nfs_bio.c revision 54480
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 54480 1999-12-12 06:09:57Z 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_page.h> 54#include <vm/vm_object.h> 55#include <vm/vm_pager.h> 56#include <vm/vnode_pager.h> 57 58#include <nfs/rpcv2.h> 59#include <nfs/nfsproto.h> 60#include <nfs/nfs.h> 61#include <nfs/nfsmount.h> 62#include <nfs/nqnfs.h> 63#include <nfs/nfsnode.h> 64 65static struct buf *nfs_getcacheblk __P((struct vnode *vp, daddr_t bn, int size, 66 struct proc *p)); 67 68extern int nfs_numasync; 69extern int nfs_pbuf_freecnt; 70extern struct nfsstats nfsstats; 71 72/* 73 * Vnode op for VM getpages. 74 */ 75int 76nfs_getpages(ap) 77 struct vop_getpages_args /* { 78 struct vnode *a_vp; 79 vm_page_t *a_m; 80 int a_count; 81 int a_reqpage; 82 vm_ooffset_t a_offset; 83 } */ *ap; 84{ 85 int i, error, nextoff, size, toff, count, npages; 86 struct uio uio; 87 struct iovec iov; 88 vm_offset_t kva; 89 struct buf *bp; 90 struct vnode *vp; 91 struct proc *p; 92 struct ucred *cred; 93 struct nfsmount *nmp; 94 vm_page_t *pages; 95 96 vp = ap->a_vp; 97 p = curproc; /* XXX */ 98 cred = curproc->p_ucred; /* XXX */ 99 nmp = VFSTONFS(vp->v_mount); 100 pages = ap->a_m; 101 count = ap->a_count; 102 103 if (vp->v_object == NULL) { 104 printf("nfs_getpages: called with non-merged cache vnode??\n"); 105 return VM_PAGER_ERROR; 106 } 107 108 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 109 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 110 (void)nfs_fsinfo(nmp, vp, cred, p); 111 112 npages = btoc(count); 113 114 /* 115 * If the requested page is partially valid, just return it and 116 * allow the pager to zero-out the blanks. Partially valid pages 117 * can only occur at the file EOF. 118 */ 119 120 { 121 vm_page_t m = pages[ap->a_reqpage]; 122 123 if (m->valid != 0) { 124 /* handled by vm_fault now */ 125 /* vm_page_zero_invalid(m, TRUE); */ 126 for (i = 0; i < npages; ++i) { 127 if (i != ap->a_reqpage) 128 vnode_pager_freepage(pages[i]); 129 } 130 return(0); 131 } 132 } 133 134 /* 135 * We use only the kva address for the buffer, but this is extremely 136 * convienient and fast. 137 */ 138 bp = getpbuf(&nfs_pbuf_freecnt); 139 140 kva = (vm_offset_t) bp->b_data; 141 pmap_qenter(kva, pages, npages); 142 143 iov.iov_base = (caddr_t) kva; 144 iov.iov_len = count; 145 uio.uio_iov = &iov; 146 uio.uio_iovcnt = 1; 147 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 148 uio.uio_resid = count; 149 uio.uio_segflg = UIO_SYSSPACE; 150 uio.uio_rw = UIO_READ; 151 uio.uio_procp = p; 152 153 error = nfs_readrpc(vp, &uio, cred); 154 pmap_qremove(kva, npages); 155 156 relpbuf(bp, &nfs_pbuf_freecnt); 157 158 if (error && (uio.uio_resid == count)) { 159 printf("nfs_getpages: error %d\n", error); 160 for (i = 0; i < npages; ++i) { 161 if (i != ap->a_reqpage) 162 vnode_pager_freepage(pages[i]); 163 } 164 return VM_PAGER_ERROR; 165 } 166 167 /* 168 * Calculate the number of bytes read and validate only that number 169 * of bytes. Note that due to pending writes, size may be 0. This 170 * does not mean that the remaining data is invalid! 171 */ 172 173 size = count - uio.uio_resid; 174 175 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 176 vm_page_t m; 177 nextoff = toff + PAGE_SIZE; 178 m = pages[i]; 179 180 m->flags &= ~PG_ZERO; 181 182 if (nextoff <= size) { 183 /* 184 * Read operation filled an entire page 185 */ 186 m->valid = VM_PAGE_BITS_ALL; 187 vm_page_undirty(m); 188 } else if (size > toff) { 189 /* 190 * Read operation filled a partial page. 191 */ 192 m->valid = 0; 193 vm_page_set_validclean(m, 0, size - toff); 194 /* handled by vm_fault now */ 195 /* vm_page_zero_invalid(m, TRUE); */ 196 } 197 198 if (i != ap->a_reqpage) { 199 /* 200 * Whether or not to leave the page activated is up in 201 * the air, but we should put the page on a page queue 202 * somewhere (it already is in the object). Result: 203 * It appears that emperical results show that 204 * deactivating pages is best. 205 */ 206 207 /* 208 * Just in case someone was asking for this page we 209 * now tell them that it is ok to use. 210 */ 211 if (!error) { 212 if (m->flags & PG_WANTED) 213 vm_page_activate(m); 214 else 215 vm_page_deactivate(m); 216 vm_page_wakeup(m); 217 } else { 218 vnode_pager_freepage(m); 219 } 220 } 221 } 222 return 0; 223} 224 225/* 226 * Vnode op for VM putpages. 227 */ 228int 229nfs_putpages(ap) 230 struct vop_putpages_args /* { 231 struct vnode *a_vp; 232 vm_page_t *a_m; 233 int a_count; 234 int a_sync; 235 int *a_rtvals; 236 vm_ooffset_t a_offset; 237 } */ *ap; 238{ 239 struct uio uio; 240 struct iovec iov; 241 vm_offset_t kva; 242 struct buf *bp; 243 int iomode, must_commit, i, error, npages, count; 244 off_t offset; 245 int *rtvals; 246 struct vnode *vp; 247 struct proc *p; 248 struct ucred *cred; 249 struct nfsmount *nmp; 250 struct nfsnode *np; 251 vm_page_t *pages; 252 253 vp = ap->a_vp; 254 np = VTONFS(vp); 255 p = curproc; /* XXX */ 256 cred = curproc->p_ucred; /* XXX */ 257 nmp = VFSTONFS(vp->v_mount); 258 pages = ap->a_m; 259 count = ap->a_count; 260 rtvals = ap->a_rtvals; 261 npages = btoc(count); 262 offset = IDX_TO_OFF(pages[0]->pindex); 263 264 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 265 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 266 (void)nfs_fsinfo(nmp, vp, cred, p); 267 268 for (i = 0; i < npages; i++) { 269 rtvals[i] = VM_PAGER_AGAIN; 270 } 271 272 /* 273 * When putting pages, do not extend file past EOF. 274 */ 275 276 if (offset + count > np->n_size) { 277 count = np->n_size - offset; 278 if (count < 0) 279 count = 0; 280 } 281 282 /* 283 * We use only the kva address for the buffer, but this is extremely 284 * convienient and fast. 285 */ 286 bp = getpbuf(&nfs_pbuf_freecnt); 287 288 kva = (vm_offset_t) bp->b_data; 289 pmap_qenter(kva, pages, npages); 290 291 iov.iov_base = (caddr_t) kva; 292 iov.iov_len = count; 293 uio.uio_iov = &iov; 294 uio.uio_iovcnt = 1; 295 uio.uio_offset = offset; 296 uio.uio_resid = count; 297 uio.uio_segflg = UIO_SYSSPACE; 298 uio.uio_rw = UIO_WRITE; 299 uio.uio_procp = p; 300 301 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 302 iomode = NFSV3WRITE_UNSTABLE; 303 else 304 iomode = NFSV3WRITE_FILESYNC; 305 306 error = nfs_writerpc(vp, &uio, cred, &iomode, &must_commit); 307 308 pmap_qremove(kva, npages); 309 relpbuf(bp, &nfs_pbuf_freecnt); 310 311 if (!error) { 312 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 313 for (i = 0; i < nwritten; i++) { 314 rtvals[i] = VM_PAGER_OK; 315 vm_page_undirty(pages[i]); 316 } 317 if (must_commit) 318 nfs_clearcommit(vp->v_mount); 319 } 320 return rtvals[0]; 321} 322 323/* 324 * Vnode op for read using bio 325 */ 326int 327nfs_bioread(vp, uio, ioflag, cred) 328 register struct vnode *vp; 329 register struct uio *uio; 330 int ioflag; 331 struct ucred *cred; 332{ 333 register struct nfsnode *np = VTONFS(vp); 334 register int biosize, i; 335 struct buf *bp = 0, *rabp; 336 struct vattr vattr; 337 struct proc *p; 338 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 339 daddr_t lbn, rabn; 340 int bcount; 341 int seqcount; 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; 353 354 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 355 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 356 (void)nfs_fsinfo(nmp, vp, cred, p); 357 if (vp->v_type != VDIR && 358 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 359 return (EFBIG); 360 biosize = vp->v_mount->mnt_stat.f_iosize; 361 seqcount = (int)((off_t)(ioflag >> 16) * biosize / BKVASIZE); 362 /* 363 * For nfs, cache consistency can only be maintained approximately. 364 * Although RFC1094 does not specify the criteria, the following is 365 * believed to be compatible with the reference port. 366 * For nqnfs, full cache consistency is maintained within the loop. 367 * For nfs: 368 * If the file's modify time on the server has changed since the 369 * last read rpc or you have written to the file, 370 * you may have lost data cache consistency with the 371 * server, so flush all of the file's data out of the cache. 372 * Then force a getattr rpc to ensure that you have up to date 373 * attributes. 374 * NB: This implies that cache data can be read when up to 375 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 376 * attributes this could be forced by setting n_attrstamp to 0 before 377 * the VOP_GETATTR() call. 378 */ 379 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) { 380 if (np->n_flag & NMODIFIED) { 381 if (vp->v_type != VREG) { 382 if (vp->v_type != VDIR) 383 panic("nfs: bioread, not dir"); 384 nfs_invaldir(vp); 385 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 386 if (error) 387 return (error); 388 } 389 np->n_attrstamp = 0; 390 error = VOP_GETATTR(vp, &vattr, cred, p); 391 if (error) 392 return (error); 393 np->n_mtime = vattr.va_mtime.tv_sec; 394 } else { 395 error = VOP_GETATTR(vp, &vattr, cred, p); 396 if (error) 397 return (error); 398 if (np->n_mtime != vattr.va_mtime.tv_sec) { 399 if (vp->v_type == VDIR) 400 nfs_invaldir(vp); 401 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 402 if (error) 403 return (error); 404 np->n_mtime = vattr.va_mtime.tv_sec; 405 } 406 } 407 } 408 do { 409 410 /* 411 * Get a valid lease. If cached data is stale, flush it. 412 */ 413 if (nmp->nm_flag & NFSMNT_NQNFS) { 414 if (NQNFS_CKINVALID(vp, np, ND_READ)) { 415 do { 416 error = nqnfs_getlease(vp, ND_READ, cred, p); 417 } while (error == NQNFS_EXPIRED); 418 if (error) 419 return (error); 420 if (np->n_lrev != np->n_brev || 421 (np->n_flag & NQNFSNONCACHE) || 422 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) { 423 if (vp->v_type == VDIR) 424 nfs_invaldir(vp); 425 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 426 if (error) 427 return (error); 428 np->n_brev = np->n_lrev; 429 } 430 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) { 431 nfs_invaldir(vp); 432 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 433 if (error) 434 return (error); 435 } 436 } 437 if (np->n_flag & NQNFSNONCACHE) { 438 switch (vp->v_type) { 439 case VREG: 440 return (nfs_readrpc(vp, uio, cred)); 441 case VLNK: 442 return (nfs_readlinkrpc(vp, uio, cred)); 443 case VDIR: 444 break; 445 default: 446 printf(" NQNFSNONCACHE: type %x unexpected\n", 447 vp->v_type); 448 }; 449 } 450 switch (vp->v_type) { 451 case VREG: 452 nfsstats.biocache_reads++; 453 lbn = uio->uio_offset / biosize; 454 on = uio->uio_offset & (biosize - 1); 455 456 /* 457 * Start the read ahead(s), as required. 458 */ 459 if (nfs_numasync > 0 && nmp->nm_readahead > 0) { 460 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 461 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 462 rabn = lbn + 1 + nra; 463 if (!incore(vp, rabn)) { 464 rabp = nfs_getcacheblk(vp, rabn, biosize, p); 465 if (!rabp) 466 return (EINTR); 467 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 468 rabp->b_flags |= (B_READ | B_ASYNC); 469 vfs_busy_pages(rabp, 0); 470 if (nfs_asyncio(rabp, cred, p)) { 471 rabp->b_flags |= B_INVAL|B_ERROR; 472 vfs_unbusy_pages(rabp); 473 brelse(rabp); 474 } 475 } else 476 brelse(rabp); 477 } 478 } 479 } 480 481 /* 482 * Obtain the buffer cache block. Figure out the buffer size 483 * when we are at EOF. nfs_getcacheblk() will also force 484 * uncached delayed-writes to be flushed to the server. 485 * 486 * Note that bcount is *not* DEV_BSIZE aligned. 487 */ 488 489 bcount = biosize; 490 if ((off_t)lbn * biosize >= np->n_size) { 491 bcount = 0; 492 } else if ((off_t)(lbn + 1) * biosize > np->n_size) { 493 bcount = np->n_size - (off_t)lbn * biosize; 494 } 495 496 bp = nfs_getcacheblk(vp, lbn, bcount, p); 497 if (!bp) 498 return (EINTR); 499 500 /* 501 * If B_CACHE is not set, we must issue the read. If this 502 * fails, we return an error. 503 */ 504 505 if ((bp->b_flags & B_CACHE) == 0) { 506 bp->b_flags |= B_READ; 507 vfs_busy_pages(bp, 0); 508 error = nfs_doio(bp, cred, p); 509 if (error) { 510 brelse(bp); 511 return (error); 512 } 513 } 514 515 /* 516 * on is the offset into the current bp. Figure out how many 517 * bytes we can copy out of the bp. Note that bcount is 518 * NOT DEV_BSIZE aligned. 519 * 520 * Then figure out how many bytes we can copy into the uio. 521 */ 522 523 n = 0; 524 if (on < bcount) 525 n = min((unsigned)(bcount - on), uio->uio_resid); 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 if (!bp) 806 return (EINTR); 807 save = bp->b_flags & B_CACHE; 808 809 np->n_size = uio->uio_offset + n; 810 np->n_flag |= NMODIFIED; 811 vnode_pager_setsize(vp, np->n_size); 812 813 bcount += n; 814 allocbuf(bp, bcount); 815 bp->b_flags |= save; 816 } else { 817 if (uio->uio_offset + n > np->n_size) { 818 np->n_size = uio->uio_offset + n; 819 np->n_flag |= NMODIFIED; 820 vnode_pager_setsize(vp, np->n_size); 821 } 822 bcount = biosize; 823 if ((off_t)(lbn + 1) * biosize > np->n_size) 824 bcount = np->n_size - (off_t)lbn * biosize; 825 bp = nfs_getcacheblk(vp, lbn, bcount, p); 826 if (!bp) 827 return (EINTR); 828 } 829 830 /* 831 * Issue a READ if B_CACHE is not set. In special-append 832 * mode, B_CACHE is based on the buffer prior to the write 833 * op and is typically set, avoiding the read. If a read 834 * is required in special append mode, the server will 835 * probably send us a short-read since we extended the file 836 * on our end, resulting in b_resid == 0 and, thusly, 837 * B_CACHE getting set. 838 * 839 * We can also avoid issuing the read if the write covers 840 * the entire buffer. We have to make sure the buffer state 841 * is reasonable in this case since we will not be initiating 842 * I/O. See the comments in kern/vfs_bio.c's getblk() for 843 * more information. 844 * 845 * B_CACHE may also be set due to the buffer being cached 846 * normally. 847 */ 848 849 if (on == 0 && n == bcount) { 850 bp->b_flags |= B_CACHE; 851 bp->b_flags &= ~(B_ERROR | B_INVAL); 852 } 853 854 if ((bp->b_flags & B_CACHE) == 0) { 855 bp->b_flags |= B_READ; 856 vfs_busy_pages(bp, 0); 857 error = nfs_doio(bp, cred, p); 858 if (error) { 859 brelse(bp); 860 return (error); 861 } 862 } 863 if (!bp) 864 return (EINTR); 865 if (bp->b_wcred == NOCRED) { 866 crhold(cred); 867 bp->b_wcred = cred; 868 } 869 np->n_flag |= NMODIFIED; 870 871 /* 872 * If dirtyend exceeds file size, chop it down. If this 873 * creates a reverse-indexed or degenerate situation with 874 * dirtyoff/end, 0 them. 875 */ 876 877 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 878 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 879 if (bp->b_dirtyoff >= bp->b_dirtyend) 880 bp->b_dirtyoff = bp->b_dirtyend = 0; 881 882 /* 883 * If the new write will leave a contiguous dirty 884 * area, just update the b_dirtyoff and b_dirtyend, 885 * otherwise force a write rpc of the old dirty area. 886 * 887 * While it is possible to merge discontiguous writes due to 888 * our having a B_CACHE buffer ( and thus valid read data 889 * for the hole), we don't because it could lead to 890 * significant cache coherency problems with multiple clients, 891 * especially if locking is implemented later on. 892 * 893 * as an optimization we could theoretically maintain 894 * a linked list of discontinuous areas, but we would still 895 * have to commit them separately so there isn't much 896 * advantage to it except perhaps a bit of asynchronization. 897 */ 898 899 if (bp->b_dirtyend > 0 && 900 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 901 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) 902 return (EINTR); 903 goto again; 904 } 905 906 /* 907 * Check for valid write lease and get one as required. 908 * In case getblk() and/or bwrite() delayed us. 909 */ 910 if ((nmp->nm_flag & NFSMNT_NQNFS) && 911 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 912 do { 913 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 914 } while (error == NQNFS_EXPIRED); 915 if (error) { 916 brelse(bp); 917 return (error); 918 } 919 if (np->n_lrev != np->n_brev || 920 (np->n_flag & NQNFSNONCACHE)) { 921 brelse(bp); 922 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 923 if (error) 924 return (error); 925 np->n_brev = np->n_lrev; 926 goto again; 927 } 928 } 929 930 error = uiomove((char *)bp->b_data + on, n, uio); 931 932 /* 933 * Since this block is being modified, it must be written 934 * again and not just committed. Since write clustering does 935 * not work for the stage 1 data write, only the stage 2 936 * commit rpc, we have to clear B_CLUSTEROK as well. 937 */ 938 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 939 940 if (error) { 941 bp->b_flags |= B_ERROR; 942 brelse(bp); 943 return (error); 944 } 945 946 /* 947 * Only update dirtyoff/dirtyend if not a degenerate 948 * condition. 949 */ 950 if (n) { 951 if (bp->b_dirtyend > 0) { 952 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 953 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 954 } else { 955 bp->b_dirtyoff = on; 956 bp->b_dirtyend = on + n; 957 } 958 vfs_bio_set_validclean(bp, on, n); 959 } 960 961 /* 962 * If the lease is non-cachable or IO_SYNC do bwrite(). 963 * 964 * IO_INVAL appears to be unused. The idea appears to be 965 * to turn off caching in this case. Very odd. XXX 966 */ 967 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 968 if (ioflag & IO_INVAL) 969 bp->b_flags |= B_NOCACHE; 970 error = VOP_BWRITE(bp->b_vp, bp); 971 if (error) 972 return (error); 973 if (np->n_flag & NQNFSNONCACHE) { 974 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 975 if (error) 976 return (error); 977 } 978 } else if ((n + on) == biosize && 979 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 980 bp->b_flags |= B_ASYNC; 981 (void)nfs_writebp(bp, 0, 0); 982 } else { 983 bdwrite(bp); 984 } 985 } while (uio->uio_resid > 0 && n > 0); 986 return (0); 987} 988 989/* 990 * Get an nfs cache block. 991 * 992 * Allocate a new one if the block isn't currently in the cache 993 * and return the block marked busy. If the calling process is 994 * interrupted by a signal for an interruptible mount point, return 995 * NULL. 996 * 997 * The caller must carefully deal with the possible B_INVAL state of 998 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 999 * indirectly), so synchronous reads can be issued without worrying about 1000 * the B_INVAL state. We have to be a little more careful when dealing 1001 * with writes (see comments in nfs_write()) when extending a file past 1002 * its EOF. 1003 */ 1004static struct buf * 1005nfs_getcacheblk(vp, bn, size, p) 1006 struct vnode *vp; 1007 daddr_t bn; 1008 int size; 1009 struct proc *p; 1010{ 1011 register struct buf *bp; 1012 struct mount *mp; 1013 struct nfsmount *nmp; 1014 1015 mp = vp->v_mount; 1016 nmp = VFSTONFS(mp); 1017 1018 if (nmp->nm_flag & NFSMNT_INT) { 1019 bp = getblk(vp, bn, size, PCATCH, 0); 1020 while (bp == (struct buf *)0) { 1021 if (nfs_sigintr(nmp, (struct nfsreq *)0, p)) 1022 return ((struct buf *)0); 1023 bp = getblk(vp, bn, size, 0, 2 * hz); 1024 } 1025 } else { 1026 bp = getblk(vp, bn, size, 0, 0); 1027 } 1028 1029 if (vp->v_type == VREG) { 1030 int biosize; 1031 1032 biosize = mp->mnt_stat.f_iosize; 1033 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1034 } 1035 return (bp); 1036} 1037 1038/* 1039 * Flush and invalidate all dirty buffers. If another process is already 1040 * doing the flush, just wait for completion. 1041 */ 1042int 1043nfs_vinvalbuf(vp, flags, cred, p, intrflg) 1044 struct vnode *vp; 1045 int flags; 1046 struct ucred *cred; 1047 struct proc *p; 1048 int intrflg; 1049{ 1050 register struct nfsnode *np = VTONFS(vp); 1051 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1052 int error = 0, slpflag, slptimeo; 1053 1054 if (vp->v_flag & VXLOCK) { 1055 return (0); 1056 } 1057 1058 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1059 intrflg = 0; 1060 if (intrflg) { 1061 slpflag = PCATCH; 1062 slptimeo = 2 * hz; 1063 } else { 1064 slpflag = 0; 1065 slptimeo = 0; 1066 } 1067 /* 1068 * First wait for any other process doing a flush to complete. 1069 */ 1070 while (np->n_flag & NFLUSHINPROG) { 1071 np->n_flag |= NFLUSHWANT; 1072 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval", 1073 slptimeo); 1074 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) 1075 return (EINTR); 1076 } 1077 1078 /* 1079 * Now, flush as required. 1080 */ 1081 np->n_flag |= NFLUSHINPROG; 1082 error = vinvalbuf(vp, flags, cred, p, slpflag, 0); 1083 while (error) { 1084 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) { 1085 np->n_flag &= ~NFLUSHINPROG; 1086 if (np->n_flag & NFLUSHWANT) { 1087 np->n_flag &= ~NFLUSHWANT; 1088 wakeup((caddr_t)&np->n_flag); 1089 } 1090 return (EINTR); 1091 } 1092 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo); 1093 } 1094 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 1095 if (np->n_flag & NFLUSHWANT) { 1096 np->n_flag &= ~NFLUSHWANT; 1097 wakeup((caddr_t)&np->n_flag); 1098 } 1099 return (0); 1100} 1101 1102/* 1103 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1104 * This is mainly to avoid queueing async I/O requests when the nfsiods 1105 * are all hung on a dead server. 1106 * 1107 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp 1108 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1109 */ 1110int 1111nfs_asyncio(bp, cred, procp) 1112 register struct buf *bp; 1113 struct ucred *cred; 1114 struct proc *procp; 1115{ 1116 struct nfsmount *nmp; 1117 int i; 1118 int gotiod; 1119 int slpflag = 0; 1120 int slptimeo = 0; 1121 int error; 1122 1123 if (nfs_numasync == 0) 1124 return (EIO); 1125 1126 nmp = VFSTONFS(bp->b_vp->v_mount); 1127again: 1128 if (nmp->nm_flag & NFSMNT_INT) 1129 slpflag = PCATCH; 1130 gotiod = FALSE; 1131 1132 /* 1133 * Find a free iod to process this request. 1134 */ 1135 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 1136 if (nfs_iodwant[i]) { 1137 /* 1138 * Found one, so wake it up and tell it which 1139 * mount to process. 1140 */ 1141 NFS_DPF(ASYNCIO, 1142 ("nfs_asyncio: waking iod %d for mount %p\n", 1143 i, nmp)); 1144 nfs_iodwant[i] = (struct proc *)0; 1145 nfs_iodmount[i] = nmp; 1146 nmp->nm_bufqiods++; 1147 wakeup((caddr_t)&nfs_iodwant[i]); 1148 gotiod = TRUE; 1149 break; 1150 } 1151 1152 /* 1153 * If none are free, we may already have an iod working on this mount 1154 * point. If so, it will process our request. 1155 */ 1156 if (!gotiod) { 1157 if (nmp->nm_bufqiods > 0) { 1158 NFS_DPF(ASYNCIO, 1159 ("nfs_asyncio: %d iods are already processing mount %p\n", 1160 nmp->nm_bufqiods, nmp)); 1161 gotiod = TRUE; 1162 } 1163 } 1164 1165 /* 1166 * If we have an iod which can process the request, then queue 1167 * the buffer. 1168 */ 1169 if (gotiod) { 1170 /* 1171 * Ensure that the queue never grows too large. 1172 */ 1173 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1174 NFS_DPF(ASYNCIO, 1175 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1176 nmp->nm_bufqwant = TRUE; 1177 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO, 1178 "nfsaio", slptimeo); 1179 if (error) { 1180 if (nfs_sigintr(nmp, NULL, procp)) 1181 return (EINTR); 1182 if (slpflag == PCATCH) { 1183 slpflag = 0; 1184 slptimeo = 2 * hz; 1185 } 1186 } 1187 /* 1188 * We might have lost our iod while sleeping, 1189 * so check and loop if nescessary. 1190 */ 1191 if (nmp->nm_bufqiods == 0) { 1192 NFS_DPF(ASYNCIO, 1193 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1194 goto again; 1195 } 1196 } 1197 1198 if (bp->b_flags & B_READ) { 1199 if (bp->b_rcred == NOCRED && cred != NOCRED) { 1200 crhold(cred); 1201 bp->b_rcred = cred; 1202 } 1203 } else { 1204 bp->b_flags |= B_WRITEINPROG; 1205 if (bp->b_wcred == NOCRED && cred != NOCRED) { 1206 crhold(cred); 1207 bp->b_wcred = cred; 1208 } 1209 } 1210 1211 BUF_KERNPROC(bp); 1212 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1213 nmp->nm_bufqlen++; 1214 return (0); 1215 } 1216 1217 /* 1218 * All the iods are busy on other mounts, so return EIO to 1219 * force the caller to process the i/o synchronously. 1220 */ 1221 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1222 return (EIO); 1223} 1224 1225/* 1226 * Do an I/O operation to/from a cache block. This may be called 1227 * synchronously or from an nfsiod. 1228 */ 1229int 1230nfs_doio(bp, cr, p) 1231 struct buf *bp; 1232 struct ucred *cr; 1233 struct proc *p; 1234{ 1235 struct uio *uiop; 1236 struct vnode *vp; 1237 struct nfsnode *np; 1238 struct nfsmount *nmp; 1239 int error = 0, iomode, must_commit = 0; 1240 struct uio uio; 1241 struct iovec io; 1242 1243 vp = bp->b_vp; 1244 np = VTONFS(vp); 1245 nmp = VFSTONFS(vp->v_mount); 1246 uiop = &uio; 1247 uiop->uio_iov = &io; 1248 uiop->uio_iovcnt = 1; 1249 uiop->uio_segflg = UIO_SYSSPACE; 1250 uiop->uio_procp = p; 1251 1252 /* 1253 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 1254 * do this here so we do not have to do it in all the code that 1255 * calls us. 1256 */ 1257 bp->b_flags &= ~(B_ERROR | B_INVAL); 1258 1259 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1260 1261 /* 1262 * Historically, paging was done with physio, but no more. 1263 */ 1264 if (bp->b_flags & B_PHYS) { 1265 /* 1266 * ...though reading /dev/drum still gets us here. 1267 */ 1268 io.iov_len = uiop->uio_resid = bp->b_bcount; 1269 /* mapping was done by vmapbuf() */ 1270 io.iov_base = bp->b_data; 1271 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1272 if (bp->b_flags & B_READ) { 1273 uiop->uio_rw = UIO_READ; 1274 nfsstats.read_physios++; 1275 error = nfs_readrpc(vp, uiop, cr); 1276 } else { 1277 int com; 1278 1279 iomode = NFSV3WRITE_DATASYNC; 1280 uiop->uio_rw = UIO_WRITE; 1281 nfsstats.write_physios++; 1282 error = nfs_writerpc(vp, uiop, cr, &iomode, &com); 1283 } 1284 if (error) { 1285 bp->b_flags |= B_ERROR; 1286 bp->b_error = error; 1287 } 1288 } else if (bp->b_flags & B_READ) { 1289 io.iov_len = uiop->uio_resid = bp->b_bcount; 1290 io.iov_base = bp->b_data; 1291 uiop->uio_rw = UIO_READ; 1292 switch (vp->v_type) { 1293 case VREG: 1294 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1295 nfsstats.read_bios++; 1296 error = nfs_readrpc(vp, uiop, cr); 1297 if (!error) { 1298 if (uiop->uio_resid) { 1299 /* 1300 * If we had a short read with no error, we must have 1301 * hit a file hole. We should zero-fill the remainder. 1302 * This can also occur if the server hits the file EOF. 1303 * 1304 * Holes used to be able to occur due to pending 1305 * writes, but that is not possible any longer. 1306 */ 1307 int nread = bp->b_bcount - uiop->uio_resid; 1308 int left = bp->b_bcount - nread; 1309 1310 if (left > 0) 1311 bzero((char *)bp->b_data + nread, left); 1312 uiop->uio_resid = 0; 1313 } 1314 } 1315 if (p && (vp->v_flag & VTEXT) && 1316 (((nmp->nm_flag & NFSMNT_NQNFS) && 1317 NQNFS_CKINVALID(vp, np, ND_READ) && 1318 np->n_lrev != np->n_brev) || 1319 (!(nmp->nm_flag & NFSMNT_NQNFS) && 1320 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) { 1321 uprintf("Process killed due to text file modification\n"); 1322 psignal(p, SIGKILL); 1323 PHOLD(p); 1324 } 1325 break; 1326 case VLNK: 1327 uiop->uio_offset = (off_t)0; 1328 nfsstats.readlink_bios++; 1329 error = nfs_readlinkrpc(vp, uiop, cr); 1330 break; 1331 case VDIR: 1332 nfsstats.readdir_bios++; 1333 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1334 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1335 error = nfs_readdirplusrpc(vp, uiop, cr); 1336 if (error == NFSERR_NOTSUPP) 1337 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1338 } 1339 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1340 error = nfs_readdirrpc(vp, uiop, cr); 1341 /* 1342 * end-of-directory sets B_INVAL but does not generate an 1343 * error. 1344 */ 1345 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1346 bp->b_flags |= B_INVAL; 1347 break; 1348 default: 1349 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1350 break; 1351 }; 1352 if (error) { 1353 bp->b_flags |= B_ERROR; 1354 bp->b_error = error; 1355 } 1356 } else { 1357 /* 1358 * If we only need to commit, try to commit 1359 */ 1360 if (bp->b_flags & B_NEEDCOMMIT) { 1361 int retv; 1362 off_t off; 1363 1364 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1365 bp->b_flags |= B_WRITEINPROG; 1366 retv = nfs_commit( 1367 bp->b_vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1368 bp->b_wcred, p); 1369 bp->b_flags &= ~B_WRITEINPROG; 1370 if (retv == 0) { 1371 bp->b_dirtyoff = bp->b_dirtyend = 0; 1372 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1373 bp->b_resid = 0; 1374 biodone(bp); 1375 return (0); 1376 } 1377 if (retv == NFSERR_STALEWRITEVERF) { 1378 nfs_clearcommit(bp->b_vp->v_mount); 1379 } 1380 } 1381 1382 /* 1383 * Setup for actual write 1384 */ 1385 1386 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1387 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1388 1389 if (bp->b_dirtyend > bp->b_dirtyoff) { 1390 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1391 - bp->b_dirtyoff; 1392 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1393 + bp->b_dirtyoff; 1394 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1395 uiop->uio_rw = UIO_WRITE; 1396 nfsstats.write_bios++; 1397 1398 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1399 iomode = NFSV3WRITE_UNSTABLE; 1400 else 1401 iomode = NFSV3WRITE_FILESYNC; 1402 1403 bp->b_flags |= B_WRITEINPROG; 1404 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit); 1405 1406 /* 1407 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1408 * to cluster the buffers needing commit. This will allow 1409 * the system to submit a single commit rpc for the whole 1410 * cluster. We can do this even if the buffer is not 100% 1411 * dirty (relative to the NFS blocksize), so we optimize the 1412 * append-to-file-case. 1413 * 1414 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1415 * cleared because write clustering only works for commit 1416 * rpc's, not for the data portion of the write). 1417 */ 1418 1419 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1420 bp->b_flags |= B_NEEDCOMMIT; 1421 if (bp->b_dirtyoff == 0 1422 && bp->b_dirtyend == bp->b_bcount) 1423 bp->b_flags |= B_CLUSTEROK; 1424 } else { 1425 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1426 } 1427 bp->b_flags &= ~B_WRITEINPROG; 1428 1429 /* 1430 * For an interrupted write, the buffer is still valid 1431 * and the write hasn't been pushed to the server yet, 1432 * so we can't set B_ERROR and report the interruption 1433 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1434 * is not relevant, so the rpc attempt is essentially 1435 * a noop. For the case of a V3 write rpc not being 1436 * committed to stable storage, the block is still 1437 * dirty and requires either a commit rpc or another 1438 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1439 * the block is reused. This is indicated by setting 1440 * the B_DELWRI and B_NEEDCOMMIT flags. 1441 * 1442 * If the buffer is marked B_PAGING, it does not reside on 1443 * the vp's paging queues so we cannot call bdirty(). The 1444 * bp in this case is not an NFS cache block so we should 1445 * be safe. XXX 1446 */ 1447 if (error == EINTR 1448 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1449 int s; 1450 1451 s = splbio(); 1452 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1453 if ((bp->b_flags & B_PAGING) == 0) { 1454 bdirty(bp); 1455 bp->b_flags &= ~B_DONE; 1456 } 1457 if (error && (bp->b_flags & B_ASYNC) == 0) 1458 bp->b_flags |= B_EINTR; 1459 splx(s); 1460 } else { 1461 if (error) { 1462 bp->b_flags |= B_ERROR; 1463 bp->b_error = np->n_error = error; 1464 np->n_flag |= NWRITEERR; 1465 } 1466 bp->b_dirtyoff = bp->b_dirtyend = 0; 1467 } 1468 } else { 1469 bp->b_resid = 0; 1470 biodone(bp); 1471 return (0); 1472 } 1473 } 1474 bp->b_resid = uiop->uio_resid; 1475 if (must_commit) 1476 nfs_clearcommit(vp->v_mount); 1477 biodone(bp); 1478 return (error); 1479} 1480