nfs_bio.c revision 51344
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: head/sys/nfsclient/nfs_bio.c 51344 1999-09-17 05:57:57Z dillon $ 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 vm_page_undirty(m); 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 vm_page_undirty(pages[i]); 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 seqcount; 343 int nra, error = 0, n = 0, on = 0; 344 345#ifdef DIAGNOSTIC 346 if (uio->uio_rw != UIO_READ) 347 panic("nfs_read mode"); 348#endif 349 if (uio->uio_resid == 0) 350 return (0); 351 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 352 return (EINVAL); 353 p = uio->uio_procp; 354 355 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 356 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 357 (void)nfs_fsinfo(nmp, vp, cred, p); 358 if (vp->v_type != VDIR && 359 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 360 return (EFBIG); 361 biosize = vp->v_mount->mnt_stat.f_iosize; 362 seqcount = (int)((off_t)(ioflag >> 16) * biosize / BKVASIZE); 363 /* 364 * For nfs, cache consistency can only be maintained approximately. 365 * Although RFC1094 does not specify the criteria, the following is 366 * believed to be compatible with the reference port. 367 * For nqnfs, full cache consistency is maintained within the loop. 368 * For nfs: 369 * If the file's modify time on the server has changed since the 370 * last read rpc or you have written to the file, 371 * you may have lost data cache consistency with the 372 * server, so flush all of the file's data out of the cache. 373 * Then force a getattr rpc to ensure that you have up to date 374 * attributes. 375 * NB: This implies that cache data can be read when up to 376 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 377 * attributes this could be forced by setting n_attrstamp to 0 before 378 * the VOP_GETATTR() call. 379 */ 380 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) { 381 if (np->n_flag & NMODIFIED) { 382 if (vp->v_type != VREG) { 383 if (vp->v_type != VDIR) 384 panic("nfs: bioread, not dir"); 385 nfs_invaldir(vp); 386 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 387 if (error) 388 return (error); 389 } 390 np->n_attrstamp = 0; 391 error = VOP_GETATTR(vp, &vattr, cred, p); 392 if (error) 393 return (error); 394 np->n_mtime = vattr.va_mtime.tv_sec; 395 } else { 396 error = VOP_GETATTR(vp, &vattr, cred, p); 397 if (error) 398 return (error); 399 if (np->n_mtime != vattr.va_mtime.tv_sec) { 400 if (vp->v_type == VDIR) 401 nfs_invaldir(vp); 402 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 403 if (error) 404 return (error); 405 np->n_mtime = vattr.va_mtime.tv_sec; 406 } 407 } 408 } 409 do { 410 411 /* 412 * Get a valid lease. If cached data is stale, flush it. 413 */ 414 if (nmp->nm_flag & NFSMNT_NQNFS) { 415 if (NQNFS_CKINVALID(vp, np, ND_READ)) { 416 do { 417 error = nqnfs_getlease(vp, ND_READ, cred, p); 418 } while (error == NQNFS_EXPIRED); 419 if (error) 420 return (error); 421 if (np->n_lrev != np->n_brev || 422 (np->n_flag & NQNFSNONCACHE) || 423 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) { 424 if (vp->v_type == VDIR) 425 nfs_invaldir(vp); 426 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 427 if (error) 428 return (error); 429 np->n_brev = np->n_lrev; 430 } 431 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) { 432 nfs_invaldir(vp); 433 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 434 if (error) 435 return (error); 436 } 437 } 438 if (np->n_flag & NQNFSNONCACHE) { 439 switch (vp->v_type) { 440 case VREG: 441 return (nfs_readrpc(vp, uio, cred)); 442 case VLNK: 443 return (nfs_readlinkrpc(vp, uio, cred)); 444 case VDIR: 445 break; 446 default: 447 printf(" NQNFSNONCACHE: type %x unexpected\n", 448 vp->v_type); 449 }; 450 } 451 switch (vp->v_type) { 452 case VREG: 453 nfsstats.biocache_reads++; 454 lbn = uio->uio_offset / biosize; 455 on = uio->uio_offset & (biosize - 1); 456 457 /* 458 * Start the read ahead(s), as required. 459 */ 460 if (nfs_numasync > 0 && nmp->nm_readahead > 0) { 461 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 462 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 463 rabn = lbn + 1 + nra; 464 if (!incore(vp, rabn)) { 465 rabp = nfs_getcacheblk(vp, rabn, biosize, p); 466 if (!rabp) 467 return (EINTR); 468 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 469 rabp->b_flags |= (B_READ | B_ASYNC); 470 vfs_busy_pages(rabp, 0); 471 if (nfs_asyncio(rabp, cred, p)) { 472 rabp->b_flags |= B_INVAL|B_ERROR; 473 vfs_unbusy_pages(rabp); 474 brelse(rabp); 475 } 476 } else 477 brelse(rabp); 478 } 479 } 480 } 481 482 /* 483 * Obtain the buffer cache block. Figure out the buffer size 484 * when we are at EOF. nfs_getcacheblk() will also force 485 * uncached delayed-writes to be flushed to the server. 486 * 487 * Note that bcount is *not* DEV_BSIZE aligned. 488 */ 489 490 bcount = biosize; 491 if ((off_t)lbn * biosize >= np->n_size) { 492 bcount = 0; 493 } else if ((off_t)(lbn + 1) * biosize > np->n_size) { 494 bcount = np->n_size - (off_t)lbn * biosize; 495 } 496 497 bp = nfs_getcacheblk(vp, lbn, bcount, p); 498 if (!bp) 499 return (EINTR); 500 501 /* 502 * If B_CACHE is not set, we must issue the read. If this 503 * fails, we return an error. 504 */ 505 506 if ((bp->b_flags & B_CACHE) == 0) { 507 bp->b_flags |= B_READ; 508 vfs_busy_pages(bp, 0); 509 error = nfs_doio(bp, cred, p); 510 if (error) { 511 brelse(bp); 512 return (error); 513 } 514 } 515 516 /* 517 * on is the offset into the current bp. Figure out how many 518 * bytes we can copy out of the bp. Note that bcount is 519 * NOT DEV_BSIZE aligned. 520 * 521 * Then figure out how many bytes we can copy into the uio. 522 */ 523 524 n = 0; 525 if (on < bcount) 526 n = min((unsigned)(bcount - on), uio->uio_resid); 527 break; 528 case VLNK: 529 nfsstats.biocache_readlinks++; 530 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p); 531 if (!bp) 532 return (EINTR); 533 if ((bp->b_flags & B_CACHE) == 0) { 534 bp->b_flags |= B_READ; 535 vfs_busy_pages(bp, 0); 536 error = nfs_doio(bp, cred, p); 537 if (error) { 538 bp->b_flags |= B_ERROR; 539 brelse(bp); 540 return (error); 541 } 542 } 543 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 544 on = 0; 545 break; 546 case VDIR: 547 nfsstats.biocache_readdirs++; 548 if (np->n_direofoffset 549 && uio->uio_offset >= np->n_direofoffset) { 550 return (0); 551 } 552 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 553 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 554 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, p); 555 if (!bp) 556 return (EINTR); 557 if ((bp->b_flags & B_CACHE) == 0) { 558 bp->b_flags |= B_READ; 559 vfs_busy_pages(bp, 0); 560 error = nfs_doio(bp, cred, p); 561 if (error) { 562 brelse(bp); 563 } 564 while (error == NFSERR_BAD_COOKIE) { 565 printf("got bad cookie vp %p bp %p\n", vp, bp); 566 nfs_invaldir(vp); 567 error = nfs_vinvalbuf(vp, 0, cred, p, 1); 568 /* 569 * Yuck! The directory has been modified on the 570 * server. The only way to get the block is by 571 * reading from the beginning to get all the 572 * offset cookies. 573 * 574 * Leave the last bp intact unless there is an error. 575 * Loop back up to the while if the error is another 576 * NFSERR_BAD_COOKIE (double yuch!). 577 */ 578 for (i = 0; i <= lbn && !error; i++) { 579 if (np->n_direofoffset 580 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 581 return (0); 582 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, p); 583 if (!bp) 584 return (EINTR); 585 if ((bp->b_flags & B_CACHE) == 0) { 586 bp->b_flags |= B_READ; 587 vfs_busy_pages(bp, 0); 588 error = nfs_doio(bp, cred, p); 589 /* 590 * no error + B_INVAL == directory EOF, 591 * use the block. 592 */ 593 if (error == 0 && (bp->b_flags & B_INVAL)) 594 break; 595 } 596 /* 597 * An error will throw away the block and the 598 * for loop will break out. If no error and this 599 * is not the block we want, we throw away the 600 * block and go for the next one via the for loop. 601 */ 602 if (error || i < lbn) 603 brelse(bp); 604 } 605 } 606 /* 607 * The above while is repeated if we hit another cookie 608 * error. If we hit an error and it wasn't a cookie error, 609 * we give up. 610 */ 611 if (error) 612 return (error); 613 } 614 615 /* 616 * If not eof and read aheads are enabled, start one. 617 * (You need the current block first, so that you have the 618 * directory offset cookie of the next block.) 619 */ 620 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 621 (bp->b_flags & B_INVAL) == 0 && 622 (np->n_direofoffset == 0 || 623 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 624 !(np->n_flag & NQNFSNONCACHE) && 625 !incore(vp, lbn + 1)) { 626 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, p); 627 if (rabp) { 628 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 629 rabp->b_flags |= (B_READ | B_ASYNC); 630 vfs_busy_pages(rabp, 0); 631 if (nfs_asyncio(rabp, cred, p)) { 632 rabp->b_flags |= B_INVAL|B_ERROR; 633 vfs_unbusy_pages(rabp); 634 brelse(rabp); 635 } 636 } else { 637 brelse(rabp); 638 } 639 } 640 } 641 /* 642 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 643 * chopped for the EOF condition, we cannot tell how large 644 * NFS directories are going to be until we hit EOF. So 645 * an NFS directory buffer is *not* chopped to its EOF. Now, 646 * it just so happens that b_resid will effectively chop it 647 * to EOF. *BUT* this information is lost if the buffer goes 648 * away and is reconstituted into a B_CACHE state ( due to 649 * being VMIO ) later. So we keep track of the directory eof 650 * in np->n_direofoffset and chop it off as an extra step 651 * right here. 652 */ 653 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 654 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 655 n = np->n_direofoffset - uio->uio_offset; 656 break; 657 default: 658 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 659 break; 660 }; 661 662 if (n > 0) { 663 error = uiomove(bp->b_data + on, (int)n, uio); 664 } 665 switch (vp->v_type) { 666 case VREG: 667 break; 668 case VLNK: 669 n = 0; 670 break; 671 case VDIR: 672 /* 673 * Invalidate buffer if caching is disabled, forcing a 674 * re-read from the remote later. 675 */ 676 if (np->n_flag & NQNFSNONCACHE) 677 bp->b_flags |= B_INVAL; 678 break; 679 default: 680 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 681 } 682 brelse(bp); 683 } while (error == 0 && uio->uio_resid > 0 && n > 0); 684 return (error); 685} 686 687/* 688 * Vnode op for write using bio 689 */ 690int 691nfs_write(ap) 692 struct vop_write_args /* { 693 struct vnode *a_vp; 694 struct uio *a_uio; 695 int a_ioflag; 696 struct ucred *a_cred; 697 } */ *ap; 698{ 699 int biosize; 700 struct uio *uio = ap->a_uio; 701 struct proc *p = uio->uio_procp; 702 struct vnode *vp = ap->a_vp; 703 struct nfsnode *np = VTONFS(vp); 704 struct ucred *cred = ap->a_cred; 705 int ioflag = ap->a_ioflag; 706 struct buf *bp; 707 struct vattr vattr; 708 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 709 daddr_t lbn; 710 int bcount; 711 int n, on, error = 0, iomode, must_commit; 712 713#ifdef DIAGNOSTIC 714 if (uio->uio_rw != UIO_WRITE) 715 panic("nfs_write mode"); 716 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc) 717 panic("nfs_write proc"); 718#endif 719 if (vp->v_type != VREG) 720 return (EIO); 721 if (np->n_flag & NWRITEERR) { 722 np->n_flag &= ~NWRITEERR; 723 return (np->n_error); 724 } 725 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 726 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 727 (void)nfs_fsinfo(nmp, vp, cred, p); 728 if (ioflag & (IO_APPEND | IO_SYNC)) { 729 if (np->n_flag & NMODIFIED) { 730 np->n_attrstamp = 0; 731 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 732 if (error) 733 return (error); 734 } 735 if (ioflag & IO_APPEND) { 736 np->n_attrstamp = 0; 737 error = VOP_GETATTR(vp, &vattr, cred, p); 738 if (error) 739 return (error); 740 uio->uio_offset = np->n_size; 741 } 742 } 743 if (uio->uio_offset < 0) 744 return (EINVAL); 745 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 746 return (EFBIG); 747 if (uio->uio_resid == 0) 748 return (0); 749 /* 750 * Maybe this should be above the vnode op call, but so long as 751 * file servers have no limits, i don't think it matters 752 */ 753 if (p && uio->uio_offset + uio->uio_resid > 754 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 755 psignal(p, SIGXFSZ); 756 return (EFBIG); 757 } 758 759 biosize = vp->v_mount->mnt_stat.f_iosize; 760 761 do { 762 /* 763 * Check for a valid write lease. 764 */ 765 if ((nmp->nm_flag & NFSMNT_NQNFS) && 766 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 767 do { 768 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 769 } while (error == NQNFS_EXPIRED); 770 if (error) 771 return (error); 772 if (np->n_lrev != np->n_brev || 773 (np->n_flag & NQNFSNONCACHE)) { 774 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 775 if (error) 776 return (error); 777 np->n_brev = np->n_lrev; 778 } 779 } 780 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 781 iomode = NFSV3WRITE_FILESYNC; 782 error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit); 783 if (must_commit) 784 nfs_clearcommit(vp->v_mount); 785 return (error); 786 } 787 nfsstats.biocache_writes++; 788 lbn = uio->uio_offset / biosize; 789 on = uio->uio_offset & (biosize-1); 790 n = min((unsigned)(biosize - on), uio->uio_resid); 791again: 792 /* 793 * Handle direct append and file extension cases, calculate 794 * unaligned buffer size. 795 */ 796 797 if (uio->uio_offset == np->n_size && n) { 798 /* 799 * special append case. Obtain buffer prior to 800 * resizing it to maintain B_CACHE. 801 */ 802 long save; 803 804 bcount = on; 805 bp = nfs_getcacheblk(vp, lbn, bcount, p); 806 if (!bp) 807 return (EINTR); 808 save = bp->b_flags & B_CACHE; 809 810 np->n_size = uio->uio_offset + n; 811 np->n_flag |= NMODIFIED; 812 vnode_pager_setsize(vp, np->n_size); 813 814 bcount += n; 815 allocbuf(bp, bcount); 816 bp->b_flags |= save; 817 } else { 818 if (uio->uio_offset + n > np->n_size) { 819 np->n_size = uio->uio_offset + n; 820 np->n_flag |= NMODIFIED; 821 vnode_pager_setsize(vp, np->n_size); 822 } 823 bcount = biosize; 824 if ((off_t)(lbn + 1) * biosize > np->n_size) 825 bcount = np->n_size - (off_t)lbn * biosize; 826 bp = nfs_getcacheblk(vp, lbn, bcount, p); 827 if (!bp) 828 return (EINTR); 829 } 830 831 /* 832 * Issue a READ if B_CACHE is not set. In special-append 833 * mode, B_CACHE is based on the buffer prior to the write 834 * op and is typically set, avoiding the read. If a read 835 * is required in special append mode, the server will 836 * probably send us a short-read since we extended the file 837 * on our end, resulting in b_resid == 0 and, thusly, 838 * B_CACHE getting set. 839 * 840 * We can also avoid issuing the read if the write covers 841 * the entire buffer. We have to make sure the buffer state 842 * is reasonable in this case since we will not be initiating 843 * I/O. See the comments in kern/vfs_bio.c's getblk() for 844 * more information. 845 * 846 * B_CACHE may also be set due to the buffer being cached 847 * normally. 848 */ 849 850 if (on == 0 && n == bcount) { 851 bp->b_flags |= B_CACHE; 852 bp->b_flags &= ~(B_ERROR | B_INVAL); 853 } 854 855 if ((bp->b_flags & B_CACHE) == 0) { 856 bp->b_flags |= B_READ; 857 vfs_busy_pages(bp, 0); 858 error = nfs_doio(bp, cred, p); 859 if (error) { 860 brelse(bp); 861 return (error); 862 } 863 } 864 if (!bp) 865 return (EINTR); 866 if (bp->b_wcred == NOCRED) { 867 crhold(cred); 868 bp->b_wcred = cred; 869 } 870 np->n_flag |= NMODIFIED; 871 872 /* 873 * If dirtyend exceeds file size, chop it down. If this 874 * creates a reverse-indexed or degenerate situation with 875 * dirtyoff/end, 0 them. 876 */ 877 878 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 879 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 880 if (bp->b_dirtyoff >= bp->b_dirtyend) 881 bp->b_dirtyoff = bp->b_dirtyend = 0; 882 883 /* 884 * If the new write will leave a contiguous dirty 885 * area, just update the b_dirtyoff and b_dirtyend, 886 * otherwise force a write rpc of the old dirty area. 887 * 888 * While it is possible to merge discontiguous writes due to 889 * our having a B_CACHE buffer ( and thus valid read data 890 * for the hole), we don't because it could lead to 891 * significant cache coherency problems with multiple clients, 892 * especially if locking is implemented later on. 893 * 894 * as an optimization we could theoretically maintain 895 * a linked list of discontinuous areas, but we would still 896 * have to commit them separately so there isn't much 897 * advantage to it except perhaps a bit of asynchronization. 898 */ 899 900 if (bp->b_dirtyend > 0 && 901 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 902 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) 903 return (EINTR); 904 goto again; 905 } 906 907 /* 908 * Check for valid write lease and get one as required. 909 * In case getblk() and/or bwrite() delayed us. 910 */ 911 if ((nmp->nm_flag & NFSMNT_NQNFS) && 912 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 913 do { 914 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 915 } while (error == NQNFS_EXPIRED); 916 if (error) { 917 brelse(bp); 918 return (error); 919 } 920 if (np->n_lrev != np->n_brev || 921 (np->n_flag & NQNFSNONCACHE)) { 922 brelse(bp); 923 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 924 if (error) 925 return (error); 926 np->n_brev = np->n_lrev; 927 goto again; 928 } 929 } 930 931 error = uiomove((char *)bp->b_data + on, n, uio); 932 bp->b_flags &= ~B_NEEDCOMMIT; 933 if (error) { 934 bp->b_flags |= B_ERROR; 935 brelse(bp); 936 return (error); 937 } 938 939 /* 940 * Only update dirtyoff/dirtyend if not a degenerate 941 * condition. 942 */ 943 if (n) { 944 if (bp->b_dirtyend > 0) { 945 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 946 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 947 } else { 948 bp->b_dirtyoff = on; 949 bp->b_dirtyend = on + n; 950 } 951 vfs_bio_set_validclean(bp, on, n); 952 } 953 954 /* 955 * Since this block is being modified, it must be written 956 * again and not just committed. 957 */ 958 bp->b_flags &= ~B_NEEDCOMMIT; 959 960 /* 961 * If the lease is non-cachable or IO_SYNC do bwrite(). 962 * 963 * IO_INVAL appears to be unused. The idea appears to be 964 * to turn off caching in this case. Very odd. XXX 965 */ 966 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 967 if (ioflag & IO_INVAL) 968 bp->b_flags |= B_NOCACHE; 969 error = VOP_BWRITE(bp->b_vp, bp); 970 if (error) 971 return (error); 972 if (np->n_flag & NQNFSNONCACHE) { 973 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 974 if (error) 975 return (error); 976 } 977 } else if ((n + on) == biosize && 978 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 979 bp->b_flags |= B_ASYNC; 980 (void)nfs_writebp(bp, 0, 0); 981 } else { 982 bdwrite(bp); 983 } 984 } while (uio->uio_resid > 0 && n > 0); 985 return (0); 986} 987 988/* 989 * Get an nfs cache block. 990 * Allocate a new one if the block isn't currently in the cache 991 * and return the block marked busy. If the calling process is 992 * interrupted by a signal for an interruptible mount point, return 993 * NULL. 994 */ 995static struct buf * 996nfs_getcacheblk(vp, bn, size, p) 997 struct vnode *vp; 998 daddr_t bn; 999 int size; 1000 struct proc *p; 1001{ 1002 register struct buf *bp; 1003 struct mount *mp; 1004 struct nfsmount *nmp; 1005 1006 mp = vp->v_mount; 1007 nmp = VFSTONFS(mp); 1008 1009 if (nmp->nm_flag & NFSMNT_INT) { 1010 bp = getblk(vp, bn, size, PCATCH, 0); 1011 while (bp == (struct buf *)0) { 1012 if (nfs_sigintr(nmp, (struct nfsreq *)0, p)) 1013 return ((struct buf *)0); 1014 bp = getblk(vp, bn, size, 0, 2 * hz); 1015 } 1016 } else { 1017 bp = getblk(vp, bn, size, 0, 0); 1018 } 1019 1020 if (vp->v_type == VREG) { 1021 int biosize; 1022 1023 biosize = mp->mnt_stat.f_iosize; 1024 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1025 } 1026 return (bp); 1027} 1028 1029/* 1030 * Flush and invalidate all dirty buffers. If another process is already 1031 * doing the flush, just wait for completion. 1032 */ 1033int 1034nfs_vinvalbuf(vp, flags, cred, p, intrflg) 1035 struct vnode *vp; 1036 int flags; 1037 struct ucred *cred; 1038 struct proc *p; 1039 int intrflg; 1040{ 1041 register struct nfsnode *np = VTONFS(vp); 1042 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1043 int error = 0, slpflag, slptimeo; 1044 1045 if (vp->v_flag & VXLOCK) { 1046 return (0); 1047 } 1048 1049 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1050 intrflg = 0; 1051 if (intrflg) { 1052 slpflag = PCATCH; 1053 slptimeo = 2 * hz; 1054 } else { 1055 slpflag = 0; 1056 slptimeo = 0; 1057 } 1058 /* 1059 * First wait for any other process doing a flush to complete. 1060 */ 1061 while (np->n_flag & NFLUSHINPROG) { 1062 np->n_flag |= NFLUSHWANT; 1063 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval", 1064 slptimeo); 1065 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) 1066 return (EINTR); 1067 } 1068 1069 /* 1070 * Now, flush as required. 1071 */ 1072 np->n_flag |= NFLUSHINPROG; 1073 error = vinvalbuf(vp, flags, cred, p, slpflag, 0); 1074 while (error) { 1075 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) { 1076 np->n_flag &= ~NFLUSHINPROG; 1077 if (np->n_flag & NFLUSHWANT) { 1078 np->n_flag &= ~NFLUSHWANT; 1079 wakeup((caddr_t)&np->n_flag); 1080 } 1081 return (EINTR); 1082 } 1083 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo); 1084 } 1085 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 1086 if (np->n_flag & NFLUSHWANT) { 1087 np->n_flag &= ~NFLUSHWANT; 1088 wakeup((caddr_t)&np->n_flag); 1089 } 1090 return (0); 1091} 1092 1093/* 1094 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1095 * This is mainly to avoid queueing async I/O requests when the nfsiods 1096 * are all hung on a dead server. 1097 * 1098 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp 1099 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1100 */ 1101int 1102nfs_asyncio(bp, cred, procp) 1103 register struct buf *bp; 1104 struct ucred *cred; 1105 struct proc *procp; 1106{ 1107 struct nfsmount *nmp; 1108 int i; 1109 int gotiod; 1110 int slpflag = 0; 1111 int slptimeo = 0; 1112 int error; 1113 1114 if (nfs_numasync == 0) 1115 return (EIO); 1116 1117 nmp = VFSTONFS(bp->b_vp->v_mount); 1118again: 1119 if (nmp->nm_flag & NFSMNT_INT) 1120 slpflag = PCATCH; 1121 gotiod = FALSE; 1122 1123 /* 1124 * Find a free iod to process this request. 1125 */ 1126 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 1127 if (nfs_iodwant[i]) { 1128 /* 1129 * Found one, so wake it up and tell it which 1130 * mount to process. 1131 */ 1132 NFS_DPF(ASYNCIO, 1133 ("nfs_asyncio: waking iod %d for mount %p\n", 1134 i, nmp)); 1135 nfs_iodwant[i] = (struct proc *)0; 1136 nfs_iodmount[i] = nmp; 1137 nmp->nm_bufqiods++; 1138 wakeup((caddr_t)&nfs_iodwant[i]); 1139 gotiod = TRUE; 1140 break; 1141 } 1142 1143 /* 1144 * If none are free, we may already have an iod working on this mount 1145 * point. If so, it will process our request. 1146 */ 1147 if (!gotiod) { 1148 if (nmp->nm_bufqiods > 0) { 1149 NFS_DPF(ASYNCIO, 1150 ("nfs_asyncio: %d iods are already processing mount %p\n", 1151 nmp->nm_bufqiods, nmp)); 1152 gotiod = TRUE; 1153 } 1154 } 1155 1156 /* 1157 * If we have an iod which can process the request, then queue 1158 * the buffer. 1159 */ 1160 if (gotiod) { 1161 /* 1162 * Ensure that the queue never grows too large. 1163 */ 1164 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1165 NFS_DPF(ASYNCIO, 1166 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1167 nmp->nm_bufqwant = TRUE; 1168 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO, 1169 "nfsaio", slptimeo); 1170 if (error) { 1171 if (nfs_sigintr(nmp, NULL, procp)) 1172 return (EINTR); 1173 if (slpflag == PCATCH) { 1174 slpflag = 0; 1175 slptimeo = 2 * hz; 1176 } 1177 } 1178 /* 1179 * We might have lost our iod while sleeping, 1180 * so check and loop if nescessary. 1181 */ 1182 if (nmp->nm_bufqiods == 0) { 1183 NFS_DPF(ASYNCIO, 1184 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1185 goto again; 1186 } 1187 } 1188 1189 if (bp->b_flags & B_READ) { 1190 if (bp->b_rcred == NOCRED && cred != NOCRED) { 1191 crhold(cred); 1192 bp->b_rcred = cred; 1193 } 1194 } else { 1195 bp->b_flags |= B_WRITEINPROG; 1196 if (bp->b_wcred == NOCRED && cred != NOCRED) { 1197 crhold(cred); 1198 bp->b_wcred = cred; 1199 } 1200 } 1201 1202 BUF_KERNPROC(bp); 1203 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1204 nmp->nm_bufqlen++; 1205 return (0); 1206 } 1207 1208 /* 1209 * All the iods are busy on other mounts, so return EIO to 1210 * force the caller to process the i/o synchronously. 1211 */ 1212 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1213 return (EIO); 1214} 1215 1216/* 1217 * Do an I/O operation to/from a cache block. This may be called 1218 * synchronously or from an nfsiod. 1219 */ 1220int 1221nfs_doio(bp, cr, p) 1222 struct buf *bp; 1223 struct ucred *cr; 1224 struct proc *p; 1225{ 1226 struct uio *uiop; 1227 struct vnode *vp; 1228 struct nfsnode *np; 1229 struct nfsmount *nmp; 1230 int error = 0, iomode, must_commit = 0; 1231 struct uio uio; 1232 struct iovec io; 1233 1234 vp = bp->b_vp; 1235 np = VTONFS(vp); 1236 nmp = VFSTONFS(vp->v_mount); 1237 uiop = &uio; 1238 uiop->uio_iov = &io; 1239 uiop->uio_iovcnt = 1; 1240 uiop->uio_segflg = UIO_SYSSPACE; 1241 uiop->uio_procp = p; 1242 1243 /* 1244 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 1245 * do this here so we do not have to do it in all the code that 1246 * calls us. 1247 */ 1248 bp->b_flags &= ~(B_ERROR | B_INVAL); 1249 1250 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1251 1252 /* 1253 * Historically, paging was done with physio, but no more. 1254 */ 1255 if (bp->b_flags & B_PHYS) { 1256 /* 1257 * ...though reading /dev/drum still gets us here. 1258 */ 1259 io.iov_len = uiop->uio_resid = bp->b_bcount; 1260 /* mapping was done by vmapbuf() */ 1261 io.iov_base = bp->b_data; 1262 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1263 if (bp->b_flags & B_READ) { 1264 uiop->uio_rw = UIO_READ; 1265 nfsstats.read_physios++; 1266 error = nfs_readrpc(vp, uiop, cr); 1267 } else { 1268 int com; 1269 1270 iomode = NFSV3WRITE_DATASYNC; 1271 uiop->uio_rw = UIO_WRITE; 1272 nfsstats.write_physios++; 1273 error = nfs_writerpc(vp, uiop, cr, &iomode, &com); 1274 } 1275 if (error) { 1276 bp->b_flags |= B_ERROR; 1277 bp->b_error = error; 1278 } 1279 } else if (bp->b_flags & B_READ) { 1280 io.iov_len = uiop->uio_resid = bp->b_bcount; 1281 io.iov_base = bp->b_data; 1282 uiop->uio_rw = UIO_READ; 1283 switch (vp->v_type) { 1284 case VREG: 1285 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1286 nfsstats.read_bios++; 1287 error = nfs_readrpc(vp, uiop, cr); 1288 if (!error) { 1289 if (uiop->uio_resid) { 1290 /* 1291 * If we had a short read with no error, we must have 1292 * hit a file hole. We should zero-fill the remainder. 1293 * This can also occur if the server hits the file EOF. 1294 * 1295 * Holes used to be able to occur due to pending 1296 * writes, but that is not possible any longer. 1297 */ 1298 int nread = bp->b_bcount - uiop->uio_resid; 1299 int left = bp->b_bcount - nread; 1300 1301 if (left > 0) 1302 bzero((char *)bp->b_data + nread, left); 1303 uiop->uio_resid = 0; 1304 } 1305 } 1306 if (p && (vp->v_flag & VTEXT) && 1307 (((nmp->nm_flag & NFSMNT_NQNFS) && 1308 NQNFS_CKINVALID(vp, np, ND_READ) && 1309 np->n_lrev != np->n_brev) || 1310 (!(nmp->nm_flag & NFSMNT_NQNFS) && 1311 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) { 1312 uprintf("Process killed due to text file modification\n"); 1313 psignal(p, SIGKILL); 1314 PHOLD(p); 1315 } 1316 break; 1317 case VLNK: 1318 uiop->uio_offset = (off_t)0; 1319 nfsstats.readlink_bios++; 1320 error = nfs_readlinkrpc(vp, uiop, cr); 1321 break; 1322 case VDIR: 1323 nfsstats.readdir_bios++; 1324 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1325 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1326 error = nfs_readdirplusrpc(vp, uiop, cr); 1327 if (error == NFSERR_NOTSUPP) 1328 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1329 } 1330 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1331 error = nfs_readdirrpc(vp, uiop, cr); 1332 /* 1333 * end-of-directory sets B_INVAL but does not generate an 1334 * error. 1335 */ 1336 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1337 bp->b_flags |= B_INVAL; 1338 break; 1339 default: 1340 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1341 break; 1342 }; 1343 if (error) { 1344 bp->b_flags |= B_ERROR; 1345 bp->b_error = error; 1346 } 1347 } else { 1348 /* 1349 * If we only need to commit, try to commit 1350 */ 1351 if (bp->b_flags & B_NEEDCOMMIT) { 1352 int retv; 1353 off_t off; 1354 1355 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1356 bp->b_flags |= B_WRITEINPROG; 1357 retv = nfs_commit( 1358 bp->b_vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1359 bp->b_wcred, p); 1360 bp->b_flags &= ~B_WRITEINPROG; 1361 if (retv == 0) { 1362 bp->b_dirtyoff = bp->b_dirtyend = 0; 1363 bp->b_flags &= ~B_NEEDCOMMIT; 1364 bp->b_resid = 0; 1365 biodone(bp); 1366 return (0); 1367 } 1368 if (retv == NFSERR_STALEWRITEVERF) { 1369 nfs_clearcommit(bp->b_vp->v_mount); 1370 } 1371 } 1372 1373 /* 1374 * Setup for actual write 1375 */ 1376 1377 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1378 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1379 1380 if (bp->b_dirtyend > bp->b_dirtyoff) { 1381 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1382 - bp->b_dirtyoff; 1383 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1384 + bp->b_dirtyoff; 1385 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1386 uiop->uio_rw = UIO_WRITE; 1387 nfsstats.write_bios++; 1388 1389 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1390 iomode = NFSV3WRITE_UNSTABLE; 1391 else 1392 iomode = NFSV3WRITE_FILESYNC; 1393 1394 bp->b_flags |= B_WRITEINPROG; 1395 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit); 1396 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1397 bp->b_flags |= B_NEEDCOMMIT; 1398 if (bp->b_dirtyoff == 0 1399 && bp->b_dirtyend == bp->b_bcount) 1400 bp->b_flags |= B_CLUSTEROK; 1401 } else { 1402 bp->b_flags &= ~B_NEEDCOMMIT; 1403 } 1404 bp->b_flags &= ~B_WRITEINPROG; 1405 1406 /* 1407 * For an interrupted write, the buffer is still valid 1408 * and the write hasn't been pushed to the server yet, 1409 * so we can't set B_ERROR and report the interruption 1410 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1411 * is not relevant, so the rpc attempt is essentially 1412 * a noop. For the case of a V3 write rpc not being 1413 * committed to stable storage, the block is still 1414 * dirty and requires either a commit rpc or another 1415 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1416 * the block is reused. This is indicated by setting 1417 * the B_DELWRI and B_NEEDCOMMIT flags. 1418 * 1419 * If the buffer is marked B_PAGING, it does not reside on 1420 * the vp's paging queues so we cannot call bdirty(). The 1421 * bp in this case is not an NFS cache block so we should 1422 * be safe. XXX 1423 */ 1424 if (error == EINTR 1425 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1426 int s; 1427 1428 s = splbio(); 1429 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1430 if ((bp->b_flags & B_PAGING) == 0) { 1431 bdirty(bp); 1432 bp->b_flags &= ~B_DONE; 1433 } 1434 if (error && (bp->b_flags & B_ASYNC) == 0) 1435 bp->b_flags |= B_EINTR; 1436 splx(s); 1437 } else { 1438 if (error) { 1439 bp->b_flags |= B_ERROR; 1440 bp->b_error = np->n_error = error; 1441 np->n_flag |= NWRITEERR; 1442 } 1443 bp->b_dirtyoff = bp->b_dirtyend = 0; 1444 } 1445 } else { 1446 bp->b_resid = 0; 1447 biodone(bp); 1448 return (0); 1449 } 1450 } 1451 bp->b_resid = uiop->uio_resid; 1452 if (must_commit) 1453 nfs_clearcommit(vp->v_mount); 1454 biodone(bp); 1455 return (error); 1456} 1457