EINTR); 899 goto again; 900 } 901 902 /* 903 * Check for valid write lease and get one as required. 904 * In case getblk() and/or bwrite() delayed us. 905 */ 906 if ((nmp->nm_flag & NFSMNT_NQNFS) && 907 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 908 do { 909 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 910 } while (error == NQNFS_EXPIRED); 911 if (error) { 912 brelse(bp); 913 return (error); 914 } 915 if (np->n_lrev != np->n_brev || 916 (np->n_flag & NQNFSNONCACHE)) { 917 brelse(bp); 918 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 919 if (error) 920 return (error); 921 np->n_brev = np->n_lrev; 922 goto again; 923 } 924 } 925 926 error = uiomove((char *)bp->b_data + on, n, uio); 927 bp->b_flags &= ~B_NEEDCOMMIT; 928 if (error) { 929 bp->b_flags |= B_ERROR; 930 brelse(bp); 931 return (error); 932 } 933 934 /* 935 * Only update dirtyoff/dirtyend if not a degenerate 936 * condition. 937 */ 938 if (n) { 939 if (bp->b_dirtyend > 0) { 940 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 941 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 942 } else { 943 bp->b_dirtyoff = on; 944 bp->b_dirtyend = on + n; 945 } 946 vfs_bio_set_validclean(bp, on, n); 947 } 948 949 /* 950 * Since this block is being modified, it must be written 951 * again and not just committed. 952 */ 953 bp->b_flags &= ~B_NEEDCOMMIT; 954 955 /* 956 * If the lease is non-cachable or IO_SYNC do bwrite(). 957 * 958 * IO_INVAL appears to be unused. The idea appears to be 959 * to turn off caching in this case. Very odd. XXX 960 */ 961 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 962 if (ioflag & IO_INVAL) 963 bp->b_flags |= B_NOCACHE; 964 error = VOP_BWRITE(bp->b_vp, bp); 965 if (error) 966 return (error); 967 if (np->n_flag & NQNFSNONCACHE) { 968 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 969 if (error) 970 return (error); 971 } 972 } else if ((n + on) == biosize && 973 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 974 bp->b_flags |= B_ASYNC; 975 (void)nfs_writebp(bp, 0, 0); 976 } else { 977 bdwrite(bp); 978 } 979 } while (uio->uio_resid > 0 && n > 0); 980 return (0); 981} 982 983/* 984 * Get an nfs cache block. 985 * Allocate a new one if the block isn't currently in the cache 986 * and return the block marked busy. If the calling process is 987 * interrupted by a signal for an interruptible mount point, return 988 * NULL. 989 */ 990static struct buf * 991nfs_getcacheblk(vp, bn, size, p) 992 struct vnode *vp; 993 daddr_t bn; 994 int size; 995 struct proc *p; 996{ 997 register struct buf *bp; 998 struct mount *mp; 999 struct nfsmount *nmp; 1000 1001 mp = vp->v_mount; 1002 nmp = VFSTONFS(mp); 1003 1004 if (nmp->nm_flag & NFSMNT_INT) { 1005 bp = getblk(vp, bn, size, PCATCH, 0); 1006 while (bp == (struct buf *)0) { 1007 if (nfs_sigintr(nmp, (struct nfsreq *)0, p)) 1008 return ((struct buf *)0); 1009 bp = getblk(vp, bn, size, 0, 2 * hz); 1010 } 1011 } else { 1012 bp = getblk(vp, bn, size, 0, 0); 1013 } 1014 1015 if (vp->v_type == VREG) { 1016 int biosize; 1017 1018 biosize = mp->mnt_stat.f_iosize; 1019 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1020 } 1021 return (bp); 1022} 1023 1024/* 1025 * Flush and invalidate all dirty buffers. If another process is already 1026 * doing the flush, just wait for completion. 1027 */ 1028int 1029nfs_vinvalbuf(vp, flags, cred, p, intrflg) 1030 struct vnode *vp; 1031 int flags; 1032 struct ucred *cred; 1033 struct proc *p; 1034 int intrflg; 1035{ 1036 register struct nfsnode *np = VTONFS(vp); 1037 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1038 int error = 0, slpflag, slptimeo; 1039 1040 if (vp->v_flag & VXLOCK) { 1041 return (0); 1042 } 1043 1044 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1045 intrflg = 0; 1046 if (intrflg) { 1047 slpflag = PCATCH; 1048 slptimeo = 2 * hz; 1049 } else { 1050 slpflag = 0; 1051 slptimeo = 0; 1052 } 1053 /* 1054 * First wait for any other process doing a flush to complete. 1055 */ 1056 while (np->n_flag & NFLUSHINPROG) { 1057 np->n_flag |= NFLUSHWANT; 1058 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval", 1059 slptimeo); 1060 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) 1061 return (EINTR); 1062 } 1063 1064 /* 1065 * Now, flush as required. 1066 */ 1067 np->n_flag |= NFLUSHINPROG; 1068 error = vinvalbuf(vp, flags, cred, p, slpflag, 0); 1069 while (error) { 1070 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) { 1071 np->n_flag &= ~NFLUSHINPROG; 1072 if (np->n_flag & NFLUSHWANT) { 1073 np->n_flag &= ~NFLUSHWANT; 1074 wakeup((caddr_t)&np->n_flag); 1075 } 1076 return (EINTR); 1077 } 1078 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo); 1079 } 1080 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 1081 if (np->n_flag & NFLUSHWANT) { 1082 np->n_flag &= ~NFLUSHWANT; 1083 wakeup((caddr_t)&np->n_flag); 1084 } 1085 return (0); 1086} 1087 1088/* 1089 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1090 * This is mainly to avoid queueing async I/O requests when the nfsiods 1091 * are all hung on a dead server. 1092 * 1093 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp 1094 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1095 */ 1096int 1097nfs_asyncio(bp, cred, procp) 1098 register struct buf *bp; 1099 struct ucred *cred; 1100 struct proc *procp; 1101{ 1102 struct nfsmount *nmp; 1103 int i; 1104 int gotiod; 1105 int slpflag = 0; 1106 int slptimeo = 0; 1107 int error; 1108 1109 if (nfs_numasync == 0) 1110 return (EIO); 1111 1112 nmp = VFSTONFS(bp->b_vp->v_mount); 1113again: 1114 if (nmp->nm_flag & NFSMNT_INT) 1115 slpflag = PCATCH; 1116 gotiod = FALSE; 1117 1118 /* 1119 * Find a free iod to process this request. 1120 */ 1121 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 1122 if (nfs_iodwant[i]) { 1123 /* 1124 * Found one, so wake it up and tell it which 1125 * mount to process. 1126 */ 1127 NFS_DPF(ASYNCIO, 1128 ("nfs_asyncio: waking iod %d for mount %p\n", 1129 i, nmp)); 1130 nfs_iodwant[i] = (struct proc *)0; 1131 nfs_iodmount[i] = nmp; 1132 nmp->nm_bufqiods++; 1133 wakeup((caddr_t)&nfs_iodwant[i]); 1134 gotiod = TRUE; 1135 break; 1136 } 1137 1138 /* 1139 * If none are free, we may already have an iod working on this mount 1140 * point. If so, it will process our request. 1141 */ 1142 if (!gotiod) { 1143 if (nmp->nm_bufqiods > 0) { 1144 NFS_DPF(ASYNCIO, 1145 ("nfs_asyncio: %d iods are already processing mount %p\n", 1146 nmp->nm_bufqiods, nmp)); 1147 gotiod = TRUE; 1148 } 1149 } 1150 1151 /* 1152 * If we have an iod which can process the request, then queue 1153 * the buffer. 1154 */ 1155 if (gotiod) { 1156 /* 1157 * Ensure that the queue never grows too large. 1158 */ 1159 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1160 NFS_DPF(ASYNCIO, 1161 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1162 nmp->nm_bufqwant = TRUE; 1163 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO, 1164 "nfsaio", slptimeo); 1165 if (error) { 1166 if (nfs_sigintr(nmp, NULL, procp)) 1167 return (EINTR); 1168 if (slpflag == PCATCH) { 1169 slpflag = 0; 1170 slptimeo = 2 * hz; 1171 } 1172 } 1173 /* 1174 * We might have lost our iod while sleeping, 1175 * so check and loop if nescessary. 1176 */ 1177 if (nmp->nm_bufqiods == 0) { 1178 NFS_DPF(ASYNCIO, 1179 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1180 goto again; 1181 } 1182 } 1183 1184 if (bp->b_flags & B_READ) { 1185 if (bp->b_rcred == NOCRED && cred != NOCRED) { 1186 crhold(cred); 1187 bp->b_rcred = cred; 1188 } 1189 } else { 1190 bp->b_flags |= B_WRITEINPROG; 1191 if (bp->b_wcred == NOCRED && cred != NOCRED) { 1192 crhold(cred); 1193 bp->b_wcred = cred; 1194 } 1195 } 1196 1197 BUF_KERNPROC(bp); 1198 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1199 nmp->nm_bufqlen++; 1200 return (0); 1201 } 1202 1203 /* 1204 * All the iods are busy on other mounts, so return EIO to 1205 * force the caller to process the i/o synchronously. 1206 */ 1207 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1208 return (EIO); 1209} 1210 1211/* 1212 * Do an I/O operation to/from a cache block. This may be called 1213 * synchronously or from an nfsiod. 1214 */ 1215int 1216nfs_doio(bp, cr, p) 1217 struct buf *bp; 1218 struct ucred *cr; 1219 struct proc *p; 1220{ 1221 struct uio *uiop; 1222 struct vnode *vp; 1223 struct nfsnode *np; 1224 struct nfsmount *nmp; 1225 int error = 0, iomode, must_commit = 0; 1226 struct uio uio; 1227 struct iovec io; 1228 1229 vp = bp->b_vp; 1230 np = VTONFS(vp); 1231 nmp = VFSTONFS(vp->v_mount); 1232 uiop = &uio; 1233 uiop->uio_iov = &io; 1234 uiop->uio_iovcnt = 1; 1235 uiop->uio_segflg = UIO_SYSSPACE; 1236 uiop->uio_procp = p; 1237 1238 /* 1239 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 1240 * do this here so we do not have to do it in all the code that 1241 * calls us. 1242 */ 1243 bp->b_flags &= ~(B_ERROR | B_INVAL); 1244 1245 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1246 1247 /* 1248 * Historically, paging was done with physio, but no more. 1249 */ 1250 if (bp->b_flags & B_PHYS) { 1251 /* 1252 * ...though reading /dev/drum still gets us here. 1253 */ 1254 io.iov_len = uiop->uio_resid = bp->b_bcount; 1255 /* mapping was done by vmapbuf() */ 1256 io.iov_base = bp->b_data; 1257 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1258 if (bp->b_flags & B_READ) { 1259 uiop->uio_rw = UIO_READ; 1260 nfsstats.read_physios++; 1261 error = nfs_readrpc(vp, uiop, cr); 1262 } else { 1263 int com; 1264 1265 iomode = NFSV3WRITE_DATASYNC; 1266 uiop->uio_rw = UIO_WRITE; 1267 nfsstats.write_physios++; 1268 error = nfs_writerpc(vp, uiop, cr, &iomode, &com); 1269 } 1270 if (error) { 1271 bp->b_flags |= B_ERROR; 1272 bp->b_error = error; 1273 } 1274 } else if (bp->b_flags & B_READ) { 1275 io.iov_len = uiop->uio_resid = bp->b_bcount; 1276 io.iov_base = bp->b_data; 1277 uiop->uio_rw = UIO_READ; 1278 switch (vp->v_type) { 1279 case VREG: 1280 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1281 nfsstats.read_bios++; 1282 error = nfs_readrpc(vp, uiop, cr); 1283 if (!error) { 1284 if (uiop->uio_resid) { 1285 /* 1286 * If we had a short read with no error, we must have 1287 * hit a file hole. We should zero-fill the remainder. 1288 * This can also occur if the server hits the file EOF. 1289 * 1290 * Holes used to be able to occur due to pending 1291 * writes, but that is not possible any longer. 1292 */ 1293 int nread = bp->b_bcount - uiop->uio_resid; 1294 int left = bp->b_bcount - nread; 1295 1296 if (left > 0) 1297 bzero((char *)bp->b_data + nread, left); 1298 uiop->uio_resid = 0; 1299 } 1300 } 1301 if (p && (vp->v_flag & VTEXT) && 1302 (((nmp->nm_flag & NFSMNT_NQNFS) && 1303 NQNFS_CKINVALID(vp, np, ND_READ) && 1304 np->n_lrev != np->n_brev) || 1305 (!(nmp->nm_flag & NFSMNT_NQNFS) && 1306 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) { 1307 uprintf("Process killed due to text file modification\n"); 1308 psignal(p, SIGKILL); 1309 PHOLD(p); 1310 } 1311 break; 1312 case VLNK: 1313 uiop->uio_offset = (off_t)0; 1314 nfsstats.readlink_bios++; 1315 error = nfs_readlinkrpc(vp, uiop, cr); 1316 break; 1317 case VDIR: 1318 nfsstats.readdir_bios++; 1319 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1320 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1321 error = nfs_readdirplusrpc(vp, uiop, cr); 1322 if (error == NFSERR_NOTSUPP) 1323 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1324 } 1325 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1326 error = nfs_readdirrpc(vp, uiop, cr); 1327 /* 1328 * end-of-directory sets B_INVAL but does not generate an 1329 * error. 1330 */ 1331 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1332 bp->b_flags |= B_INVAL; 1333 break; 1334 default: 1335 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1336 break; 1337 }; 1338 if (error) { 1339 bp->b_flags |= B_ERROR; 1340 bp->b_error = error; 1341 } 1342 } else { 1343 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1344 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1345 1346 if (bp->b_dirtyend > bp->b_dirtyoff) { 1347 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1348 - bp->b_dirtyoff; 1349 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1350 + bp->b_dirtyoff; 1351 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1352 uiop->uio_rw = UIO_WRITE; 1353 nfsstats.write_bios++; 1354 1355 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1356 iomode = NFSV3WRITE_UNSTABLE; 1357 else 1358 iomode = NFSV3WRITE_FILESYNC; 1359 1360 bp->b_flags |= B_WRITEINPROG; 1361 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit); 1362 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1363 bp->b_flags |= B_NEEDCOMMIT; 1364 if (bp->b_dirtyoff == 0 1365 && bp->b_dirtyend == bp->b_bcount) 1366 bp->b_flags |= B_CLUSTEROK; 1367 } else { 1368 bp->b_flags &= ~B_NEEDCOMMIT; 1369 } 1370 bp->b_flags &= ~B_WRITEINPROG; 1371 1372 /* 1373 * For an interrupted write, the buffer is still valid 1374 * and the write hasn't been pushed to the server yet, 1375 * so we can't set B_ERROR and report the interruption 1376 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1377 * is not relevant, so the rpc attempt is essentially 1378 * a noop. For the case of a V3 write rpc not being 1379 * committed to stable storage, the block is still 1380 * dirty and requires either a commit rpc or another 1381 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1382 * the block is reused. This is indicated by setting 1383 * the B_DELWRI and B_NEEDCOMMIT flags. 1384 * 1385 * If the buffer is marked B_PAGING, it does not reside on 1386 * the vp's paging queues so we cannot call bdirty(). The 1387 * bp in this case is not an NFS cache block so we should 1388 * be safe. XXX 1389 */ 1390 if (error == EINTR 1391 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1392 int s; 1393 1394 s = splbio(); 1395 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1396 if ((bp->b_flags & B_PAGING) == 0) { 1397 bdirty(bp); 1398 bp->b_flags &= ~B_DONE; 1399 } 1400 if (error && (bp->b_flags & B_ASYNC) == 0) 1401 bp->b_flags |= B_EINTR; 1402 splx(s); 1403 } else { 1404 if (error) { 1405 bp->b_flags |= B_ERROR; 1406 bp->b_error = np->n_error = error; 1407 np->n_flag |= NWRITEERR; 1408 } 1409 bp->b_dirtyoff = bp->b_dirtyend = 0; 1410 } 1411 } else { 1412 bp->b_resid = 0; 1413 biodone(bp); 1414 return (0); 1415 } 1416 } 1417 bp->b_resid = uiop->uio_resid; 1418 if (must_commit) 1419 nfs_clearcommit(vp->v_mount); 1420 biodone(bp); 1421 return (error); 1422} 1423