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 1.20 2005/03/04 05:21:17 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 char *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_map_pagelist(&msf, pages, npages, 0); 152 KKASSERT(msf); 153 kva = msf_buf_kva(msf); 154 155 iov.iov_base = kva; 156 iov.iov_len = count; 157 uio.uio_iov = &iov; 158 uio.uio_iovcnt = 1; 159 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 160 uio.uio_resid = count; 161 uio.uio_segflg = UIO_SYSSPACE; 162 uio.uio_rw = UIO_READ; 163 uio.uio_td = td; 164 165 error = nfs_readrpc(vp, &uio); 166 msf_buf_free(msf); 167 168 if (error && (uio.uio_resid == count)) { 169 printf("nfs_getpages: error %d\n", error); 170 for (i = 0; i < npages; ++i) { 171 if (i != ap->a_reqpage) 172 vnode_pager_freepage(pages[i]); 173 } 174 return VM_PAGER_ERROR; 175 } 176 177 /* 178 * Calculate the number of bytes read and validate only that number 179 * of bytes. Note that due to pending writes, size may be 0. This 180 * does not mean that the remaining data is invalid! 181 */ 182 183 size = count - uio.uio_resid; 184 185 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 186 nextoff = toff + PAGE_SIZE; 187 m = pages[i]; 188 189 m->flags &= ~PG_ZERO; 190 191 if (nextoff <= size) { 192 /* 193 * Read operation filled an entire page 194 */ 195 m->valid = VM_PAGE_BITS_ALL; 196 vm_page_undirty(m); 197 } else if (size > toff) { 198 /* 199 * Read operation filled a partial page. 200 */ 201 m->valid = 0; 202 vm_page_set_validclean(m, 0, size - toff); 203 /* handled by vm_fault now */ 204 /* vm_page_zero_invalid(m, TRUE); */ 205 } else { 206 /* 207 * Read operation was short. If no error occured 208 * we may have hit a zero-fill section. We simply 209 * leave valid set to 0. 210 */ 211 ; 212 } 213 if (i != ap->a_reqpage) { 214 /* 215 * Whether or not to leave the page activated is up in 216 * the air, but we should put the page on a page queue 217 * somewhere (it already is in the object). Result: 218 * It appears that emperical results show that 219 * deactivating pages is best. 220 */ 221 222 /* 223 * Just in case someone was asking for this page we 224 * now tell them that it is ok to use. 225 */ 226 if (!error) { 227 if (m->flags & PG_WANTED) 228 vm_page_activate(m); 229 else 230 vm_page_deactivate(m); 231 vm_page_wakeup(m); 232 } else { 233 vnode_pager_freepage(m); 234 } 235 } 236 } 237 return 0; 238 } 239 240 /* 241 * Vnode op for VM putpages. 242 * 243 * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync, 244 * int *a_rtvals, vm_ooffset_t a_offset) 245 */ 246 int 247 nfs_putpages(struct vop_putpages_args *ap) 248 { 249 struct thread *td = curthread; 250 struct uio uio; 251 struct iovec iov; 252 char *kva; 253 int iomode, must_commit, i, error, npages, count; 254 off_t offset; 255 int *rtvals; 256 struct vnode *vp; 257 struct nfsmount *nmp; 258 struct nfsnode *np; 259 vm_page_t *pages; 260 struct msf_buf *msf; 261 262 vp = ap->a_vp; 263 np = VTONFS(vp); 264 nmp = VFSTONFS(vp->v_mount); 265 pages = ap->a_m; 266 count = ap->a_count; 267 rtvals = ap->a_rtvals; 268 npages = btoc(count); 269 offset = IDX_TO_OFF(pages[0]->pindex); 270 271 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 272 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 273 (void)nfs_fsinfo(nmp, vp, td); 274 275 for (i = 0; i < npages; i++) { 276 rtvals[i] = VM_PAGER_AGAIN; 277 } 278 279 /* 280 * When putting pages, do not extend file past EOF. 281 */ 282 283 if (offset + count > np->n_size) { 284 count = np->n_size - offset; 285 if (count < 0) 286 count = 0; 287 } 288 289 /* 290 * Use an MSF_BUF as a medium to retrieve data from the pages. 291 */ 292 msf_map_pagelist(&msf, pages, npages, 0); 293 KKASSERT(msf); 294 kva = msf_buf_kva(msf); 295 296 iov.iov_base = kva; 297 iov.iov_len = count; 298 uio.uio_iov = &iov; 299 uio.uio_iovcnt = 1; 300 uio.uio_offset = offset; 301 uio.uio_resid = count; 302 uio.uio_segflg = UIO_SYSSPACE; 303 uio.uio_rw = UIO_WRITE; 304 uio.uio_td = td; 305 306 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 307 iomode = NFSV3WRITE_UNSTABLE; 308 else 309 iomode = NFSV3WRITE_FILESYNC; 310 311 error = nfs_writerpc(vp, &uio, &iomode, &must_commit); 312 313 msf_buf_free(msf); 314 315 if (!error) { 316 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 317 for (i = 0; i < nwritten; i++) { 318 rtvals[i] = VM_PAGER_OK; 319 vm_page_undirty(pages[i]); 320 } 321 if (must_commit) 322 nfs_clearcommit(vp->v_mount); 323 } 324 return rtvals[0]; 325 } 326 327 /* 328 * Vnode op for read using bio 329 */ 330 int 331 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag) 332 { 333 struct nfsnode *np = VTONFS(vp); 334 int biosize, i; 335 struct buf *bp = 0, *rabp; 336 struct vattr vattr; 337 struct thread *td; 338 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 339 daddr_t lbn, rabn; 340 int bcount; 341 int seqcount; 342 int nra, error = 0, n = 0, on = 0; 343 344 #ifdef DIAGNOSTIC 345 if (uio->uio_rw != UIO_READ) 346 panic("nfs_read mode"); 347 #endif 348 if (uio->uio_resid == 0) 349 return (0); 350 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 351 return (EINVAL); 352 td = uio->uio_td; 353 354 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 355 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 356 (void)nfs_fsinfo(nmp, vp, td); 357 if (vp->v_type != VDIR && 358 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 359 return (EFBIG); 360 biosize = vp->v_mount->mnt_stat.f_iosize; 361 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 362 /* 363 * For nfs, cache consistency can only be maintained approximately. 364 * Although RFC1094 does not specify the criteria, the following is 365 * believed to be compatible with the reference port. 366 * For nqnfs, full cache consistency is maintained within the loop. 367 * For nfs: 368 * If the file's modify time on the server has changed since the 369 * last read rpc or you have written to the file, 370 * you may have lost data cache consistency with the 371 * server, so flush all of the file's data out of the cache. 372 * Then force a getattr rpc to ensure that you have up to date 373 * attributes. 374 * NB: This implies that cache data can be read when up to 375 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 376 * attributes this could be forced by setting n_attrstamp to 0 before 377 * the VOP_GETATTR() call. 378 */ 379 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) { 380 if (np->n_flag & NMODIFIED) { 381 if (vp->v_type != VREG) { 382 if (vp->v_type != VDIR) 383 panic("nfs: bioread, not dir"); 384 nfs_invaldir(vp); 385 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 386 if (error) 387 return (error); 388 } 389 np->n_attrstamp = 0; 390 error = VOP_GETATTR(vp, &vattr, td); 391 if (error) 392 return (error); 393 np->n_mtime = vattr.va_mtime.tv_sec; 394 } else { 395 error = VOP_GETATTR(vp, &vattr, td); 396 if (error) 397 return (error); 398 if ((np->n_flag & NSIZECHANGED) 399 || 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, td, 1); 403 if (error) 404 return (error); 405 np->n_mtime = vattr.va_mtime.tv_sec; 406 np->n_flag &= ~NSIZECHANGED; 407 } 408 } 409 } 410 do { 411 412 /* 413 * Get a valid lease. If cached data is stale, flush it. 414 */ 415 if (nmp->nm_flag & NFSMNT_NQNFS) { 416 if (NQNFS_CKINVALID(vp, np, ND_READ)) { 417 do { 418 error = nqnfs_getlease(vp, ND_READ, td); 419 } while (error == NQNFS_EXPIRED); 420 if (error) 421 return (error); 422 if (np->n_lrev != np->n_brev || 423 (np->n_flag & NQNFSNONCACHE) || 424 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) { 425 if (vp->v_type == VDIR) 426 nfs_invaldir(vp); 427 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 428 if (error) 429 return (error); 430 np->n_brev = np->n_lrev; 431 } 432 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) { 433 nfs_invaldir(vp); 434 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 435 if (error) 436 return (error); 437 } 438 } 439 if (np->n_flag & NQNFSNONCACHE) { 440 switch (vp->v_type) { 441 case VREG: 442 return (nfs_readrpc(vp, uio)); 443 case VLNK: 444 return (nfs_readlinkrpc(vp, uio)); 445 case VDIR: 446 break; 447 default: 448 printf(" NQNFSNONCACHE: type %x unexpected\n", 449 vp->v_type); 450 }; 451 } 452 switch (vp->v_type) { 453 case VREG: 454 nfsstats.biocache_reads++; 455 lbn = uio->uio_offset / biosize; 456 on = uio->uio_offset & (biosize - 1); 457 458 /* 459 * Start the read ahead(s), as required. 460 */ 461 if (nfs_numasync > 0 && nmp->nm_readahead > 0) { 462 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 463 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 464 rabn = lbn + 1 + nra; 465 if (!incore(vp, rabn)) { 466 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 467 if (!rabp) 468 return (EINTR); 469 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 470 rabp->b_flags |= (B_READ | B_ASYNC); 471 vfs_busy_pages(rabp, 0); 472 if (nfs_asyncio(rabp, td)) { 473 rabp->b_flags |= B_INVAL|B_ERROR; 474 vfs_unbusy_pages(rabp); 475 brelse(rabp); 476 break; 477 } 478 } else { 479 brelse(rabp); 480 } 481 } 482 } 483 } 484 485 /* 486 * Obtain the buffer cache block. Figure out the buffer size 487 * when we are at EOF. If we are modifying the size of the 488 * buffer based on an EOF condition we need to hold 489 * nfs_rslock() through obtaining the buffer to prevent 490 * a potential writer-appender from messing with n_size. 491 * Otherwise we may accidently truncate the buffer and 492 * lose dirty data. 493 * 494 * Note that bcount is *not* DEV_BSIZE aligned. 495 */ 496 497 again: 498 bcount = biosize; 499 if ((off_t)lbn * biosize >= np->n_size) { 500 bcount = 0; 501 } else if ((off_t)(lbn + 1) * biosize > np->n_size) { 502 bcount = np->n_size - (off_t)lbn * biosize; 503 } 504 if (bcount != biosize) { 505 switch(nfs_rslock(np, td)) { 506 case ENOLCK: 507 goto again; 508 /* not reached */ 509 case EINTR: 510 case ERESTART: 511 return(EINTR); 512 /* not reached */ 513 default: 514 break; 515 } 516 } 517 518 bp = nfs_getcacheblk(vp, lbn, bcount, td); 519 520 if (bcount != biosize) 521 nfs_rsunlock(np, td); 522 if (!bp) 523 return (EINTR); 524 525 /* 526 * If B_CACHE is not set, we must issue the read. If this 527 * fails, we return an error. 528 */ 529 530 if ((bp->b_flags & B_CACHE) == 0) { 531 bp->b_flags |= B_READ; 532 vfs_busy_pages(bp, 0); 533 error = nfs_doio(bp, td); 534 if (error) { 535 brelse(bp); 536 return (error); 537 } 538 } 539 540 /* 541 * on is the offset into the current bp. Figure out how many 542 * bytes we can copy out of the bp. Note that bcount is 543 * NOT DEV_BSIZE aligned. 544 * 545 * Then figure out how many bytes we can copy into the uio. 546 */ 547 548 n = 0; 549 if (on < bcount) 550 n = min((unsigned)(bcount - on), uio->uio_resid); 551 break; 552 case VLNK: 553 nfsstats.biocache_readlinks++; 554 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 555 if (!bp) 556 return (EINTR); 557 if ((bp->b_flags & B_CACHE) == 0) { 558 bp->b_flags |= B_READ; 559 vfs_busy_pages(bp, 0); 560 error = nfs_doio(bp, td); 561 if (error) { 562 bp->b_flags |= B_ERROR; 563 brelse(bp); 564 return (error); 565 } 566 } 567 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 568 on = 0; 569 break; 570 case VDIR: 571 nfsstats.biocache_readdirs++; 572 if (np->n_direofoffset 573 && uio->uio_offset >= np->n_direofoffset) { 574 return (0); 575 } 576 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 577 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 578 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 579 if (!bp) 580 return (EINTR); 581 if ((bp->b_flags & B_CACHE) == 0) { 582 bp->b_flags |= B_READ; 583 vfs_busy_pages(bp, 0); 584 error = nfs_doio(bp, td); 585 if (error) { 586 brelse(bp); 587 } 588 while (error == NFSERR_BAD_COOKIE) { 589 printf("got bad cookie vp %p bp %p\n", vp, bp); 590 nfs_invaldir(vp); 591 error = nfs_vinvalbuf(vp, 0, td, 1); 592 /* 593 * Yuck! The directory has been modified on the 594 * server. The only way to get the block is by 595 * reading from the beginning to get all the 596 * offset cookies. 597 * 598 * Leave the last bp intact unless there is an error. 599 * Loop back up to the while if the error is another 600 * NFSERR_BAD_COOKIE (double yuch!). 601 */ 602 for (i = 0; i <= lbn && !error; i++) { 603 if (np->n_direofoffset 604 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 605 return (0); 606 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 607 if (!bp) 608 return (EINTR); 609 if ((bp->b_flags & B_CACHE) == 0) { 610 bp->b_flags |= B_READ; 611 vfs_busy_pages(bp, 0); 612 error = nfs_doio(bp, td); 613 /* 614 * no error + B_INVAL == directory EOF, 615 * use the block. 616 */ 617 if (error == 0 && (bp->b_flags & B_INVAL)) 618 break; 619 } 620 /* 621 * An error will throw away the block and the 622 * for loop will break out. If no error and this 623 * is not the block we want, we throw away the 624 * block and go for the next one via the for loop. 625 */ 626 if (error || i < lbn) 627 brelse(bp); 628 } 629 } 630 /* 631 * The above while is repeated if we hit another cookie 632 * error. If we hit an error and it wasn't a cookie error, 633 * we give up. 634 */ 635 if (error) 636 return (error); 637 } 638 639 /* 640 * If not eof and read aheads are enabled, start one. 641 * (You need the current block first, so that you have the 642 * directory offset cookie of the next block.) 643 */ 644 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 645 (bp->b_flags & B_INVAL) == 0 && 646 (np->n_direofoffset == 0 || 647 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 648 !(np->n_flag & NQNFSNONCACHE) && 649 !incore(vp, lbn + 1)) { 650 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 651 if (rabp) { 652 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 653 rabp->b_flags |= (B_READ | B_ASYNC); 654 vfs_busy_pages(rabp, 0); 655 if (nfs_asyncio(rabp, td)) { 656 rabp->b_flags |= B_INVAL|B_ERROR; 657 vfs_unbusy_pages(rabp); 658 brelse(rabp); 659 } 660 } else { 661 brelse(rabp); 662 } 663 } 664 } 665 /* 666 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 667 * chopped for the EOF condition, we cannot tell how large 668 * NFS directories are going to be until we hit EOF. So 669 * an NFS directory buffer is *not* chopped to its EOF. Now, 670 * it just so happens that b_resid will effectively chop it 671 * to EOF. *BUT* this information is lost if the buffer goes 672 * away and is reconstituted into a B_CACHE state ( due to 673 * being VMIO ) later. So we keep track of the directory eof 674 * in np->n_direofoffset and chop it off as an extra step 675 * right here. 676 */ 677 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 678 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 679 n = np->n_direofoffset - uio->uio_offset; 680 break; 681 default: 682 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 683 break; 684 }; 685 686 if (n > 0) { 687 error = uiomove(bp->b_data + on, (int)n, uio); 688 } 689 switch (vp->v_type) { 690 case VREG: 691 break; 692 case VLNK: 693 n = 0; 694 break; 695 case VDIR: 696 /* 697 * Invalidate buffer if caching is disabled, forcing a 698 * re-read from the remote later. 699 */ 700 if (np->n_flag & NQNFSNONCACHE) 701 bp->b_flags |= B_INVAL; 702 break; 703 default: 704 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 705 } 706 brelse(bp); 707 } while (error == 0 && uio->uio_resid > 0 && n > 0); 708 return (error); 709 } 710 711 /* 712 * Vnode op for write using bio 713 * 714 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag, 715 * struct ucred *a_cred) 716 */ 717 int 718 nfs_write(struct vop_write_args *ap) 719 { 720 int biosize; 721 struct uio *uio = ap->a_uio; 722 struct thread *td = uio->uio_td; 723 struct vnode *vp = ap->a_vp; 724 struct nfsnode *np = VTONFS(vp); 725 int ioflag = ap->a_ioflag; 726 struct buf *bp; 727 struct vattr vattr; 728 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 729 daddr_t lbn; 730 int bcount; 731 int n, on, error = 0, iomode, must_commit; 732 int haverslock = 0; 733 734 #ifdef DIAGNOSTIC 735 if (uio->uio_rw != UIO_WRITE) 736 panic("nfs_write mode"); 737 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 738 panic("nfs_write proc"); 739 #endif 740 if (vp->v_type != VREG) 741 return (EIO); 742 if (np->n_flag & NWRITEERR) { 743 np->n_flag &= ~NWRITEERR; 744 return (np->n_error); 745 } 746 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 747 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 748 (void)nfs_fsinfo(nmp, vp, td); 749 750 /* 751 * Synchronously flush pending buffers if we are in synchronous 752 * mode or if we are appending. 753 */ 754 if (ioflag & (IO_APPEND | IO_SYNC)) { 755 if (np->n_flag & NMODIFIED) { 756 np->n_attrstamp = 0; 757 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 758 if (error) 759 return (error); 760 } 761 } 762 763 /* 764 * If IO_APPEND then load uio_offset. We restart here if we cannot 765 * get the append lock. 766 */ 767 restart: 768 if (ioflag & IO_APPEND) { 769 np->n_attrstamp = 0; 770 error = VOP_GETATTR(vp, &vattr, td); 771 if (error) 772 return (error); 773 uio->uio_offset = np->n_size; 774 } 775 776 if (uio->uio_offset < 0) 777 return (EINVAL); 778 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 779 return (EFBIG); 780 if (uio->uio_resid == 0) 781 return (0); 782 783 /* 784 * We need to obtain the rslock if we intend to modify np->n_size 785 * in order to guarentee the append point with multiple contending 786 * writers, to guarentee that no other appenders modify n_size 787 * while we are trying to obtain a truncated buffer (i.e. to avoid 788 * accidently truncating data written by another appender due to 789 * the race), and to ensure that the buffer is populated prior to 790 * our extending of the file. We hold rslock through the entire 791 * operation. 792 * 793 * Note that we do not synchronize the case where someone truncates 794 * the file while we are appending to it because attempting to lock 795 * this case may deadlock other parts of the system unexpectedly. 796 */ 797 if ((ioflag & IO_APPEND) || 798 uio->uio_offset + uio->uio_resid > np->n_size) { 799 switch(nfs_rslock(np, td)) { 800 case ENOLCK: 801 goto restart; 802 /* not reached */ 803 case EINTR: 804 case ERESTART: 805 return(EINTR); 806 /* not reached */ 807 default: 808 break; 809 } 810 haverslock = 1; 811 } 812 813 /* 814 * Maybe this should be above the vnode op call, but so long as 815 * file servers have no limits, i don't think it matters 816 */ 817 if (td->td_proc && uio->uio_offset + uio->uio_resid > 818 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 819 psignal(td->td_proc, SIGXFSZ); 820 if (haverslock) 821 nfs_rsunlock(np, td); 822 return (EFBIG); 823 } 824 825 biosize = vp->v_mount->mnt_stat.f_iosize; 826 827 do { 828 /* 829 * Check for a valid write lease. 830 */ 831 if ((nmp->nm_flag & NFSMNT_NQNFS) && 832 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 833 do { 834 error = nqnfs_getlease(vp, ND_WRITE, td); 835 } while (error == NQNFS_EXPIRED); 836 if (error) 837 break; 838 if (np->n_lrev != np->n_brev || 839 (np->n_flag & NQNFSNONCACHE)) { 840 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 841 if (error) 842 break; 843 np->n_brev = np->n_lrev; 844 } 845 } 846 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 847 iomode = NFSV3WRITE_FILESYNC; 848 error = nfs_writerpc(vp, uio, &iomode, &must_commit); 849 if (must_commit) 850 nfs_clearcommit(vp->v_mount); 851 break; 852 } 853 nfsstats.biocache_writes++; 854 lbn = uio->uio_offset / biosize; 855 on = uio->uio_offset & (biosize-1); 856 n = min((unsigned)(biosize - on), uio->uio_resid); 857 again: 858 /* 859 * Handle direct append and file extension cases, calculate 860 * unaligned buffer size. 861 */ 862 863 if (uio->uio_offset == np->n_size && n) { 864 /* 865 * Get the buffer (in its pre-append state to maintain 866 * B_CACHE if it was previously set). Resize the 867 * nfsnode after we have locked the buffer to prevent 868 * readers from reading garbage. 869 */ 870 bcount = on; 871 bp = nfs_getcacheblk(vp, lbn, bcount, td); 872 873 if (bp != NULL) { 874 long save; 875 876 np->n_size = uio->uio_offset + n; 877 np->n_flag |= NMODIFIED; 878 vnode_pager_setsize(vp, np->n_size); 879 880 save = bp->b_flags & B_CACHE; 881 bcount += n; 882 allocbuf(bp, bcount); 883 bp->b_flags |= save; 884 } 885 } else { 886 /* 887 * Obtain the locked cache block first, and then 888 * adjust the file's size as appropriate. 889 */ 890 bcount = on + n; 891 if ((off_t)lbn * biosize + bcount < np->n_size) { 892 if ((off_t)(lbn + 1) * biosize < np->n_size) 893 bcount = biosize; 894 else 895 bcount = np->n_size - (off_t)lbn * biosize; 896 } 897 bp = nfs_getcacheblk(vp, lbn, bcount, td); 898 if (uio->uio_offset + n > np->n_size) { 899 np->n_size = uio->uio_offset + n; 900 np->n_flag |= NMODIFIED; 901 vnode_pager_setsize(vp, np->n_size); 902 } 903 } 904 905 if (!bp) { 906 error = EINTR; 907 break; 908 } 909 910 /* 911 * Issue a READ if B_CACHE is not set. In special-append 912 * mode, B_CACHE is based on the buffer prior to the write 913 * op and is typically set, avoiding the read. If a read 914 * is required in special append mode, the server will 915 * probably send us a short-read since we extended the file 916 * on our end, resulting in b_resid == 0 and, thusly, 917 * B_CACHE getting set. 918 * 919 * We can also avoid issuing the read if the write covers 920 * the entire buffer. We have to make sure the buffer state 921 * is reasonable in this case since we will not be initiating 922 * I/O. See the comments in kern/vfs_bio.c's getblk() for 923 * more information. 924 * 925 * B_CACHE may also be set due to the buffer being cached 926 * normally. 927 */ 928 929 if (on == 0 && n == bcount) { 930 bp->b_flags |= B_CACHE; 931 bp->b_flags &= ~(B_ERROR | B_INVAL); 932 } 933 934 if ((bp->b_flags & B_CACHE) == 0) { 935 bp->b_flags |= B_READ; 936 vfs_busy_pages(bp, 0); 937 error = nfs_doio(bp, td); 938 if (error) { 939 brelse(bp); 940 break; 941 } 942 } 943 if (!bp) { 944 error = EINTR; 945 break; 946 } 947 np->n_flag |= NMODIFIED; 948 949 /* 950 * If dirtyend exceeds file size, chop it down. This should 951 * not normally occur but there is an append race where it 952 * might occur XXX, so we log it. 953 * 954 * If the chopping creates a reverse-indexed or degenerate 955 * situation with dirtyoff/end, we 0 both of them. 956 */ 957 958 if (bp->b_dirtyend > bcount) { 959 printf("NFS append race @%lx:%d\n", 960 (long)bp->b_blkno * DEV_BSIZE, 961 bp->b_dirtyend - bcount); 962 bp->b_dirtyend = bcount; 963 } 964 965 if (bp->b_dirtyoff >= bp->b_dirtyend) 966 bp->b_dirtyoff = bp->b_dirtyend = 0; 967 968 /* 969 * If the new write will leave a contiguous dirty 970 * area, just update the b_dirtyoff and b_dirtyend, 971 * otherwise force a write rpc of the old dirty area. 972 * 973 * While it is possible to merge discontiguous writes due to 974 * our having a B_CACHE buffer ( and thus valid read data 975 * for the hole), we don't because it could lead to 976 * significant cache coherency problems with multiple clients, 977 * especially if locking is implemented later on. 978 * 979 * as an optimization we could theoretically maintain 980 * a linked list of discontinuous areas, but we would still 981 * have to commit them separately so there isn't much 982 * advantage to it except perhaps a bit of asynchronization. 983 */ 984 985 if (bp->b_dirtyend > 0 && 986 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 987 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) { 988 error = EINTR; 989 break; 990 } 991 goto again; 992 } 993 994 /* 995 * Check for valid write lease and get one as required. 996 * In case getblk() and/or bwrite() delayed us. 997 */ 998 if ((nmp->nm_flag & NFSMNT_NQNFS) && 999 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 1000 do { 1001 error = nqnfs_getlease(vp, ND_WRITE, td); 1002 } while (error == NQNFS_EXPIRED); 1003 if (error) { 1004 brelse(bp); 1005 break; 1006 } 1007 if (np->n_lrev != np->n_brev || 1008 (np->n_flag & NQNFSNONCACHE)) { 1009 brelse(bp); 1010 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 1011 if (error) 1012 break; 1013 np->n_brev = np->n_lrev; 1014 goto again; 1015 } 1016 } 1017 1018 error = uiomove((char *)bp->b_data + on, n, uio); 1019 1020 /* 1021 * Since this block is being modified, it must be written 1022 * again and not just committed. Since write clustering does 1023 * not work for the stage 1 data write, only the stage 2 1024 * commit rpc, we have to clear B_CLUSTEROK as well. 1025 */ 1026 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1027 1028 if (error) { 1029 bp->b_flags |= B_ERROR; 1030 brelse(bp); 1031 break; 1032 } 1033 1034 /* 1035 * Only update dirtyoff/dirtyend if not a degenerate 1036 * condition. 1037 */ 1038 if (n) { 1039 if (bp->b_dirtyend > 0) { 1040 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1041 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1042 } else { 1043 bp->b_dirtyoff = on; 1044 bp->b_dirtyend = on + n; 1045 } 1046 vfs_bio_set_validclean(bp, on, n); 1047 } 1048 /* 1049 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN 1050 * filesystem). XXX also use for loopback NFS mounts. 1051 */ 1052 if (ioflag & IO_NOWDRAIN) 1053 bp->b_flags |= B_NOWDRAIN; 1054 1055 /* 1056 * If the lease is non-cachable or IO_SYNC do bwrite(). 1057 * 1058 * IO_INVAL appears to be unused. The idea appears to be 1059 * to turn off caching in this case. Very odd. XXX 1060 */ 1061 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 1062 if (ioflag & IO_INVAL) 1063 bp->b_flags |= B_NOCACHE; 1064 error = VOP_BWRITE(bp->b_vp, bp); 1065 if (error) 1066 break; 1067 if (np->n_flag & NQNFSNONCACHE) { 1068 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 1069 if (error) 1070 break; 1071 } 1072 } else if ((n + on) == biosize && 1073 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 1074 bp->b_flags |= B_ASYNC; 1075 (void)nfs_writebp(bp, 0, 0); 1076 } else { 1077 bdwrite(bp); 1078 } 1079 } while (uio->uio_resid > 0 && n > 0); 1080 1081 if (haverslock) 1082 nfs_rsunlock(np, td); 1083 1084 return (error); 1085 } 1086 1087 /* 1088 * Get an nfs cache block. 1089 * 1090 * Allocate a new one if the block isn't currently in the cache 1091 * and return the block marked busy. If the calling process is 1092 * interrupted by a signal for an interruptible mount point, return 1093 * NULL. 1094 * 1095 * The caller must carefully deal with the possible B_INVAL state of 1096 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1097 * indirectly), so synchronous reads can be issued without worrying about 1098 * the B_INVAL state. We have to be a little more careful when dealing 1099 * with writes (see comments in nfs_write()) when extending a file past 1100 * its EOF. 1101 */ 1102 static struct buf * 1103 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1104 { 1105 struct buf *bp; 1106 struct mount *mp; 1107 struct nfsmount *nmp; 1108 1109 mp = vp->v_mount; 1110 nmp = VFSTONFS(mp); 1111 1112 if (nmp->nm_flag & NFSMNT_INT) { 1113 bp = getblk(vp, bn, size, PCATCH, 0); 1114 while (bp == (struct buf *)0) { 1115 if (nfs_sigintr(nmp, (struct nfsreq *)0, td)) 1116 return ((struct buf *)0); 1117 bp = getblk(vp, bn, size, 0, 2 * hz); 1118 } 1119 } else { 1120 bp = getblk(vp, bn, size, 0, 0); 1121 } 1122 1123 if (vp->v_type == VREG) { 1124 int biosize; 1125 1126 biosize = mp->mnt_stat.f_iosize; 1127 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1128 } 1129 return (bp); 1130 } 1131 1132 /* 1133 * Flush and invalidate all dirty buffers. If another process is already 1134 * doing the flush, just wait for completion. 1135 */ 1136 int 1137 nfs_vinvalbuf(struct vnode *vp, int flags, 1138 struct thread *td, int intrflg) 1139 { 1140 struct nfsnode *np = VTONFS(vp); 1141 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1142 int error = 0, slpflag, slptimeo; 1143 1144 if (vp->v_flag & VRECLAIMED) 1145 return (0); 1146 1147 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1148 intrflg = 0; 1149 if (intrflg) { 1150 slpflag = PCATCH; 1151 slptimeo = 2 * hz; 1152 } else { 1153 slpflag = 0; 1154 slptimeo = 0; 1155 } 1156 /* 1157 * First wait for any other process doing a flush to complete. 1158 */ 1159 while (np->n_flag & NFLUSHINPROG) { 1160 np->n_flag |= NFLUSHWANT; 1161 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo); 1162 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) 1163 return (EINTR); 1164 } 1165 1166 /* 1167 * Now, flush as required. 1168 */ 1169 np->n_flag |= NFLUSHINPROG; 1170 error = vinvalbuf(vp, flags, td, slpflag, 0); 1171 while (error) { 1172 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) { 1173 np->n_flag &= ~NFLUSHINPROG; 1174 if (np->n_flag & NFLUSHWANT) { 1175 np->n_flag &= ~NFLUSHWANT; 1176 wakeup((caddr_t)&np->n_flag); 1177 } 1178 return (EINTR); 1179 } 1180 error = vinvalbuf(vp, flags, td, 0, slptimeo); 1181 } 1182 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 1183 if (np->n_flag & NFLUSHWANT) { 1184 np->n_flag &= ~NFLUSHWANT; 1185 wakeup((caddr_t)&np->n_flag); 1186 } 1187 return (0); 1188 } 1189 1190 /* 1191 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1192 * This is mainly to avoid queueing async I/O requests when the nfsiods 1193 * are all hung on a dead server. 1194 * 1195 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp 1196 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1197 */ 1198 int 1199 nfs_asyncio(struct buf *bp, struct thread *td) 1200 { 1201 struct nfsmount *nmp; 1202 int i; 1203 int gotiod; 1204 int slpflag = 0; 1205 int slptimeo = 0; 1206 int error; 1207 1208 /* 1209 * If no async daemons then return EIO to force caller to run the rpc 1210 * synchronously. 1211 */ 1212 if (nfs_numasync == 0) 1213 return (EIO); 1214 1215 nmp = VFSTONFS(bp->b_vp->v_mount); 1216 1217 /* 1218 * Commits are usually short and sweet so lets save some cpu and 1219 * leave the async daemons for more important rpc's (such as reads 1220 * and writes). 1221 */ 1222 if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT && 1223 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1224 return(EIO); 1225 } 1226 1227 again: 1228 if (nmp->nm_flag & NFSMNT_INT) 1229 slpflag = PCATCH; 1230 gotiod = FALSE; 1231 1232 /* 1233 * Find a free iod to process this request. 1234 */ 1235 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 1236 if (nfs_iodwant[i]) { 1237 /* 1238 * Found one, so wake it up and tell it which 1239 * mount to process. 1240 */ 1241 NFS_DPF(ASYNCIO, 1242 ("nfs_asyncio: waking iod %d for mount %p\n", 1243 i, nmp)); 1244 nfs_iodwant[i] = NULL; 1245 nfs_iodmount[i] = nmp; 1246 nmp->nm_bufqiods++; 1247 wakeup((caddr_t)&nfs_iodwant[i]); 1248 gotiod = TRUE; 1249 break; 1250 } 1251 1252 /* 1253 * If none are free, we may already have an iod working on this mount 1254 * point. If so, it will process our request. 1255 */ 1256 if (!gotiod) { 1257 if (nmp->nm_bufqiods > 0) { 1258 NFS_DPF(ASYNCIO, 1259 ("nfs_asyncio: %d iods are already processing mount %p\n", 1260 nmp->nm_bufqiods, nmp)); 1261 gotiod = TRUE; 1262 } 1263 } 1264 1265 /* 1266 * If we have an iod which can process the request, then queue 1267 * the buffer. 1268 */ 1269 if (gotiod) { 1270 /* 1271 * Ensure that the queue never grows too large. We still want 1272 * to asynchronize so we block rather then return EIO. 1273 */ 1274 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1275 NFS_DPF(ASYNCIO, 1276 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1277 nmp->nm_bufqwant = TRUE; 1278 error = tsleep(&nmp->nm_bufq, slpflag, 1279 "nfsaio", slptimeo); 1280 if (error) { 1281 if (nfs_sigintr(nmp, NULL, td)) 1282 return (EINTR); 1283 if (slpflag == PCATCH) { 1284 slpflag = 0; 1285 slptimeo = 2 * hz; 1286 } 1287 } 1288 /* 1289 * We might have lost our iod while sleeping, 1290 * so check and loop if nescessary. 1291 */ 1292 if (nmp->nm_bufqiods == 0) { 1293 NFS_DPF(ASYNCIO, 1294 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1295 goto again; 1296 } 1297 } 1298 1299 if ((bp->b_flags & B_READ) == 0) 1300 bp->b_flags |= B_WRITEINPROG; 1301 1302 BUF_KERNPROC(bp); 1303 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1304 nmp->nm_bufqlen++; 1305 return (0); 1306 } 1307 1308 /* 1309 * All the iods are busy on other mounts, so return EIO to 1310 * force the caller to process the i/o synchronously. 1311 */ 1312 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1313 return (EIO); 1314 } 1315 1316 /* 1317 * Do an I/O operation to/from a cache block. This may be called 1318 * synchronously or from an nfsiod. 1319 * 1320 * NOTE! TD MIGHT BE NULL 1321 */ 1322 int 1323 nfs_doio(struct buf *bp, struct thread *td) 1324 { 1325 struct uio *uiop; 1326 struct vnode *vp; 1327 struct nfsnode *np; 1328 struct nfsmount *nmp; 1329 int error = 0, iomode, must_commit = 0; 1330 struct uio uio; 1331 struct iovec io; 1332 1333 vp = bp->b_vp; 1334 np = VTONFS(vp); 1335 nmp = VFSTONFS(vp->v_mount); 1336 uiop = &uio; 1337 uiop->uio_iov = &io; 1338 uiop->uio_iovcnt = 1; 1339 uiop->uio_segflg = UIO_SYSSPACE; 1340 uiop->uio_td = td; 1341 1342 /* 1343 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 1344 * do this here so we do not have to do it in all the code that 1345 * calls us. 1346 */ 1347 bp->b_flags &= ~(B_ERROR | B_INVAL); 1348 1349 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1350 1351 /* 1352 * Historically, paging was done with physio, but no more. 1353 */ 1354 if (bp->b_flags & B_PHYS) { 1355 /* 1356 * ...though reading /dev/drum still gets us here. 1357 */ 1358 io.iov_len = uiop->uio_resid = bp->b_bcount; 1359 /* mapping was done by vmapbuf() */ 1360 io.iov_base = bp->b_data; 1361 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1362 if (bp->b_flags & B_READ) { 1363 uiop->uio_rw = UIO_READ; 1364 nfsstats.read_physios++; 1365 error = nfs_readrpc(vp, uiop); 1366 } else { 1367 int com; 1368 1369 iomode = NFSV3WRITE_DATASYNC; 1370 uiop->uio_rw = UIO_WRITE; 1371 nfsstats.write_physios++; 1372 error = nfs_writerpc(vp, uiop, &iomode, &com); 1373 } 1374 if (error) { 1375 bp->b_flags |= B_ERROR; 1376 bp->b_error = error; 1377 } 1378 } else if (bp->b_flags & B_READ) { 1379 io.iov_len = uiop->uio_resid = bp->b_bcount; 1380 io.iov_base = bp->b_data; 1381 uiop->uio_rw = UIO_READ; 1382 1383 switch (vp->v_type) { 1384 case VREG: 1385 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1386 nfsstats.read_bios++; 1387 error = nfs_readrpc(vp, uiop); 1388 1389 if (!error) { 1390 if (uiop->uio_resid) { 1391 /* 1392 * If we had a short read with no error, we must have 1393 * hit a file hole. We should zero-fill the remainder. 1394 * This can also occur if the server hits the file EOF. 1395 * 1396 * Holes used to be able to occur due to pending 1397 * writes, but that is not possible any longer. 1398 */ 1399 int nread = bp->b_bcount - uiop->uio_resid; 1400 int left = uiop->uio_resid; 1401 1402 if (left > 0) 1403 bzero((char *)bp->b_data + nread, left); 1404 uiop->uio_resid = 0; 1405 } 1406 } 1407 if (td && td->td_proc && (vp->v_flag & VTEXT) && 1408 (((nmp->nm_flag & NFSMNT_NQNFS) && 1409 NQNFS_CKINVALID(vp, np, ND_READ) && 1410 np->n_lrev != np->n_brev) || 1411 (!(nmp->nm_flag & NFSMNT_NQNFS) && 1412 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) { 1413 uprintf("Process killed due to text file modification\n"); 1414 psignal(td->td_proc, SIGKILL); 1415 PHOLD(td->td_proc); 1416 } 1417 break; 1418 case VLNK: 1419 uiop->uio_offset = (off_t)0; 1420 nfsstats.readlink_bios++; 1421 error = nfs_readlinkrpc(vp, uiop); 1422 break; 1423 case VDIR: 1424 nfsstats.readdir_bios++; 1425 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1426 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1427 error = nfs_readdirplusrpc(vp, uiop); 1428 if (error == NFSERR_NOTSUPP) 1429 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1430 } 1431 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1432 error = nfs_readdirrpc(vp, uiop); 1433 /* 1434 * end-of-directory sets B_INVAL but does not generate an 1435 * error. 1436 */ 1437 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1438 bp->b_flags |= B_INVAL; 1439 break; 1440 default: 1441 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1442 break; 1443 }; 1444 if (error) { 1445 bp->b_flags |= B_ERROR; 1446 bp->b_error = error; 1447 } 1448 } else { 1449 /* 1450 * If we only need to commit, try to commit 1451 */ 1452 if (bp->b_flags & B_NEEDCOMMIT) { 1453 int retv; 1454 off_t off; 1455 1456 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1457 bp->b_flags |= B_WRITEINPROG; 1458 retv = nfs_commit(bp->b_vp, off, 1459 bp->b_dirtyend - bp->b_dirtyoff, td); 1460 bp->b_flags &= ~B_WRITEINPROG; 1461 if (retv == 0) { 1462 bp->b_dirtyoff = bp->b_dirtyend = 0; 1463 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1464 bp->b_resid = 0; 1465 biodone(bp); 1466 return (0); 1467 } 1468 if (retv == NFSERR_STALEWRITEVERF) { 1469 nfs_clearcommit(bp->b_vp->v_mount); 1470 } 1471 } 1472 1473 /* 1474 * Setup for actual write 1475 */ 1476 1477 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1478 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1479 1480 if (bp->b_dirtyend > bp->b_dirtyoff) { 1481 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1482 - bp->b_dirtyoff; 1483 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1484 + bp->b_dirtyoff; 1485 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1486 uiop->uio_rw = UIO_WRITE; 1487 nfsstats.write_bios++; 1488 1489 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1490 iomode = NFSV3WRITE_UNSTABLE; 1491 else 1492 iomode = NFSV3WRITE_FILESYNC; 1493 1494 bp->b_flags |= B_WRITEINPROG; 1495 error = nfs_writerpc(vp, uiop, &iomode, &must_commit); 1496 1497 /* 1498 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1499 * to cluster the buffers needing commit. This will allow 1500 * the system to submit a single commit rpc for the whole 1501 * cluster. We can do this even if the buffer is not 100% 1502 * dirty (relative to the NFS blocksize), so we optimize the 1503 * append-to-file-case. 1504 * 1505 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1506 * cleared because write clustering only works for commit 1507 * rpc's, not for the data portion of the write). 1508 */ 1509 1510 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1511 bp->b_flags |= B_NEEDCOMMIT; 1512 if (bp->b_dirtyoff == 0 1513 && bp->b_dirtyend == bp->b_bcount) 1514 bp->b_flags |= B_CLUSTEROK; 1515 } else { 1516 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1517 } 1518 bp->b_flags &= ~B_WRITEINPROG; 1519 1520 /* 1521 * For an interrupted write, the buffer is still valid 1522 * and the write hasn't been pushed to the server yet, 1523 * so we can't set B_ERROR and report the interruption 1524 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1525 * is not relevant, so the rpc attempt is essentially 1526 * a noop. For the case of a V3 write rpc not being 1527 * committed to stable storage, the block is still 1528 * dirty and requires either a commit rpc or another 1529 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1530 * the block is reused. This is indicated by setting 1531 * the B_DELWRI and B_NEEDCOMMIT flags. 1532 * 1533 * If the buffer is marked B_PAGING, it does not reside on 1534 * the vp's paging queues so we cannot call bdirty(). The 1535 * bp in this case is not an NFS cache block so we should 1536 * be safe. XXX 1537 */ 1538 if (error == EINTR 1539 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1540 int s; 1541 1542 s = splbio(); 1543 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1544 if ((bp->b_flags & B_PAGING) == 0) { 1545 bdirty(bp); 1546 bp->b_flags &= ~B_DONE; 1547 } 1548 if (error && (bp->b_flags & B_ASYNC) == 0) 1549 bp->b_flags |= B_EINTR; 1550 splx(s); 1551 } else { 1552 if (error) { 1553 bp->b_flags |= B_ERROR; 1554 bp->b_error = np->n_error = error; 1555 np->n_flag |= NWRITEERR; 1556 } 1557 bp->b_dirtyoff = bp->b_dirtyend = 0; 1558 } 1559 } else { 1560 bp->b_resid = 0; 1561 biodone(bp); 1562 return (0); 1563 } 1564 } 1565 bp->b_resid = uiop->uio_resid; 1566 if (must_commit) 1567 nfs_clearcommit(vp->v_mount); 1568 biodone(bp); 1569 return (error); 1570 } 1571 1572 /* 1573 * Used to aid in handling ftruncate() operations on the NFS client side. 1574 * Truncation creates a number of special problems for NFS. We have to 1575 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1576 * we have to properly handle VM pages or (potentially dirty) buffers 1577 * that straddle the truncation point. 1578 */ 1579 1580 int 1581 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize) 1582 { 1583 struct nfsnode *np = VTONFS(vp); 1584 u_quad_t tsize = np->n_size; 1585 int biosize = vp->v_mount->mnt_stat.f_iosize; 1586 int error = 0; 1587 1588 np->n_size = nsize; 1589 1590 if (np->n_size < tsize) { 1591 struct buf *bp; 1592 daddr_t lbn; 1593 int bufsize; 1594 1595 /* 1596 * vtruncbuf() doesn't get the buffer overlapping the 1597 * truncation point. We may have a B_DELWRI and/or B_CACHE 1598 * buffer that now needs to be truncated. 1599 */ 1600 error = vtruncbuf(vp, td, nsize, biosize); 1601 lbn = nsize / biosize; 1602 bufsize = nsize & (biosize - 1); 1603 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1604 if (bp->b_dirtyoff > bp->b_bcount) 1605 bp->b_dirtyoff = bp->b_bcount; 1606 if (bp->b_dirtyend > bp->b_bcount) 1607 bp->b_dirtyend = bp->b_bcount; 1608 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1609 brelse(bp); 1610 } else { 1611 vnode_pager_setsize(vp, nsize); 1612 } 1613 return(error); 1614 } 1615 1616