nfs_bio.c revision 79224
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 79224 2001-07-04 16:20:28Z 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/bio.h> 47#include <sys/buf.h> 48#include <sys/vnode.h> 49#include <sys/mount.h> 50#include <sys/kernel.h> 51 52#include <vm/vm.h> 53#include <vm/vm_extern.h> 54#include <vm/vm_page.h> 55#include <vm/vm_object.h> 56#include <vm/vm_pager.h> 57#include <vm/vnode_pager.h> 58 59#include <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 66/* 67 * Just call nfs_writebp() with the force argument set to 1. 68 * 69 * NOTE: B_DONE may or may not be set in a_bp on call. 70 */ 71static int 72nfs_bwrite(struct buf *bp) 73{ 74 return (nfs_writebp(bp, 1, curproc)); 75} 76 77struct buf_ops buf_ops_nfs = { 78 "buf_ops_nfs", 79 nfs_bwrite 80}; 81 82 83static struct buf *nfs_getcacheblk __P((struct vnode *vp, daddr_t bn, int size, 84 struct proc *p)); 85 86extern int nfs_numasync; 87extern int nfs_pbuf_freecnt; 88extern struct nfsstats nfsstats; 89 90/* 91 * Vnode op for VM getpages. 92 */ 93int 94nfs_getpages(ap) 95 struct vop_getpages_args /* { 96 struct vnode *a_vp; 97 vm_page_t *a_m; 98 int a_count; 99 int a_reqpage; 100 vm_ooffset_t a_offset; 101 } */ *ap; 102{ 103 int i, error, nextoff, size, toff, count, npages; 104 struct uio uio; 105 struct iovec iov; 106 vm_offset_t kva; 107 struct buf *bp; 108 struct vnode *vp; 109 struct proc *p; 110 struct ucred *cred; 111 struct nfsmount *nmp; 112 vm_page_t *pages; 113 114 GIANT_REQUIRED; 115 116 vp = ap->a_vp; 117 p = curproc; /* XXX */ 118 cred = curproc->p_ucred; /* XXX */ 119 nmp = VFSTONFS(vp->v_mount); 120 pages = ap->a_m; 121 count = ap->a_count; 122 123 if (vp->v_object == NULL) { 124 printf("nfs_getpages: called with non-merged cache vnode??\n"); 125 return VM_PAGER_ERROR; 126 } 127 128 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 129 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 130 (void)nfs_fsinfo(nmp, vp, cred, p); 131 } 132 133 npages = btoc(count); 134 135 /* 136 * If the requested page is partially valid, just return it and 137 * allow the pager to zero-out the blanks. Partially valid pages 138 * can only occur at the file EOF. 139 */ 140 141 { 142 vm_page_t m = pages[ap->a_reqpage]; 143 144 if (m->valid != 0) { 145 /* handled by vm_fault now */ 146 /* vm_page_zero_invalid(m, TRUE); */ 147 for (i = 0; i < npages; ++i) { 148 if (i != ap->a_reqpage) 149 vm_page_free(pages[i]); 150 } 151 return(0); 152 } 153 } 154 155 /* 156 * We use only the kva address for the buffer, but this is extremely 157 * convienient and fast. 158 */ 159 bp = getpbuf(&nfs_pbuf_freecnt); 160 161 kva = (vm_offset_t) bp->b_data; 162 pmap_qenter(kva, pages, npages); 163 164 iov.iov_base = (caddr_t) kva; 165 iov.iov_len = count; 166 uio.uio_iov = &iov; 167 uio.uio_iovcnt = 1; 168 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 169 uio.uio_resid = count; 170 uio.uio_segflg = UIO_SYSSPACE; 171 uio.uio_rw = UIO_READ; 172 uio.uio_procp = p; 173 174 error = nfs_readrpc(vp, &uio, cred); 175 pmap_qremove(kva, npages); 176 177 relpbuf(bp, &nfs_pbuf_freecnt); 178 179 if (error && (uio.uio_resid == count)) { 180 printf("nfs_getpages: error %d\n", error); 181 for (i = 0; i < npages; ++i) { 182 if (i != ap->a_reqpage) 183 vm_page_free(pages[i]); 184 } 185 return VM_PAGER_ERROR; 186 } 187 188 /* 189 * Calculate the number of bytes read and validate only that number 190 * of bytes. Note that due to pending writes, size may be 0. This 191 * does not mean that the remaining data is invalid! 192 */ 193 194 size = count - uio.uio_resid; 195 196 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 197 vm_page_t m; 198 nextoff = toff + PAGE_SIZE; 199 m = pages[i]; 200 201 m->flags &= ~PG_ZERO; 202 203 if (nextoff <= size) { 204 /* 205 * Read operation filled an entire page 206 */ 207 m->valid = VM_PAGE_BITS_ALL; 208 vm_page_undirty(m); 209 } else if (size > toff) { 210 /* 211 * Read operation filled a partial page. 212 */ 213 m->valid = 0; 214 vm_page_set_validclean(m, 0, size - toff); 215 /* handled by vm_fault now */ 216 /* vm_page_zero_invalid(m, TRUE); */ 217 } 218 219 if (i != ap->a_reqpage) { 220 /* 221 * Whether or not to leave the page activated is up in 222 * the air, but we should put the page on a page queue 223 * somewhere (it already is in the object). Result: 224 * It appears that emperical results show that 225 * deactivating pages is best. 226 */ 227 228 /* 229 * Just in case someone was asking for this page we 230 * now tell them that it is ok to use. 231 */ 232 if (!error) { 233 if (m->flags & PG_WANTED) 234 vm_page_activate(m); 235 else 236 vm_page_deactivate(m); 237 vm_page_wakeup(m); 238 } else { 239 vm_page_free(m); 240 } 241 } 242 } 243 return 0; 244} 245 246/* 247 * Vnode op for VM putpages. 248 */ 249int 250nfs_putpages(ap) 251 struct vop_putpages_args /* { 252 struct vnode *a_vp; 253 vm_page_t *a_m; 254 int a_count; 255 int a_sync; 256 int *a_rtvals; 257 vm_ooffset_t a_offset; 258 } */ *ap; 259{ 260 struct uio uio; 261 struct iovec iov; 262 vm_offset_t kva; 263 struct buf *bp; 264 int iomode, must_commit, i, error, npages, count; 265 off_t offset; 266 int *rtvals; 267 struct vnode *vp; 268 struct proc *p; 269 struct ucred *cred; 270 struct nfsmount *nmp; 271 struct nfsnode *np; 272 vm_page_t *pages; 273 274 GIANT_REQUIRED; 275 276 vp = ap->a_vp; 277 np = VTONFS(vp); 278 p = curproc; /* XXX */ 279 cred = curproc->p_ucred; /* XXX */ 280 nmp = VFSTONFS(vp->v_mount); 281 pages = ap->a_m; 282 count = ap->a_count; 283 rtvals = ap->a_rtvals; 284 npages = btoc(count); 285 offset = IDX_TO_OFF(pages[0]->pindex); 286 287 GIANT_REQUIRED; 288 289 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 290 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 291 (void)nfs_fsinfo(nmp, vp, cred, p); 292 } 293 294 for (i = 0; i < npages; i++) { 295 rtvals[i] = VM_PAGER_AGAIN; 296 } 297 298 /* 299 * When putting pages, do not extend file past EOF. 300 */ 301 302 if (offset + count > np->n_size) { 303 count = np->n_size - offset; 304 if (count < 0) 305 count = 0; 306 } 307 308 /* 309 * We use only the kva address for the buffer, but this is extremely 310 * convienient and fast. 311 */ 312 bp = getpbuf(&nfs_pbuf_freecnt); 313 314 kva = (vm_offset_t) bp->b_data; 315 pmap_qenter(kva, pages, npages); 316 317 iov.iov_base = (caddr_t) kva; 318 iov.iov_len = count; 319 uio.uio_iov = &iov; 320 uio.uio_iovcnt = 1; 321 uio.uio_offset = offset; 322 uio.uio_resid = count; 323 uio.uio_segflg = UIO_SYSSPACE; 324 uio.uio_rw = UIO_WRITE; 325 uio.uio_procp = p; 326 327 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 328 iomode = NFSV3WRITE_UNSTABLE; 329 else 330 iomode = NFSV3WRITE_FILESYNC; 331 332 error = nfs_writerpc(vp, &uio, cred, &iomode, &must_commit); 333 334 pmap_qremove(kva, npages); 335 relpbuf(bp, &nfs_pbuf_freecnt); 336 337 if (!error) { 338 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 339 for (i = 0; i < nwritten; i++) { 340 rtvals[i] = VM_PAGER_OK; 341 vm_page_undirty(pages[i]); 342 } 343 if (must_commit) { 344 nfs_clearcommit(vp->v_mount); 345 } 346 } 347 return rtvals[0]; 348} 349 350/* 351 * Vnode op for read using bio 352 */ 353int 354nfs_bioread(vp, uio, ioflag, cred) 355 register struct vnode *vp; 356 register struct uio *uio; 357 int ioflag; 358 struct ucred *cred; 359{ 360 register struct nfsnode *np = VTONFS(vp); 361 register int biosize, i; 362 struct buf *bp = 0, *rabp; 363 struct vattr vattr; 364 struct proc *p; 365 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 366 daddr_t lbn, rabn; 367 int bcount; 368 int seqcount; 369 int nra, error = 0, n = 0, on = 0; 370 371#ifdef DIAGNOSTIC 372 if (uio->uio_rw != UIO_READ) 373 panic("nfs_read mode"); 374#endif 375 if (uio->uio_resid == 0) 376 return (0); 377 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 378 return (EINVAL); 379 p = uio->uio_procp; 380 381 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 382 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 383 (void)nfs_fsinfo(nmp, vp, cred, p); 384 if (vp->v_type != VDIR && 385 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 386 return (EFBIG); 387 biosize = vp->v_mount->mnt_stat.f_iosize; 388 seqcount = (int)((off_t)(ioflag >> 16) * biosize / BKVASIZE); 389 /* 390 * For nfs, cache consistency can only be maintained approximately. 391 * Although RFC1094 does not specify the criteria, the following is 392 * believed to be compatible with the reference port. 393 * For nqnfs, full cache consistency is maintained within the loop. 394 * For nfs: 395 * If the file's modify time on the server has changed since the 396 * last read rpc or you have written to the file, 397 * you may have lost data cache consistency with the 398 * server, so flush all of the file's data out of the cache. 399 * Then force a getattr rpc to ensure that you have up to date 400 * attributes. 401 * NB: This implies that cache data can be read when up to 402 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 403 * attributes this could be forced by setting n_attrstamp to 0 before 404 * the VOP_GETATTR() call. 405 */ 406 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) { 407 if (np->n_flag & NMODIFIED) { 408 if (vp->v_type != VREG) { 409 if (vp->v_type != VDIR) 410 panic("nfs: bioread, not dir"); 411 nfs_invaldir(vp); 412 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 413 if (error) 414 return (error); 415 } 416 np->n_attrstamp = 0; 417 error = VOP_GETATTR(vp, &vattr, cred, p); 418 if (error) 419 return (error); 420 np->n_mtime = vattr.va_mtime.tv_sec; 421 } else { 422 error = VOP_GETATTR(vp, &vattr, cred, p); 423 if (error) 424 return (error); 425 if (np->n_mtime != vattr.va_mtime.tv_sec) { 426 if (vp->v_type == VDIR) 427 nfs_invaldir(vp); 428 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 429 if (error) 430 return (error); 431 np->n_mtime = vattr.va_mtime.tv_sec; 432 } 433 } 434 } 435 do { 436 437 /* 438 * Get a valid lease. If cached data is stale, flush it. 439 */ 440 if (nmp->nm_flag & NFSMNT_NQNFS) { 441 if (NQNFS_CKINVALID(vp, np, ND_READ)) { 442 do { 443 error = nqnfs_getlease(vp, ND_READ, cred, p); 444 } while (error == NQNFS_EXPIRED); 445 if (error) 446 return (error); 447 if (np->n_lrev != np->n_brev || 448 (np->n_flag & NQNFSNONCACHE) || 449 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) { 450 if (vp->v_type == VDIR) 451 nfs_invaldir(vp); 452 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 453 if (error) 454 return (error); 455 np->n_brev = np->n_lrev; 456 } 457 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) { 458 nfs_invaldir(vp); 459 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 460 if (error) 461 return (error); 462 } 463 } 464 if (np->n_flag & NQNFSNONCACHE) { 465 switch (vp->v_type) { 466 case VREG: 467 return (nfs_readrpc(vp, uio, cred)); 468 case VLNK: 469 return (nfs_readlinkrpc(vp, uio, cred)); 470 case VDIR: 471 break; 472 default: 473 printf(" NQNFSNONCACHE: type %x unexpected\n", 474 vp->v_type); 475 }; 476 } 477 switch (vp->v_type) { 478 case VREG: 479 nfsstats.biocache_reads++; 480 lbn = uio->uio_offset / biosize; 481 on = uio->uio_offset & (biosize - 1); 482 483 /* 484 * Start the read ahead(s), as required. 485 */ 486 if (nfs_numasync > 0 && nmp->nm_readahead > 0) { 487 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 488 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 489 rabn = lbn + 1 + nra; 490 if (!incore(vp, rabn)) { 491 rabp = nfs_getcacheblk(vp, rabn, biosize, p); 492 if (!rabp) 493 return (EINTR); 494 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 495 rabp->b_flags |= B_ASYNC; 496 rabp->b_iocmd = BIO_READ; 497 vfs_busy_pages(rabp, 0); 498 if (nfs_asyncio(rabp, cred, p)) { 499 rabp->b_flags |= B_INVAL; 500 rabp->b_ioflags |= BIO_ERROR; 501 vfs_unbusy_pages(rabp); 502 brelse(rabp); 503 break; 504 } 505 } else { 506 brelse(rabp); 507 } 508 } 509 } 510 } 511 512 /* 513 * Obtain the buffer cache block. Figure out the buffer size 514 * when we are at EOF. If we are modifying the size of the 515 * buffer based on an EOF condition we need to hold 516 * nfs_rslock() through obtaining the buffer to prevent 517 * a potential writer-appender from messing with n_size. 518 * Otherwise we may accidently truncate the buffer and 519 * lose dirty data. 520 * 521 * Note that bcount is *not* DEV_BSIZE aligned. 522 */ 523 524again: 525 bcount = biosize; 526 if ((off_t)lbn * biosize >= np->n_size) { 527 bcount = 0; 528 } else if ((off_t)(lbn + 1) * biosize > np->n_size) { 529 bcount = np->n_size - (off_t)lbn * biosize; 530 } 531 if (bcount != biosize) { 532 switch(nfs_rslock(np, p)) { 533 case ENOLCK: 534 goto again; 535 /* not reached */ 536 case EINTR: 537 case ERESTART: 538 return(EINTR); 539 /* not reached */ 540 default: 541 break; 542 } 543 } 544 545 bp = nfs_getcacheblk(vp, lbn, bcount, p); 546 547 if (bcount != biosize) 548 nfs_rsunlock(np, p); 549 if (!bp) 550 return (EINTR); 551 552 /* 553 * If B_CACHE is not set, we must issue the read. If this 554 * fails, we return an error. 555 */ 556 557 if ((bp->b_flags & B_CACHE) == 0) { 558 bp->b_iocmd = BIO_READ; 559 vfs_busy_pages(bp, 0); 560 error = nfs_doio(bp, cred, p); 561 if (error) { 562 brelse(bp); 563 return (error); 564 } 565 } 566 567 /* 568 * on is the offset into the current bp. Figure out how many 569 * bytes we can copy out of the bp. Note that bcount is 570 * NOT DEV_BSIZE aligned. 571 * 572 * Then figure out how many bytes we can copy into the uio. 573 */ 574 575 n = 0; 576 if (on < bcount) 577 n = min((unsigned)(bcount - on), uio->uio_resid); 578 break; 579 case VLNK: 580 nfsstats.biocache_readlinks++; 581 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p); 582 if (!bp) 583 return (EINTR); 584 if ((bp->b_flags & B_CACHE) == 0) { 585 bp->b_iocmd = BIO_READ; 586 vfs_busy_pages(bp, 0); 587 error = nfs_doio(bp, cred, p); 588 if (error) { 589 bp->b_ioflags |= BIO_ERROR; 590 brelse(bp); 591 return (error); 592 } 593 } 594 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 595 on = 0; 596 break; 597 case VDIR: 598 nfsstats.biocache_readdirs++; 599 if (np->n_direofoffset 600 && uio->uio_offset >= np->n_direofoffset) { 601 return (0); 602 } 603 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 604 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 605 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, p); 606 if (!bp) 607 return (EINTR); 608 if ((bp->b_flags & B_CACHE) == 0) { 609 bp->b_iocmd = BIO_READ; 610 vfs_busy_pages(bp, 0); 611 error = nfs_doio(bp, cred, p); 612 if (error) { 613 brelse(bp); 614 } 615 while (error == NFSERR_BAD_COOKIE) { 616 printf("got bad cookie vp %p bp %p\n", vp, bp); 617 nfs_invaldir(vp); 618 error = nfs_vinvalbuf(vp, 0, cred, p, 1); 619 /* 620 * Yuck! The directory has been modified on the 621 * server. The only way to get the block is by 622 * reading from the beginning to get all the 623 * offset cookies. 624 * 625 * Leave the last bp intact unless there is an error. 626 * Loop back up to the while if the error is another 627 * NFSERR_BAD_COOKIE (double yuch!). 628 */ 629 for (i = 0; i <= lbn && !error; i++) { 630 if (np->n_direofoffset 631 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 632 return (0); 633 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, p); 634 if (!bp) 635 return (EINTR); 636 if ((bp->b_flags & B_CACHE) == 0) { 637 bp->b_iocmd = BIO_READ; 638 vfs_busy_pages(bp, 0); 639 error = nfs_doio(bp, cred, p); 640 /* 641 * no error + B_INVAL == directory EOF, 642 * use the block. 643 */ 644 if (error == 0 && (bp->b_flags & B_INVAL)) 645 break; 646 } 647 /* 648 * An error will throw away the block and the 649 * for loop will break out. If no error and this 650 * is not the block we want, we throw away the 651 * block and go for the next one via the for loop. 652 */ 653 if (error || i < lbn) 654 brelse(bp); 655 } 656 } 657 /* 658 * The above while is repeated if we hit another cookie 659 * error. If we hit an error and it wasn't a cookie error, 660 * we give up. 661 */ 662 if (error) 663 return (error); 664 } 665 666 /* 667 * If not eof and read aheads are enabled, start one. 668 * (You need the current block first, so that you have the 669 * directory offset cookie of the next block.) 670 */ 671 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 672 (bp->b_flags & B_INVAL) == 0 && 673 (np->n_direofoffset == 0 || 674 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 675 !(np->n_flag & NQNFSNONCACHE) && 676 !incore(vp, lbn + 1)) { 677 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, p); 678 if (rabp) { 679 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 680 rabp->b_flags |= B_ASYNC; 681 rabp->b_iocmd = BIO_READ; 682 vfs_busy_pages(rabp, 0); 683 if (nfs_asyncio(rabp, cred, p)) { 684 rabp->b_flags |= B_INVAL; 685 rabp->b_ioflags |= BIO_ERROR; 686 vfs_unbusy_pages(rabp); 687 brelse(rabp); 688 } 689 } else { 690 brelse(rabp); 691 } 692 } 693 } 694 /* 695 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 696 * chopped for the EOF condition, we cannot tell how large 697 * NFS directories are going to be until we hit EOF. So 698 * an NFS directory buffer is *not* chopped to its EOF. Now, 699 * it just so happens that b_resid will effectively chop it 700 * to EOF. *BUT* this information is lost if the buffer goes 701 * away and is reconstituted into a B_CACHE state ( due to 702 * being VMIO ) later. So we keep track of the directory eof 703 * in np->n_direofoffset and chop it off as an extra step 704 * right here. 705 */ 706 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 707 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 708 n = np->n_direofoffset - uio->uio_offset; 709 break; 710 default: 711 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 712 break; 713 }; 714 715 if (n > 0) { 716 error = uiomove(bp->b_data + on, (int)n, uio); 717 } 718 switch (vp->v_type) { 719 case VREG: 720 break; 721 case VLNK: 722 n = 0; 723 break; 724 case VDIR: 725 /* 726 * Invalidate buffer if caching is disabled, forcing a 727 * re-read from the remote later. 728 */ 729 if (np->n_flag & NQNFSNONCACHE) 730 bp->b_flags |= B_INVAL; 731 break; 732 default: 733 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 734 } 735 brelse(bp); 736 } while (error == 0 && uio->uio_resid > 0 && n > 0); 737 return (error); 738} 739 740/* 741 * Vnode op for write using bio 742 */ 743int 744nfs_write(ap) 745 struct vop_write_args /* { 746 struct vnode *a_vp; 747 struct uio *a_uio; 748 int a_ioflag; 749 struct ucred *a_cred; 750 } */ *ap; 751{ 752 int biosize; 753 struct uio *uio = ap->a_uio; 754 struct proc *p = uio->uio_procp; 755 struct vnode *vp = ap->a_vp; 756 struct nfsnode *np = VTONFS(vp); 757 struct ucred *cred = ap->a_cred; 758 int ioflag = ap->a_ioflag; 759 struct buf *bp; 760 struct vattr vattr; 761 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 762 daddr_t lbn; 763 int bcount; 764 int n, on, error = 0, iomode, must_commit; 765 int haverslock = 0; 766 767 GIANT_REQUIRED; 768 769#ifdef DIAGNOSTIC 770 if (uio->uio_rw != UIO_WRITE) 771 panic("nfs_write mode"); 772 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc) 773 panic("nfs_write proc"); 774#endif 775 if (vp->v_type != VREG) 776 return (EIO); 777 if (np->n_flag & NWRITEERR) { 778 np->n_flag &= ~NWRITEERR; 779 return (np->n_error); 780 } 781 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 782 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 783 (void)nfs_fsinfo(nmp, vp, cred, p); 784 785 /* 786 * Synchronously flush pending buffers if we are in synchronous 787 * mode or if we are appending. 788 */ 789 if (ioflag & (IO_APPEND | IO_SYNC)) { 790 if (np->n_flag & NMODIFIED) { 791 np->n_attrstamp = 0; 792 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 793 if (error) 794 return (error); 795 } 796 } 797 798 /* 799 * If IO_APPEND then load uio_offset. We restart here if we cannot 800 * get the append lock. 801 */ 802restart: 803 if (ioflag & IO_APPEND) { 804 np->n_attrstamp = 0; 805 error = VOP_GETATTR(vp, &vattr, cred, p); 806 if (error) 807 return (error); 808 uio->uio_offset = np->n_size; 809 } 810 811 if (uio->uio_offset < 0) 812 return (EINVAL); 813 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 814 return (EFBIG); 815 if (uio->uio_resid == 0) 816 return (0); 817 818 /* 819 * We need to obtain the rslock if we intend to modify np->n_size 820 * in order to guarentee the append point with multiple contending 821 * writers, to guarentee that no other appenders modify n_size 822 * while we are trying to obtain a truncated buffer (i.e. to avoid 823 * accidently truncating data written by another appender due to 824 * the race), and to ensure that the buffer is populated prior to 825 * our extending of the file. We hold rslock through the entire 826 * operation. 827 * 828 * Note that we do not synchronize the case where someone truncates 829 * the file while we are appending to it because attempting to lock 830 * this case may deadlock other parts of the system unexpectedly. 831 */ 832 if ((ioflag & IO_APPEND) || 833 uio->uio_offset + uio->uio_resid > np->n_size) { 834 switch(nfs_rslock(np, p)) { 835 case ENOLCK: 836 goto restart; 837 /* not reached */ 838 case EINTR: 839 case ERESTART: 840 return(EINTR); 841 /* not reached */ 842 default: 843 break; 844 } 845 haverslock = 1; 846 } 847 848 /* 849 * Maybe this should be above the vnode op call, but so long as 850 * file servers have no limits, i don't think it matters 851 */ 852 if (p && uio->uio_offset + uio->uio_resid > 853 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 854 PROC_LOCK(p); 855 psignal(p, SIGXFSZ); 856 PROC_UNLOCK(p); 857 if (haverslock) 858 nfs_rsunlock(np, p); 859 return (EFBIG); 860 } 861 862 biosize = vp->v_mount->mnt_stat.f_iosize; 863 864 do { 865 /* 866 * Check for a valid write lease. 867 */ 868 if ((nmp->nm_flag & NFSMNT_NQNFS) && 869 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 870 do { 871 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 872 } while (error == NQNFS_EXPIRED); 873 if (error) 874 break; 875 if (np->n_lrev != np->n_brev || 876 (np->n_flag & NQNFSNONCACHE)) { 877 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 878 if (error) 879 break; 880 np->n_brev = np->n_lrev; 881 } 882 } 883 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 884 iomode = NFSV3WRITE_FILESYNC; 885 error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit); 886 if (must_commit) 887 nfs_clearcommit(vp->v_mount); 888 break; 889 } 890 nfsstats.biocache_writes++; 891 lbn = uio->uio_offset / biosize; 892 on = uio->uio_offset & (biosize-1); 893 n = min((unsigned)(biosize - on), uio->uio_resid); 894again: 895 /* 896 * Handle direct append and file extension cases, calculate 897 * unaligned buffer size. 898 */ 899 900 if (uio->uio_offset == np->n_size && n) { 901 /* 902 * Get the buffer (in its pre-append state to maintain 903 * B_CACHE if it was previously set). Resize the 904 * nfsnode after we have locked the buffer to prevent 905 * readers from reading garbage. 906 */ 907 bcount = on; 908 bp = nfs_getcacheblk(vp, lbn, bcount, p); 909 910 if (bp != NULL) { 911 long save; 912 913 np->n_size = uio->uio_offset + n; 914 np->n_flag |= NMODIFIED; 915 vnode_pager_setsize(vp, np->n_size); 916 917 save = bp->b_flags & B_CACHE; 918 bcount += n; 919 allocbuf(bp, bcount); 920 bp->b_flags |= save; 921 bp->b_magic = B_MAGIC_NFS; 922 bp->b_op = &buf_ops_nfs; 923 } 924 } else { 925 /* 926 * Obtain the locked cache block first, and then 927 * adjust the file's size as appropriate. 928 */ 929 bcount = on + n; 930 if ((off_t)lbn * biosize + bcount < np->n_size) { 931 if ((off_t)(lbn + 1) * biosize < np->n_size) 932 bcount = biosize; 933 else 934 bcount = np->n_size - (off_t)lbn * biosize; 935 } 936 937 bp = nfs_getcacheblk(vp, lbn, bcount, p); 938 939 if (uio->uio_offset + n > np->n_size) { 940 np->n_size = uio->uio_offset + n; 941 np->n_flag |= NMODIFIED; 942 vnode_pager_setsize(vp, np->n_size); 943 } 944 } 945 946 if (!bp) { 947 error = EINTR; 948 break; 949 } 950 951 /* 952 * Issue a READ if B_CACHE is not set. In special-append 953 * mode, B_CACHE is based on the buffer prior to the write 954 * op and is typically set, avoiding the read. If a read 955 * is required in special append mode, the server will 956 * probably send us a short-read since we extended the file 957 * on our end, resulting in b_resid == 0 and, thusly, 958 * B_CACHE getting set. 959 * 960 * We can also avoid issuing the read if the write covers 961 * the entire buffer. We have to make sure the buffer state 962 * is reasonable in this case since we will not be initiating 963 * I/O. See the comments in kern/vfs_bio.c's getblk() for 964 * more information. 965 * 966 * B_CACHE may also be set due to the buffer being cached 967 * normally. 968 */ 969 970 if (on == 0 && n == bcount) { 971 bp->b_flags |= B_CACHE; 972 bp->b_flags &= ~B_INVAL; 973 bp->b_ioflags &= ~BIO_ERROR; 974 } 975 976 if ((bp->b_flags & B_CACHE) == 0) { 977 bp->b_iocmd = BIO_READ; 978 vfs_busy_pages(bp, 0); 979 error = nfs_doio(bp, cred, p); 980 if (error) { 981 brelse(bp); 982 break; 983 } 984 } 985 if (!bp) { 986 error = EINTR; 987 break; 988 } 989 if (bp->b_wcred == NOCRED) { 990 crhold(cred); 991 bp->b_wcred = cred; 992 } 993 np->n_flag |= NMODIFIED; 994 995 /* 996 * If dirtyend exceeds file size, chop it down. This should 997 * not normally occur but there is an append race where it 998 * might occur XXX, so we log it. 999 * 1000 * If the chopping creates a reverse-indexed or degenerate 1001 * situation with dirtyoff/end, we 0 both of them. 1002 */ 1003 1004 if (bp->b_dirtyend > bcount) { 1005 printf("NFS append race @%lx:%d\n", 1006 (long)bp->b_blkno * DEV_BSIZE, 1007 bp->b_dirtyend - bcount); 1008 bp->b_dirtyend = bcount; 1009 } 1010 1011 if (bp->b_dirtyoff >= bp->b_dirtyend) 1012 bp->b_dirtyoff = bp->b_dirtyend = 0; 1013 1014 /* 1015 * If the new write will leave a contiguous dirty 1016 * area, just update the b_dirtyoff and b_dirtyend, 1017 * otherwise force a write rpc of the old dirty area. 1018 * 1019 * While it is possible to merge discontiguous writes due to 1020 * our having a B_CACHE buffer ( and thus valid read data 1021 * for the hole), we don't because it could lead to 1022 * significant cache coherency problems with multiple clients, 1023 * especially if locking is implemented later on. 1024 * 1025 * as an optimization we could theoretically maintain 1026 * a linked list of discontinuous areas, but we would still 1027 * have to commit them separately so there isn't much 1028 * advantage to it except perhaps a bit of asynchronization. 1029 */ 1030 1031 if (bp->b_dirtyend > 0 && 1032 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1033 if (BUF_WRITE(bp) == EINTR) 1034 return (EINTR); 1035 goto again; 1036 } 1037 1038 /* 1039 * Check for valid write lease and get one as required. 1040 * In case getblk() and/or bwrite() delayed us. 1041 */ 1042 if ((nmp->nm_flag & NFSMNT_NQNFS) && 1043 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 1044 do { 1045 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 1046 } while (error == NQNFS_EXPIRED); 1047 if (error) { 1048 brelse(bp); 1049 break; 1050 } 1051 if (np->n_lrev != np->n_brev || 1052 (np->n_flag & NQNFSNONCACHE)) { 1053 brelse(bp); 1054 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 1055 if (error) 1056 break; 1057 np->n_brev = np->n_lrev; 1058 goto again; 1059 } 1060 } 1061 1062 error = uiomove((char *)bp->b_data + on, n, uio); 1063 1064 /* 1065 * Since this block is being modified, it must be written 1066 * again and not just committed. Since write clustering does 1067 * not work for the stage 1 data write, only the stage 2 1068 * commit rpc, we have to clear B_CLUSTEROK as well. 1069 */ 1070 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1071 1072 if (error) { 1073 bp->b_ioflags |= BIO_ERROR; 1074 brelse(bp); 1075 break; 1076 } 1077 1078 /* 1079 * Only update dirtyoff/dirtyend if not a degenerate 1080 * condition. 1081 */ 1082 if (n) { 1083 if (bp->b_dirtyend > 0) { 1084 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1085 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1086 } else { 1087 bp->b_dirtyoff = on; 1088 bp->b_dirtyend = on + n; 1089 } 1090 vfs_bio_set_validclean(bp, on, n); 1091 } 1092 1093 /* 1094 * If the lease is non-cachable or IO_SYNC do bwrite(). 1095 * 1096 * IO_INVAL appears to be unused. The idea appears to be 1097 * to turn off caching in this case. Very odd. XXX 1098 */ 1099 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 1100 if (ioflag & IO_INVAL) 1101 bp->b_flags |= B_NOCACHE; 1102 error = BUF_WRITE(bp); 1103 if (error) 1104 break; 1105 if (np->n_flag & NQNFSNONCACHE) { 1106 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 1107 if (error) 1108 break; 1109 } 1110 } else if ((n + on) == biosize && 1111 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 1112 bp->b_flags |= B_ASYNC; 1113 (void)nfs_writebp(bp, 0, 0); 1114 } else { 1115 bdwrite(bp); 1116 } 1117 } while (uio->uio_resid > 0 && n > 0); 1118 1119 if (haverslock) 1120 nfs_rsunlock(np, p); 1121 1122 return (error); 1123} 1124 1125/* 1126 * Get an nfs cache block. 1127 * 1128 * Allocate a new one if the block isn't currently in the cache 1129 * and return the block marked busy. If the calling process is 1130 * interrupted by a signal for an interruptible mount point, return 1131 * NULL. 1132 * 1133 * The caller must carefully deal with the possible B_INVAL state of 1134 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1135 * indirectly), so synchronous reads can be issued without worrying about 1136 * the B_INVAL state. We have to be a little more careful when dealing 1137 * with writes (see comments in nfs_write()) when extending a file past 1138 * its EOF. 1139 */ 1140static struct buf * 1141nfs_getcacheblk(vp, bn, size, p) 1142 struct vnode *vp; 1143 daddr_t bn; 1144 int size; 1145 struct proc *p; 1146{ 1147 register struct buf *bp; 1148 struct mount *mp; 1149 struct nfsmount *nmp; 1150 1151 mp = vp->v_mount; 1152 nmp = VFSTONFS(mp); 1153 1154 if (nmp->nm_flag & NFSMNT_INT) { 1155 bp = getblk(vp, bn, size, PCATCH, 0); 1156 while (bp == (struct buf *)0) { 1157 if (nfs_sigintr(nmp, (struct nfsreq *)0, p)) 1158 return ((struct buf *)0); 1159 bp = getblk(vp, bn, size, 0, 2 * hz); 1160 } 1161 } else { 1162 bp = getblk(vp, bn, size, 0, 0); 1163 } 1164 1165 if (vp->v_type == VREG) { 1166 int biosize; 1167 1168 biosize = mp->mnt_stat.f_iosize; 1169 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1170 } 1171 return (bp); 1172} 1173 1174/* 1175 * Flush and invalidate all dirty buffers. If another process is already 1176 * doing the flush, just wait for completion. 1177 */ 1178int 1179nfs_vinvalbuf(vp, flags, cred, p, intrflg) 1180 struct vnode *vp; 1181 int flags; 1182 struct ucred *cred; 1183 struct proc *p; 1184 int intrflg; 1185{ 1186 register struct nfsnode *np = VTONFS(vp); 1187 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1188 int error = 0, slpflag, slptimeo; 1189 1190 if (vp->v_flag & VXLOCK) { 1191 return (0); 1192 } 1193 1194 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1195 intrflg = 0; 1196 if (intrflg) { 1197 slpflag = PCATCH; 1198 slptimeo = 2 * hz; 1199 } else { 1200 slpflag = 0; 1201 slptimeo = 0; 1202 } 1203 /* 1204 * First wait for any other process doing a flush to complete. 1205 */ 1206 while (np->n_flag & NFLUSHINPROG) { 1207 np->n_flag |= NFLUSHWANT; 1208 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval", 1209 slptimeo); 1210 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) 1211 return (EINTR); 1212 } 1213 1214 /* 1215 * Now, flush as required. 1216 */ 1217 np->n_flag |= NFLUSHINPROG; 1218 error = vinvalbuf(vp, flags, cred, p, slpflag, 0); 1219 while (error) { 1220 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) { 1221 np->n_flag &= ~NFLUSHINPROG; 1222 if (np->n_flag & NFLUSHWANT) { 1223 np->n_flag &= ~NFLUSHWANT; 1224 wakeup((caddr_t)&np->n_flag); 1225 } 1226 return (EINTR); 1227 } 1228 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo); 1229 } 1230 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 1231 if (np->n_flag & NFLUSHWANT) { 1232 np->n_flag &= ~NFLUSHWANT; 1233 wakeup((caddr_t)&np->n_flag); 1234 } 1235 return (0); 1236} 1237 1238/* 1239 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1240 * This is mainly to avoid queueing async I/O requests when the nfsiods 1241 * are all hung on a dead server. 1242 * 1243 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1244 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1245 */ 1246int 1247nfs_asyncio(bp, cred, procp) 1248 register struct buf *bp; 1249 struct ucred *cred; 1250 struct proc *procp; 1251{ 1252 struct nfsmount *nmp; 1253 int i; 1254 int gotiod; 1255 int slpflag = 0; 1256 int slptimeo = 0; 1257 int error; 1258 1259 /* 1260 * If no async daemons then return EIO to force caller to run the rpc 1261 * synchronously. 1262 */ 1263 if (nfs_numasync == 0) 1264 return (EIO); 1265 1266 nmp = VFSTONFS(bp->b_vp->v_mount); 1267 1268 /* 1269 * Commits are usually short and sweet so lets save some cpu and 1270 * leave the async daemons for more important rpc's (such as reads 1271 * and writes). 1272 */ 1273 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1274 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1275 return(EIO); 1276 } 1277 1278again: 1279 if (nmp->nm_flag & NFSMNT_INT) 1280 slpflag = PCATCH; 1281 gotiod = FALSE; 1282 1283 /* 1284 * Find a free iod to process this request. 1285 */ 1286 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 1287 if (nfs_iodwant[i]) { 1288 /* 1289 * Found one, so wake it up and tell it which 1290 * mount to process. 1291 */ 1292 NFS_DPF(ASYNCIO, 1293 ("nfs_asyncio: waking iod %d for mount %p\n", 1294 i, nmp)); 1295 nfs_iodwant[i] = (struct proc *)0; 1296 nfs_iodmount[i] = nmp; 1297 nmp->nm_bufqiods++; 1298 wakeup((caddr_t)&nfs_iodwant[i]); 1299 gotiod = TRUE; 1300 break; 1301 } 1302 1303 /* 1304 * If none are free, we may already have an iod working on this mount 1305 * point. If so, it will process our request. 1306 */ 1307 if (!gotiod) { 1308 if (nmp->nm_bufqiods > 0) { 1309 NFS_DPF(ASYNCIO, 1310 ("nfs_asyncio: %d iods are already processing mount %p\n", 1311 nmp->nm_bufqiods, nmp)); 1312 gotiod = TRUE; 1313 } 1314 } 1315 1316 /* 1317 * If we have an iod which can process the request, then queue 1318 * the buffer. 1319 */ 1320 if (gotiod) { 1321 /* 1322 * Ensure that the queue never grows too large. We still want 1323 * to asynchronize so we block rather then return EIO. 1324 */ 1325 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1326 NFS_DPF(ASYNCIO, 1327 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1328 nmp->nm_bufqwant = TRUE; 1329 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO, 1330 "nfsaio", slptimeo); 1331 if (error) { 1332 if (nfs_sigintr(nmp, NULL, procp)) 1333 return (EINTR); 1334 if (slpflag == PCATCH) { 1335 slpflag = 0; 1336 slptimeo = 2 * hz; 1337 } 1338 } 1339 /* 1340 * We might have lost our iod while sleeping, 1341 * so check and loop if nescessary. 1342 */ 1343 if (nmp->nm_bufqiods == 0) { 1344 NFS_DPF(ASYNCIO, 1345 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1346 goto again; 1347 } 1348 } 1349 1350 if (bp->b_iocmd == BIO_READ) { 1351 if (bp->b_rcred == NOCRED && cred != NOCRED) { 1352 crhold(cred); 1353 bp->b_rcred = cred; 1354 } 1355 } else { 1356 bp->b_flags |= B_WRITEINPROG; 1357 if (bp->b_wcred == NOCRED && cred != NOCRED) { 1358 crhold(cred); 1359 bp->b_wcred = cred; 1360 } 1361 } 1362 1363 BUF_KERNPROC(bp); 1364 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1365 nmp->nm_bufqlen++; 1366 return (0); 1367 } 1368 1369 /* 1370 * All the iods are busy on other mounts, so return EIO to 1371 * force the caller to process the i/o synchronously. 1372 */ 1373 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1374 return (EIO); 1375} 1376 1377/* 1378 * Do an I/O operation to/from a cache block. This may be called 1379 * synchronously or from an nfsiod. 1380 */ 1381int 1382nfs_doio(bp, cr, p) 1383 struct buf *bp; 1384 struct ucred *cr; 1385 struct proc *p; 1386{ 1387 struct uio *uiop; 1388 struct vnode *vp; 1389 struct nfsnode *np; 1390 struct nfsmount *nmp; 1391 int error = 0, iomode, must_commit = 0; 1392 struct uio uio; 1393 struct iovec io; 1394 1395 vp = bp->b_vp; 1396 np = VTONFS(vp); 1397 nmp = VFSTONFS(vp->v_mount); 1398 uiop = &uio; 1399 uiop->uio_iov = &io; 1400 uiop->uio_iovcnt = 1; 1401 uiop->uio_segflg = UIO_SYSSPACE; 1402 uiop->uio_procp = p; 1403 1404 /* 1405 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1406 * do this here so we do not have to do it in all the code that 1407 * calls us. 1408 */ 1409 bp->b_flags &= ~B_INVAL; 1410 bp->b_ioflags &= ~BIO_ERROR; 1411 1412 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1413 1414 /* 1415 * Historically, paging was done with physio, but no more. 1416 */ 1417 if (bp->b_flags & B_PHYS) { 1418 /* 1419 * ...though reading /dev/drum still gets us here. 1420 */ 1421 io.iov_len = uiop->uio_resid = bp->b_bcount; 1422 /* mapping was done by vmapbuf() */ 1423 io.iov_base = bp->b_data; 1424 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1425 if (bp->b_iocmd == BIO_READ) { 1426 uiop->uio_rw = UIO_READ; 1427 nfsstats.read_physios++; 1428 error = nfs_readrpc(vp, uiop, cr); 1429 } else { 1430 int com; 1431 1432 iomode = NFSV3WRITE_DATASYNC; 1433 uiop->uio_rw = UIO_WRITE; 1434 nfsstats.write_physios++; 1435 error = nfs_writerpc(vp, uiop, cr, &iomode, &com); 1436 } 1437 if (error) { 1438 bp->b_ioflags |= BIO_ERROR; 1439 bp->b_error = error; 1440 } 1441 } else if (bp->b_iocmd == BIO_READ) { 1442 io.iov_len = uiop->uio_resid = bp->b_bcount; 1443 io.iov_base = bp->b_data; 1444 uiop->uio_rw = UIO_READ; 1445 switch (vp->v_type) { 1446 case VREG: 1447 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1448 nfsstats.read_bios++; 1449 error = nfs_readrpc(vp, uiop, cr); 1450 if (!error) { 1451 if (uiop->uio_resid) { 1452 /* 1453 * If we had a short read with no error, we must have 1454 * hit a file hole. We should zero-fill the remainder. 1455 * This can also occur if the server hits the file EOF. 1456 * 1457 * Holes used to be able to occur due to pending 1458 * writes, but that is not possible any longer. 1459 */ 1460 int nread = bp->b_bcount - uiop->uio_resid; 1461 int left = bp->b_bcount - nread; 1462 1463 if (left > 0) 1464 bzero((char *)bp->b_data + nread, left); 1465 uiop->uio_resid = 0; 1466 } 1467 } 1468 if (p && (vp->v_flag & VTEXT) && 1469 (((nmp->nm_flag & NFSMNT_NQNFS) && 1470 NQNFS_CKINVALID(vp, np, ND_READ) && 1471 np->n_lrev != np->n_brev) || 1472 (!(nmp->nm_flag & NFSMNT_NQNFS) && 1473 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) { 1474 uprintf("Process killed due to text file modification\n"); 1475 PROC_LOCK(p); 1476 psignal(p, SIGKILL); 1477 _PHOLD(p); 1478 PROC_UNLOCK(p); 1479 } 1480 break; 1481 case VLNK: 1482 uiop->uio_offset = (off_t)0; 1483 nfsstats.readlink_bios++; 1484 error = nfs_readlinkrpc(vp, uiop, cr); 1485 break; 1486 case VDIR: 1487 nfsstats.readdir_bios++; 1488 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1489 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1490 error = nfs_readdirplusrpc(vp, uiop, cr); 1491 if (error == NFSERR_NOTSUPP) 1492 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1493 } 1494 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1495 error = nfs_readdirrpc(vp, uiop, cr); 1496 /* 1497 * end-of-directory sets B_INVAL but does not generate an 1498 * error. 1499 */ 1500 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1501 bp->b_flags |= B_INVAL; 1502 break; 1503 default: 1504 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1505 break; 1506 }; 1507 if (error) { 1508 bp->b_ioflags |= BIO_ERROR; 1509 bp->b_error = error; 1510 } 1511 } else { 1512 /* 1513 * If we only need to commit, try to commit 1514 */ 1515 if (bp->b_flags & B_NEEDCOMMIT) { 1516 int retv; 1517 off_t off; 1518 1519 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1520 bp->b_flags |= B_WRITEINPROG; 1521 retv = nfs_commit( 1522 bp->b_vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1523 bp->b_wcred, p); 1524 bp->b_flags &= ~B_WRITEINPROG; 1525 if (retv == 0) { 1526 bp->b_dirtyoff = bp->b_dirtyend = 0; 1527 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1528 bp->b_resid = 0; 1529 bufdone(bp); 1530 return (0); 1531 } 1532 if (retv == NFSERR_STALEWRITEVERF) { 1533 nfs_clearcommit(bp->b_vp->v_mount); 1534 } 1535 } 1536 1537 /* 1538 * Setup for actual write 1539 */ 1540 1541 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1542 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1543 1544 if (bp->b_dirtyend > bp->b_dirtyoff) { 1545 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1546 - bp->b_dirtyoff; 1547 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1548 + bp->b_dirtyoff; 1549 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1550 uiop->uio_rw = UIO_WRITE; 1551 nfsstats.write_bios++; 1552 1553 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1554 iomode = NFSV3WRITE_UNSTABLE; 1555 else 1556 iomode = NFSV3WRITE_FILESYNC; 1557 1558 bp->b_flags |= B_WRITEINPROG; 1559 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit); 1560 1561 /* 1562 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1563 * to cluster the buffers needing commit. This will allow 1564 * the system to submit a single commit rpc for the whole 1565 * cluster. We can do this even if the buffer is not 100% 1566 * dirty (relative to the NFS blocksize), so we optimize the 1567 * append-to-file-case. 1568 * 1569 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1570 * cleared because write clustering only works for commit 1571 * rpc's, not for the data portion of the write). 1572 */ 1573 1574 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1575 bp->b_flags |= B_NEEDCOMMIT; 1576 if (bp->b_dirtyoff == 0 1577 && bp->b_dirtyend == bp->b_bcount) 1578 bp->b_flags |= B_CLUSTEROK; 1579 } else { 1580 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1581 } 1582 bp->b_flags &= ~B_WRITEINPROG; 1583 1584 /* 1585 * For an interrupted write, the buffer is still valid 1586 * and the write hasn't been pushed to the server yet, 1587 * so we can't set BIO_ERROR and report the interruption 1588 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1589 * is not relevant, so the rpc attempt is essentially 1590 * a noop. For the case of a V3 write rpc not being 1591 * committed to stable storage, the block is still 1592 * dirty and requires either a commit rpc or another 1593 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1594 * the block is reused. This is indicated by setting 1595 * the B_DELWRI and B_NEEDCOMMIT flags. 1596 * 1597 * If the buffer is marked B_PAGING, it does not reside on 1598 * the vp's paging queues so we cannot call bdirty(). The 1599 * bp in this case is not an NFS cache block so we should 1600 * be safe. XXX 1601 */ 1602 if (error == EINTR 1603 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1604 int s; 1605 1606 s = splbio(); 1607 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1608 if ((bp->b_flags & B_PAGING) == 0) { 1609 bdirty(bp); 1610 bp->b_flags &= ~B_DONE; 1611 } 1612 if (error && (bp->b_flags & B_ASYNC) == 0) 1613 bp->b_flags |= B_EINTR; 1614 splx(s); 1615 } else { 1616 if (error) { 1617 bp->b_ioflags |= BIO_ERROR; 1618 bp->b_error = np->n_error = error; 1619 np->n_flag |= NWRITEERR; 1620 } 1621 bp->b_dirtyoff = bp->b_dirtyend = 0; 1622 } 1623 } else { 1624 bp->b_resid = 0; 1625 bufdone(bp); 1626 return (0); 1627 } 1628 } 1629 bp->b_resid = uiop->uio_resid; 1630 if (must_commit) 1631 nfs_clearcommit(vp->v_mount); 1632 bufdone(bp); 1633 return (error); 1634} 1635