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: /repoman/r/ncvs/src/sys/nfsclient/nfs_bio.c,v 1.130 2004/04/14 23:23:55 peadar Exp $ 38 * $DragonFly: src/sys/vfs/nfs/nfs_bio.c,v 1.18 2004/10/12 19:21:01 dillon Exp $ 39 */ 40 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/resourcevar.h> 45 #include <sys/signalvar.h> 46 #include <sys/proc.h> 47 #include <sys/buf.h> 48 #include <sys/vnode.h> 49 #include <sys/mount.h> 50 #include <sys/kernel.h> 51 #include <sys/buf2.h> 52 #include <sys/msfbuf.h> 53 54 #include <vm/vm.h> 55 #include <vm/vm_extern.h> 56 #include <vm/vm_page.h> 57 #include <vm/vm_object.h> 58 #include <vm/vm_pager.h> 59 #include <vm/vnode_pager.h> 60 61 #include "rpcv2.h" 62 #include "nfsproto.h" 63 #include "nfs.h" 64 #include "nfsmount.h" 65 #include "nqnfs.h" 66 #include "nfsnode.h" 67 68 static struct buf *nfs_getcacheblk (struct vnode *vp, daddr_t bn, int size, 69 struct thread *td); 70 71 extern int nfs_numasync; 72 extern int nfs_pbuf_freecnt; 73 extern struct nfsstats nfsstats; 74 75 /* 76 * Vnode op for VM getpages. 77 * 78 * nfs_getpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, 79 * int a_reqpage, vm_ooffset_t a_offset) 80 */ 81 int 82 nfs_getpages(struct vop_getpages_args *ap) 83 { 84 struct thread *td = curthread; /* XXX */ 85 int i, error, nextoff, size, toff, count, npages; 86 struct uio uio; 87 struct iovec iov; 88 vm_offset_t kva; 89 struct vnode *vp; 90 struct nfsmount *nmp; 91 vm_page_t *pages; 92 vm_page_t m; 93 struct msf_buf *msf; 94 95 vp = ap->a_vp; 96 nmp = VFSTONFS(vp->v_mount); 97 pages = ap->a_m; 98 count = ap->a_count; 99 100 if (vp->v_object == NULL) { 101 printf("nfs_getpages: called with non-merged cache vnode??\n"); 102 return VM_PAGER_ERROR; 103 } 104 105 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 106 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 107 (void)nfs_fsinfo(nmp, vp, td); 108 109 npages = btoc(count); 110 111 /* 112 * NOTE that partially valid pages may occur in cases other 113 * then file EOF, such as when a file is partially written and 114 * ftruncate()-extended to a larger size. It is also possible 115 * for the valid bits to be set on garbage beyond the file EOF and 116 * clear in the area before EOF (e.g. m->valid == 0xfc), which can 117 * occur due to vtruncbuf() and the buffer cache's handling of 118 * pages which 'straddle' buffers or when b_bufsize is not a 119 * multiple of PAGE_SIZE.... the buffer cache cannot normally 120 * clear the extra bits. This kind of situation occurs when you 121 * make a small write() (m->valid == 0x03) and then mmap() and 122 * fault in the buffer(m->valid = 0xFF). When NFS flushes the 123 * buffer (vinvalbuf() m->valid = 0xFC) we are left with a mess. 124 * 125 * This is combined with the possibility that the pages are partially 126 * dirty or that there is a buffer backing the pages that is dirty 127 * (even if m->dirty is 0). 128 * 129 * To solve this problem several hacks have been made: (1) NFS 130 * guarentees that the IO block size is a multiple of PAGE_SIZE and 131 * (2) The buffer cache, when invalidating an NFS buffer, will 132 * disregard the buffer's fragmentory b_bufsize and invalidate 133 * the whole page rather then just the piece the buffer owns. 134 * 135 * This allows us to assume that a partially valid page found here 136 * is fully valid (vm_fault will zero'd out areas of the page not 137 * marked as valid). 138 */ 139 m = pages[ap->a_reqpage]; 140 if (m->valid != 0) { 141 for (i = 0; i < npages; ++i) { 142 if (i != ap->a_reqpage) 143 vnode_pager_freepage(pages[i]); 144 } 145 return(0); 146 } 147 148 /* 149 * Use an MSF_BUF as a medium to retrieve data from the pages. 150 */ 151 msf = msf_buf_alloc(pages, npages, 0); 152 kva = msf_buf_kva(msf); 153 154 iov.iov_base = (caddr_t) kva; 155 iov.iov_len = count; 156 uio.uio_iov = &iov; 157 uio.uio_iovcnt = 1; 158 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 159 uio.uio_resid = count; 160 uio.uio_segflg = UIO_SYSSPACE; 161 uio.uio_rw = UIO_READ; 162 uio.uio_td = td; 163 164 error = nfs_readrpc(vp, &uio); 165 msf_buf_free(msf); 166 167 if (error && (uio.uio_resid == count)) { 168 printf("nfs_getpages: error %d\n", error); 169 for (i = 0; i < npages; ++i) { 170 if (i != ap->a_reqpage) 171 vnode_pager_freepage(pages[i]); 172 } 173 return VM_PAGER_ERROR; 174 } 175 176 /* 177 * Calculate the number of bytes read and validate only that number 178 * of bytes. Note that due to pending writes, size may be 0. This 179 * does not mean that the remaining data is invalid! 180 */ 181 182 size = count - uio.uio_resid; 183 184 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 185 nextoff = toff + PAGE_SIZE; 186 m = pages[i]; 187 188 m->flags &= ~PG_ZERO; 189 190 if (nextoff <= size) { 191 /* 192 * Read operation filled an entire page 193 */ 194 m->valid = VM_PAGE_BITS_ALL; 195 vm_page_undirty(m); 196 } else if (size > toff) { 197 /* 198 * Read operation filled a partial page. 199 */ 200 m->valid = 0; 201 vm_page_set_validclean(m, 0, size - toff); 202 /* handled by vm_fault now */ 203 /* vm_page_zero_invalid(m, TRUE); */ 204 } else { 205 /* 206 * Read operation was short. If no error occured 207 * we may have hit a zero-fill section. We simply 208 * leave valid set to 0. 209 */ 210 ; 211 } 212 if (i != ap->a_reqpage) { 213 /* 214 * Whether or not to leave the page activated is up in 215 * the air, but we should put the page on a page queue 216 * somewhere (it already is in the object). Result: 217 * It appears that emperical results show that 218 * deactivating pages is best. 219 */ 220 221 /* 222 * Just in case someone was asking for this page we 223 * now tell them that it is ok to use. 224 */ 225 if (!error) { 226 if (m->flags & PG_WANTED) 227 vm_page_activate(m); 228 else 229 vm_page_deactivate(m); 230 vm_page_wakeup(m); 231 } else { 232 vnode_pager_freepage(m); 233 } 234 } 235 } 236 return 0; 237 } 238 239 /* 240 * Vnode op for VM putpages. 241 * 242 * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync, 243 * int *a_rtvals, vm_ooffset_t a_offset) 244 */ 245 int 246 nfs_putpages(struct vop_putpages_args *ap) 247 { 248 struct thread *td = curthread; 249 struct uio uio; 250 struct iovec iov; 251 vm_offset_t kva; 252 int iomode, must_commit, i, error, npages, count; 253 off_t offset; 254 int *rtvals; 255 struct vnode *vp; 256 struct nfsmount *nmp; 257 struct nfsnode *np; 258 vm_page_t *pages; 259 struct msf_buf *msf; 260 261 vp = ap->a_vp; 262 np = VTONFS(vp); 263 nmp = VFSTONFS(vp->v_mount); 264 pages = ap->a_m; 265 count = ap->a_count; 266 rtvals = ap->a_rtvals; 267 npages = btoc(count); 268 offset = IDX_TO_OFF(pages[0]->pindex); 269 270 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 271 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 272 (void)nfs_fsinfo(nmp, vp, td); 273 274 for (i = 0; i < npages; i++) { 275 rtvals[i] = VM_PAGER_AGAIN; 276 } 277 278 /* 279 * When putting pages, do not extend file past EOF. 280 */ 281 282 if (offset + count > np->n_size) { 283 count = np->n_size - offset; 284 if (count < 0) 285 count = 0; 286 } 287 288 /* 289 * Use an MSF_BUF as a medium to retrieve data from the pages. 290 */ 291 msf = msf_buf_alloc(pages, npages, 0); 292 kva = msf_buf_kva(msf); 293 294 iov.iov_base = (caddr_t) kva; 295 iov.iov_len = count; 296 uio.uio_iov = &iov; 297 uio.uio_iovcnt = 1; 298 uio.uio_offset = offset; 299 uio.uio_resid = count; 300 uio.uio_segflg = UIO_SYSSPACE; 301 uio.uio_rw = UIO_WRITE; 302 uio.uio_td = td; 303 304 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 305 iomode = NFSV3WRITE_UNSTABLE; 306 else 307 iomode = NFSV3WRITE_FILESYNC; 308 309 error = nfs_writerpc(vp, &uio, &iomode, &must_commit); 310 311 msf_buf_free(msf); 312 313 if (!error) { 314 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 315 for (i = 0; i < nwritten; i++) { 316 rtvals[i] = VM_PAGER_OK; 317 vm_page_undirty(pages[i]); 318 } 319 if (must_commit) 320 nfs_clearcommit(vp->v_mount); 321 } 322 return rtvals[0]; 323 } 324 325 /* 326 * Vnode op for read using bio 327 */ 328 int 329 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag) 330 { 331 struct nfsnode *np = VTONFS(vp); 332 int biosize, i; 333 struct buf *bp = 0, *rabp; 334 struct vattr vattr; 335 struct thread *td; 336 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 337 daddr_t lbn, rabn; 338 int bcount; 339 int seqcount; 340 int nra, error = 0, n = 0, on = 0; 341 342 #ifdef DIAGNOSTIC 343 if (uio->uio_rw != UIO_READ) 344 panic("nfs_read mode"); 345 #endif 346 if (uio->uio_resid == 0) 347 return (0); 348 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 349 return (EINVAL); 350 td = uio->uio_td; 351 352 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 353 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 354 (void)nfs_fsinfo(nmp, vp, td); 355 if (vp->v_type != VDIR && 356 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 357 return (EFBIG); 358 biosize = vp->v_mount->mnt_stat.f_iosize; 359 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 360 /* 361 * For nfs, cache consistency can only be maintained approximately. 362 * Although RFC1094 does not specify the criteria, the following is 363 * believed to be compatible with the reference port. 364 * For nqnfs, full cache consistency is maintained within the loop. 365 * For nfs: 366 * If the file's modify time on the server has changed since the 367 * last read rpc or you have written to the file, 368 * you may have lost data cache consistency with the 369 * server, so flush all of the file's data out of the cache. 370 * Then force a getattr rpc to ensure that you have up to date 371 * attributes. 372 * NB: This implies that cache data can be read when up to 373 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 374 * attributes this could be forced by setting n_attrstamp to 0 before 375 * the VOP_GETATTR() call. 376 */ 377 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) { 378 if (np->n_flag & NMODIFIED) { 379 if (vp->v_type != VREG) { 380 if (vp->v_type != VDIR) 381 panic("nfs: bioread, not dir"); 382 nfs_invaldir(vp); 383 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 384 if (error) 385 return (error); 386 } 387 np->n_attrstamp = 0; 388 error = VOP_GETATTR(vp, &vattr, td); 389 if (error) 390 return (error); 391 np->n_mtime = vattr.va_mtime.tv_sec; 392 } else { 393 error = VOP_GETATTR(vp, &vattr, td); 394 if (error) 395 return (error); 396 if ((np->n_flag & NSIZECHANGED) 397 || np->n_mtime != vattr.va_mtime.tv_sec) { 398 if (vp->v_type == VDIR) 399 nfs_invaldir(vp); 400 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 401 if (error) 402 return (error); 403 np->n_mtime = vattr.va_mtime.tv_sec; 404 np->n_flag &= ~NSIZECHANGED; 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, td); 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, td, 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, td, 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)); 441 case VLNK: 442 return (nfs_readlinkrpc(vp, uio)); 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, td); 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, td)) { 471 rabp->b_flags |= B_INVAL|B_ERROR; 472 vfs_unbusy_pages(rabp); 473 brelse(rabp); 474 break; 475 } 476 } else { 477 brelse(rabp); 478 } 479 } 480 } 481 } 482 483 /* 484 * Obtain the buffer cache block. Figure out the buffer size 485 * when we are at EOF. If we are modifying the size of the 486 * buffer based on an EOF condition we need to hold 487 * nfs_rslock() through obtaining the buffer to prevent 488 * a potential writer-appender from messing with n_size. 489 * Otherwise we may accidently truncate the buffer and 490 * lose dirty data. 491 * 492 * Note that bcount is *not* DEV_BSIZE aligned. 493 */ 494 495 again: 496 bcount = biosize; 497 if ((off_t)lbn * biosize >= np->n_size) { 498 bcount = 0; 499 } else if ((off_t)(lbn + 1) * biosize > np->n_size) { 500 bcount = np->n_size - (off_t)lbn * biosize; 501 } 502 if (bcount != biosize) { 503 switch(nfs_rslock(np, td)) { 504 case ENOLCK: 505 goto again; 506 /* not reached */ 507 case EINTR: 508 case ERESTART: 509 return(EINTR); 510 /* not reached */ 511 default: 512 break; 513 } 514 } 515 516 bp = nfs_getcacheblk(vp, lbn, bcount, td); 517 518 if (bcount != biosize) 519 nfs_rsunlock(np, td); 520 if (!bp) 521 return (EINTR); 522 523 /* 524 * If B_CACHE is not set, we must issue the read. If this 525 * fails, we return an error. 526 */ 527 528 if ((bp->b_flags & B_CACHE) == 0) { 529 bp->b_flags |= B_READ; 530 vfs_busy_pages(bp, 0); 531 error = nfs_doio(bp, td); 532 if (error) { 533 brelse(bp); 534 return (error); 535 } 536 } 537 538 /* 539 * on is the offset into the current bp. Figure out how many 540 * bytes we can copy out of the bp. Note that bcount is 541 * NOT DEV_BSIZE aligned. 542 * 543 * Then figure out how many bytes we can copy into the uio. 544 */ 545 546 n = 0; 547 if (on < bcount) 548 n = min((unsigned)(bcount - on), uio->uio_resid); 549 break; 550 case VLNK: 551 nfsstats.biocache_readlinks++; 552 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 553 if (!bp) 554 return (EINTR); 555 if ((bp->b_flags & B_CACHE) == 0) { 556 bp->b_flags |= B_READ; 557 vfs_busy_pages(bp, 0); 558 error = nfs_doio(bp, td); 559 if (error) { 560 bp->b_flags |= B_ERROR; 561 brelse(bp); 562 return (error); 563 } 564 } 565 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 566 on = 0; 567 break; 568 case VDIR: 569 nfsstats.biocache_readdirs++; 570 if (np->n_direofoffset 571 && uio->uio_offset >= np->n_direofoffset) { 572 return (0); 573 } 574 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 575 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 576 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 577 if (!bp) 578 return (EINTR); 579 if ((bp->b_flags & B_CACHE) == 0) { 580 bp->b_flags |= B_READ; 581 vfs_busy_pages(bp, 0); 582 error = nfs_doio(bp, td); 583 if (error) { 584 brelse(bp); 585 } 586 while (error == NFSERR_BAD_COOKIE) { 587 printf("got bad cookie vp %p bp %p\n", vp, bp); 588 nfs_invaldir(vp); 589 error = nfs_vinvalbuf(vp, 0, td, 1); 590 /* 591 * Yuck! The directory has been modified on the 592 * server. The only way to get the block is by 593 * reading from the beginning to get all the 594 * offset cookies. 595 * 596 * Leave the last bp intact unless there is an error. 597 * Loop back up to the while if the error is another 598 * NFSERR_BAD_COOKIE (double yuch!). 599 */ 600 for (i = 0; i <= lbn && !error; i++) { 601 if (np->n_direofoffset 602 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 603 return (0); 604 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 605 if (!bp) 606 return (EINTR); 607 if ((bp->b_flags & B_CACHE) == 0) { 608 bp->b_flags |= B_READ; 609 vfs_busy_pages(bp, 0); 610 error = nfs_doio(bp, td); 611 /* 612 * no error + B_INVAL == directory EOF, 613 * use the block. 614 */ 615 if (error == 0 && (bp->b_flags & B_INVAL)) 616 break; 617 } 618 /* 619 * An error will throw away the block and the 620 * for loop will break out. If no error and this 621 * is not the block we want, we throw away the 622 * block and go for the next one via the for loop. 623 */ 624 if (error || i < lbn) 625 brelse(bp); 626 } 627 } 628 /* 629 * The above while is repeated if we hit another cookie 630 * error. If we hit an error and it wasn't a cookie error, 631 * we give up. 632 */ 633 if (error) 634 return (error); 635 } 636 637 /* 638 * If not eof and read aheads are enabled, start one. 639 * (You need the current block first, so that you have the 640 * directory offset cookie of the next block.) 641 */ 642 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 643 (bp->b_flags & B_INVAL) == 0 && 644 (np->n_direofoffset == 0 || 645 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 646 !(np->n_flag & NQNFSNONCACHE) && 647 !incore(vp, lbn + 1)) { 648 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 649 if (rabp) { 650 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 651 rabp->b_flags |= (B_READ | B_ASYNC); 652 vfs_busy_pages(rabp, 0); 653 if (nfs_asyncio(rabp, td)) { 654 rabp->b_flags |= B_INVAL|B_ERROR; 655 vfs_unbusy_pages(rabp); 656 brelse(rabp); 657 } 658 } else { 659 brelse(rabp); 660 } 661 } 662 } 663 /* 664 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 665 * chopped for the EOF condition, we cannot tell how large 666 * NFS directories are going to be until we hit EOF. So 667 * an NFS directory buffer is *not* chopped to its EOF. Now, 668 * it just so happens that b_resid will effectively chop it 669 * to EOF. *BUT* this information is lost if the buffer goes 670 * away and is reconstituted into a B_CACHE state ( due to 671 * being VMIO ) later. So we keep track of the directory eof 672 * in np->n_direofoffset and chop it off as an extra step 673 * right here. 674 */ 675 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 676 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 677 n = np->n_direofoffset - uio->uio_offset; 678 break; 679 default: 680 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 681 break; 682 }; 683 684 if (n > 0) { 685 error = uiomove(bp->b_data + on, (int)n, uio); 686 } 687 switch (vp->v_type) { 688 case VREG: 689 break; 690 case VLNK: 691 n = 0; 692 break; 693 case VDIR: 694 /* 695 * Invalidate buffer if caching is disabled, forcing a 696 * re-read from the remote later. 697 */ 698 if (np->n_flag & NQNFSNONCACHE) 699 bp->b_flags |= B_INVAL; 700 break; 701 default: 702 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 703 } 704 brelse(bp); 705 } while (error == 0 && uio->uio_resid > 0 && n > 0); 706 return (error); 707 } 708 709 /* 710 * Vnode op for write using bio 711 * 712 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag, 713 * struct ucred *a_cred) 714 */ 715 int 716 nfs_write(struct vop_write_args *ap) 717 { 718 int biosize; 719 struct uio *uio = ap->a_uio; 720 struct thread *td = uio->uio_td; 721 struct vnode *vp = ap->a_vp; 722 struct nfsnode *np = VTONFS(vp); 723 int ioflag = ap->a_ioflag; 724 struct buf *bp; 725 struct vattr vattr; 726 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 727 daddr_t lbn; 728 int bcount; 729 int n, on, error = 0, iomode, must_commit; 730 int haverslock = 0; 731 732 #ifdef DIAGNOSTIC 733 if (uio->uio_rw != UIO_WRITE) 734 panic("nfs_write mode"); 735 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 736 panic("nfs_write proc"); 737 #endif 738 if (vp->v_type != VREG) 739 return (EIO); 740 if (np->n_flag & NWRITEERR) { 741 np->n_flag &= ~NWRITEERR; 742 return (np->n_error); 743 } 744 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 745 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 746 (void)nfs_fsinfo(nmp, vp, td); 747 748 /* 749 * Synchronously flush pending buffers if we are in synchronous 750 * mode or if we are appending. 751 */ 752 if (ioflag & (IO_APPEND | IO_SYNC)) { 753 if (np->n_flag & NMODIFIED) { 754 np->n_attrstamp = 0; 755 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 756 if (error) 757 return (error); 758 } 759 } 760 761 /* 762 * If IO_APPEND then load uio_offset. We restart here if we cannot 763 * get the append lock. 764 */ 765 restart: 766 if (ioflag & IO_APPEND) { 767 np->n_attrstamp = 0; 768 error = VOP_GETATTR(vp, &vattr, td); 769 if (error) 770 return (error); 771 uio->uio_offset = np->n_size; 772 } 773 774 if (uio->uio_offset < 0) 775 return (EINVAL); 776 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 777 return (EFBIG); 778 if (uio->uio_resid == 0) 779 return (0); 780 781 /* 782 * We need to obtain the rslock if we intend to modify np->n_size 783 * in order to guarentee the append point with multiple contending 784 * writers, to guarentee that no other appenders modify n_size 785 * while we are trying to obtain a truncated buffer (i.e. to avoid 786 * accidently truncating data written by another appender due to 787 * the race), and to ensure that the buffer is populated prior to 788 * our extending of the file. We hold rslock through the entire 789 * operation. 790 * 791 * Note that we do not synchronize the case where someone truncates 792 * the file while we are appending to it because attempting to lock 793 * this case may deadlock other parts of the system unexpectedly. 794 */ 795 if ((ioflag & IO_APPEND) || 796 uio->uio_offset + uio->uio_resid > np->n_size) { 797 switch(nfs_rslock(np, td)) { 798 case ENOLCK: 799 goto restart; 800 /* not reached */ 801 case EINTR: 802 case ERESTART: 803 return(EINTR); 804 /* not reached */ 805 default: 806 break; 807 } 808 haverslock = 1; 809 } 810 811 /* 812 * Maybe this should be above the vnode op call, but so long as 813 * file servers have no limits, i don't think it matters 814 */ 815 if (td->td_proc && uio->uio_offset + uio->uio_resid > 816 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 817 psignal(td->td_proc, SIGXFSZ); 818 if (haverslock) 819 nfs_rsunlock(np, td); 820 return (EFBIG); 821 } 822 823 biosize = vp->v_mount->mnt_stat.f_iosize; 824 825 do { 826 /* 827 * Check for a valid write lease. 828 */ 829 if ((nmp->nm_flag & NFSMNT_NQNFS) && 830 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 831 do { 832 error = nqnfs_getlease(vp, ND_WRITE, td); 833 } while (error == NQNFS_EXPIRED); 834 if (error) 835 break; 836 if (np->n_lrev != np->n_brev || 837 (np->n_flag & NQNFSNONCACHE)) { 838 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 839 if (error) 840 break; 841 np->n_brev = np->n_lrev; 842 } 843 } 844 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 845 iomode = NFSV3WRITE_FILESYNC; 846 error = nfs_writerpc(vp, uio, &iomode, &must_commit); 847 if (must_commit) 848 nfs_clearcommit(vp->v_mount); 849 break; 850 } 851 nfsstats.biocache_writes++; 852 lbn = uio->uio_offset / biosize; 853 on = uio->uio_offset & (biosize-1); 854 n = min((unsigned)(biosize - on), uio->uio_resid); 855 again: 856 /* 857 * Handle direct append and file extension cases, calculate 858 * unaligned buffer size. 859 */ 860 861 if (uio->uio_offset == np->n_size && n) { 862 /* 863 * Get the buffer (in its pre-append state to maintain 864 * B_CACHE if it was previously set). Resize the 865 * nfsnode after we have locked the buffer to prevent 866 * readers from reading garbage. 867 */ 868 bcount = on; 869 bp = nfs_getcacheblk(vp, lbn, bcount, td); 870 871 if (bp != NULL) { 872 long save; 873 874 np->n_size = uio->uio_offset + n; 875 np->n_flag |= NMODIFIED; 876 vnode_pager_setsize(vp, np->n_size); 877 878 save = bp->b_flags & B_CACHE; 879 bcount += n; 880 allocbuf(bp, bcount); 881 bp->b_flags |= save; 882 } 883 } else { 884 /* 885 * Obtain the locked cache block first, and then 886 * adjust the file's size as appropriate. 887 */ 888 bcount = on + n; 889 if ((off_t)lbn * biosize + bcount < np->n_size) { 890 if ((off_t)(lbn + 1) * biosize < np->n_size) 891 bcount = biosize; 892 else 893 bcount = np->n_size - (off_t)lbn * biosize; 894 } 895 bp = nfs_getcacheblk(vp, lbn, bcount, td); 896 if (uio->uio_offset + n > np->n_size) { 897 np->n_size = uio->uio_offset + n; 898 np->n_flag |= NMODIFIED; 899 vnode_pager_setsize(vp, np->n_size); 900 } 901 } 902 903 if (!bp) { 904 error = EINTR; 905 break; 906 } 907 908 /* 909 * Issue a READ if B_CACHE is not set. In special-append 910 * mode, B_CACHE is based on the buffer prior to the write 911 * op and is typically set, avoiding the read. If a read 912 * is required in special append mode, the server will 913 * probably send us a short-read since we extended the file 914 * on our end, resulting in b_resid == 0 and, thusly, 915 * B_CACHE getting set. 916 * 917 * We can also avoid issuing the read if the write covers 918 * the entire buffer. We have to make sure the buffer state 919 * is reasonable in this case since we will not be initiating 920 * I/O. See the comments in kern/vfs_bio.c's getblk() for 921 * more information. 922 * 923 * B_CACHE may also be set due to the buffer being cached 924 * normally. 925 */ 926 927 if (on == 0 && n == bcount) { 928 bp->b_flags |= B_CACHE; 929 bp->b_flags &= ~(B_ERROR | B_INVAL); 930 } 931 932 if ((bp->b_flags & B_CACHE) == 0) { 933 bp->b_flags |= B_READ; 934 vfs_busy_pages(bp, 0); 935 error = nfs_doio(bp, td); 936 if (error) { 937 brelse(bp); 938 break; 939 } 940 } 941 if (!bp) { 942 error = EINTR; 943 break; 944 } 945 np->n_flag |= NMODIFIED; 946 947 /* 948 * If dirtyend exceeds file size, chop it down. This should 949 * not normally occur but there is an append race where it 950 * might occur XXX, so we log it. 951 * 952 * If the chopping creates a reverse-indexed or degenerate 953 * situation with dirtyoff/end, we 0 both of them. 954 */ 955 956 if (bp->b_dirtyend > bcount) { 957 printf("NFS append race @%lx:%d\n", 958 (long)bp->b_blkno * DEV_BSIZE, 959 bp->b_dirtyend - bcount); 960 bp->b_dirtyend = bcount; 961 } 962 963 if (bp->b_dirtyoff >= bp->b_dirtyend) 964 bp->b_dirtyoff = bp->b_dirtyend = 0; 965 966 /* 967 * If the new write will leave a contiguous dirty 968 * area, just update the b_dirtyoff and b_dirtyend, 969 * otherwise force a write rpc of the old dirty area. 970 * 971 * While it is possible to merge discontiguous writes due to 972 * our having a B_CACHE buffer ( and thus valid read data 973 * for the hole), we don't because it could lead to 974 * significant cache coherency problems with multiple clients, 975 * especially if locking is implemented later on. 976 * 977 * as an optimization we could theoretically maintain 978 * a linked list of discontinuous areas, but we would still 979 * have to commit them separately so there isn't much 980 * advantage to it except perhaps a bit of asynchronization. 981 */ 982 983 if (bp->b_dirtyend > 0 && 984 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 985 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) { 986 error = EINTR; 987 break; 988 } 989 goto again; 990 } 991 992 /* 993 * Check for valid write lease and get one as required. 994 * In case getblk() and/or bwrite() delayed us. 995 */ 996 if ((nmp->nm_flag & NFSMNT_NQNFS) && 997 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 998 do { 999 error = nqnfs_getlease(vp, ND_WRITE, td); 1000 } while (error == NQNFS_EXPIRED); 1001 if (error) { 1002 brelse(bp); 1003 break; 1004 } 1005 if (np->n_lrev != np->n_brev || 1006 (np->n_flag & NQNFSNONCACHE)) { 1007 brelse(bp); 1008 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 1009 if (error) 1010 break; 1011 np->n_brev = np->n_lrev; 1012 goto again; 1013 } 1014 } 1015 1016 error = uiomove((char *)bp->b_data + on, n, uio); 1017 1018 /* 1019 * Since this block is being modified, it must be written 1020 * again and not just committed. Since write clustering does 1021 * not work for the stage 1 data write, only the stage 2 1022 * commit rpc, we have to clear B_CLUSTEROK as well. 1023 */ 1024 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1025 1026 if (error) { 1027 bp->b_flags |= B_ERROR; 1028 brelse(bp); 1029 break; 1030 } 1031 1032 /* 1033 * Only update dirtyoff/dirtyend if not a degenerate 1034 * condition. 1035 */ 1036 if (n) { 1037 if (bp->b_dirtyend > 0) { 1038 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1039 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1040 } else { 1041 bp->b_dirtyoff = on; 1042 bp->b_dirtyend = on + n; 1043 } 1044 vfs_bio_set_validclean(bp, on, n); 1045 } 1046 /* 1047 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN 1048 * filesystem). XXX also use for loopback NFS mounts. 1049 */ 1050 if (ioflag & IO_NOWDRAIN) 1051 bp->b_flags |= B_NOWDRAIN; 1052 1053 /* 1054 * If the lease is non-cachable or IO_SYNC do bwrite(). 1055 * 1056 * IO_INVAL appears to be unused. The idea appears to be 1057 * to turn off caching in this case. Very odd. XXX 1058 */ 1059 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 1060 if (ioflag & IO_INVAL) 1061 bp->b_flags |= B_NOCACHE; 1062 error = VOP_BWRITE(bp->b_vp, bp); 1063 if (error) 1064 break; 1065 if (np->n_flag & NQNFSNONCACHE) { 1066 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 1067 if (error) 1068 break; 1069 } 1070 } else if ((n + on) == biosize && 1071 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 1072 bp->b_flags |= B_ASYNC; 1073 (void)nfs_writebp(bp, 0, 0); 1074 } else { 1075 bdwrite(bp); 1076 } 1077 } while (uio->uio_resid > 0 && n > 0); 1078 1079 if (haverslock) 1080 nfs_rsunlock(np, td); 1081 1082 return (error); 1083 } 1084 1085 /* 1086 * Get an nfs cache block. 1087 * 1088 * Allocate a new one if the block isn't currently in the cache 1089 * and return the block marked busy. If the calling process is 1090 * interrupted by a signal for an interruptible mount point, return 1091 * NULL. 1092 * 1093 * The caller must carefully deal with the possible B_INVAL state of 1094 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1095 * indirectly), so synchronous reads can be issued without worrying about 1096 * the B_INVAL state. We have to be a little more careful when dealing 1097 * with writes (see comments in nfs_write()) when extending a file past 1098 * its EOF. 1099 */ 1100 static struct buf * 1101 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1102 { 1103 struct buf *bp; 1104 struct mount *mp; 1105 struct nfsmount *nmp; 1106 1107 mp = vp->v_mount; 1108 nmp = VFSTONFS(mp); 1109 1110 if (nmp->nm_flag & NFSMNT_INT) { 1111 bp = getblk(vp, bn, size, PCATCH, 0); 1112 while (bp == (struct buf *)0) { 1113 if (nfs_sigintr(nmp, (struct nfsreq *)0, td)) 1114 return ((struct buf *)0); 1115 bp = getblk(vp, bn, size, 0, 2 * hz); 1116 } 1117 } else { 1118 bp = getblk(vp, bn, size, 0, 0); 1119 } 1120 1121 if (vp->v_type == VREG) { 1122 int biosize; 1123 1124 biosize = mp->mnt_stat.f_iosize; 1125 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1126 } 1127 return (bp); 1128 } 1129 1130 /* 1131 * Flush and invalidate all dirty buffers. If another process is already 1132 * doing the flush, just wait for completion. 1133 */ 1134 int 1135 nfs_vinvalbuf(struct vnode *vp, int flags, 1136 struct thread *td, int intrflg) 1137 { 1138 struct nfsnode *np = VTONFS(vp); 1139 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1140 int error = 0, slpflag, slptimeo; 1141 1142 if (vp->v_flag & VRECLAIMED) 1143 return (0); 1144 1145 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1146 intrflg = 0; 1147 if (intrflg) { 1148 slpflag = PCATCH; 1149 slptimeo = 2 * hz; 1150 } else { 1151 slpflag = 0; 1152 slptimeo = 0; 1153 } 1154 /* 1155 * First wait for any other process doing a flush to complete. 1156 */ 1157 while (np->n_flag & NFLUSHINPROG) { 1158 np->n_flag |= NFLUSHWANT; 1159 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo); 1160 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) 1161 return (EINTR); 1162 } 1163 1164 /* 1165 * Now, flush as required. 1166 */ 1167 np->n_flag |= NFLUSHINPROG; 1168 error = vinvalbuf(vp, flags, td, slpflag, 0); 1169 while (error) { 1170 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) { 1171 np->n_flag &= ~NFLUSHINPROG; 1172 if (np->n_flag & NFLUSHWANT) { 1173 np->n_flag &= ~NFLUSHWANT; 1174 wakeup((caddr_t)&np->n_flag); 1175 } 1176 return (EINTR); 1177 } 1178 error = vinvalbuf(vp, flags, td, 0, slptimeo); 1179 } 1180 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 1181 if (np->n_flag & NFLUSHWANT) { 1182 np->n_flag &= ~NFLUSHWANT; 1183 wakeup((caddr_t)&np->n_flag); 1184 } 1185 return (0); 1186 } 1187 1188 /* 1189 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1190 * This is mainly to avoid queueing async I/O requests when the nfsiods 1191 * are all hung on a dead server. 1192 * 1193 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp 1194 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1195 */ 1196 int 1197 nfs_asyncio(struct buf *bp, struct thread *td) 1198 { 1199 struct nfsmount *nmp; 1200 int i; 1201 int gotiod; 1202 int slpflag = 0; 1203 int slptimeo = 0; 1204 int error; 1205 1206 /* 1207 * If no async daemons then return EIO to force caller to run the rpc 1208 * synchronously. 1209 */ 1210 if (nfs_numasync == 0) 1211 return (EIO); 1212 1213 nmp = VFSTONFS(bp->b_vp->v_mount); 1214 1215 /* 1216 * Commits are usually short and sweet so lets save some cpu and 1217 * leave the async daemons for more important rpc's (such as reads 1218 * and writes). 1219 */ 1220 if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT && 1221 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1222 return(EIO); 1223 } 1224 1225 again: 1226 if (nmp->nm_flag & NFSMNT_INT) 1227 slpflag = PCATCH; 1228 gotiod = FALSE; 1229 1230 /* 1231 * Find a free iod to process this request. 1232 */ 1233 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 1234 if (nfs_iodwant[i]) { 1235 /* 1236 * Found one, so wake it up and tell it which 1237 * mount to process. 1238 */ 1239 NFS_DPF(ASYNCIO, 1240 ("nfs_asyncio: waking iod %d for mount %p\n", 1241 i, nmp)); 1242 nfs_iodwant[i] = NULL; 1243 nfs_iodmount[i] = nmp; 1244 nmp->nm_bufqiods++; 1245 wakeup((caddr_t)&nfs_iodwant[i]); 1246 gotiod = TRUE; 1247 break; 1248 } 1249 1250 /* 1251 * If none are free, we may already have an iod working on this mount 1252 * point. If so, it will process our request. 1253 */ 1254 if (!gotiod) { 1255 if (nmp->nm_bufqiods > 0) { 1256 NFS_DPF(ASYNCIO, 1257 ("nfs_asyncio: %d iods are already processing mount %p\n", 1258 nmp->nm_bufqiods, nmp)); 1259 gotiod = TRUE; 1260 } 1261 } 1262 1263 /* 1264 * If we have an iod which can process the request, then queue 1265 * the buffer. 1266 */ 1267 if (gotiod) { 1268 /* 1269 * Ensure that the queue never grows too large. We still want 1270 * to asynchronize so we block rather then return EIO. 1271 */ 1272 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1273 NFS_DPF(ASYNCIO, 1274 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1275 nmp->nm_bufqwant = TRUE; 1276 error = tsleep(&nmp->nm_bufq, slpflag, 1277 "nfsaio", slptimeo); 1278 if (error) { 1279 if (nfs_sigintr(nmp, NULL, td)) 1280 return (EINTR); 1281 if (slpflag == PCATCH) { 1282 slpflag = 0; 1283 slptimeo = 2 * hz; 1284 } 1285 } 1286 /* 1287 * We might have lost our iod while sleeping, 1288 * so check and loop if nescessary. 1289 */ 1290 if (nmp->nm_bufqiods == 0) { 1291 NFS_DPF(ASYNCIO, 1292 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1293 goto again; 1294 } 1295 } 1296 1297 if ((bp->b_flags & B_READ) == 0) 1298 bp->b_flags |= B_WRITEINPROG; 1299 1300 BUF_KERNPROC(bp); 1301 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1302 nmp->nm_bufqlen++; 1303 return (0); 1304 } 1305 1306 /* 1307 * All the iods are busy on other mounts, so return EIO to 1308 * force the caller to process the i/o synchronously. 1309 */ 1310 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1311 return (EIO); 1312 } 1313 1314 /* 1315 * Do an I/O operation to/from a cache block. This may be called 1316 * synchronously or from an nfsiod. 1317 * 1318 * NOTE! TD MIGHT BE NULL 1319 */ 1320 int 1321 nfs_doio(struct buf *bp, struct thread *td) 1322 { 1323 struct uio *uiop; 1324 struct vnode *vp; 1325 struct nfsnode *np; 1326 struct nfsmount *nmp; 1327 int error = 0, iomode, must_commit = 0; 1328 struct uio uio; 1329 struct iovec io; 1330 1331 vp = bp->b_vp; 1332 np = VTONFS(vp); 1333 nmp = VFSTONFS(vp->v_mount); 1334 uiop = &uio; 1335 uiop->uio_iov = &io; 1336 uiop->uio_iovcnt = 1; 1337 uiop->uio_segflg = UIO_SYSSPACE; 1338 uiop->uio_td = td; 1339 1340 /* 1341 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 1342 * do this here so we do not have to do it in all the code that 1343 * calls us. 1344 */ 1345 bp->b_flags &= ~(B_ERROR | B_INVAL); 1346 1347 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1348 1349 /* 1350 * Historically, paging was done with physio, but no more. 1351 */ 1352 if (bp->b_flags & B_PHYS) { 1353 /* 1354 * ...though reading /dev/drum still gets us here. 1355 */ 1356 io.iov_len = uiop->uio_resid = bp->b_bcount; 1357 /* mapping was done by vmapbuf() */ 1358 io.iov_base = bp->b_data; 1359 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1360 if (bp->b_flags & B_READ) { 1361 uiop->uio_rw = UIO_READ; 1362 nfsstats.read_physios++; 1363 error = nfs_readrpc(vp, uiop); 1364 } else { 1365 int com; 1366 1367 iomode = NFSV3WRITE_DATASYNC; 1368 uiop->uio_rw = UIO_WRITE; 1369 nfsstats.write_physios++; 1370 error = nfs_writerpc(vp, uiop, &iomode, &com); 1371 } 1372 if (error) { 1373 bp->b_flags |= B_ERROR; 1374 bp->b_error = error; 1375 } 1376 } else if (bp->b_flags & B_READ) { 1377 io.iov_len = uiop->uio_resid = bp->b_bcount; 1378 io.iov_base = bp->b_data; 1379 uiop->uio_rw = UIO_READ; 1380 1381 switch (vp->v_type) { 1382 case VREG: 1383 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1384 nfsstats.read_bios++; 1385 error = nfs_readrpc(vp, uiop); 1386 1387 if (!error) { 1388 if (uiop->uio_resid) { 1389 /* 1390 * If we had a short read with no error, we must have 1391 * hit a file hole. We should zero-fill the remainder. 1392 * This can also occur if the server hits the file EOF. 1393 * 1394 * Holes used to be able to occur due to pending 1395 * writes, but that is not possible any longer. 1396 */ 1397 int nread = bp->b_bcount - uiop->uio_resid; 1398 int left = uiop->uio_resid; 1399 1400 if (left > 0) 1401 bzero((char *)bp->b_data + nread, left); 1402 uiop->uio_resid = 0; 1403 } 1404 } 1405 if (td && td->td_proc && (vp->v_flag & VTEXT) && 1406 (((nmp->nm_flag & NFSMNT_NQNFS) && 1407 NQNFS_CKINVALID(vp, np, ND_READ) && 1408 np->n_lrev != np->n_brev) || 1409 (!(nmp->nm_flag & NFSMNT_NQNFS) && 1410 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) { 1411 uprintf("Process killed due to text file modification\n"); 1412 psignal(td->td_proc, SIGKILL); 1413 PHOLD(td->td_proc); 1414 } 1415 break; 1416 case VLNK: 1417 uiop->uio_offset = (off_t)0; 1418 nfsstats.readlink_bios++; 1419 error = nfs_readlinkrpc(vp, uiop); 1420 break; 1421 case VDIR: 1422 nfsstats.readdir_bios++; 1423 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1424 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1425 error = nfs_readdirplusrpc(vp, uiop); 1426 if (error == NFSERR_NOTSUPP) 1427 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1428 } 1429 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1430 error = nfs_readdirrpc(vp, uiop); 1431 /* 1432 * end-of-directory sets B_INVAL but does not generate an 1433 * error. 1434 */ 1435 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1436 bp->b_flags |= B_INVAL; 1437 break; 1438 default: 1439 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1440 break; 1441 }; 1442 if (error) { 1443 bp->b_flags |= B_ERROR; 1444 bp->b_error = error; 1445 } 1446 } else { 1447 /* 1448 * If we only need to commit, try to commit 1449 */ 1450 if (bp->b_flags & B_NEEDCOMMIT) { 1451 int retv; 1452 off_t off; 1453 1454 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1455 bp->b_flags |= B_WRITEINPROG; 1456 retv = nfs_commit(bp->b_vp, off, 1457 bp->b_dirtyend - bp->b_dirtyoff, td); 1458 bp->b_flags &= ~B_WRITEINPROG; 1459 if (retv == 0) { 1460 bp->b_dirtyoff = bp->b_dirtyend = 0; 1461 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1462 bp->b_resid = 0; 1463 biodone(bp); 1464 return (0); 1465 } 1466 if (retv == NFSERR_STALEWRITEVERF) { 1467 nfs_clearcommit(bp->b_vp->v_mount); 1468 } 1469 } 1470 1471 /* 1472 * Setup for actual write 1473 */ 1474 1475 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1476 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1477 1478 if (bp->b_dirtyend > bp->b_dirtyoff) { 1479 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1480 - bp->b_dirtyoff; 1481 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1482 + bp->b_dirtyoff; 1483 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1484 uiop->uio_rw = UIO_WRITE; 1485 nfsstats.write_bios++; 1486 1487 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1488 iomode = NFSV3WRITE_UNSTABLE; 1489 else 1490 iomode = NFSV3WRITE_FILESYNC; 1491 1492 bp->b_flags |= B_WRITEINPROG; 1493 error = nfs_writerpc(vp, uiop, &iomode, &must_commit); 1494 1495 /* 1496 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1497 * to cluster the buffers needing commit. This will allow 1498 * the system to submit a single commit rpc for the whole 1499 * cluster. We can do this even if the buffer is not 100% 1500 * dirty (relative to the NFS blocksize), so we optimize the 1501 * append-to-file-case. 1502 * 1503 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1504 * cleared because write clustering only works for commit 1505 * rpc's, not for the data portion of the write). 1506 */ 1507 1508 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1509 bp->b_flags |= B_NEEDCOMMIT; 1510 if (bp->b_dirtyoff == 0 1511 && bp->b_dirtyend == bp->b_bcount) 1512 bp->b_flags |= B_CLUSTEROK; 1513 } else { 1514 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1515 } 1516 bp->b_flags &= ~B_WRITEINPROG; 1517 1518 /* 1519 * For an interrupted write, the buffer is still valid 1520 * and the write hasn't been pushed to the server yet, 1521 * so we can't set B_ERROR and report the interruption 1522 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1523 * is not relevant, so the rpc attempt is essentially 1524 * a noop. For the case of a V3 write rpc not being 1525 * committed to stable storage, the block is still 1526 * dirty and requires either a commit rpc or another 1527 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1528 * the block is reused. This is indicated by setting 1529 * the B_DELWRI and B_NEEDCOMMIT flags. 1530 * 1531 * If the buffer is marked B_PAGING, it does not reside on 1532 * the vp's paging queues so we cannot call bdirty(). The 1533 * bp in this case is not an NFS cache block so we should 1534 * be safe. XXX 1535 */ 1536 if (error == EINTR 1537 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1538 int s; 1539 1540 s = splbio(); 1541 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1542 if ((bp->b_flags & B_PAGING) == 0) { 1543 bdirty(bp); 1544 bp->b_flags &= ~B_DONE; 1545 } 1546 if (error && (bp->b_flags & B_ASYNC) == 0) 1547 bp->b_flags |= B_EINTR; 1548 splx(s); 1549 } else { 1550 if (error) { 1551 bp->b_flags |= B_ERROR; 1552 bp->b_error = np->n_error = error; 1553 np->n_flag |= NWRITEERR; 1554 } 1555 bp->b_dirtyoff = bp->b_dirtyend = 0; 1556 } 1557 } else { 1558 bp->b_resid = 0; 1559 biodone(bp); 1560 return (0); 1561 } 1562 } 1563 bp->b_resid = uiop->uio_resid; 1564 if (must_commit) 1565 nfs_clearcommit(vp->v_mount); 1566 biodone(bp); 1567 return (error); 1568 } 1569 1570 /* 1571 * Used to aid in handling ftruncate() operations on the NFS client side. 1572 * Truncation creates a number of special problems for NFS. We have to 1573 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1574 * we have to properly handle VM pages or (potentially dirty) buffers 1575 * that straddle the truncation point. 1576 */ 1577 1578 int 1579 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize) 1580 { 1581 struct nfsnode *np = VTONFS(vp); 1582 u_quad_t tsize = np->n_size; 1583 int biosize = vp->v_mount->mnt_stat.f_iosize; 1584 int error = 0; 1585 1586 np->n_size = nsize; 1587 1588 if (np->n_size < tsize) { 1589 struct buf *bp; 1590 daddr_t lbn; 1591 int bufsize; 1592 1593 /* 1594 * vtruncbuf() doesn't get the buffer overlapping the 1595 * truncation point. We may have a B_DELWRI and/or B_CACHE 1596 * buffer that now needs to be truncated. 1597 */ 1598 error = vtruncbuf(vp, td, nsize, biosize); 1599 lbn = nsize / biosize; 1600 bufsize = nsize & (biosize - 1); 1601 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1602 if (bp->b_dirtyoff > bp->b_bcount) 1603 bp->b_dirtyoff = bp->b_bcount; 1604 if (bp->b_dirtyend > bp->b_bcount) 1605 bp->b_dirtyend = bp->b_bcount; 1606 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1607 brelse(bp); 1608 } else { 1609 vnode_pager_setsize(vp, nsize); 1610 } 1611 return(error); 1612 } 1613 1614