1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1989, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * Rick Macklem at The University of Guelph. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 35 */ 36 37 #include <sys/cdefs.h> 38 __FBSDID("$FreeBSD$"); 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/bio.h> 43 #include <sys/buf.h> 44 #include <sys/kernel.h> 45 #include <sys/mount.h> 46 #include <sys/rwlock.h> 47 #include <sys/vmmeter.h> 48 #include <sys/vnode.h> 49 50 #include <vm/vm.h> 51 #include <vm/vm_param.h> 52 #include <vm/vm_extern.h> 53 #include <vm/vm_page.h> 54 #include <vm/vm_object.h> 55 #include <vm/vm_pager.h> 56 #include <vm/vnode_pager.h> 57 58 #include <fs/nfs/nfsport.h> 59 #include <fs/nfsclient/nfsmount.h> 60 #include <fs/nfsclient/nfs.h> 61 #include <fs/nfsclient/nfsnode.h> 62 #include <fs/nfsclient/nfs_kdtrace.h> 63 64 extern int newnfs_directio_allow_mmap; 65 extern struct nfsstatsv1 nfsstatsv1; 66 extern struct mtx ncl_iod_mutex; 67 extern int ncl_numasync; 68 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON]; 69 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON]; 70 extern int newnfs_directio_enable; 71 extern int nfs_keep_dirty_on_error; 72 73 uma_zone_t ncl_pbuf_zone; 74 75 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 76 struct thread *td); 77 static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 78 struct ucred *cred, int ioflag); 79 80 /* 81 * Vnode op for VM getpages. 82 */ 83 SYSCTL_DECL(_vfs_nfs); 84 static int use_buf_pager = 1; 85 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN, 86 &use_buf_pager, 0, 87 "Use buffer pager instead of direct readrpc call"); 88 89 static daddr_t 90 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off) 91 { 92 93 return (off / vp->v_bufobj.bo_bsize); 94 } 95 96 static int 97 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn, long *sz) 98 { 99 struct nfsnode *np; 100 u_quad_t nsize; 101 int biosize, bcount; 102 103 np = VTONFS(vp); 104 NFSLOCKNODE(np); 105 nsize = np->n_size; 106 NFSUNLOCKNODE(np); 107 108 biosize = vp->v_bufobj.bo_bsize; 109 bcount = biosize; 110 if ((off_t)lbn * biosize >= nsize) 111 bcount = 0; 112 else if ((off_t)(lbn + 1) * biosize > nsize) 113 bcount = nsize - (off_t)lbn * biosize; 114 *sz = bcount; 115 return (0); 116 } 117 118 int 119 ncl_getpages(struct vop_getpages_args *ap) 120 { 121 int i, error, nextoff, size, toff, count, npages; 122 struct uio uio; 123 struct iovec iov; 124 vm_offset_t kva; 125 struct buf *bp; 126 struct vnode *vp; 127 struct thread *td; 128 struct ucred *cred; 129 struct nfsmount *nmp; 130 vm_object_t object; 131 vm_page_t *pages; 132 struct nfsnode *np; 133 134 vp = ap->a_vp; 135 np = VTONFS(vp); 136 td = curthread; 137 cred = curthread->td_ucred; 138 nmp = VFSTONFS(vp->v_mount); 139 pages = ap->a_m; 140 npages = ap->a_count; 141 142 if ((object = vp->v_object) == NULL) { 143 printf("ncl_getpages: called with non-merged cache vnode\n"); 144 return (VM_PAGER_ERROR); 145 } 146 147 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) { 148 NFSLOCKNODE(np); 149 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 150 NFSUNLOCKNODE(np); 151 printf("ncl_getpages: called on non-cacheable vnode\n"); 152 return (VM_PAGER_ERROR); 153 } else 154 NFSUNLOCKNODE(np); 155 } 156 157 mtx_lock(&nmp->nm_mtx); 158 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 159 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 160 mtx_unlock(&nmp->nm_mtx); 161 /* We'll never get here for v4, because we always have fsinfo */ 162 (void)ncl_fsinfo(nmp, vp, cred, td); 163 } else 164 mtx_unlock(&nmp->nm_mtx); 165 166 if (use_buf_pager) 167 return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind, 168 ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz)); 169 170 /* 171 * If the requested page is partially valid, just return it and 172 * allow the pager to zero-out the blanks. Partially valid pages 173 * can only occur at the file EOF. 174 * 175 * XXXGL: is that true for NFS, where short read can occur??? 176 */ 177 VM_OBJECT_WLOCK(object); 178 if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0) 179 goto out; 180 VM_OBJECT_WUNLOCK(object); 181 182 /* 183 * We use only the kva address for the buffer, but this is extremely 184 * convenient and fast. 185 */ 186 bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK); 187 188 kva = (vm_offset_t) bp->b_data; 189 pmap_qenter(kva, pages, npages); 190 VM_CNT_INC(v_vnodein); 191 VM_CNT_ADD(v_vnodepgsin, npages); 192 193 count = npages << PAGE_SHIFT; 194 iov.iov_base = (caddr_t) kva; 195 iov.iov_len = count; 196 uio.uio_iov = &iov; 197 uio.uio_iovcnt = 1; 198 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 199 uio.uio_resid = count; 200 uio.uio_segflg = UIO_SYSSPACE; 201 uio.uio_rw = UIO_READ; 202 uio.uio_td = td; 203 204 error = ncl_readrpc(vp, &uio, cred); 205 pmap_qremove(kva, npages); 206 207 uma_zfree(ncl_pbuf_zone, bp); 208 209 if (error && (uio.uio_resid == count)) { 210 printf("ncl_getpages: error %d\n", error); 211 return (VM_PAGER_ERROR); 212 } 213 214 /* 215 * Calculate the number of bytes read and validate only that number 216 * of bytes. Note that due to pending writes, size may be 0. This 217 * does not mean that the remaining data is invalid! 218 */ 219 220 size = count - uio.uio_resid; 221 VM_OBJECT_WLOCK(object); 222 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 223 vm_page_t m; 224 nextoff = toff + PAGE_SIZE; 225 m = pages[i]; 226 227 if (nextoff <= size) { 228 /* 229 * Read operation filled an entire page 230 */ 231 vm_page_valid(m); 232 KASSERT(m->dirty == 0, 233 ("nfs_getpages: page %p is dirty", m)); 234 } else if (size > toff) { 235 /* 236 * Read operation filled a partial page. 237 */ 238 vm_page_invalid(m); 239 vm_page_set_valid_range(m, 0, size - toff); 240 KASSERT(m->dirty == 0, 241 ("nfs_getpages: page %p is dirty", m)); 242 } else { 243 /* 244 * Read operation was short. If no error 245 * occurred we may have hit a zero-fill 246 * section. We leave valid set to 0, and page 247 * is freed by vm_page_readahead_finish() if 248 * its index is not equal to requested, or 249 * page is zeroed and set valid by 250 * vm_pager_get_pages() for requested page. 251 */ 252 ; 253 } 254 } 255 out: 256 VM_OBJECT_WUNLOCK(object); 257 if (ap->a_rbehind) 258 *ap->a_rbehind = 0; 259 if (ap->a_rahead) 260 *ap->a_rahead = 0; 261 return (VM_PAGER_OK); 262 } 263 264 /* 265 * Vnode op for VM putpages. 266 */ 267 int 268 ncl_putpages(struct vop_putpages_args *ap) 269 { 270 struct uio uio; 271 struct iovec iov; 272 int i, error, npages, count; 273 off_t offset; 274 int *rtvals; 275 struct vnode *vp; 276 struct thread *td; 277 struct ucred *cred; 278 struct nfsmount *nmp; 279 struct nfsnode *np; 280 vm_page_t *pages; 281 282 vp = ap->a_vp; 283 np = VTONFS(vp); 284 td = curthread; /* XXX */ 285 /* Set the cred to n_writecred for the write rpcs. */ 286 if (np->n_writecred != NULL) 287 cred = crhold(np->n_writecred); 288 else 289 cred = crhold(curthread->td_ucred); /* XXX */ 290 nmp = VFSTONFS(vp->v_mount); 291 pages = ap->a_m; 292 count = ap->a_count; 293 rtvals = ap->a_rtvals; 294 npages = btoc(count); 295 offset = IDX_TO_OFF(pages[0]->pindex); 296 297 mtx_lock(&nmp->nm_mtx); 298 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 299 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 300 mtx_unlock(&nmp->nm_mtx); 301 (void)ncl_fsinfo(nmp, vp, cred, td); 302 } else 303 mtx_unlock(&nmp->nm_mtx); 304 305 NFSLOCKNODE(np); 306 if (newnfs_directio_enable && !newnfs_directio_allow_mmap && 307 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 308 NFSUNLOCKNODE(np); 309 printf("ncl_putpages: called on noncache-able vnode\n"); 310 NFSLOCKNODE(np); 311 } 312 /* 313 * When putting pages, do not extend file past EOF. 314 */ 315 if (offset + count > np->n_size) { 316 count = np->n_size - offset; 317 if (count < 0) 318 count = 0; 319 } 320 NFSUNLOCKNODE(np); 321 322 for (i = 0; i < npages; i++) 323 rtvals[i] = VM_PAGER_ERROR; 324 325 VM_CNT_INC(v_vnodeout); 326 VM_CNT_ADD(v_vnodepgsout, count); 327 328 iov.iov_base = unmapped_buf; 329 iov.iov_len = count; 330 uio.uio_iov = &iov; 331 uio.uio_iovcnt = 1; 332 uio.uio_offset = offset; 333 uio.uio_resid = count; 334 uio.uio_segflg = UIO_NOCOPY; 335 uio.uio_rw = UIO_WRITE; 336 uio.uio_td = td; 337 338 error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync), 339 cred); 340 crfree(cred); 341 342 if (error == 0 || !nfs_keep_dirty_on_error) { 343 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid, 344 np->n_size - offset, npages * PAGE_SIZE); 345 } 346 return (rtvals[0]); 347 } 348 349 /* 350 * For nfs, cache consistency can only be maintained approximately. 351 * Although RFC1094 does not specify the criteria, the following is 352 * believed to be compatible with the reference port. 353 * For nfs: 354 * If the file's modify time on the server has changed since the 355 * last read rpc or you have written to the file, 356 * you may have lost data cache consistency with the 357 * server, so flush all of the file's data out of the cache. 358 * Then force a getattr rpc to ensure that you have up to date 359 * attributes. 360 * NB: This implies that cache data can be read when up to 361 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 362 * attributes this could be forced by setting n_attrstamp to 0 before 363 * the VOP_GETATTR() call. 364 */ 365 static inline int 366 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 367 { 368 int error = 0; 369 struct vattr vattr; 370 struct nfsnode *np = VTONFS(vp); 371 bool old_lock; 372 373 /* 374 * Ensure the exclusove access to the node before checking 375 * whether the cache is consistent. 376 */ 377 old_lock = ncl_excl_start(vp); 378 NFSLOCKNODE(np); 379 if (np->n_flag & NMODIFIED) { 380 NFSUNLOCKNODE(np); 381 if (vp->v_type != VREG) { 382 if (vp->v_type != VDIR) 383 panic("nfs: bioread, not dir"); 384 ncl_invaldir(vp); 385 error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1); 386 if (error != 0) 387 goto out; 388 } 389 np->n_attrstamp = 0; 390 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 391 error = VOP_GETATTR(vp, &vattr, cred); 392 if (error) 393 goto out; 394 NFSLOCKNODE(np); 395 np->n_mtime = vattr.va_mtime; 396 NFSUNLOCKNODE(np); 397 } else { 398 NFSUNLOCKNODE(np); 399 error = VOP_GETATTR(vp, &vattr, cred); 400 if (error) 401 goto out; 402 NFSLOCKNODE(np); 403 if ((np->n_flag & NSIZECHANGED) 404 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 405 NFSUNLOCKNODE(np); 406 if (vp->v_type == VDIR) 407 ncl_invaldir(vp); 408 error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1); 409 if (error != 0) 410 goto out; 411 NFSLOCKNODE(np); 412 np->n_mtime = vattr.va_mtime; 413 np->n_flag &= ~NSIZECHANGED; 414 } 415 NFSUNLOCKNODE(np); 416 } 417 out: 418 ncl_excl_finish(vp, old_lock); 419 return (error); 420 } 421 422 /* 423 * Vnode op for read using bio 424 */ 425 int 426 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 427 { 428 struct nfsnode *np = VTONFS(vp); 429 struct buf *bp, *rabp; 430 struct thread *td; 431 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 432 daddr_t lbn, rabn; 433 int biosize, bcount, error, i, n, nra, on, save2, seqcount; 434 off_t tmp_off; 435 436 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode")); 437 if (uio->uio_resid == 0) 438 return (0); 439 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 440 return (EINVAL); 441 td = uio->uio_td; 442 443 mtx_lock(&nmp->nm_mtx); 444 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 445 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 446 mtx_unlock(&nmp->nm_mtx); 447 (void)ncl_fsinfo(nmp, vp, cred, td); 448 mtx_lock(&nmp->nm_mtx); 449 } 450 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0) 451 (void) newnfs_iosize(nmp); 452 453 tmp_off = uio->uio_offset + uio->uio_resid; 454 if (vp->v_type != VDIR && 455 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) { 456 mtx_unlock(&nmp->nm_mtx); 457 return (EFBIG); 458 } 459 mtx_unlock(&nmp->nm_mtx); 460 461 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 462 /* No caching/ no readaheads. Just read data into the user buffer */ 463 return ncl_readrpc(vp, uio, cred); 464 465 n = 0; 466 on = 0; 467 biosize = vp->v_bufobj.bo_bsize; 468 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 469 470 error = nfs_bioread_check_cons(vp, td, cred); 471 if (error) 472 return error; 473 474 save2 = curthread_pflags2_set(TDP2_SBPAGES); 475 do { 476 u_quad_t nsize; 477 478 NFSLOCKNODE(np); 479 nsize = np->n_size; 480 NFSUNLOCKNODE(np); 481 482 switch (vp->v_type) { 483 case VREG: 484 NFSINCRGLOBAL(nfsstatsv1.biocache_reads); 485 lbn = uio->uio_offset / biosize; 486 on = uio->uio_offset - (lbn * biosize); 487 488 /* 489 * Start the read ahead(s), as required. 490 */ 491 if (nmp->nm_readahead > 0) { 492 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 493 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 494 rabn = lbn + 1 + nra; 495 if (incore(&vp->v_bufobj, rabn) == NULL) { 496 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 497 if (!rabp) { 498 error = newnfs_sigintr(nmp, td); 499 if (error == 0) 500 error = EINTR; 501 goto out; 502 } 503 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 504 rabp->b_flags |= B_ASYNC; 505 rabp->b_iocmd = BIO_READ; 506 vfs_busy_pages(rabp, 0); 507 if (ncl_asyncio(nmp, rabp, cred, td)) { 508 rabp->b_flags |= B_INVAL; 509 rabp->b_ioflags |= BIO_ERROR; 510 vfs_unbusy_pages(rabp); 511 brelse(rabp); 512 break; 513 } 514 } else { 515 brelse(rabp); 516 } 517 } 518 } 519 } 520 521 /* Note that bcount is *not* DEV_BSIZE aligned. */ 522 bcount = biosize; 523 if ((off_t)lbn * biosize >= nsize) { 524 bcount = 0; 525 } else if ((off_t)(lbn + 1) * biosize > nsize) { 526 bcount = nsize - (off_t)lbn * biosize; 527 } 528 bp = nfs_getcacheblk(vp, lbn, bcount, td); 529 530 if (!bp) { 531 error = newnfs_sigintr(nmp, td); 532 if (error == 0) 533 error = EINTR; 534 goto out; 535 } 536 537 /* 538 * If B_CACHE is not set, we must issue the read. If this 539 * fails, we return an error. 540 */ 541 542 if ((bp->b_flags & B_CACHE) == 0) { 543 bp->b_iocmd = BIO_READ; 544 vfs_busy_pages(bp, 0); 545 error = ncl_doio(vp, bp, cred, td, 0); 546 if (error) { 547 brelse(bp); 548 goto out; 549 } 550 } 551 552 /* 553 * on is the offset into the current bp. Figure out how many 554 * bytes we can copy out of the bp. Note that bcount is 555 * NOT DEV_BSIZE aligned. 556 * 557 * Then figure out how many bytes we can copy into the uio. 558 */ 559 560 n = 0; 561 if (on < bcount) 562 n = MIN((unsigned)(bcount - on), uio->uio_resid); 563 break; 564 case VLNK: 565 NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks); 566 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 567 if (!bp) { 568 error = newnfs_sigintr(nmp, td); 569 if (error == 0) 570 error = EINTR; 571 goto out; 572 } 573 if ((bp->b_flags & B_CACHE) == 0) { 574 bp->b_iocmd = BIO_READ; 575 vfs_busy_pages(bp, 0); 576 error = ncl_doio(vp, bp, cred, td, 0); 577 if (error) { 578 bp->b_ioflags |= BIO_ERROR; 579 brelse(bp); 580 goto out; 581 } 582 } 583 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 584 on = 0; 585 break; 586 case VDIR: 587 NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs); 588 NFSLOCKNODE(np); 589 if (np->n_direofoffset 590 && uio->uio_offset >= np->n_direofoffset) { 591 NFSUNLOCKNODE(np); 592 error = 0; 593 goto out; 594 } 595 NFSUNLOCKNODE(np); 596 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 597 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 598 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 599 if (!bp) { 600 error = newnfs_sigintr(nmp, td); 601 if (error == 0) 602 error = EINTR; 603 goto out; 604 } 605 if ((bp->b_flags & B_CACHE) == 0) { 606 bp->b_iocmd = BIO_READ; 607 vfs_busy_pages(bp, 0); 608 error = ncl_doio(vp, bp, cred, td, 0); 609 if (error) { 610 brelse(bp); 611 } 612 while (error == NFSERR_BAD_COOKIE) { 613 ncl_invaldir(vp); 614 error = ncl_vinvalbuf(vp, 0, td, 1); 615 616 /* 617 * Yuck! The directory has been modified on the 618 * server. The only way to get the block is by 619 * reading from the beginning to get all the 620 * offset cookies. 621 * 622 * Leave the last bp intact unless there is an error. 623 * Loop back up to the while if the error is another 624 * NFSERR_BAD_COOKIE (double yuch!). 625 */ 626 for (i = 0; i <= lbn && !error; i++) { 627 NFSLOCKNODE(np); 628 if (np->n_direofoffset 629 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) { 630 NFSUNLOCKNODE(np); 631 error = 0; 632 goto out; 633 } 634 NFSUNLOCKNODE(np); 635 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 636 if (!bp) { 637 error = newnfs_sigintr(nmp, td); 638 if (error == 0) 639 error = EINTR; 640 goto out; 641 } 642 if ((bp->b_flags & B_CACHE) == 0) { 643 bp->b_iocmd = BIO_READ; 644 vfs_busy_pages(bp, 0); 645 error = ncl_doio(vp, bp, cred, td, 0); 646 /* 647 * no error + B_INVAL == directory EOF, 648 * use the block. 649 */ 650 if (error == 0 && (bp->b_flags & B_INVAL)) 651 break; 652 } 653 /* 654 * An error will throw away the block and the 655 * for loop will break out. If no error and this 656 * is not the block we want, we throw away the 657 * block and go for the next one via the for loop. 658 */ 659 if (error || i < lbn) 660 brelse(bp); 661 } 662 } 663 /* 664 * The above while is repeated if we hit another cookie 665 * error. If we hit an error and it wasn't a cookie error, 666 * we give up. 667 */ 668 if (error) 669 goto out; 670 } 671 672 /* 673 * If not eof and read aheads are enabled, start one. 674 * (You need the current block first, so that you have the 675 * directory offset cookie of the next block.) 676 */ 677 NFSLOCKNODE(np); 678 if (nmp->nm_readahead > 0 && 679 (bp->b_flags & B_INVAL) == 0 && 680 (np->n_direofoffset == 0 || 681 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 682 incore(&vp->v_bufobj, lbn + 1) == NULL) { 683 NFSUNLOCKNODE(np); 684 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 685 if (rabp) { 686 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 687 rabp->b_flags |= B_ASYNC; 688 rabp->b_iocmd = BIO_READ; 689 vfs_busy_pages(rabp, 0); 690 if (ncl_asyncio(nmp, rabp, cred, td)) { 691 rabp->b_flags |= B_INVAL; 692 rabp->b_ioflags |= BIO_ERROR; 693 vfs_unbusy_pages(rabp); 694 brelse(rabp); 695 } 696 } else { 697 brelse(rabp); 698 } 699 } 700 NFSLOCKNODE(np); 701 } 702 /* 703 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 704 * chopped for the EOF condition, we cannot tell how large 705 * NFS directories are going to be until we hit EOF. So 706 * an NFS directory buffer is *not* chopped to its EOF. Now, 707 * it just so happens that b_resid will effectively chop it 708 * to EOF. *BUT* this information is lost if the buffer goes 709 * away and is reconstituted into a B_CACHE state ( due to 710 * being VMIO ) later. So we keep track of the directory eof 711 * in np->n_direofoffset and chop it off as an extra step 712 * right here. 713 */ 714 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 715 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 716 n = np->n_direofoffset - uio->uio_offset; 717 NFSUNLOCKNODE(np); 718 break; 719 default: 720 printf(" ncl_bioread: type %x unexpected\n", vp->v_type); 721 bp = NULL; 722 break; 723 } 724 725 if (n > 0) { 726 error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio); 727 } 728 if (vp->v_type == VLNK) 729 n = 0; 730 if (bp != NULL) 731 brelse(bp); 732 } while (error == 0 && uio->uio_resid > 0 && n > 0); 733 out: 734 curthread_pflags2_restore(save2); 735 if ((curthread->td_pflags2 & TDP2_SBPAGES) == 0) { 736 NFSLOCKNODE(np); 737 ncl_pager_setsize(vp, NULL); 738 } 739 return (error); 740 } 741 742 /* 743 * The NFS write path cannot handle iovecs with len > 1. So we need to 744 * break up iovecs accordingly (restricting them to wsize). 745 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 746 * For the ASYNC case, 2 copies are needed. The first a copy from the 747 * user buffer to a staging buffer and then a second copy from the staging 748 * buffer to mbufs. This can be optimized by copying from the user buffer 749 * directly into mbufs and passing the chain down, but that requires a 750 * fair amount of re-working of the relevant codepaths (and can be done 751 * later). 752 */ 753 static int 754 nfs_directio_write(vp, uiop, cred, ioflag) 755 struct vnode *vp; 756 struct uio *uiop; 757 struct ucred *cred; 758 int ioflag; 759 { 760 int error; 761 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 762 struct thread *td = uiop->uio_td; 763 int size; 764 int wsize; 765 766 mtx_lock(&nmp->nm_mtx); 767 wsize = nmp->nm_wsize; 768 mtx_unlock(&nmp->nm_mtx); 769 if (ioflag & IO_SYNC) { 770 int iomode, must_commit; 771 struct uio uio; 772 struct iovec iov; 773 do_sync: 774 while (uiop->uio_resid > 0) { 775 size = MIN(uiop->uio_resid, wsize); 776 size = MIN(uiop->uio_iov->iov_len, size); 777 iov.iov_base = uiop->uio_iov->iov_base; 778 iov.iov_len = size; 779 uio.uio_iov = &iov; 780 uio.uio_iovcnt = 1; 781 uio.uio_offset = uiop->uio_offset; 782 uio.uio_resid = size; 783 uio.uio_segflg = UIO_USERSPACE; 784 uio.uio_rw = UIO_WRITE; 785 uio.uio_td = td; 786 iomode = NFSWRITE_FILESYNC; 787 error = ncl_writerpc(vp, &uio, cred, &iomode, 788 &must_commit, 0); 789 KASSERT((must_commit == 0), 790 ("ncl_directio_write: Did not commit write")); 791 if (error) 792 return (error); 793 uiop->uio_offset += size; 794 uiop->uio_resid -= size; 795 if (uiop->uio_iov->iov_len <= size) { 796 uiop->uio_iovcnt--; 797 uiop->uio_iov++; 798 } else { 799 uiop->uio_iov->iov_base = 800 (char *)uiop->uio_iov->iov_base + size; 801 uiop->uio_iov->iov_len -= size; 802 } 803 } 804 } else { 805 struct uio *t_uio; 806 struct iovec *t_iov; 807 struct buf *bp; 808 809 /* 810 * Break up the write into blocksize chunks and hand these 811 * over to nfsiod's for write back. 812 * Unfortunately, this incurs a copy of the data. Since 813 * the user could modify the buffer before the write is 814 * initiated. 815 * 816 * The obvious optimization here is that one of the 2 copies 817 * in the async write path can be eliminated by copying the 818 * data here directly into mbufs and passing the mbuf chain 819 * down. But that will require a fair amount of re-working 820 * of the code and can be done if there's enough interest 821 * in NFS directio access. 822 */ 823 while (uiop->uio_resid > 0) { 824 size = MIN(uiop->uio_resid, wsize); 825 size = MIN(uiop->uio_iov->iov_len, size); 826 bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK); 827 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 828 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 829 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 830 t_iov->iov_len = size; 831 t_uio->uio_iov = t_iov; 832 t_uio->uio_iovcnt = 1; 833 t_uio->uio_offset = uiop->uio_offset; 834 t_uio->uio_resid = size; 835 t_uio->uio_segflg = UIO_SYSSPACE; 836 t_uio->uio_rw = UIO_WRITE; 837 t_uio->uio_td = td; 838 KASSERT(uiop->uio_segflg == UIO_USERSPACE || 839 uiop->uio_segflg == UIO_SYSSPACE, 840 ("nfs_directio_write: Bad uio_segflg")); 841 if (uiop->uio_segflg == UIO_USERSPACE) { 842 error = copyin(uiop->uio_iov->iov_base, 843 t_iov->iov_base, size); 844 if (error != 0) 845 goto err_free; 846 } else 847 /* 848 * UIO_SYSSPACE may never happen, but handle 849 * it just in case it does. 850 */ 851 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, 852 size); 853 bp->b_flags |= B_DIRECT; 854 bp->b_iocmd = BIO_WRITE; 855 if (cred != NOCRED) { 856 crhold(cred); 857 bp->b_wcred = cred; 858 } else 859 bp->b_wcred = NOCRED; 860 bp->b_caller1 = (void *)t_uio; 861 bp->b_vp = vp; 862 error = ncl_asyncio(nmp, bp, NOCRED, td); 863 err_free: 864 if (error) { 865 free(t_iov->iov_base, M_NFSDIRECTIO); 866 free(t_iov, M_NFSDIRECTIO); 867 free(t_uio, M_NFSDIRECTIO); 868 bp->b_vp = NULL; 869 uma_zfree(ncl_pbuf_zone, bp); 870 if (error == EINTR) 871 return (error); 872 goto do_sync; 873 } 874 uiop->uio_offset += size; 875 uiop->uio_resid -= size; 876 if (uiop->uio_iov->iov_len <= size) { 877 uiop->uio_iovcnt--; 878 uiop->uio_iov++; 879 } else { 880 uiop->uio_iov->iov_base = 881 (char *)uiop->uio_iov->iov_base + size; 882 uiop->uio_iov->iov_len -= size; 883 } 884 } 885 } 886 return (0); 887 } 888 889 /* 890 * Vnode op for write using bio 891 */ 892 int 893 ncl_write(struct vop_write_args *ap) 894 { 895 int biosize; 896 struct uio *uio = ap->a_uio; 897 struct thread *td = uio->uio_td; 898 struct vnode *vp = ap->a_vp; 899 struct nfsnode *np = VTONFS(vp); 900 struct ucred *cred = ap->a_cred; 901 int ioflag = ap->a_ioflag; 902 struct buf *bp; 903 struct vattr vattr; 904 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 905 daddr_t lbn; 906 int bcount, noncontig_write, obcount; 907 int bp_cached, n, on, error = 0, error1, save2, wouldcommit; 908 size_t orig_resid, local_resid; 909 off_t orig_size, tmp_off; 910 911 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode")); 912 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, 913 ("ncl_write proc")); 914 if (vp->v_type != VREG) 915 return (EIO); 916 NFSLOCKNODE(np); 917 if (np->n_flag & NWRITEERR) { 918 np->n_flag &= ~NWRITEERR; 919 NFSUNLOCKNODE(np); 920 return (np->n_error); 921 } else 922 NFSUNLOCKNODE(np); 923 mtx_lock(&nmp->nm_mtx); 924 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 925 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 926 mtx_unlock(&nmp->nm_mtx); 927 (void)ncl_fsinfo(nmp, vp, cred, td); 928 mtx_lock(&nmp->nm_mtx); 929 } 930 if (nmp->nm_wsize == 0) 931 (void) newnfs_iosize(nmp); 932 mtx_unlock(&nmp->nm_mtx); 933 934 /* 935 * Synchronously flush pending buffers if we are in synchronous 936 * mode or if we are appending. 937 */ 938 if (ioflag & (IO_APPEND | IO_SYNC)) { 939 NFSLOCKNODE(np); 940 if (np->n_flag & NMODIFIED) { 941 NFSUNLOCKNODE(np); 942 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 943 /* 944 * Require non-blocking, synchronous writes to 945 * dirty files to inform the program it needs 946 * to fsync(2) explicitly. 947 */ 948 if (ioflag & IO_NDELAY) 949 return (EAGAIN); 950 #endif 951 np->n_attrstamp = 0; 952 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 953 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag & 954 IO_VMIO) != 0 ? V_VMIO : 0), td, 1); 955 if (error != 0) 956 return (error); 957 } else 958 NFSUNLOCKNODE(np); 959 } 960 961 orig_resid = uio->uio_resid; 962 NFSLOCKNODE(np); 963 orig_size = np->n_size; 964 NFSUNLOCKNODE(np); 965 966 /* 967 * If IO_APPEND then load uio_offset. We restart here if we cannot 968 * get the append lock. 969 */ 970 if (ioflag & IO_APPEND) { 971 np->n_attrstamp = 0; 972 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 973 error = VOP_GETATTR(vp, &vattr, cred); 974 if (error) 975 return (error); 976 NFSLOCKNODE(np); 977 uio->uio_offset = np->n_size; 978 NFSUNLOCKNODE(np); 979 } 980 981 if (uio->uio_offset < 0) 982 return (EINVAL); 983 tmp_off = uio->uio_offset + uio->uio_resid; 984 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset) 985 return (EFBIG); 986 if (uio->uio_resid == 0) 987 return (0); 988 989 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 990 return nfs_directio_write(vp, uio, cred, ioflag); 991 992 /* 993 * Maybe this should be above the vnode op call, but so long as 994 * file servers have no limits, i don't think it matters 995 */ 996 if (vn_rlimit_fsize(vp, uio, td)) 997 return (EFBIG); 998 999 save2 = curthread_pflags2_set(TDP2_SBPAGES); 1000 biosize = vp->v_bufobj.bo_bsize; 1001 /* 1002 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 1003 * would exceed the local maximum per-file write commit size when 1004 * combined with those, we must decide whether to flush, 1005 * go synchronous, or return error. We don't bother checking 1006 * IO_UNIT -- we just make all writes atomic anyway, as there's 1007 * no point optimizing for something that really won't ever happen. 1008 */ 1009 wouldcommit = 0; 1010 if (!(ioflag & IO_SYNC)) { 1011 int nflag; 1012 1013 NFSLOCKNODE(np); 1014 nflag = np->n_flag; 1015 NFSUNLOCKNODE(np); 1016 if (nflag & NMODIFIED) { 1017 BO_LOCK(&vp->v_bufobj); 1018 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 1019 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 1020 b_bobufs) { 1021 if (bp->b_flags & B_NEEDCOMMIT) 1022 wouldcommit += bp->b_bcount; 1023 } 1024 } 1025 BO_UNLOCK(&vp->v_bufobj); 1026 } 1027 } 1028 1029 do { 1030 if (!(ioflag & IO_SYNC)) { 1031 wouldcommit += biosize; 1032 if (wouldcommit > nmp->nm_wcommitsize) { 1033 np->n_attrstamp = 0; 1034 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1035 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag & 1036 IO_VMIO) != 0 ? V_VMIO : 0), td, 1); 1037 if (error != 0) 1038 goto out; 1039 wouldcommit = biosize; 1040 } 1041 } 1042 1043 NFSINCRGLOBAL(nfsstatsv1.biocache_writes); 1044 lbn = uio->uio_offset / biosize; 1045 on = uio->uio_offset - (lbn * biosize); 1046 n = MIN((unsigned)(biosize - on), uio->uio_resid); 1047 again: 1048 /* 1049 * Handle direct append and file extension cases, calculate 1050 * unaligned buffer size. 1051 */ 1052 NFSLOCKNODE(np); 1053 if ((np->n_flag & NHASBEENLOCKED) == 0 && 1054 (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0) 1055 noncontig_write = 1; 1056 else 1057 noncontig_write = 0; 1058 if ((uio->uio_offset == np->n_size || 1059 (noncontig_write != 0 && 1060 lbn == (np->n_size / biosize) && 1061 uio->uio_offset + n > np->n_size)) && n) { 1062 NFSUNLOCKNODE(np); 1063 /* 1064 * Get the buffer (in its pre-append state to maintain 1065 * B_CACHE if it was previously set). Resize the 1066 * nfsnode after we have locked the buffer to prevent 1067 * readers from reading garbage. 1068 */ 1069 obcount = np->n_size - (lbn * biosize); 1070 bp = nfs_getcacheblk(vp, lbn, obcount, td); 1071 1072 if (bp != NULL) { 1073 long save; 1074 1075 NFSLOCKNODE(np); 1076 np->n_size = uio->uio_offset + n; 1077 np->n_flag |= NMODIFIED; 1078 np->n_flag &= ~NVNSETSZSKIP; 1079 vnode_pager_setsize(vp, np->n_size); 1080 NFSUNLOCKNODE(np); 1081 1082 save = bp->b_flags & B_CACHE; 1083 bcount = on + n; 1084 allocbuf(bp, bcount); 1085 bp->b_flags |= save; 1086 if (noncontig_write != 0 && on > obcount) 1087 vfs_bio_bzero_buf(bp, obcount, on - 1088 obcount); 1089 } 1090 } else { 1091 /* 1092 * Obtain the locked cache block first, and then 1093 * adjust the file's size as appropriate. 1094 */ 1095 bcount = on + n; 1096 if ((off_t)lbn * biosize + bcount < np->n_size) { 1097 if ((off_t)(lbn + 1) * biosize < np->n_size) 1098 bcount = biosize; 1099 else 1100 bcount = np->n_size - (off_t)lbn * biosize; 1101 } 1102 NFSUNLOCKNODE(np); 1103 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1104 NFSLOCKNODE(np); 1105 if (uio->uio_offset + n > np->n_size) { 1106 np->n_size = uio->uio_offset + n; 1107 np->n_flag |= NMODIFIED; 1108 np->n_flag &= ~NVNSETSZSKIP; 1109 vnode_pager_setsize(vp, np->n_size); 1110 } 1111 NFSUNLOCKNODE(np); 1112 } 1113 1114 if (!bp) { 1115 error = newnfs_sigintr(nmp, td); 1116 if (!error) 1117 error = EINTR; 1118 break; 1119 } 1120 1121 /* 1122 * Issue a READ if B_CACHE is not set. In special-append 1123 * mode, B_CACHE is based on the buffer prior to the write 1124 * op and is typically set, avoiding the read. If a read 1125 * is required in special append mode, the server will 1126 * probably send us a short-read since we extended the file 1127 * on our end, resulting in b_resid == 0 and, thusly, 1128 * B_CACHE getting set. 1129 * 1130 * We can also avoid issuing the read if the write covers 1131 * the entire buffer. We have to make sure the buffer state 1132 * is reasonable in this case since we will not be initiating 1133 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1134 * more information. 1135 * 1136 * B_CACHE may also be set due to the buffer being cached 1137 * normally. 1138 */ 1139 1140 bp_cached = 1; 1141 if (on == 0 && n == bcount) { 1142 if ((bp->b_flags & B_CACHE) == 0) 1143 bp_cached = 0; 1144 bp->b_flags |= B_CACHE; 1145 bp->b_flags &= ~B_INVAL; 1146 bp->b_ioflags &= ~BIO_ERROR; 1147 } 1148 1149 if ((bp->b_flags & B_CACHE) == 0) { 1150 bp->b_iocmd = BIO_READ; 1151 vfs_busy_pages(bp, 0); 1152 error = ncl_doio(vp, bp, cred, td, 0); 1153 if (error) { 1154 brelse(bp); 1155 break; 1156 } 1157 } 1158 if (bp->b_wcred == NOCRED) 1159 bp->b_wcred = crhold(cred); 1160 NFSLOCKNODE(np); 1161 np->n_flag |= NMODIFIED; 1162 NFSUNLOCKNODE(np); 1163 1164 /* 1165 * If dirtyend exceeds file size, chop it down. This should 1166 * not normally occur but there is an append race where it 1167 * might occur XXX, so we log it. 1168 * 1169 * If the chopping creates a reverse-indexed or degenerate 1170 * situation with dirtyoff/end, we 0 both of them. 1171 */ 1172 1173 if (bp->b_dirtyend > bcount) { 1174 printf("NFS append race @%lx:%d\n", 1175 (long)bp->b_blkno * DEV_BSIZE, 1176 bp->b_dirtyend - bcount); 1177 bp->b_dirtyend = bcount; 1178 } 1179 1180 if (bp->b_dirtyoff >= bp->b_dirtyend) 1181 bp->b_dirtyoff = bp->b_dirtyend = 0; 1182 1183 /* 1184 * If the new write will leave a contiguous dirty 1185 * area, just update the b_dirtyoff and b_dirtyend, 1186 * otherwise force a write rpc of the old dirty area. 1187 * 1188 * If there has been a file lock applied to this file 1189 * or vfs.nfs.old_noncontig_writing is set, do the following: 1190 * While it is possible to merge discontiguous writes due to 1191 * our having a B_CACHE buffer ( and thus valid read data 1192 * for the hole), we don't because it could lead to 1193 * significant cache coherency problems with multiple clients, 1194 * especially if locking is implemented later on. 1195 * 1196 * If vfs.nfs.old_noncontig_writing is not set and there has 1197 * not been file locking done on this file: 1198 * Relax coherency a bit for the sake of performance and 1199 * expand the current dirty region to contain the new 1200 * write even if it means we mark some non-dirty data as 1201 * dirty. 1202 */ 1203 1204 if (noncontig_write == 0 && bp->b_dirtyend > 0 && 1205 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1206 if (bwrite(bp) == EINTR) { 1207 error = EINTR; 1208 break; 1209 } 1210 goto again; 1211 } 1212 1213 local_resid = uio->uio_resid; 1214 error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio); 1215 1216 if (error != 0 && !bp_cached) { 1217 /* 1218 * This block has no other content then what 1219 * possibly was written by the faulty uiomove. 1220 * Release it, forgetting the data pages, to 1221 * prevent the leak of uninitialized data to 1222 * usermode. 1223 */ 1224 bp->b_ioflags |= BIO_ERROR; 1225 brelse(bp); 1226 uio->uio_offset -= local_resid - uio->uio_resid; 1227 uio->uio_resid = local_resid; 1228 break; 1229 } 1230 1231 /* 1232 * Since this block is being modified, it must be written 1233 * again and not just committed. Since write clustering does 1234 * not work for the stage 1 data write, only the stage 2 1235 * commit rpc, we have to clear B_CLUSTEROK as well. 1236 */ 1237 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1238 1239 /* 1240 * Get the partial update on the progress made from 1241 * uiomove, if an error occurred. 1242 */ 1243 if (error != 0) 1244 n = local_resid - uio->uio_resid; 1245 1246 /* 1247 * Only update dirtyoff/dirtyend if not a degenerate 1248 * condition. 1249 */ 1250 if (n > 0) { 1251 if (bp->b_dirtyend > 0) { 1252 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1253 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1254 } else { 1255 bp->b_dirtyoff = on; 1256 bp->b_dirtyend = on + n; 1257 } 1258 vfs_bio_set_valid(bp, on, n); 1259 } 1260 1261 /* 1262 * If IO_SYNC do bwrite(). 1263 * 1264 * IO_INVAL appears to be unused. The idea appears to be 1265 * to turn off caching in this case. Very odd. XXX 1266 */ 1267 if ((ioflag & IO_SYNC)) { 1268 if (ioflag & IO_INVAL) 1269 bp->b_flags |= B_NOCACHE; 1270 error1 = bwrite(bp); 1271 if (error1 != 0) { 1272 if (error == 0) 1273 error = error1; 1274 break; 1275 } 1276 } else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) { 1277 bp->b_flags |= B_ASYNC; 1278 (void) ncl_writebp(bp, 0, NULL); 1279 } else { 1280 bdwrite(bp); 1281 } 1282 1283 if (error != 0) 1284 break; 1285 } while (uio->uio_resid > 0 && n > 0); 1286 1287 if (error != 0) { 1288 if (ioflag & IO_UNIT) { 1289 VATTR_NULL(&vattr); 1290 vattr.va_size = orig_size; 1291 /* IO_SYNC is handled implicitely */ 1292 (void)VOP_SETATTR(vp, &vattr, cred); 1293 uio->uio_offset -= orig_resid - uio->uio_resid; 1294 uio->uio_resid = orig_resid; 1295 } 1296 } 1297 1298 out: 1299 curthread_pflags2_restore(save2); 1300 return (error); 1301 } 1302 1303 /* 1304 * Get an nfs cache block. 1305 * 1306 * Allocate a new one if the block isn't currently in the cache 1307 * and return the block marked busy. If the calling process is 1308 * interrupted by a signal for an interruptible mount point, return 1309 * NULL. 1310 * 1311 * The caller must carefully deal with the possible B_INVAL state of 1312 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it 1313 * indirectly), so synchronous reads can be issued without worrying about 1314 * the B_INVAL state. We have to be a little more careful when dealing 1315 * with writes (see comments in nfs_write()) when extending a file past 1316 * its EOF. 1317 */ 1318 static struct buf * 1319 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1320 { 1321 struct buf *bp; 1322 struct mount *mp; 1323 struct nfsmount *nmp; 1324 1325 mp = vp->v_mount; 1326 nmp = VFSTONFS(mp); 1327 1328 if (nmp->nm_flag & NFSMNT_INT) { 1329 sigset_t oldset; 1330 1331 newnfs_set_sigmask(td, &oldset); 1332 bp = getblk(vp, bn, size, PCATCH, 0, 0); 1333 newnfs_restore_sigmask(td, &oldset); 1334 while (bp == NULL) { 1335 if (newnfs_sigintr(nmp, td)) 1336 return (NULL); 1337 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1338 } 1339 } else { 1340 bp = getblk(vp, bn, size, 0, 0, 0); 1341 } 1342 1343 if (vp->v_type == VREG) 1344 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE); 1345 return (bp); 1346 } 1347 1348 /* 1349 * Flush and invalidate all dirty buffers. If another process is already 1350 * doing the flush, just wait for completion. 1351 */ 1352 int 1353 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1354 { 1355 struct nfsnode *np = VTONFS(vp); 1356 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1357 int error = 0, slpflag, slptimeo; 1358 bool old_lock; 1359 1360 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf"); 1361 1362 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1363 intrflg = 0; 1364 if (NFSCL_FORCEDISM(nmp->nm_mountp)) 1365 intrflg = 1; 1366 if (intrflg) { 1367 slpflag = PCATCH; 1368 slptimeo = 2 * hz; 1369 } else { 1370 slpflag = 0; 1371 slptimeo = 0; 1372 } 1373 1374 old_lock = ncl_excl_start(vp); 1375 if (old_lock) 1376 flags |= V_ALLOWCLEAN; 1377 1378 /* 1379 * Now, flush as required. 1380 */ 1381 if ((flags & (V_SAVE | V_VMIO)) == V_SAVE && 1382 vp->v_bufobj.bo_object != NULL) { 1383 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object); 1384 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1385 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object); 1386 /* 1387 * If the page clean was interrupted, fail the invalidation. 1388 * Not doing so, we run the risk of losing dirty pages in the 1389 * vinvalbuf() call below. 1390 */ 1391 if (intrflg && (error = newnfs_sigintr(nmp, td))) 1392 goto out; 1393 } 1394 1395 error = vinvalbuf(vp, flags, slpflag, 0); 1396 while (error) { 1397 if (intrflg && (error = newnfs_sigintr(nmp, td))) 1398 goto out; 1399 error = vinvalbuf(vp, flags, 0, slptimeo); 1400 } 1401 if (NFSHASPNFS(nmp)) { 1402 nfscl_layoutcommit(vp, td); 1403 /* 1404 * Invalidate the attribute cache, since writes to a DS 1405 * won't update the size attribute. 1406 */ 1407 NFSLOCKNODE(np); 1408 np->n_attrstamp = 0; 1409 } else 1410 NFSLOCKNODE(np); 1411 if (np->n_directio_asyncwr == 0) 1412 np->n_flag &= ~NMODIFIED; 1413 NFSUNLOCKNODE(np); 1414 out: 1415 ncl_excl_finish(vp, old_lock); 1416 return error; 1417 } 1418 1419 /* 1420 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1421 * This is mainly to avoid queueing async I/O requests when the nfsiods 1422 * are all hung on a dead server. 1423 * 1424 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1425 * is eventually dequeued by the async daemon, ncl_doio() *will*. 1426 */ 1427 int 1428 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1429 { 1430 int iod; 1431 int gotiod; 1432 int slpflag = 0; 1433 int slptimeo = 0; 1434 int error, error2; 1435 1436 /* 1437 * Commits are usually short and sweet so lets save some cpu and 1438 * leave the async daemons for more important rpc's (such as reads 1439 * and writes). 1440 * 1441 * Readdirplus RPCs do vget()s to acquire the vnodes for entries 1442 * in the directory in order to update attributes. This can deadlock 1443 * with another thread that is waiting for async I/O to be done by 1444 * an nfsiod thread while holding a lock on one of these vnodes. 1445 * To avoid this deadlock, don't allow the async nfsiod threads to 1446 * perform Readdirplus RPCs. 1447 */ 1448 NFSLOCKIOD(); 1449 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1450 (nmp->nm_bufqiods > ncl_numasync / 2)) || 1451 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) { 1452 NFSUNLOCKIOD(); 1453 return(EIO); 1454 } 1455 again: 1456 if (nmp->nm_flag & NFSMNT_INT) 1457 slpflag = PCATCH; 1458 gotiod = FALSE; 1459 1460 /* 1461 * Find a free iod to process this request. 1462 */ 1463 for (iod = 0; iod < ncl_numasync; iod++) 1464 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) { 1465 gotiod = TRUE; 1466 break; 1467 } 1468 1469 /* 1470 * Try to create one if none are free. 1471 */ 1472 if (!gotiod) 1473 ncl_nfsiodnew(); 1474 else { 1475 /* 1476 * Found one, so wake it up and tell it which 1477 * mount to process. 1478 */ 1479 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n", 1480 iod, nmp)); 1481 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE; 1482 ncl_iodmount[iod] = nmp; 1483 nmp->nm_bufqiods++; 1484 wakeup(&ncl_iodwant[iod]); 1485 } 1486 1487 /* 1488 * If none are free, we may already have an iod working on this mount 1489 * point. If so, it will process our request. 1490 */ 1491 if (!gotiod) { 1492 if (nmp->nm_bufqiods > 0) { 1493 NFS_DPF(ASYNCIO, 1494 ("ncl_asyncio: %d iods are already processing mount %p\n", 1495 nmp->nm_bufqiods, nmp)); 1496 gotiod = TRUE; 1497 } 1498 } 1499 1500 /* 1501 * If we have an iod which can process the request, then queue 1502 * the buffer. 1503 */ 1504 if (gotiod) { 1505 /* 1506 * Ensure that the queue never grows too large. We still want 1507 * to asynchronize so we block rather then return EIO. 1508 */ 1509 while (nmp->nm_bufqlen >= 2*ncl_numasync) { 1510 NFS_DPF(ASYNCIO, 1511 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp)); 1512 nmp->nm_bufqwant = TRUE; 1513 error = newnfs_msleep(td, &nmp->nm_bufq, 1514 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio", 1515 slptimeo); 1516 if (error) { 1517 error2 = newnfs_sigintr(nmp, td); 1518 if (error2) { 1519 NFSUNLOCKIOD(); 1520 return (error2); 1521 } 1522 if (slpflag == PCATCH) { 1523 slpflag = 0; 1524 slptimeo = 2 * hz; 1525 } 1526 } 1527 /* 1528 * We might have lost our iod while sleeping, 1529 * so check and loop if necessary. 1530 */ 1531 goto again; 1532 } 1533 1534 /* We might have lost our nfsiod */ 1535 if (nmp->nm_bufqiods == 0) { 1536 NFS_DPF(ASYNCIO, 1537 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1538 goto again; 1539 } 1540 1541 if (bp->b_iocmd == BIO_READ) { 1542 if (bp->b_rcred == NOCRED && cred != NOCRED) 1543 bp->b_rcred = crhold(cred); 1544 } else { 1545 if (bp->b_wcred == NOCRED && cred != NOCRED) 1546 bp->b_wcred = crhold(cred); 1547 } 1548 1549 if (bp->b_flags & B_REMFREE) 1550 bremfreef(bp); 1551 BUF_KERNPROC(bp); 1552 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1553 nmp->nm_bufqlen++; 1554 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1555 NFSLOCKNODE(VTONFS(bp->b_vp)); 1556 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1557 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1558 NFSUNLOCKNODE(VTONFS(bp->b_vp)); 1559 } 1560 NFSUNLOCKIOD(); 1561 return (0); 1562 } 1563 1564 NFSUNLOCKIOD(); 1565 1566 /* 1567 * All the iods are busy on other mounts, so return EIO to 1568 * force the caller to process the i/o synchronously. 1569 */ 1570 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n")); 1571 return (EIO); 1572 } 1573 1574 void 1575 ncl_doio_directwrite(struct buf *bp) 1576 { 1577 int iomode, must_commit; 1578 struct uio *uiop = (struct uio *)bp->b_caller1; 1579 char *iov_base = uiop->uio_iov->iov_base; 1580 1581 iomode = NFSWRITE_FILESYNC; 1582 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1583 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0); 1584 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write")); 1585 free(iov_base, M_NFSDIRECTIO); 1586 free(uiop->uio_iov, M_NFSDIRECTIO); 1587 free(uiop, M_NFSDIRECTIO); 1588 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1589 struct nfsnode *np = VTONFS(bp->b_vp); 1590 NFSLOCKNODE(np); 1591 if (NFSHASPNFS(VFSTONFS(bp->b_vp->v_mount))) { 1592 /* 1593 * Invalidate the attribute cache, since writes to a DS 1594 * won't update the size attribute. 1595 */ 1596 np->n_attrstamp = 0; 1597 } 1598 np->n_directio_asyncwr--; 1599 if (np->n_directio_asyncwr == 0) { 1600 np->n_flag &= ~NMODIFIED; 1601 if ((np->n_flag & NFSYNCWAIT)) { 1602 np->n_flag &= ~NFSYNCWAIT; 1603 wakeup((caddr_t)&np->n_directio_asyncwr); 1604 } 1605 } 1606 NFSUNLOCKNODE(np); 1607 } 1608 bp->b_vp = NULL; 1609 uma_zfree(ncl_pbuf_zone, bp); 1610 } 1611 1612 /* 1613 * Do an I/O operation to/from a cache block. This may be called 1614 * synchronously or from an nfsiod. 1615 */ 1616 int 1617 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td, 1618 int called_from_strategy) 1619 { 1620 struct uio *uiop; 1621 struct nfsnode *np; 1622 struct nfsmount *nmp; 1623 int error = 0, iomode, must_commit = 0; 1624 struct uio uio; 1625 struct iovec io; 1626 struct proc *p = td ? td->td_proc : NULL; 1627 uint8_t iocmd; 1628 1629 np = VTONFS(vp); 1630 nmp = VFSTONFS(vp->v_mount); 1631 uiop = &uio; 1632 uiop->uio_iov = &io; 1633 uiop->uio_iovcnt = 1; 1634 uiop->uio_segflg = UIO_SYSSPACE; 1635 uiop->uio_td = td; 1636 1637 /* 1638 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1639 * do this here so we do not have to do it in all the code that 1640 * calls us. 1641 */ 1642 bp->b_flags &= ~B_INVAL; 1643 bp->b_ioflags &= ~BIO_ERROR; 1644 1645 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp)); 1646 iocmd = bp->b_iocmd; 1647 if (iocmd == BIO_READ) { 1648 io.iov_len = uiop->uio_resid = bp->b_bcount; 1649 io.iov_base = bp->b_data; 1650 uiop->uio_rw = UIO_READ; 1651 1652 switch (vp->v_type) { 1653 case VREG: 1654 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1655 NFSINCRGLOBAL(nfsstatsv1.read_bios); 1656 error = ncl_readrpc(vp, uiop, cr); 1657 1658 if (!error) { 1659 if (uiop->uio_resid) { 1660 /* 1661 * If we had a short read with no error, we must have 1662 * hit a file hole. We should zero-fill the remainder. 1663 * This can also occur if the server hits the file EOF. 1664 * 1665 * Holes used to be able to occur due to pending 1666 * writes, but that is not possible any longer. 1667 */ 1668 int nread = bp->b_bcount - uiop->uio_resid; 1669 ssize_t left = uiop->uio_resid; 1670 1671 if (left > 0) 1672 bzero((char *)bp->b_data + nread, left); 1673 uiop->uio_resid = 0; 1674 } 1675 } 1676 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */ 1677 if (p && vp->v_writecount <= -1) { 1678 NFSLOCKNODE(np); 1679 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) { 1680 NFSUNLOCKNODE(np); 1681 PROC_LOCK(p); 1682 killproc(p, "text file modification"); 1683 PROC_UNLOCK(p); 1684 } else 1685 NFSUNLOCKNODE(np); 1686 } 1687 break; 1688 case VLNK: 1689 uiop->uio_offset = (off_t)0; 1690 NFSINCRGLOBAL(nfsstatsv1.readlink_bios); 1691 error = ncl_readlinkrpc(vp, uiop, cr); 1692 break; 1693 case VDIR: 1694 NFSINCRGLOBAL(nfsstatsv1.readdir_bios); 1695 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1696 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1697 error = ncl_readdirplusrpc(vp, uiop, cr, td); 1698 if (error == NFSERR_NOTSUPP) 1699 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1700 } 1701 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1702 error = ncl_readdirrpc(vp, uiop, cr, td); 1703 /* 1704 * end-of-directory sets B_INVAL but does not generate an 1705 * error. 1706 */ 1707 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1708 bp->b_flags |= B_INVAL; 1709 break; 1710 default: 1711 printf("ncl_doio: type %x unexpected\n", vp->v_type); 1712 break; 1713 } 1714 if (error) { 1715 bp->b_ioflags |= BIO_ERROR; 1716 bp->b_error = error; 1717 } 1718 } else { 1719 /* 1720 * If we only need to commit, try to commit 1721 */ 1722 if (bp->b_flags & B_NEEDCOMMIT) { 1723 int retv; 1724 off_t off; 1725 1726 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1727 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1728 bp->b_wcred, td); 1729 if (retv == 0) { 1730 bp->b_dirtyoff = bp->b_dirtyend = 0; 1731 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1732 bp->b_resid = 0; 1733 bufdone(bp); 1734 return (0); 1735 } 1736 if (retv == NFSERR_STALEWRITEVERF) { 1737 ncl_clearcommit(vp->v_mount); 1738 } 1739 } 1740 1741 /* 1742 * Setup for actual write 1743 */ 1744 NFSLOCKNODE(np); 1745 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1746 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1747 NFSUNLOCKNODE(np); 1748 1749 if (bp->b_dirtyend > bp->b_dirtyoff) { 1750 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1751 - bp->b_dirtyoff; 1752 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1753 + bp->b_dirtyoff; 1754 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1755 uiop->uio_rw = UIO_WRITE; 1756 NFSINCRGLOBAL(nfsstatsv1.write_bios); 1757 1758 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1759 iomode = NFSWRITE_UNSTABLE; 1760 else 1761 iomode = NFSWRITE_FILESYNC; 1762 1763 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit, 1764 called_from_strategy); 1765 1766 /* 1767 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1768 * to cluster the buffers needing commit. This will allow 1769 * the system to submit a single commit rpc for the whole 1770 * cluster. We can do this even if the buffer is not 100% 1771 * dirty (relative to the NFS blocksize), so we optimize the 1772 * append-to-file-case. 1773 * 1774 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1775 * cleared because write clustering only works for commit 1776 * rpc's, not for the data portion of the write). 1777 */ 1778 1779 if (!error && iomode == NFSWRITE_UNSTABLE) { 1780 bp->b_flags |= B_NEEDCOMMIT; 1781 if (bp->b_dirtyoff == 0 1782 && bp->b_dirtyend == bp->b_bcount) 1783 bp->b_flags |= B_CLUSTEROK; 1784 } else { 1785 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1786 } 1787 1788 /* 1789 * For an interrupted write, the buffer is still valid 1790 * and the write hasn't been pushed to the server yet, 1791 * so we can't set BIO_ERROR and report the interruption 1792 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1793 * is not relevant, so the rpc attempt is essentially 1794 * a noop. For the case of a V3 write rpc not being 1795 * committed to stable storage, the block is still 1796 * dirty and requires either a commit rpc or another 1797 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1798 * the block is reused. This is indicated by setting 1799 * the B_DELWRI and B_NEEDCOMMIT flags. 1800 * 1801 * EIO is returned by ncl_writerpc() to indicate a recoverable 1802 * write error and is handled as above, except that 1803 * B_EINTR isn't set. One cause of this is a stale stateid 1804 * error for the RPC that indicates recovery is required, 1805 * when called with called_from_strategy != 0. 1806 * 1807 * If the buffer is marked B_PAGING, it does not reside on 1808 * the vp's paging queues so we cannot call bdirty(). The 1809 * bp in this case is not an NFS cache block so we should 1810 * be safe. XXX 1811 * 1812 * The logic below breaks up errors into recoverable and 1813 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1814 * and keep the buffer around for potential write retries. 1815 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1816 * and save the error in the nfsnode. This is less than ideal 1817 * but necessary. Keeping such buffers around could potentially 1818 * cause buffer exhaustion eventually (they can never be written 1819 * out, so will get constantly be re-dirtied). It also causes 1820 * all sorts of vfs panics. For non-recoverable write errors, 1821 * also invalidate the attrcache, so we'll be forced to go over 1822 * the wire for this object, returning an error to user on next 1823 * call (most of the time). 1824 */ 1825 if (error == EINTR || error == EIO || error == ETIMEDOUT 1826 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1827 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1828 if ((bp->b_flags & B_PAGING) == 0) { 1829 bdirty(bp); 1830 bp->b_flags &= ~B_DONE; 1831 } 1832 if ((error == EINTR || error == ETIMEDOUT) && 1833 (bp->b_flags & B_ASYNC) == 0) 1834 bp->b_flags |= B_EINTR; 1835 } else { 1836 if (error) { 1837 bp->b_ioflags |= BIO_ERROR; 1838 bp->b_flags |= B_INVAL; 1839 bp->b_error = np->n_error = error; 1840 NFSLOCKNODE(np); 1841 np->n_flag |= NWRITEERR; 1842 np->n_attrstamp = 0; 1843 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1844 NFSUNLOCKNODE(np); 1845 } 1846 bp->b_dirtyoff = bp->b_dirtyend = 0; 1847 } 1848 } else { 1849 bp->b_resid = 0; 1850 bufdone(bp); 1851 return (0); 1852 } 1853 } 1854 bp->b_resid = uiop->uio_resid; 1855 if (must_commit) 1856 ncl_clearcommit(vp->v_mount); 1857 bufdone(bp); 1858 return (error); 1859 } 1860 1861 /* 1862 * Used to aid in handling ftruncate() operations on the NFS client side. 1863 * Truncation creates a number of special problems for NFS. We have to 1864 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1865 * we have to properly handle VM pages or (potentially dirty) buffers 1866 * that straddle the truncation point. 1867 */ 1868 1869 int 1870 ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize) 1871 { 1872 struct nfsnode *np = VTONFS(vp); 1873 u_quad_t tsize; 1874 int biosize = vp->v_bufobj.bo_bsize; 1875 int error = 0; 1876 1877 NFSLOCKNODE(np); 1878 tsize = np->n_size; 1879 np->n_size = nsize; 1880 NFSUNLOCKNODE(np); 1881 1882 if (nsize < tsize) { 1883 struct buf *bp; 1884 daddr_t lbn; 1885 int bufsize; 1886 1887 /* 1888 * vtruncbuf() doesn't get the buffer overlapping the 1889 * truncation point. We may have a B_DELWRI and/or B_CACHE 1890 * buffer that now needs to be truncated. 1891 */ 1892 error = vtruncbuf(vp, nsize, biosize); 1893 lbn = nsize / biosize; 1894 bufsize = nsize - (lbn * biosize); 1895 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1896 if (!bp) 1897 return EINTR; 1898 if (bp->b_dirtyoff > bp->b_bcount) 1899 bp->b_dirtyoff = bp->b_bcount; 1900 if (bp->b_dirtyend > bp->b_bcount) 1901 bp->b_dirtyend = bp->b_bcount; 1902 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1903 brelse(bp); 1904 } else { 1905 vnode_pager_setsize(vp, nsize); 1906 } 1907 return(error); 1908 } 1909