nfs_bio.c revision 73929
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 73929 2001-03-07 03:37:06Z jhb $ 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 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_ASYNC; 470 rabp->b_iocmd = BIO_READ; 471 vfs_busy_pages(rabp, 0); 472 if (nfs_asyncio(rabp, cred, p)) { 473 rabp->b_flags |= B_INVAL; 474 rabp->b_ioflags |= BIO_ERROR; 475 vfs_unbusy_pages(rabp); 476 brelse(rabp); 477 break; 478 } 479 } else { 480 brelse(rabp); 481 } 482 } 483 } 484 } 485 486 /* 487 * Obtain the buffer cache block. Figure out the buffer size 488 * when we are at EOF. If we are modifying the size of the 489 * buffer based on an EOF condition we need to hold 490 * nfs_rslock() through obtaining the buffer to prevent 491 * a potential writer-appender from messing with n_size. 492 * Otherwise we may accidently truncate the buffer and 493 * lose dirty data. 494 * 495 * Note that bcount is *not* DEV_BSIZE aligned. 496 */ 497 498again: 499 bcount = biosize; 500 if ((off_t)lbn * biosize >= np->n_size) { 501 bcount = 0; 502 } else if ((off_t)(lbn + 1) * biosize > np->n_size) { 503 bcount = np->n_size - (off_t)lbn * biosize; 504 } 505 if (bcount != biosize) { 506 switch(nfs_rslock(np, p)) { 507 case ENOLCK: 508 goto again; 509 /* not reached */ 510 case EINTR: 511 case ERESTART: 512 return(EINTR); 513 /* not reached */ 514 default: 515 break; 516 } 517 } 518 519 bp = nfs_getcacheblk(vp, lbn, bcount, p); 520 521 if (bcount != biosize) 522 nfs_rsunlock(np, p); 523 if (!bp) 524 return (EINTR); 525 526 /* 527 * If B_CACHE is not set, we must issue the read. If this 528 * fails, we return an error. 529 */ 530 531 if ((bp->b_flags & B_CACHE) == 0) { 532 bp->b_iocmd = BIO_READ; 533 vfs_busy_pages(bp, 0); 534 error = nfs_doio(bp, cred, p); 535 if (error) { 536 brelse(bp); 537 return (error); 538 } 539 } 540 541 /* 542 * on is the offset into the current bp. Figure out how many 543 * bytes we can copy out of the bp. Note that bcount is 544 * NOT DEV_BSIZE aligned. 545 * 546 * Then figure out how many bytes we can copy into the uio. 547 */ 548 549 n = 0; 550 if (on < bcount) 551 n = min((unsigned)(bcount - on), uio->uio_resid); 552 break; 553 case VLNK: 554 nfsstats.biocache_readlinks++; 555 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p); 556 if (!bp) 557 return (EINTR); 558 if ((bp->b_flags & B_CACHE) == 0) { 559 bp->b_iocmd = BIO_READ; 560 vfs_busy_pages(bp, 0); 561 error = nfs_doio(bp, cred, p); 562 if (error) { 563 bp->b_ioflags |= BIO_ERROR; 564 brelse(bp); 565 return (error); 566 } 567 } 568 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 569 on = 0; 570 break; 571 case VDIR: 572 nfsstats.biocache_readdirs++; 573 if (np->n_direofoffset 574 && uio->uio_offset >= np->n_direofoffset) { 575 return (0); 576 } 577 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 578 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 579 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, p); 580 if (!bp) 581 return (EINTR); 582 if ((bp->b_flags & B_CACHE) == 0) { 583 bp->b_iocmd = BIO_READ; 584 vfs_busy_pages(bp, 0); 585 error = nfs_doio(bp, cred, p); 586 if (error) { 587 brelse(bp); 588 } 589 while (error == NFSERR_BAD_COOKIE) { 590 printf("got bad cookie vp %p bp %p\n", vp, bp); 591 nfs_invaldir(vp); 592 error = nfs_vinvalbuf(vp, 0, cred, p, 1); 593 /* 594 * Yuck! The directory has been modified on the 595 * server. The only way to get the block is by 596 * reading from the beginning to get all the 597 * offset cookies. 598 * 599 * Leave the last bp intact unless there is an error. 600 * Loop back up to the while if the error is another 601 * NFSERR_BAD_COOKIE (double yuch!). 602 */ 603 for (i = 0; i <= lbn && !error; i++) { 604 if (np->n_direofoffset 605 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 606 return (0); 607 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, p); 608 if (!bp) 609 return (EINTR); 610 if ((bp->b_flags & B_CACHE) == 0) { 611 bp->b_iocmd = BIO_READ; 612 vfs_busy_pages(bp, 0); 613 error = nfs_doio(bp, cred, p); 614 /* 615 * no error + B_INVAL == directory EOF, 616 * use the block. 617 */ 618 if (error == 0 && (bp->b_flags & B_INVAL)) 619 break; 620 } 621 /* 622 * An error will throw away the block and the 623 * for loop will break out. If no error and this 624 * is not the block we want, we throw away the 625 * block and go for the next one via the for loop. 626 */ 627 if (error || i < lbn) 628 brelse(bp); 629 } 630 } 631 /* 632 * The above while is repeated if we hit another cookie 633 * error. If we hit an error and it wasn't a cookie error, 634 * we give up. 635 */ 636 if (error) 637 return (error); 638 } 639 640 /* 641 * If not eof and read aheads are enabled, start one. 642 * (You need the current block first, so that you have the 643 * directory offset cookie of the next block.) 644 */ 645 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 646 (bp->b_flags & B_INVAL) == 0 && 647 (np->n_direofoffset == 0 || 648 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 649 !(np->n_flag & NQNFSNONCACHE) && 650 !incore(vp, lbn + 1)) { 651 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, p); 652 if (rabp) { 653 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 654 rabp->b_flags |= B_ASYNC; 655 rabp->b_iocmd = BIO_READ; 656 vfs_busy_pages(rabp, 0); 657 if (nfs_asyncio(rabp, cred, p)) { 658 rabp->b_flags |= B_INVAL; 659 rabp->b_ioflags |= BIO_ERROR; 660 vfs_unbusy_pages(rabp); 661 brelse(rabp); 662 } 663 } else { 664 brelse(rabp); 665 } 666 } 667 } 668 /* 669 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 670 * chopped for the EOF condition, we cannot tell how large 671 * NFS directories are going to be until we hit EOF. So 672 * an NFS directory buffer is *not* chopped to its EOF. Now, 673 * it just so happens that b_resid will effectively chop it 674 * to EOF. *BUT* this information is lost if the buffer goes 675 * away and is reconstituted into a B_CACHE state ( due to 676 * being VMIO ) later. So we keep track of the directory eof 677 * in np->n_direofoffset and chop it off as an extra step 678 * right here. 679 */ 680 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 681 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 682 n = np->n_direofoffset - uio->uio_offset; 683 break; 684 default: 685 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 686 break; 687 }; 688 689 if (n > 0) { 690 error = uiomove(bp->b_data + on, (int)n, uio); 691 } 692 switch (vp->v_type) { 693 case VREG: 694 break; 695 case VLNK: 696 n = 0; 697 break; 698 case VDIR: 699 /* 700 * Invalidate buffer if caching is disabled, forcing a 701 * re-read from the remote later. 702 */ 703 if (np->n_flag & NQNFSNONCACHE) 704 bp->b_flags |= B_INVAL; 705 break; 706 default: 707 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 708 } 709 brelse(bp); 710 } while (error == 0 && uio->uio_resid > 0 && n > 0); 711 return (error); 712} 713 714/* 715 * Vnode op for write using bio 716 */ 717int 718nfs_write(ap) 719 struct vop_write_args /* { 720 struct vnode *a_vp; 721 struct uio *a_uio; 722 int a_ioflag; 723 struct ucred *a_cred; 724 } */ *ap; 725{ 726 int biosize; 727 struct uio *uio = ap->a_uio; 728 struct proc *p = uio->uio_procp; 729 struct vnode *vp = ap->a_vp; 730 struct nfsnode *np = VTONFS(vp); 731 struct ucred *cred = ap->a_cred; 732 int ioflag = ap->a_ioflag; 733 struct buf *bp; 734 struct vattr vattr; 735 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 736 daddr_t lbn; 737 int bcount; 738 int n, on, error = 0, iomode, must_commit; 739 int haverslock = 0; 740 741#ifdef DIAGNOSTIC 742 if (uio->uio_rw != UIO_WRITE) 743 panic("nfs_write mode"); 744 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc) 745 panic("nfs_write proc"); 746#endif 747 if (vp->v_type != VREG) 748 return (EIO); 749 if (np->n_flag & NWRITEERR) { 750 np->n_flag &= ~NWRITEERR; 751 return (np->n_error); 752 } 753 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 754 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 755 (void)nfs_fsinfo(nmp, vp, cred, p); 756 757 /* 758 * Synchronously flush pending buffers if we are in synchronous 759 * mode or if we are appending. 760 */ 761 if (ioflag & (IO_APPEND | IO_SYNC)) { 762 if (np->n_flag & NMODIFIED) { 763 np->n_attrstamp = 0; 764 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 765 if (error) 766 return (error); 767 } 768 } 769 770 /* 771 * If IO_APPEND then load uio_offset. We restart here if we cannot 772 * get the append lock. 773 */ 774restart: 775 if (ioflag & IO_APPEND) { 776 np->n_attrstamp = 0; 777 error = VOP_GETATTR(vp, &vattr, cred, p); 778 if (error) 779 return (error); 780 uio->uio_offset = np->n_size; 781 } 782 783 if (uio->uio_offset < 0) 784 return (EINVAL); 785 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 786 return (EFBIG); 787 if (uio->uio_resid == 0) 788 return (0); 789 790 /* 791 * We need to obtain the rslock if we intend to modify np->n_size 792 * in order to guarentee the append point with multiple contending 793 * writers, to guarentee that no other appenders modify n_size 794 * while we are trying to obtain a truncated buffer (i.e. to avoid 795 * accidently truncating data written by another appender due to 796 * the race), and to ensure that the buffer is populated prior to 797 * our extending of the file. We hold rslock through the entire 798 * operation. 799 * 800 * Note that we do not synchronize the case where someone truncates 801 * the file while we are appending to it because attempting to lock 802 * this case may deadlock other parts of the system unexpectedly. 803 */ 804 if ((ioflag & IO_APPEND) || 805 uio->uio_offset + uio->uio_resid > np->n_size) { 806 switch(nfs_rslock(np, p)) { 807 case ENOLCK: 808 goto restart; 809 /* not reached */ 810 case EINTR: 811 case ERESTART: 812 return(EINTR); 813 /* not reached */ 814 default: 815 break; 816 } 817 haverslock = 1; 818 } 819 820 /* 821 * Maybe this should be above the vnode op call, but so long as 822 * file servers have no limits, i don't think it matters 823 */ 824 if (p && uio->uio_offset + uio->uio_resid > 825 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 826 PROC_LOCK(p); 827 psignal(p, SIGXFSZ); 828 PROC_UNLOCK(p); 829 if (haverslock) 830 nfs_rsunlock(np, p); 831 return (EFBIG); 832 } 833 834 biosize = vp->v_mount->mnt_stat.f_iosize; 835 836 do { 837 /* 838 * Check for a valid write lease. 839 */ 840 if ((nmp->nm_flag & NFSMNT_NQNFS) && 841 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 842 do { 843 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 844 } while (error == NQNFS_EXPIRED); 845 if (error) 846 break; 847 if (np->n_lrev != np->n_brev || 848 (np->n_flag & NQNFSNONCACHE)) { 849 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 850 if (error) 851 break; 852 np->n_brev = np->n_lrev; 853 } 854 } 855 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 856 iomode = NFSV3WRITE_FILESYNC; 857 error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit); 858 if (must_commit) 859 nfs_clearcommit(vp->v_mount); 860 break; 861 } 862 nfsstats.biocache_writes++; 863 lbn = uio->uio_offset / biosize; 864 on = uio->uio_offset & (biosize-1); 865 n = min((unsigned)(biosize - on), uio->uio_resid); 866again: 867 /* 868 * Handle direct append and file extension cases, calculate 869 * unaligned buffer size. 870 */ 871 872 if (uio->uio_offset == np->n_size && n) { 873 /* 874 * Get the buffer (in its pre-append state to maintain 875 * B_CACHE if it was previously set). Resize the 876 * nfsnode after we have locked the buffer to prevent 877 * readers from reading garbage. 878 */ 879 bcount = on; 880 bp = nfs_getcacheblk(vp, lbn, bcount, p); 881 882 if (bp != NULL) { 883 long save; 884 885 np->n_size = uio->uio_offset + n; 886 np->n_flag |= NMODIFIED; 887 vnode_pager_setsize(vp, np->n_size); 888 889 save = bp->b_flags & B_CACHE; 890 bcount += n; 891 allocbuf(bp, bcount); 892 bp->b_flags |= save; 893 } 894 } else { 895 /* 896 * Obtain the locked cache block first, and then 897 * adjust the file's size as appropriate. 898 */ 899 bcount = on + n; 900 if ((off_t)lbn * biosize + bcount < np->n_size) { 901 if ((off_t)(lbn + 1) * biosize < np->n_size) 902 bcount = biosize; 903 else 904 bcount = np->n_size - (off_t)lbn * biosize; 905 } 906 907 bp = nfs_getcacheblk(vp, lbn, bcount, p); 908 909 if (uio->uio_offset + n > np->n_size) { 910 np->n_size = uio->uio_offset + n; 911 np->n_flag |= NMODIFIED; 912 vnode_pager_setsize(vp, np->n_size); 913 } 914 } 915 916 if (!bp) { 917 error = EINTR; 918 break; 919 } 920 921 /* 922 * Issue a READ if B_CACHE is not set. In special-append 923 * mode, B_CACHE is based on the buffer prior to the write 924 * op and is typically set, avoiding the read. If a read 925 * is required in special append mode, the server will 926 * probably send us a short-read since we extended the file 927 * on our end, resulting in b_resid == 0 and, thusly, 928 * B_CACHE getting set. 929 * 930 * We can also avoid issuing the read if the write covers 931 * the entire buffer. We have to make sure the buffer state 932 * is reasonable in this case since we will not be initiating 933 * I/O. See the comments in kern/vfs_bio.c's getblk() for 934 * more information. 935 * 936 * B_CACHE may also be set due to the buffer being cached 937 * normally. 938 */ 939 940 if (on == 0 && n == bcount) { 941 bp->b_flags |= B_CACHE; 942 bp->b_flags &= ~B_INVAL; 943 bp->b_ioflags &= ~BIO_ERROR; 944 } 945 946 if ((bp->b_flags & B_CACHE) == 0) { 947 bp->b_iocmd = BIO_READ; 948 vfs_busy_pages(bp, 0); 949 error = nfs_doio(bp, cred, p); 950 if (error) { 951 brelse(bp); 952 break; 953 } 954 } 955 if (!bp) { 956 error = EINTR; 957 break; 958 } 959 if (bp->b_wcred == NOCRED) { 960 crhold(cred); 961 bp->b_wcred = cred; 962 } 963 np->n_flag |= NMODIFIED; 964 965 /* 966 * If dirtyend exceeds file size, chop it down. This should 967 * not normally occur but there is an append race where it 968 * might occur XXX, so we log it. 969 * 970 * If the chopping creates a reverse-indexed or degenerate 971 * situation with dirtyoff/end, we 0 both of them. 972 */ 973 974 if (bp->b_dirtyend > bcount) { 975 printf("NFS append race @%lx:%d\n", 976 (long)bp->b_blkno * DEV_BSIZE, 977 bp->b_dirtyend - bcount); 978 bp->b_dirtyend = bcount; 979 } 980 981 if (bp->b_dirtyoff >= bp->b_dirtyend) 982 bp->b_dirtyoff = bp->b_dirtyend = 0; 983 984 /* 985 * If the new write will leave a contiguous dirty 986 * area, just update the b_dirtyoff and b_dirtyend, 987 * otherwise force a write rpc of the old dirty area. 988 * 989 * While it is possible to merge discontiguous writes due to 990 * our having a B_CACHE buffer ( and thus valid read data 991 * for the hole), we don't because it could lead to 992 * significant cache coherency problems with multiple clients, 993 * especially if locking is implemented later on. 994 * 995 * as an optimization we could theoretically maintain 996 * a linked list of discontinuous areas, but we would still 997 * have to commit them separately so there isn't much 998 * advantage to it except perhaps a bit of asynchronization. 999 */ 1000 1001 if (bp->b_dirtyend > 0 && 1002 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1003 if (BUF_WRITE(bp) == EINTR) 1004 return (EINTR); 1005 goto again; 1006 } 1007 1008 /* 1009 * Check for valid write lease and get one as required. 1010 * In case getblk() and/or bwrite() delayed us. 1011 */ 1012 if ((nmp->nm_flag & NFSMNT_NQNFS) && 1013 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 1014 do { 1015 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 1016 } while (error == NQNFS_EXPIRED); 1017 if (error) { 1018 brelse(bp); 1019 break; 1020 } 1021 if (np->n_lrev != np->n_brev || 1022 (np->n_flag & NQNFSNONCACHE)) { 1023 brelse(bp); 1024 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 1025 if (error) 1026 break; 1027 np->n_brev = np->n_lrev; 1028 goto again; 1029 } 1030 } 1031 1032 error = uiomove((char *)bp->b_data + on, n, uio); 1033 1034 /* 1035 * Since this block is being modified, it must be written 1036 * again and not just committed. Since write clustering does 1037 * not work for the stage 1 data write, only the stage 2 1038 * commit rpc, we have to clear B_CLUSTEROK as well. 1039 */ 1040 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1041 1042 if (error) { 1043 bp->b_ioflags |= BIO_ERROR; 1044 brelse(bp); 1045 break; 1046 } 1047 1048 /* 1049 * Only update dirtyoff/dirtyend if not a degenerate 1050 * condition. 1051 */ 1052 if (n) { 1053 if (bp->b_dirtyend > 0) { 1054 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1055 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1056 } else { 1057 bp->b_dirtyoff = on; 1058 bp->b_dirtyend = on + n; 1059 } 1060 vfs_bio_set_validclean(bp, on, n); 1061 } 1062 1063 /* 1064 * If the lease is non-cachable or IO_SYNC do bwrite(). 1065 * 1066 * IO_INVAL appears to be unused. The idea appears to be 1067 * to turn off caching in this case. Very odd. XXX 1068 */ 1069 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 1070 if (ioflag & IO_INVAL) 1071 bp->b_flags |= B_NOCACHE; 1072 error = BUF_WRITE(bp); 1073 if (error) 1074 break; 1075 if (np->n_flag & NQNFSNONCACHE) { 1076 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 1077 if (error) 1078 break; 1079 } 1080 } else if ((n + on) == biosize && 1081 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 1082 bp->b_flags |= B_ASYNC; 1083 (void)nfs_writebp(bp, 0, 0); 1084 } else { 1085 bdwrite(bp); 1086 } 1087 } while (uio->uio_resid > 0 && n > 0); 1088 1089 if (haverslock) 1090 nfs_rsunlock(np, p); 1091 1092 return (error); 1093} 1094 1095/* 1096 * Get an nfs cache block. 1097 * 1098 * Allocate a new one if the block isn't currently in the cache 1099 * and return the block marked busy. If the calling process is 1100 * interrupted by a signal for an interruptible mount point, return 1101 * NULL. 1102 * 1103 * The caller must carefully deal with the possible B_INVAL state of 1104 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1105 * indirectly), so synchronous reads can be issued without worrying about 1106 * the B_INVAL state. We have to be a little more careful when dealing 1107 * with writes (see comments in nfs_write()) when extending a file past 1108 * its EOF. 1109 */ 1110static struct buf * 1111nfs_getcacheblk(vp, bn, size, p) 1112 struct vnode *vp; 1113 daddr_t bn; 1114 int size; 1115 struct proc *p; 1116{ 1117 register struct buf *bp; 1118 struct mount *mp; 1119 struct nfsmount *nmp; 1120 1121 mp = vp->v_mount; 1122 nmp = VFSTONFS(mp); 1123 1124 if (nmp->nm_flag & NFSMNT_INT) { 1125 bp = getblk(vp, bn, size, PCATCH, 0); 1126 while (bp == (struct buf *)0) { 1127 if (nfs_sigintr(nmp, (struct nfsreq *)0, p)) 1128 return ((struct buf *)0); 1129 bp = getblk(vp, bn, size, 0, 2 * hz); 1130 } 1131 } else { 1132 bp = getblk(vp, bn, size, 0, 0); 1133 } 1134 1135 if (vp->v_type == VREG) { 1136 int biosize; 1137 1138 biosize = mp->mnt_stat.f_iosize; 1139 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1140 } 1141 return (bp); 1142} 1143 1144/* 1145 * Flush and invalidate all dirty buffers. If another process is already 1146 * doing the flush, just wait for completion. 1147 */ 1148int 1149nfs_vinvalbuf(vp, flags, cred, p, intrflg) 1150 struct vnode *vp; 1151 int flags; 1152 struct ucred *cred; 1153 struct proc *p; 1154 int intrflg; 1155{ 1156 register struct nfsnode *np = VTONFS(vp); 1157 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1158 int error = 0, slpflag, slptimeo; 1159 1160 if (vp->v_flag & VXLOCK) { 1161 return (0); 1162 } 1163 1164 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1165 intrflg = 0; 1166 if (intrflg) { 1167 slpflag = PCATCH; 1168 slptimeo = 2 * hz; 1169 } else { 1170 slpflag = 0; 1171 slptimeo = 0; 1172 } 1173 /* 1174 * First wait for any other process doing a flush to complete. 1175 */ 1176 while (np->n_flag & NFLUSHINPROG) { 1177 np->n_flag |= NFLUSHWANT; 1178 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval", 1179 slptimeo); 1180 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) 1181 return (EINTR); 1182 } 1183 1184 /* 1185 * Now, flush as required. 1186 */ 1187 np->n_flag |= NFLUSHINPROG; 1188 error = vinvalbuf(vp, flags, cred, p, slpflag, 0); 1189 while (error) { 1190 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) { 1191 np->n_flag &= ~NFLUSHINPROG; 1192 if (np->n_flag & NFLUSHWANT) { 1193 np->n_flag &= ~NFLUSHWANT; 1194 wakeup((caddr_t)&np->n_flag); 1195 } 1196 return (EINTR); 1197 } 1198 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo); 1199 } 1200 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 1201 if (np->n_flag & NFLUSHWANT) { 1202 np->n_flag &= ~NFLUSHWANT; 1203 wakeup((caddr_t)&np->n_flag); 1204 } 1205 return (0); 1206} 1207 1208/* 1209 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1210 * This is mainly to avoid queueing async I/O requests when the nfsiods 1211 * are all hung on a dead server. 1212 * 1213 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1214 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1215 */ 1216int 1217nfs_asyncio(bp, cred, procp) 1218 register struct buf *bp; 1219 struct ucred *cred; 1220 struct proc *procp; 1221{ 1222 struct nfsmount *nmp; 1223 int i; 1224 int gotiod; 1225 int slpflag = 0; 1226 int slptimeo = 0; 1227 int error; 1228 1229 /* 1230 * If no async daemons then return EIO to force caller to run the rpc 1231 * synchronously. 1232 */ 1233 if (nfs_numasync == 0) 1234 return (EIO); 1235 1236 nmp = VFSTONFS(bp->b_vp->v_mount); 1237 1238 /* 1239 * Commits are usually short and sweet so lets save some cpu and 1240 * leave the async daemons for more important rpc's (such as reads 1241 * and writes). 1242 */ 1243 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1244 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1245 return(EIO); 1246 } 1247 1248again: 1249 if (nmp->nm_flag & NFSMNT_INT) 1250 slpflag = PCATCH; 1251 gotiod = FALSE; 1252 1253 /* 1254 * Find a free iod to process this request. 1255 */ 1256 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 1257 if (nfs_iodwant[i]) { 1258 /* 1259 * Found one, so wake it up and tell it which 1260 * mount to process. 1261 */ 1262 NFS_DPF(ASYNCIO, 1263 ("nfs_asyncio: waking iod %d for mount %p\n", 1264 i, nmp)); 1265 nfs_iodwant[i] = (struct proc *)0; 1266 nfs_iodmount[i] = nmp; 1267 nmp->nm_bufqiods++; 1268 wakeup((caddr_t)&nfs_iodwant[i]); 1269 gotiod = TRUE; 1270 break; 1271 } 1272 1273 /* 1274 * If none are free, we may already have an iod working on this mount 1275 * point. If so, it will process our request. 1276 */ 1277 if (!gotiod) { 1278 if (nmp->nm_bufqiods > 0) { 1279 NFS_DPF(ASYNCIO, 1280 ("nfs_asyncio: %d iods are already processing mount %p\n", 1281 nmp->nm_bufqiods, nmp)); 1282 gotiod = TRUE; 1283 } 1284 } 1285 1286 /* 1287 * If we have an iod which can process the request, then queue 1288 * the buffer. 1289 */ 1290 if (gotiod) { 1291 /* 1292 * Ensure that the queue never grows too large. We still want 1293 * to asynchronize so we block rather then return EIO. 1294 */ 1295 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1296 NFS_DPF(ASYNCIO, 1297 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1298 nmp->nm_bufqwant = TRUE; 1299 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO, 1300 "nfsaio", slptimeo); 1301 if (error) { 1302 if (nfs_sigintr(nmp, NULL, procp)) 1303 return (EINTR); 1304 if (slpflag == PCATCH) { 1305 slpflag = 0; 1306 slptimeo = 2 * hz; 1307 } 1308 } 1309 /* 1310 * We might have lost our iod while sleeping, 1311 * so check and loop if nescessary. 1312 */ 1313 if (nmp->nm_bufqiods == 0) { 1314 NFS_DPF(ASYNCIO, 1315 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1316 goto again; 1317 } 1318 } 1319 1320 if (bp->b_iocmd == BIO_READ) { 1321 if (bp->b_rcred == NOCRED && cred != NOCRED) { 1322 crhold(cred); 1323 bp->b_rcred = cred; 1324 } 1325 } else { 1326 bp->b_flags |= B_WRITEINPROG; 1327 if (bp->b_wcred == NOCRED && cred != NOCRED) { 1328 crhold(cred); 1329 bp->b_wcred = cred; 1330 } 1331 } 1332 1333 BUF_KERNPROC(bp); 1334 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1335 nmp->nm_bufqlen++; 1336 return (0); 1337 } 1338 1339 /* 1340 * All the iods are busy on other mounts, so return EIO to 1341 * force the caller to process the i/o synchronously. 1342 */ 1343 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1344 return (EIO); 1345} 1346 1347/* 1348 * Do an I/O operation to/from a cache block. This may be called 1349 * synchronously or from an nfsiod. 1350 */ 1351int 1352nfs_doio(bp, cr, p) 1353 struct buf *bp; 1354 struct ucred *cr; 1355 struct proc *p; 1356{ 1357 struct uio *uiop; 1358 struct vnode *vp; 1359 struct nfsnode *np; 1360 struct nfsmount *nmp; 1361 int error = 0, iomode, must_commit = 0; 1362 struct uio uio; 1363 struct iovec io; 1364 1365 vp = bp->b_vp; 1366 np = VTONFS(vp); 1367 nmp = VFSTONFS(vp->v_mount); 1368 uiop = &uio; 1369 uiop->uio_iov = &io; 1370 uiop->uio_iovcnt = 1; 1371 uiop->uio_segflg = UIO_SYSSPACE; 1372 uiop->uio_procp = p; 1373 1374 /* 1375 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1376 * do this here so we do not have to do it in all the code that 1377 * calls us. 1378 */ 1379 bp->b_flags &= ~B_INVAL; 1380 bp->b_ioflags &= ~BIO_ERROR; 1381 1382 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1383 1384 /* 1385 * Historically, paging was done with physio, but no more. 1386 */ 1387 if (bp->b_flags & B_PHYS) { 1388 /* 1389 * ...though reading /dev/drum still gets us here. 1390 */ 1391 io.iov_len = uiop->uio_resid = bp->b_bcount; 1392 /* mapping was done by vmapbuf() */ 1393 io.iov_base = bp->b_data; 1394 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1395 if (bp->b_iocmd == BIO_READ) { 1396 uiop->uio_rw = UIO_READ; 1397 nfsstats.read_physios++; 1398 error = nfs_readrpc(vp, uiop, cr); 1399 } else { 1400 int com; 1401 1402 iomode = NFSV3WRITE_DATASYNC; 1403 uiop->uio_rw = UIO_WRITE; 1404 nfsstats.write_physios++; 1405 error = nfs_writerpc(vp, uiop, cr, &iomode, &com); 1406 } 1407 if (error) { 1408 bp->b_ioflags |= BIO_ERROR; 1409 bp->b_error = error; 1410 } 1411 } else if (bp->b_iocmd == BIO_READ) { 1412 io.iov_len = uiop->uio_resid = bp->b_bcount; 1413 io.iov_base = bp->b_data; 1414 uiop->uio_rw = UIO_READ; 1415 switch (vp->v_type) { 1416 case VREG: 1417 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1418 nfsstats.read_bios++; 1419 error = nfs_readrpc(vp, uiop, cr); 1420 if (!error) { 1421 if (uiop->uio_resid) { 1422 /* 1423 * If we had a short read with no error, we must have 1424 * hit a file hole. We should zero-fill the remainder. 1425 * This can also occur if the server hits the file EOF. 1426 * 1427 * Holes used to be able to occur due to pending 1428 * writes, but that is not possible any longer. 1429 */ 1430 int nread = bp->b_bcount - uiop->uio_resid; 1431 int left = bp->b_bcount - nread; 1432 1433 if (left > 0) 1434 bzero((char *)bp->b_data + nread, left); 1435 uiop->uio_resid = 0; 1436 } 1437 } 1438 if (p && (vp->v_flag & VTEXT) && 1439 (((nmp->nm_flag & NFSMNT_NQNFS) && 1440 NQNFS_CKINVALID(vp, np, ND_READ) && 1441 np->n_lrev != np->n_brev) || 1442 (!(nmp->nm_flag & NFSMNT_NQNFS) && 1443 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) { 1444 uprintf("Process killed due to text file modification\n"); 1445 PROC_LOCK(p); 1446 psignal(p, SIGKILL); 1447 _PHOLD(p); 1448 PROC_UNLOCK(p); 1449 } 1450 break; 1451 case VLNK: 1452 uiop->uio_offset = (off_t)0; 1453 nfsstats.readlink_bios++; 1454 error = nfs_readlinkrpc(vp, uiop, cr); 1455 break; 1456 case VDIR: 1457 nfsstats.readdir_bios++; 1458 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1459 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1460 error = nfs_readdirplusrpc(vp, uiop, cr); 1461 if (error == NFSERR_NOTSUPP) 1462 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1463 } 1464 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1465 error = nfs_readdirrpc(vp, uiop, cr); 1466 /* 1467 * end-of-directory sets B_INVAL but does not generate an 1468 * error. 1469 */ 1470 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1471 bp->b_flags |= B_INVAL; 1472 break; 1473 default: 1474 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1475 break; 1476 }; 1477 if (error) { 1478 bp->b_ioflags |= BIO_ERROR; 1479 bp->b_error = error; 1480 } 1481 } else { 1482 /* 1483 * If we only need to commit, try to commit 1484 */ 1485 if (bp->b_flags & B_NEEDCOMMIT) { 1486 int retv; 1487 off_t off; 1488 1489 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1490 bp->b_flags |= B_WRITEINPROG; 1491 retv = nfs_commit( 1492 bp->b_vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1493 bp->b_wcred, p); 1494 bp->b_flags &= ~B_WRITEINPROG; 1495 if (retv == 0) { 1496 bp->b_dirtyoff = bp->b_dirtyend = 0; 1497 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1498 bp->b_resid = 0; 1499 bufdone(bp); 1500 return (0); 1501 } 1502 if (retv == NFSERR_STALEWRITEVERF) { 1503 nfs_clearcommit(bp->b_vp->v_mount); 1504 } 1505 } 1506 1507 /* 1508 * Setup for actual write 1509 */ 1510 1511 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1512 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1513 1514 if (bp->b_dirtyend > bp->b_dirtyoff) { 1515 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1516 - bp->b_dirtyoff; 1517 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1518 + bp->b_dirtyoff; 1519 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1520 uiop->uio_rw = UIO_WRITE; 1521 nfsstats.write_bios++; 1522 1523 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1524 iomode = NFSV3WRITE_UNSTABLE; 1525 else 1526 iomode = NFSV3WRITE_FILESYNC; 1527 1528 bp->b_flags |= B_WRITEINPROG; 1529 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit); 1530 1531 /* 1532 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1533 * to cluster the buffers needing commit. This will allow 1534 * the system to submit a single commit rpc for the whole 1535 * cluster. We can do this even if the buffer is not 100% 1536 * dirty (relative to the NFS blocksize), so we optimize the 1537 * append-to-file-case. 1538 * 1539 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1540 * cleared because write clustering only works for commit 1541 * rpc's, not for the data portion of the write). 1542 */ 1543 1544 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1545 bp->b_flags |= B_NEEDCOMMIT; 1546 if (bp->b_dirtyoff == 0 1547 && bp->b_dirtyend == bp->b_bcount) 1548 bp->b_flags |= B_CLUSTEROK; 1549 } else { 1550 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1551 } 1552 bp->b_flags &= ~B_WRITEINPROG; 1553 1554 /* 1555 * For an interrupted write, the buffer is still valid 1556 * and the write hasn't been pushed to the server yet, 1557 * so we can't set BIO_ERROR and report the interruption 1558 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1559 * is not relevant, so the rpc attempt is essentially 1560 * a noop. For the case of a V3 write rpc not being 1561 * committed to stable storage, the block is still 1562 * dirty and requires either a commit rpc or another 1563 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1564 * the block is reused. This is indicated by setting 1565 * the B_DELWRI and B_NEEDCOMMIT flags. 1566 * 1567 * If the buffer is marked B_PAGING, it does not reside on 1568 * the vp's paging queues so we cannot call bdirty(). The 1569 * bp in this case is not an NFS cache block so we should 1570 * be safe. XXX 1571 */ 1572 if (error == EINTR 1573 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1574 int s; 1575 1576 s = splbio(); 1577 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1578 if ((bp->b_flags & B_PAGING) == 0) { 1579 bdirty(bp); 1580 bp->b_flags &= ~B_DONE; 1581 } 1582 if (error && (bp->b_flags & B_ASYNC) == 0) 1583 bp->b_flags |= B_EINTR; 1584 splx(s); 1585 } else { 1586 if (error) { 1587 bp->b_ioflags |= BIO_ERROR; 1588 bp->b_error = np->n_error = error; 1589 np->n_flag |= NWRITEERR; 1590 } 1591 bp->b_dirtyoff = bp->b_dirtyend = 0; 1592 } 1593 } else { 1594 bp->b_resid = 0; 1595 bufdone(bp); 1596 return (0); 1597 } 1598 } 1599 bp->b_resid = uiop->uio_resid; 1600 if (must_commit) 1601 nfs_clearcommit(vp->v_mount); 1602 bufdone(bp); 1603 return (error); 1604} 1605