1 /* $NetBSD: nfs_clbio.c,v 1.2 2013/12/09 09:35:17 wiz Exp $ */ 2 /*- 3 * Copyright (c) 1989, 1993 4 * The Regents of the University of California. All rights reserved. 5 * 6 * This code is derived from software contributed to Berkeley by 7 * Rick Macklem at The University of Guelph. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 4. Neither the name of the University nor the names of its contributors 18 * may be used to endorse or promote products derived from this software 19 * without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 * 33 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 34 */ 35 36 #include <sys/cdefs.h> 37 /* __FBSDID("FreeBSD: head/sys/fs/nfsclient/nfs_clbio.c 252072 2013-06-21 22:26:18Z rmacklem "); */ 38 __RCSID("$NetBSD: nfs_clbio.c,v 1.2 2013/12/09 09:35:17 wiz Exp $"); 39 40 #include "opt_kdtrace.h" 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/bio.h> 45 #include <sys/buf.h> 46 #include <sys/kernel.h> 47 #include <sys/mount.h> 48 #include <sys/rwlock.h> 49 #include <sys/vmmeter.h> 50 #include <sys/vnode.h> 51 52 #include <vm/vm.h> 53 #include <vm/vm_param.h> 54 #include <vm/vm_extern.h> 55 #include <vm/vm_page.h> 56 #include <vm/vm_object.h> 57 #include <vm/vm_pager.h> 58 #include <vm/vnode_pager.h> 59 60 #include <fs/nfs/nfsport.h> 61 #include <fs/nfsclient/nfsmount.h> 62 #include <fs/nfsclient/nfs.h> 63 #include <fs/nfsclient/nfsnode.h> 64 #include <fs/nfsclient/nfs_kdtrace.h> 65 66 extern int newnfs_directio_allow_mmap; 67 extern struct nfsstats newnfsstats; 68 extern struct mtx ncl_iod_mutex; 69 extern int ncl_numasync; 70 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON]; 71 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON]; 72 extern int newnfs_directio_enable; 73 extern int nfs_keep_dirty_on_error; 74 75 int ncl_pbuf_freecnt = -1; /* start out unlimited */ 76 77 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 78 struct thread *td); 79 static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 80 struct ucred *cred, int ioflag); 81 82 /* 83 * Vnode op for VM getpages. 84 */ 85 int 86 ncl_getpages(struct vop_getpages_args *ap) 87 { 88 int i, error, nextoff, size, toff, count, npages; 89 struct uio uio; 90 struct iovec iov; 91 vm_offset_t kva; 92 struct buf *bp; 93 struct vnode *vp; 94 struct thread *td; 95 struct ucred *cred; 96 struct nfsmount *nmp; 97 vm_object_t object; 98 vm_page_t *pages; 99 struct nfsnode *np; 100 101 vp = ap->a_vp; 102 np = VTONFS(vp); 103 td = curthread; /* XXX */ 104 cred = curthread->td_ucred; /* XXX */ 105 nmp = VFSTONFS(vp->v_mount); 106 pages = ap->a_m; 107 count = ap->a_count; 108 109 if ((object = vp->v_object) == NULL) { 110 ncl_printf("nfs_getpages: called with non-merged cache vnode??\n"); 111 return (VM_PAGER_ERROR); 112 } 113 114 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) { 115 mtx_lock(&np->n_mtx); 116 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 117 mtx_unlock(&np->n_mtx); 118 ncl_printf("nfs_getpages: called on non-cacheable vnode??\n"); 119 return (VM_PAGER_ERROR); 120 } else 121 mtx_unlock(&np->n_mtx); 122 } 123 124 mtx_lock(&nmp->nm_mtx); 125 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 126 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 127 mtx_unlock(&nmp->nm_mtx); 128 /* We'll never get here for v4, because we always have fsinfo */ 129 (void)ncl_fsinfo(nmp, vp, cred, td); 130 } else 131 mtx_unlock(&nmp->nm_mtx); 132 133 npages = btoc(count); 134 135 /* 136 * If the requested page is partially valid, just return it and 137 * allow the pager to zero-out the blanks. Partially valid pages 138 * can only occur at the file EOF. 139 */ 140 VM_OBJECT_WLOCK(object); 141 if (pages[ap->a_reqpage]->valid != 0) { 142 for (i = 0; i < npages; ++i) { 143 if (i != ap->a_reqpage) { 144 vm_page_lock(pages[i]); 145 vm_page_free(pages[i]); 146 vm_page_unlock(pages[i]); 147 } 148 } 149 VM_OBJECT_WUNLOCK(object); 150 return (0); 151 } 152 VM_OBJECT_WUNLOCK(object); 153 154 /* 155 * We use only the kva address for the buffer, but this is extremely 156 * convienient and fast. 157 */ 158 bp = getpbuf(&ncl_pbuf_freecnt); 159 160 kva = (vm_offset_t) bp->b_data; 161 pmap_qenter(kva, pages, npages); 162 PCPU_INC(cnt.v_vnodein); 163 PCPU_ADD(cnt.v_vnodepgsin, npages); 164 165 iov.iov_base = (caddr_t) kva; 166 iov.iov_len = count; 167 uio.uio_iov = &iov; 168 uio.uio_iovcnt = 1; 169 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 170 uio.uio_resid = count; 171 uio.uio_segflg = UIO_SYSSPACE; 172 uio.uio_rw = UIO_READ; 173 uio.uio_td = td; 174 175 error = ncl_readrpc(vp, &uio, cred); 176 pmap_qremove(kva, npages); 177 178 relpbuf(bp, &ncl_pbuf_freecnt); 179 180 if (error && (uio.uio_resid == count)) { 181 ncl_printf("nfs_getpages: error %d\n", error); 182 VM_OBJECT_WLOCK(object); 183 for (i = 0; i < npages; ++i) { 184 if (i != ap->a_reqpage) { 185 vm_page_lock(pages[i]); 186 vm_page_free(pages[i]); 187 vm_page_unlock(pages[i]); 188 } 189 } 190 VM_OBJECT_WUNLOCK(object); 191 return (VM_PAGER_ERROR); 192 } 193 194 /* 195 * Calculate the number of bytes read and validate only that number 196 * of bytes. Note that due to pending writes, size may be 0. This 197 * does not mean that the remaining data is invalid! 198 */ 199 200 size = count - uio.uio_resid; 201 VM_OBJECT_WLOCK(object); 202 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 203 vm_page_t m; 204 nextoff = toff + PAGE_SIZE; 205 m = pages[i]; 206 207 if (nextoff <= size) { 208 /* 209 * Read operation filled an entire page 210 */ 211 m->valid = VM_PAGE_BITS_ALL; 212 KASSERT(m->dirty == 0, 213 ("nfs_getpages: page %p is dirty", m)); 214 } else if (size > toff) { 215 /* 216 * Read operation filled a partial page. 217 */ 218 m->valid = 0; 219 vm_page_set_valid_range(m, 0, size - toff); 220 KASSERT(m->dirty == 0, 221 ("nfs_getpages: page %p is dirty", m)); 222 } else { 223 /* 224 * Read operation was short. If no error 225 * occured we may have hit a zero-fill 226 * section. We leave valid set to 0, and page 227 * is freed by vm_page_readahead_finish() if 228 * its index is not equal to requested, or 229 * page is zeroed and set valid by 230 * vm_pager_get_pages() for requested page. 231 */ 232 ; 233 } 234 if (i != ap->a_reqpage) 235 vm_page_readahead_finish(m); 236 } 237 VM_OBJECT_WUNLOCK(object); 238 return (0); 239 } 240 241 /* 242 * Vnode op for VM putpages. 243 */ 244 int 245 ncl_putpages(struct vop_putpages_args *ap) 246 { 247 struct uio uio; 248 struct iovec iov; 249 vm_offset_t kva; 250 struct buf *bp; 251 int iomode, must_commit, i, error, npages, count; 252 off_t offset; 253 int *rtvals; 254 struct vnode *vp; 255 struct thread *td; 256 struct ucred *cred; 257 struct nfsmount *nmp; 258 struct nfsnode *np; 259 vm_page_t *pages; 260 261 vp = ap->a_vp; 262 np = VTONFS(vp); 263 td = curthread; /* XXX */ 264 /* Set the cred to n_writecred for the write rpcs. */ 265 if (np->n_writecred != NULL) 266 cred = crhold(np->n_writecred); 267 else 268 cred = crhold(curthread->td_ucred); /* XXX */ 269 nmp = VFSTONFS(vp->v_mount); 270 pages = ap->a_m; 271 count = ap->a_count; 272 rtvals = ap->a_rtvals; 273 npages = btoc(count); 274 offset = IDX_TO_OFF(pages[0]->pindex); 275 276 mtx_lock(&nmp->nm_mtx); 277 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 278 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 279 mtx_unlock(&nmp->nm_mtx); 280 (void)ncl_fsinfo(nmp, vp, cred, td); 281 } else 282 mtx_unlock(&nmp->nm_mtx); 283 284 mtx_lock(&np->n_mtx); 285 if (newnfs_directio_enable && !newnfs_directio_allow_mmap && 286 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 287 mtx_unlock(&np->n_mtx); 288 ncl_printf("ncl_putpages: called on noncache-able vnode??\n"); 289 mtx_lock(&np->n_mtx); 290 } 291 292 for (i = 0; i < npages; i++) 293 rtvals[i] = VM_PAGER_ERROR; 294 295 /* 296 * When putting pages, do not extend file past EOF. 297 */ 298 if (offset + count > np->n_size) { 299 count = np->n_size - offset; 300 if (count < 0) 301 count = 0; 302 } 303 mtx_unlock(&np->n_mtx); 304 305 /* 306 * We use only the kva address for the buffer, but this is extremely 307 * convienient and fast. 308 */ 309 bp = getpbuf(&ncl_pbuf_freecnt); 310 311 kva = (vm_offset_t) bp->b_data; 312 pmap_qenter(kva, pages, npages); 313 PCPU_INC(cnt.v_vnodeout); 314 PCPU_ADD(cnt.v_vnodepgsout, count); 315 316 iov.iov_base = (caddr_t) kva; 317 iov.iov_len = count; 318 uio.uio_iov = &iov; 319 uio.uio_iovcnt = 1; 320 uio.uio_offset = offset; 321 uio.uio_resid = count; 322 uio.uio_segflg = UIO_SYSSPACE; 323 uio.uio_rw = UIO_WRITE; 324 uio.uio_td = td; 325 326 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 327 iomode = NFSWRITE_UNSTABLE; 328 else 329 iomode = NFSWRITE_FILESYNC; 330 331 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0); 332 crfree(cred); 333 334 pmap_qremove(kva, npages); 335 relpbuf(bp, &ncl_pbuf_freecnt); 336 337 if (error == 0 || !nfs_keep_dirty_on_error) { 338 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid); 339 if (must_commit) 340 ncl_clearcommit(vp->v_mount); 341 } 342 return rtvals[0]; 343 } 344 345 /* 346 * For nfs, cache consistency can only be maintained approximately. 347 * Although RFC1094 does not specify the criteria, the following is 348 * believed to be compatible with the reference port. 349 * For nfs: 350 * If the file's modify time on the server has changed since the 351 * last read rpc or you have written to the file, 352 * you may have lost data cache consistency with the 353 * server, so flush all of the file's data out of the cache. 354 * Then force a getattr rpc to ensure that you have up to date 355 * attributes. 356 * NB: This implies that cache data can be read when up to 357 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 358 * attributes this could be forced by setting n_attrstamp to 0 before 359 * the VOP_GETATTR() call. 360 */ 361 static inline int 362 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 363 { 364 int error = 0; 365 struct vattr vattr; 366 struct nfsnode *np = VTONFS(vp); 367 int old_lock; 368 369 /* 370 * Grab the exclusive lock before checking whether the cache is 371 * consistent. 372 * XXX - We can make this cheaper later (by acquiring cheaper locks). 373 * But for now, this suffices. 374 */ 375 old_lock = ncl_upgrade_vnlock(vp); 376 if (vp->v_iflag & VI_DOOMED) { 377 ncl_downgrade_vnlock(vp, old_lock); 378 return (EBADF); 379 } 380 381 mtx_lock(&np->n_mtx); 382 if (np->n_flag & NMODIFIED) { 383 mtx_unlock(&np->n_mtx); 384 if (vp->v_type != VREG) { 385 if (vp->v_type != VDIR) 386 panic("nfs: bioread, not dir"); 387 ncl_invaldir(vp); 388 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 389 if (error) 390 goto out; 391 } 392 np->n_attrstamp = 0; 393 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 394 error = VOP_GETATTR(vp, &vattr, cred); 395 if (error) 396 goto out; 397 mtx_lock(&np->n_mtx); 398 np->n_mtime = vattr.va_mtime; 399 mtx_unlock(&np->n_mtx); 400 } else { 401 mtx_unlock(&np->n_mtx); 402 error = VOP_GETATTR(vp, &vattr, cred); 403 if (error) 404 return (error); 405 mtx_lock(&np->n_mtx); 406 if ((np->n_flag & NSIZECHANGED) 407 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 408 mtx_unlock(&np->n_mtx); 409 if (vp->v_type == VDIR) 410 ncl_invaldir(vp); 411 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 412 if (error) 413 goto out; 414 mtx_lock(&np->n_mtx); 415 np->n_mtime = vattr.va_mtime; 416 np->n_flag &= ~NSIZECHANGED; 417 } 418 mtx_unlock(&np->n_mtx); 419 } 420 out: 421 ncl_downgrade_vnlock(vp, old_lock); 422 return error; 423 } 424 425 /* 426 * Vnode op for read using bio 427 */ 428 int 429 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 430 { 431 struct nfsnode *np = VTONFS(vp); 432 int biosize, i; 433 struct buf *bp, *rabp; 434 struct thread *td; 435 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 436 daddr_t lbn, rabn; 437 int bcount; 438 int seqcount; 439 int nra, error = 0, n = 0, on = 0; 440 off_t tmp_off; 441 442 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode")); 443 if (uio->uio_resid == 0) 444 return (0); 445 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 446 return (EINVAL); 447 td = uio->uio_td; 448 449 mtx_lock(&nmp->nm_mtx); 450 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 451 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 452 mtx_unlock(&nmp->nm_mtx); 453 (void)ncl_fsinfo(nmp, vp, cred, td); 454 mtx_lock(&nmp->nm_mtx); 455 } 456 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0) 457 (void) newnfs_iosize(nmp); 458 459 tmp_off = uio->uio_offset + uio->uio_resid; 460 if (vp->v_type != VDIR && 461 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) { 462 mtx_unlock(&nmp->nm_mtx); 463 return (EFBIG); 464 } 465 mtx_unlock(&nmp->nm_mtx); 466 467 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 468 /* No caching/ no readaheads. Just read data into the user buffer */ 469 return ncl_readrpc(vp, uio, cred); 470 471 biosize = vp->v_bufobj.bo_bsize; 472 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 473 474 error = nfs_bioread_check_cons(vp, td, cred); 475 if (error) 476 return error; 477 478 do { 479 u_quad_t nsize; 480 481 mtx_lock(&np->n_mtx); 482 nsize = np->n_size; 483 mtx_unlock(&np->n_mtx); 484 485 switch (vp->v_type) { 486 case VREG: 487 NFSINCRGLOBAL(newnfsstats.biocache_reads); 488 lbn = uio->uio_offset / biosize; 489 on = uio->uio_offset - (lbn * biosize); 490 491 /* 492 * Start the read ahead(s), as required. 493 */ 494 if (nmp->nm_readahead > 0) { 495 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 496 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 497 rabn = lbn + 1 + nra; 498 if (incore(&vp->v_bufobj, rabn) == NULL) { 499 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 500 if (!rabp) { 501 error = newnfs_sigintr(nmp, td); 502 return (error ? error : EINTR); 503 } 504 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 505 rabp->b_flags |= B_ASYNC; 506 rabp->b_iocmd = BIO_READ; 507 vfs_busy_pages(rabp, 0); 508 if (ncl_asyncio(nmp, rabp, cred, td)) { 509 rabp->b_flags |= B_INVAL; 510 rabp->b_ioflags |= BIO_ERROR; 511 vfs_unbusy_pages(rabp); 512 brelse(rabp); 513 break; 514 } 515 } else { 516 brelse(rabp); 517 } 518 } 519 } 520 } 521 522 /* Note that bcount is *not* DEV_BSIZE aligned. */ 523 bcount = biosize; 524 if ((off_t)lbn * biosize >= nsize) { 525 bcount = 0; 526 } else if ((off_t)(lbn + 1) * biosize > nsize) { 527 bcount = nsize - (off_t)lbn * biosize; 528 } 529 bp = nfs_getcacheblk(vp, lbn, bcount, td); 530 531 if (!bp) { 532 error = newnfs_sigintr(nmp, td); 533 return (error ? error : EINTR); 534 } 535 536 /* 537 * If B_CACHE is not set, we must issue the read. If this 538 * fails, we return an error. 539 */ 540 541 if ((bp->b_flags & B_CACHE) == 0) { 542 bp->b_iocmd = BIO_READ; 543 vfs_busy_pages(bp, 0); 544 error = ncl_doio(vp, bp, cred, td, 0); 545 if (error) { 546 brelse(bp); 547 return (error); 548 } 549 } 550 551 /* 552 * on is the offset into the current bp. Figure out how many 553 * bytes we can copy out of the bp. Note that bcount is 554 * NOT DEV_BSIZE aligned. 555 * 556 * Then figure out how many bytes we can copy into the uio. 557 */ 558 559 n = 0; 560 if (on < bcount) 561 n = MIN((unsigned)(bcount - on), uio->uio_resid); 562 break; 563 case VLNK: 564 NFSINCRGLOBAL(newnfsstats.biocache_readlinks); 565 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 566 if (!bp) { 567 error = newnfs_sigintr(nmp, td); 568 return (error ? error : EINTR); 569 } 570 if ((bp->b_flags & B_CACHE) == 0) { 571 bp->b_iocmd = BIO_READ; 572 vfs_busy_pages(bp, 0); 573 error = ncl_doio(vp, bp, cred, td, 0); 574 if (error) { 575 bp->b_ioflags |= BIO_ERROR; 576 brelse(bp); 577 return (error); 578 } 579 } 580 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 581 on = 0; 582 break; 583 case VDIR: 584 NFSINCRGLOBAL(newnfsstats.biocache_readdirs); 585 if (np->n_direofoffset 586 && uio->uio_offset >= np->n_direofoffset) { 587 return (0); 588 } 589 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 590 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 591 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 592 if (!bp) { 593 error = newnfs_sigintr(nmp, td); 594 return (error ? error : EINTR); 595 } 596 if ((bp->b_flags & B_CACHE) == 0) { 597 bp->b_iocmd = BIO_READ; 598 vfs_busy_pages(bp, 0); 599 error = ncl_doio(vp, bp, cred, td, 0); 600 if (error) { 601 brelse(bp); 602 } 603 while (error == NFSERR_BAD_COOKIE) { 604 ncl_invaldir(vp); 605 error = ncl_vinvalbuf(vp, 0, td, 1); 606 /* 607 * Yuck! The directory has been modified on the 608 * server. The only way to get the block is by 609 * reading from the beginning to get all the 610 * offset cookies. 611 * 612 * Leave the last bp intact unless there is an error. 613 * Loop back up to the while if the error is another 614 * NFSERR_BAD_COOKIE (double yuch!). 615 */ 616 for (i = 0; i <= lbn && !error; i++) { 617 if (np->n_direofoffset 618 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 619 return (0); 620 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 621 if (!bp) { 622 error = newnfs_sigintr(nmp, td); 623 return (error ? error : EINTR); 624 } 625 if ((bp->b_flags & B_CACHE) == 0) { 626 bp->b_iocmd = BIO_READ; 627 vfs_busy_pages(bp, 0); 628 error = ncl_doio(vp, bp, cred, td, 0); 629 /* 630 * no error + B_INVAL == directory EOF, 631 * use the block. 632 */ 633 if (error == 0 && (bp->b_flags & B_INVAL)) 634 break; 635 } 636 /* 637 * An error will throw away the block and the 638 * for loop will break out. If no error and this 639 * is not the block we want, we throw away the 640 * block and go for the next one via the for loop. 641 */ 642 if (error || i < lbn) 643 brelse(bp); 644 } 645 } 646 /* 647 * The above while is repeated if we hit another cookie 648 * error. If we hit an error and it wasn't a cookie error, 649 * we give up. 650 */ 651 if (error) 652 return (error); 653 } 654 655 /* 656 * If not eof and read aheads are enabled, start one. 657 * (You need the current block first, so that you have the 658 * directory offset cookie of the next block.) 659 */ 660 if (nmp->nm_readahead > 0 && 661 (bp->b_flags & B_INVAL) == 0 && 662 (np->n_direofoffset == 0 || 663 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 664 incore(&vp->v_bufobj, lbn + 1) == NULL) { 665 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 666 if (rabp) { 667 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 668 rabp->b_flags |= B_ASYNC; 669 rabp->b_iocmd = BIO_READ; 670 vfs_busy_pages(rabp, 0); 671 if (ncl_asyncio(nmp, rabp, cred, td)) { 672 rabp->b_flags |= B_INVAL; 673 rabp->b_ioflags |= BIO_ERROR; 674 vfs_unbusy_pages(rabp); 675 brelse(rabp); 676 } 677 } else { 678 brelse(rabp); 679 } 680 } 681 } 682 /* 683 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 684 * chopped for the EOF condition, we cannot tell how large 685 * NFS directories are going to be until we hit EOF. So 686 * an NFS directory buffer is *not* chopped to its EOF. Now, 687 * it just so happens that b_resid will effectively chop it 688 * to EOF. *BUT* this information is lost if the buffer goes 689 * away and is reconstituted into a B_CACHE state ( due to 690 * being VMIO ) later. So we keep track of the directory eof 691 * in np->n_direofoffset and chop it off as an extra step 692 * right here. 693 */ 694 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 695 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 696 n = np->n_direofoffset - uio->uio_offset; 697 break; 698 default: 699 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type); 700 bp = NULL; 701 break; 702 }; 703 704 if (n > 0) { 705 error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio); 706 } 707 if (vp->v_type == VLNK) 708 n = 0; 709 if (bp != NULL) 710 brelse(bp); 711 } while (error == 0 && uio->uio_resid > 0 && n > 0); 712 return (error); 713 } 714 715 /* 716 * The NFS write path cannot handle iovecs with len > 1. So we need to 717 * break up iovecs accordingly (restricting them to wsize). 718 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 719 * For the ASYNC case, 2 copies are needed. The first a copy from the 720 * user buffer to a staging buffer and then a second copy from the staging 721 * buffer to mbufs. This can be optimized by copying from the user buffer 722 * directly into mbufs and passing the chain down, but that requires a 723 * fair amount of re-working of the relevant codepaths (and can be done 724 * later). 725 */ 726 static int 727 nfs_directio_write(vp, uiop, cred, ioflag) 728 struct vnode *vp; 729 struct uio *uiop; 730 struct ucred *cred; 731 int ioflag; 732 { 733 int error; 734 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 735 struct thread *td = uiop->uio_td; 736 int size; 737 int wsize; 738 739 mtx_lock(&nmp->nm_mtx); 740 wsize = nmp->nm_wsize; 741 mtx_unlock(&nmp->nm_mtx); 742 if (ioflag & IO_SYNC) { 743 int iomode, must_commit; 744 struct uio uio; 745 struct iovec iov; 746 do_sync: 747 while (uiop->uio_resid > 0) { 748 size = MIN(uiop->uio_resid, wsize); 749 size = MIN(uiop->uio_iov->iov_len, size); 750 iov.iov_base = uiop->uio_iov->iov_base; 751 iov.iov_len = size; 752 uio.uio_iov = &iov; 753 uio.uio_iovcnt = 1; 754 uio.uio_offset = uiop->uio_offset; 755 uio.uio_resid = size; 756 uio.uio_segflg = UIO_USERSPACE; 757 uio.uio_rw = UIO_WRITE; 758 uio.uio_td = td; 759 iomode = NFSWRITE_FILESYNC; 760 error = ncl_writerpc(vp, &uio, cred, &iomode, 761 &must_commit, 0); 762 KASSERT((must_commit == 0), 763 ("ncl_directio_write: Did not commit write")); 764 if (error) 765 return (error); 766 uiop->uio_offset += size; 767 uiop->uio_resid -= size; 768 if (uiop->uio_iov->iov_len <= size) { 769 uiop->uio_iovcnt--; 770 uiop->uio_iov++; 771 } else { 772 uiop->uio_iov->iov_base = 773 (char *)uiop->uio_iov->iov_base + size; 774 uiop->uio_iov->iov_len -= size; 775 } 776 } 777 } else { 778 struct uio *t_uio; 779 struct iovec *t_iov; 780 struct buf *bp; 781 782 /* 783 * Break up the write into blocksize chunks and hand these 784 * over to nfsiod's for write back. 785 * Unfortunately, this incurs a copy of the data. Since 786 * the user could modify the buffer before the write is 787 * initiated. 788 * 789 * The obvious optimization here is that one of the 2 copies 790 * in the async write path can be eliminated by copying the 791 * data here directly into mbufs and passing the mbuf chain 792 * down. But that will require a fair amount of re-working 793 * of the code and can be done if there's enough interest 794 * in NFS directio access. 795 */ 796 while (uiop->uio_resid > 0) { 797 size = MIN(uiop->uio_resid, wsize); 798 size = MIN(uiop->uio_iov->iov_len, size); 799 bp = getpbuf(&ncl_pbuf_freecnt); 800 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 801 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 802 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 803 t_iov->iov_len = size; 804 t_uio->uio_iov = t_iov; 805 t_uio->uio_iovcnt = 1; 806 t_uio->uio_offset = uiop->uio_offset; 807 t_uio->uio_resid = size; 808 t_uio->uio_segflg = UIO_SYSSPACE; 809 t_uio->uio_rw = UIO_WRITE; 810 t_uio->uio_td = td; 811 KASSERT(uiop->uio_segflg == UIO_USERSPACE || 812 uiop->uio_segflg == UIO_SYSSPACE, 813 ("nfs_directio_write: Bad uio_segflg")); 814 if (uiop->uio_segflg == UIO_USERSPACE) { 815 error = copyin(uiop->uio_iov->iov_base, 816 t_iov->iov_base, size); 817 if (error != 0) 818 goto err_free; 819 } else 820 /* 821 * UIO_SYSSPACE may never happen, but handle 822 * it just in case it does. 823 */ 824 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, 825 size); 826 bp->b_flags |= B_DIRECT; 827 bp->b_iocmd = BIO_WRITE; 828 if (cred != NOCRED) { 829 crhold(cred); 830 bp->b_wcred = cred; 831 } else 832 bp->b_wcred = NOCRED; 833 bp->b_caller1 = (void *)t_uio; 834 bp->b_vp = vp; 835 error = ncl_asyncio(nmp, bp, NOCRED, td); 836 err_free: 837 if (error) { 838 free(t_iov->iov_base, M_NFSDIRECTIO); 839 free(t_iov, M_NFSDIRECTIO); 840 free(t_uio, M_NFSDIRECTIO); 841 bp->b_vp = NULL; 842 relpbuf(bp, &ncl_pbuf_freecnt); 843 if (error == EINTR) 844 return (error); 845 goto do_sync; 846 } 847 uiop->uio_offset += size; 848 uiop->uio_resid -= size; 849 if (uiop->uio_iov->iov_len <= size) { 850 uiop->uio_iovcnt--; 851 uiop->uio_iov++; 852 } else { 853 uiop->uio_iov->iov_base = 854 (char *)uiop->uio_iov->iov_base + size; 855 uiop->uio_iov->iov_len -= size; 856 } 857 } 858 } 859 return (0); 860 } 861 862 /* 863 * Vnode op for write using bio 864 */ 865 int 866 ncl_write(struct vop_write_args *ap) 867 { 868 int biosize; 869 struct uio *uio = ap->a_uio; 870 struct thread *td = uio->uio_td; 871 struct vnode *vp = ap->a_vp; 872 struct nfsnode *np = VTONFS(vp); 873 struct ucred *cred = ap->a_cred; 874 int ioflag = ap->a_ioflag; 875 struct buf *bp; 876 struct vattr vattr; 877 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 878 daddr_t lbn; 879 int bcount; 880 int bp_cached, n, on, error = 0, error1; 881 size_t orig_resid, local_resid; 882 off_t orig_size, tmp_off; 883 884 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode")); 885 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, 886 ("ncl_write proc")); 887 if (vp->v_type != VREG) 888 return (EIO); 889 mtx_lock(&np->n_mtx); 890 if (np->n_flag & NWRITEERR) { 891 np->n_flag &= ~NWRITEERR; 892 mtx_unlock(&np->n_mtx); 893 return (np->n_error); 894 } else 895 mtx_unlock(&np->n_mtx); 896 mtx_lock(&nmp->nm_mtx); 897 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 898 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 899 mtx_unlock(&nmp->nm_mtx); 900 (void)ncl_fsinfo(nmp, vp, cred, td); 901 mtx_lock(&nmp->nm_mtx); 902 } 903 if (nmp->nm_wsize == 0) 904 (void) newnfs_iosize(nmp); 905 mtx_unlock(&nmp->nm_mtx); 906 907 /* 908 * Synchronously flush pending buffers if we are in synchronous 909 * mode or if we are appending. 910 */ 911 if (ioflag & (IO_APPEND | IO_SYNC)) { 912 mtx_lock(&np->n_mtx); 913 if (np->n_flag & NMODIFIED) { 914 mtx_unlock(&np->n_mtx); 915 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 916 /* 917 * Require non-blocking, synchronous writes to 918 * dirty files to inform the program it needs 919 * to fsync(2) explicitly. 920 */ 921 if (ioflag & IO_NDELAY) 922 return (EAGAIN); 923 #endif 924 flush_and_restart: 925 np->n_attrstamp = 0; 926 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 927 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 928 if (error) 929 return (error); 930 } else 931 mtx_unlock(&np->n_mtx); 932 } 933 934 orig_resid = uio->uio_resid; 935 mtx_lock(&np->n_mtx); 936 orig_size = np->n_size; 937 mtx_unlock(&np->n_mtx); 938 939 /* 940 * If IO_APPEND then load uio_offset. We restart here if we cannot 941 * get the append lock. 942 */ 943 if (ioflag & IO_APPEND) { 944 np->n_attrstamp = 0; 945 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 946 error = VOP_GETATTR(vp, &vattr, cred); 947 if (error) 948 return (error); 949 mtx_lock(&np->n_mtx); 950 uio->uio_offset = np->n_size; 951 mtx_unlock(&np->n_mtx); 952 } 953 954 if (uio->uio_offset < 0) 955 return (EINVAL); 956 tmp_off = uio->uio_offset + uio->uio_resid; 957 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset) 958 return (EFBIG); 959 if (uio->uio_resid == 0) 960 return (0); 961 962 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 963 return nfs_directio_write(vp, uio, cred, ioflag); 964 965 /* 966 * Maybe this should be above the vnode op call, but so long as 967 * file servers have no limits, i don't think it matters 968 */ 969 if (vn_rlimit_fsize(vp, uio, td)) 970 return (EFBIG); 971 972 biosize = vp->v_bufobj.bo_bsize; 973 /* 974 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 975 * would exceed the local maximum per-file write commit size when 976 * combined with those, we must decide whether to flush, 977 * go synchronous, or return error. We don't bother checking 978 * IO_UNIT -- we just make all writes atomic anyway, as there's 979 * no point optimizing for something that really won't ever happen. 980 */ 981 if (!(ioflag & IO_SYNC)) { 982 int nflag; 983 984 mtx_lock(&np->n_mtx); 985 nflag = np->n_flag; 986 mtx_unlock(&np->n_mtx); 987 int needrestart = 0; 988 if (nmp->nm_wcommitsize < uio->uio_resid) { 989 /* 990 * If this request could not possibly be completed 991 * without exceeding the maximum outstanding write 992 * commit size, see if we can convert it into a 993 * synchronous write operation. 994 */ 995 if (ioflag & IO_NDELAY) 996 return (EAGAIN); 997 ioflag |= IO_SYNC; 998 if (nflag & NMODIFIED) 999 needrestart = 1; 1000 } else if (nflag & NMODIFIED) { 1001 int wouldcommit = 0; 1002 BO_LOCK(&vp->v_bufobj); 1003 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 1004 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 1005 b_bobufs) { 1006 if (bp->b_flags & B_NEEDCOMMIT) 1007 wouldcommit += bp->b_bcount; 1008 } 1009 } 1010 BO_UNLOCK(&vp->v_bufobj); 1011 /* 1012 * Since we're not operating synchronously and 1013 * bypassing the buffer cache, we are in a commit 1014 * and holding all of these buffers whether 1015 * transmitted or not. If not limited, this 1016 * will lead to the buffer cache deadlocking, 1017 * as no one else can flush our uncommitted buffers. 1018 */ 1019 wouldcommit += uio->uio_resid; 1020 /* 1021 * If we would initially exceed the maximum 1022 * outstanding write commit size, flush and restart. 1023 */ 1024 if (wouldcommit > nmp->nm_wcommitsize) 1025 needrestart = 1; 1026 } 1027 if (needrestart) 1028 goto flush_and_restart; 1029 } 1030 1031 do { 1032 NFSINCRGLOBAL(newnfsstats.biocache_writes); 1033 lbn = uio->uio_offset / biosize; 1034 on = uio->uio_offset - (lbn * biosize); 1035 n = MIN((unsigned)(biosize - on), uio->uio_resid); 1036 again: 1037 /* 1038 * Handle direct append and file extension cases, calculate 1039 * unaligned buffer size. 1040 */ 1041 mtx_lock(&np->n_mtx); 1042 if (uio->uio_offset == np->n_size && n) { 1043 mtx_unlock(&np->n_mtx); 1044 /* 1045 * Get the buffer (in its pre-append state to maintain 1046 * B_CACHE if it was previously set). Resize the 1047 * nfsnode after we have locked the buffer to prevent 1048 * readers from reading garbage. 1049 */ 1050 bcount = on; 1051 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1052 1053 if (bp != NULL) { 1054 long save; 1055 1056 mtx_lock(&np->n_mtx); 1057 np->n_size = uio->uio_offset + n; 1058 np->n_flag |= NMODIFIED; 1059 vnode_pager_setsize(vp, np->n_size); 1060 mtx_unlock(&np->n_mtx); 1061 1062 save = bp->b_flags & B_CACHE; 1063 bcount += n; 1064 allocbuf(bp, bcount); 1065 bp->b_flags |= save; 1066 } 1067 } else { 1068 /* 1069 * Obtain the locked cache block first, and then 1070 * adjust the file's size as appropriate. 1071 */ 1072 bcount = on + n; 1073 if ((off_t)lbn * biosize + bcount < np->n_size) { 1074 if ((off_t)(lbn + 1) * biosize < np->n_size) 1075 bcount = biosize; 1076 else 1077 bcount = np->n_size - (off_t)lbn * biosize; 1078 } 1079 mtx_unlock(&np->n_mtx); 1080 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1081 mtx_lock(&np->n_mtx); 1082 if (uio->uio_offset + n > np->n_size) { 1083 np->n_size = uio->uio_offset + n; 1084 np->n_flag |= NMODIFIED; 1085 vnode_pager_setsize(vp, np->n_size); 1086 } 1087 mtx_unlock(&np->n_mtx); 1088 } 1089 1090 if (!bp) { 1091 error = newnfs_sigintr(nmp, td); 1092 if (!error) 1093 error = EINTR; 1094 break; 1095 } 1096 1097 /* 1098 * Issue a READ if B_CACHE is not set. In special-append 1099 * mode, B_CACHE is based on the buffer prior to the write 1100 * op and is typically set, avoiding the read. If a read 1101 * is required in special append mode, the server will 1102 * probably send us a short-read since we extended the file 1103 * on our end, resulting in b_resid == 0 and, thusly, 1104 * B_CACHE getting set. 1105 * 1106 * We can also avoid issuing the read if the write covers 1107 * the entire buffer. We have to make sure the buffer state 1108 * is reasonable in this case since we will not be initiating 1109 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1110 * more information. 1111 * 1112 * B_CACHE may also be set due to the buffer being cached 1113 * normally. 1114 */ 1115 1116 bp_cached = 1; 1117 if (on == 0 && n == bcount) { 1118 if ((bp->b_flags & B_CACHE) == 0) 1119 bp_cached = 0; 1120 bp->b_flags |= B_CACHE; 1121 bp->b_flags &= ~B_INVAL; 1122 bp->b_ioflags &= ~BIO_ERROR; 1123 } 1124 1125 if ((bp->b_flags & B_CACHE) == 0) { 1126 bp->b_iocmd = BIO_READ; 1127 vfs_busy_pages(bp, 0); 1128 error = ncl_doio(vp, bp, cred, td, 0); 1129 if (error) { 1130 brelse(bp); 1131 break; 1132 } 1133 } 1134 if (bp->b_wcred == NOCRED) 1135 bp->b_wcred = crhold(cred); 1136 mtx_lock(&np->n_mtx); 1137 np->n_flag |= NMODIFIED; 1138 mtx_unlock(&np->n_mtx); 1139 1140 /* 1141 * If dirtyend exceeds file size, chop it down. This should 1142 * not normally occur but there is an append race where it 1143 * might occur XXX, so we log it. 1144 * 1145 * If the chopping creates a reverse-indexed or degenerate 1146 * situation with dirtyoff/end, we 0 both of them. 1147 */ 1148 1149 if (bp->b_dirtyend > bcount) { 1150 ncl_printf("NFS append race @%lx:%d\n", 1151 (long)bp->b_blkno * DEV_BSIZE, 1152 bp->b_dirtyend - bcount); 1153 bp->b_dirtyend = bcount; 1154 } 1155 1156 if (bp->b_dirtyoff >= bp->b_dirtyend) 1157 bp->b_dirtyoff = bp->b_dirtyend = 0; 1158 1159 /* 1160 * If the new write will leave a contiguous dirty 1161 * area, just update the b_dirtyoff and b_dirtyend, 1162 * otherwise force a write rpc of the old dirty area. 1163 * 1164 * While it is possible to merge discontiguous writes due to 1165 * our having a B_CACHE buffer ( and thus valid read data 1166 * for the hole), we don't because it could lead to 1167 * significant cache coherency problems with multiple clients, 1168 * especially if locking is implemented later on. 1169 * 1170 * As an optimization we could theoretically maintain 1171 * a linked list of discontinuous areas, but we would still 1172 * have to commit them separately so there isn't much 1173 * advantage to it except perhaps a bit of asynchronization. 1174 */ 1175 1176 if (bp->b_dirtyend > 0 && 1177 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1178 if (bwrite(bp) == EINTR) { 1179 error = EINTR; 1180 break; 1181 } 1182 goto again; 1183 } 1184 1185 local_resid = uio->uio_resid; 1186 error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio); 1187 1188 if (error != 0 && !bp_cached) { 1189 /* 1190 * This block has no other content than what 1191 * possibly was written by the faulty uiomove. 1192 * Release it, forgetting the data pages, to 1193 * prevent the leak of uninitialized data to 1194 * usermode. 1195 */ 1196 bp->b_ioflags |= BIO_ERROR; 1197 brelse(bp); 1198 uio->uio_offset -= local_resid - uio->uio_resid; 1199 uio->uio_resid = local_resid; 1200 break; 1201 } 1202 1203 /* 1204 * Since this block is being modified, it must be written 1205 * again and not just committed. Since write clustering does 1206 * not work for the stage 1 data write, only the stage 2 1207 * commit rpc, we have to clear B_CLUSTEROK as well. 1208 */ 1209 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1210 1211 /* 1212 * Get the partial update on the progress made from 1213 * uiomove, if an error occured. 1214 */ 1215 if (error != 0) 1216 n = local_resid - uio->uio_resid; 1217 1218 /* 1219 * Only update dirtyoff/dirtyend if not a degenerate 1220 * condition. 1221 */ 1222 if (n > 0) { 1223 if (bp->b_dirtyend > 0) { 1224 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1225 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1226 } else { 1227 bp->b_dirtyoff = on; 1228 bp->b_dirtyend = on + n; 1229 } 1230 vfs_bio_set_valid(bp, on, n); 1231 } 1232 1233 /* 1234 * If IO_SYNC do bwrite(). 1235 * 1236 * IO_INVAL appears to be unused. The idea appears to be 1237 * to turn off caching in this case. Very odd. XXX 1238 */ 1239 if ((ioflag & IO_SYNC)) { 1240 if (ioflag & IO_INVAL) 1241 bp->b_flags |= B_NOCACHE; 1242 error1 = bwrite(bp); 1243 if (error1 != 0) { 1244 if (error == 0) 1245 error = error1; 1246 break; 1247 } 1248 } else if ((n + on) == biosize) { 1249 bp->b_flags |= B_ASYNC; 1250 (void) ncl_writebp(bp, 0, NULL); 1251 } else { 1252 bdwrite(bp); 1253 } 1254 1255 if (error != 0) 1256 break; 1257 } while (uio->uio_resid > 0 && n > 0); 1258 1259 if (error != 0) { 1260 if (ioflag & IO_UNIT) { 1261 VATTR_NULL(&vattr); 1262 vattr.va_size = orig_size; 1263 /* IO_SYNC is handled implicitely */ 1264 (void)VOP_SETATTR(vp, &vattr, cred); 1265 uio->uio_offset -= orig_resid - uio->uio_resid; 1266 uio->uio_resid = orig_resid; 1267 } 1268 } 1269 1270 return (error); 1271 } 1272 1273 /* 1274 * Get an nfs cache block. 1275 * 1276 * Allocate a new one if the block isn't currently in the cache 1277 * and return the block marked busy. If the calling process is 1278 * interrupted by a signal for an interruptible mount point, return 1279 * NULL. 1280 * 1281 * The caller must carefully deal with the possible B_INVAL state of 1282 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it 1283 * indirectly), so synchronous reads can be issued without worrying about 1284 * the B_INVAL state. We have to be a little more careful when dealing 1285 * with writes (see comments in nfs_write()) when extending a file past 1286 * its EOF. 1287 */ 1288 static struct buf * 1289 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1290 { 1291 struct buf *bp; 1292 struct mount *mp; 1293 struct nfsmount *nmp; 1294 1295 mp = vp->v_mount; 1296 nmp = VFSTONFS(mp); 1297 1298 if (nmp->nm_flag & NFSMNT_INT) { 1299 sigset_t oldset; 1300 1301 newnfs_set_sigmask(td, &oldset); 1302 bp = getblk(vp, bn, size, PCATCH, 0, 0); 1303 newnfs_restore_sigmask(td, &oldset); 1304 while (bp == NULL) { 1305 if (newnfs_sigintr(nmp, td)) 1306 return (NULL); 1307 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1308 } 1309 } else { 1310 bp = getblk(vp, bn, size, 0, 0, 0); 1311 } 1312 1313 if (vp->v_type == VREG) 1314 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE); 1315 return (bp); 1316 } 1317 1318 /* 1319 * Flush and invalidate all dirty buffers. If another process is already 1320 * doing the flush, just wait for completion. 1321 */ 1322 int 1323 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1324 { 1325 struct nfsnode *np = VTONFS(vp); 1326 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1327 int error = 0, slpflag, slptimeo; 1328 int old_lock = 0; 1329 1330 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf"); 1331 1332 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1333 intrflg = 0; 1334 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)) 1335 intrflg = 1; 1336 if (intrflg) { 1337 slpflag = PCATCH; 1338 slptimeo = 2 * hz; 1339 } else { 1340 slpflag = 0; 1341 slptimeo = 0; 1342 } 1343 1344 old_lock = ncl_upgrade_vnlock(vp); 1345 if (vp->v_iflag & VI_DOOMED) { 1346 /* 1347 * Since vgonel() uses the generic vinvalbuf() to flush 1348 * dirty buffers and it does not call this function, it 1349 * is safe to just return OK when VI_DOOMED is set. 1350 */ 1351 ncl_downgrade_vnlock(vp, old_lock); 1352 return (0); 1353 } 1354 1355 /* 1356 * Now, flush as required. 1357 */ 1358 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1359 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object); 1360 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1361 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object); 1362 /* 1363 * If the page clean was interrupted, fail the invalidation. 1364 * Not doing so, we run the risk of losing dirty pages in the 1365 * vinvalbuf() call below. 1366 */ 1367 if (intrflg && (error = newnfs_sigintr(nmp, td))) 1368 goto out; 1369 } 1370 1371 error = vinvalbuf(vp, flags, slpflag, 0); 1372 while (error) { 1373 if (intrflg && (error = newnfs_sigintr(nmp, td))) 1374 goto out; 1375 error = vinvalbuf(vp, flags, 0, slptimeo); 1376 } 1377 if (NFSHASPNFS(nmp)) { 1378 nfscl_layoutcommit(vp, td); 1379 /* 1380 * Invalidate the attribute cache, since writes to a DS 1381 * won't update the size attribute. 1382 */ 1383 mtx_lock(&np->n_mtx); 1384 np->n_attrstamp = 0; 1385 } else 1386 mtx_lock(&np->n_mtx); 1387 if (np->n_directio_asyncwr == 0) 1388 np->n_flag &= ~NMODIFIED; 1389 mtx_unlock(&np->n_mtx); 1390 out: 1391 ncl_downgrade_vnlock(vp, old_lock); 1392 return error; 1393 } 1394 1395 /* 1396 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1397 * This is mainly to avoid queueing async I/O requests when the nfsiods 1398 * are all hung on a dead server. 1399 * 1400 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1401 * is eventually dequeued by the async daemon, ncl_doio() *will*. 1402 */ 1403 int 1404 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1405 { 1406 int iod; 1407 int gotiod; 1408 int slpflag = 0; 1409 int slptimeo = 0; 1410 int error, error2; 1411 1412 /* 1413 * Commits are usually short and sweet so lets save some cpu and 1414 * leave the async daemons for more important rpc's (such as reads 1415 * and writes). 1416 * 1417 * Readdirplus RPCs do vget()s to acquire the vnodes for entries 1418 * in the directory in order to update attributes. This can deadlock 1419 * with another thread that is waiting for async I/O to be done by 1420 * an nfsiod thread while holding a lock on one of these vnodes. 1421 * To avoid this deadlock, don't allow the async nfsiod threads to 1422 * perform Readdirplus RPCs. 1423 */ 1424 mtx_lock(&ncl_iod_mutex); 1425 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1426 (nmp->nm_bufqiods > ncl_numasync / 2)) || 1427 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) { 1428 mtx_unlock(&ncl_iod_mutex); 1429 return(EIO); 1430 } 1431 again: 1432 if (nmp->nm_flag & NFSMNT_INT) 1433 slpflag = PCATCH; 1434 gotiod = FALSE; 1435 1436 /* 1437 * Find a free iod to process this request. 1438 */ 1439 for (iod = 0; iod < ncl_numasync; iod++) 1440 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) { 1441 gotiod = TRUE; 1442 break; 1443 } 1444 1445 /* 1446 * Try to create one if none are free. 1447 */ 1448 if (!gotiod) 1449 ncl_nfsiodnew(); 1450 else { 1451 /* 1452 * Found one, so wake it up and tell it which 1453 * mount to process. 1454 */ 1455 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n", 1456 iod, nmp)); 1457 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE; 1458 ncl_iodmount[iod] = nmp; 1459 nmp->nm_bufqiods++; 1460 wakeup(&ncl_iodwant[iod]); 1461 } 1462 1463 /* 1464 * If none are free, we may already have an iod working on this mount 1465 * point. If so, it will process our request. 1466 */ 1467 if (!gotiod) { 1468 if (nmp->nm_bufqiods > 0) { 1469 NFS_DPF(ASYNCIO, 1470 ("ncl_asyncio: %d iods are already processing mount %p\n", 1471 nmp->nm_bufqiods, nmp)); 1472 gotiod = TRUE; 1473 } 1474 } 1475 1476 /* 1477 * If we have an iod which can process the request, then queue 1478 * the buffer. 1479 */ 1480 if (gotiod) { 1481 /* 1482 * Ensure that the queue never grows too large. We still want 1483 * to asynchronize so we block rather than return EIO. 1484 */ 1485 while (nmp->nm_bufqlen >= 2*ncl_numasync) { 1486 NFS_DPF(ASYNCIO, 1487 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp)); 1488 nmp->nm_bufqwant = TRUE; 1489 error = newnfs_msleep(td, &nmp->nm_bufq, 1490 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio", 1491 slptimeo); 1492 if (error) { 1493 error2 = newnfs_sigintr(nmp, td); 1494 if (error2) { 1495 mtx_unlock(&ncl_iod_mutex); 1496 return (error2); 1497 } 1498 if (slpflag == PCATCH) { 1499 slpflag = 0; 1500 slptimeo = 2 * hz; 1501 } 1502 } 1503 /* 1504 * We might have lost our iod while sleeping, 1505 * so check and loop if nescessary. 1506 */ 1507 goto again; 1508 } 1509 1510 /* We might have lost our nfsiod */ 1511 if (nmp->nm_bufqiods == 0) { 1512 NFS_DPF(ASYNCIO, 1513 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1514 goto again; 1515 } 1516 1517 if (bp->b_iocmd == BIO_READ) { 1518 if (bp->b_rcred == NOCRED && cred != NOCRED) 1519 bp->b_rcred = crhold(cred); 1520 } else { 1521 if (bp->b_wcred == NOCRED && cred != NOCRED) 1522 bp->b_wcred = crhold(cred); 1523 } 1524 1525 if (bp->b_flags & B_REMFREE) 1526 bremfreef(bp); 1527 BUF_KERNPROC(bp); 1528 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1529 nmp->nm_bufqlen++; 1530 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1531 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1532 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1533 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1534 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1535 } 1536 mtx_unlock(&ncl_iod_mutex); 1537 return (0); 1538 } 1539 1540 mtx_unlock(&ncl_iod_mutex); 1541 1542 /* 1543 * All the iods are busy on other mounts, so return EIO to 1544 * force the caller to process the i/o synchronously. 1545 */ 1546 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n")); 1547 return (EIO); 1548 } 1549 1550 void 1551 ncl_doio_directwrite(struct buf *bp) 1552 { 1553 int iomode, must_commit; 1554 struct uio *uiop = (struct uio *)bp->b_caller1; 1555 char *iov_base = uiop->uio_iov->iov_base; 1556 1557 iomode = NFSWRITE_FILESYNC; 1558 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1559 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0); 1560 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write")); 1561 free(iov_base, M_NFSDIRECTIO); 1562 free(uiop->uio_iov, M_NFSDIRECTIO); 1563 free(uiop, M_NFSDIRECTIO); 1564 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1565 struct nfsnode *np = VTONFS(bp->b_vp); 1566 mtx_lock(&np->n_mtx); 1567 if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) { 1568 /* 1569 * Invalidate the attribute cache, since writes to a DS 1570 * won't update the size attribute. 1571 */ 1572 np->n_attrstamp = 0; 1573 } 1574 np->n_directio_asyncwr--; 1575 if (np->n_directio_asyncwr == 0) { 1576 np->n_flag &= ~NMODIFIED; 1577 if ((np->n_flag & NFSYNCWAIT)) { 1578 np->n_flag &= ~NFSYNCWAIT; 1579 wakeup((caddr_t)&np->n_directio_asyncwr); 1580 } 1581 } 1582 mtx_unlock(&np->n_mtx); 1583 } 1584 bp->b_vp = NULL; 1585 relpbuf(bp, &ncl_pbuf_freecnt); 1586 } 1587 1588 /* 1589 * Do an I/O operation to/from a cache block. This may be called 1590 * synchronously or from an nfsiod. 1591 */ 1592 int 1593 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td, 1594 int called_from_strategy) 1595 { 1596 struct uio *uiop; 1597 struct nfsnode *np; 1598 struct nfsmount *nmp; 1599 int error = 0, iomode, must_commit = 0; 1600 struct uio uio; 1601 struct iovec io; 1602 struct proc *p = td ? td->td_proc : NULL; 1603 uint8_t iocmd; 1604 1605 np = VTONFS(vp); 1606 nmp = VFSTONFS(vp->v_mount); 1607 uiop = &uio; 1608 uiop->uio_iov = &io; 1609 uiop->uio_iovcnt = 1; 1610 uiop->uio_segflg = UIO_SYSSPACE; 1611 uiop->uio_td = td; 1612 1613 /* 1614 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1615 * do this here so we do not have to do it in all the code that 1616 * calls us. 1617 */ 1618 bp->b_flags &= ~B_INVAL; 1619 bp->b_ioflags &= ~BIO_ERROR; 1620 1621 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp)); 1622 iocmd = bp->b_iocmd; 1623 if (iocmd == BIO_READ) { 1624 io.iov_len = uiop->uio_resid = bp->b_bcount; 1625 io.iov_base = bp->b_data; 1626 uiop->uio_rw = UIO_READ; 1627 1628 switch (vp->v_type) { 1629 case VREG: 1630 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1631 NFSINCRGLOBAL(newnfsstats.read_bios); 1632 error = ncl_readrpc(vp, uiop, cr); 1633 1634 if (!error) { 1635 if (uiop->uio_resid) { 1636 /* 1637 * If we had a short read with no error, we must have 1638 * hit a file hole. We should zero-fill the remainder. 1639 * This can also occur if the server hits the file EOF. 1640 * 1641 * Holes used to be able to occur due to pending 1642 * writes, but that is not possible any longer. 1643 */ 1644 int nread = bp->b_bcount - uiop->uio_resid; 1645 ssize_t left = uiop->uio_resid; 1646 1647 if (left > 0) 1648 bzero((char *)bp->b_data + nread, left); 1649 uiop->uio_resid = 0; 1650 } 1651 } 1652 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */ 1653 if (p && (vp->v_vflag & VV_TEXT)) { 1654 mtx_lock(&np->n_mtx); 1655 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) { 1656 mtx_unlock(&np->n_mtx); 1657 PROC_LOCK(p); 1658 killproc(p, "text file modification"); 1659 PROC_UNLOCK(p); 1660 } else 1661 mtx_unlock(&np->n_mtx); 1662 } 1663 break; 1664 case VLNK: 1665 uiop->uio_offset = (off_t)0; 1666 NFSINCRGLOBAL(newnfsstats.readlink_bios); 1667 error = ncl_readlinkrpc(vp, uiop, cr); 1668 break; 1669 case VDIR: 1670 NFSINCRGLOBAL(newnfsstats.readdir_bios); 1671 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1672 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1673 error = ncl_readdirplusrpc(vp, uiop, cr, td); 1674 if (error == NFSERR_NOTSUPP) 1675 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1676 } 1677 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1678 error = ncl_readdirrpc(vp, uiop, cr, td); 1679 /* 1680 * end-of-directory sets B_INVAL but does not generate an 1681 * error. 1682 */ 1683 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1684 bp->b_flags |= B_INVAL; 1685 break; 1686 default: 1687 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type); 1688 break; 1689 }; 1690 if (error) { 1691 bp->b_ioflags |= BIO_ERROR; 1692 bp->b_error = error; 1693 } 1694 } else { 1695 /* 1696 * If we only need to commit, try to commit 1697 */ 1698 if (bp->b_flags & B_NEEDCOMMIT) { 1699 int retv; 1700 off_t off; 1701 1702 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1703 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1704 bp->b_wcred, td); 1705 if (retv == 0) { 1706 bp->b_dirtyoff = bp->b_dirtyend = 0; 1707 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1708 bp->b_resid = 0; 1709 bufdone(bp); 1710 return (0); 1711 } 1712 if (retv == NFSERR_STALEWRITEVERF) { 1713 ncl_clearcommit(vp->v_mount); 1714 } 1715 } 1716 1717 /* 1718 * Setup for actual write 1719 */ 1720 mtx_lock(&np->n_mtx); 1721 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1722 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1723 mtx_unlock(&np->n_mtx); 1724 1725 if (bp->b_dirtyend > bp->b_dirtyoff) { 1726 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1727 - bp->b_dirtyoff; 1728 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1729 + bp->b_dirtyoff; 1730 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1731 uiop->uio_rw = UIO_WRITE; 1732 NFSINCRGLOBAL(newnfsstats.write_bios); 1733 1734 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1735 iomode = NFSWRITE_UNSTABLE; 1736 else 1737 iomode = NFSWRITE_FILESYNC; 1738 1739 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit, 1740 called_from_strategy); 1741 1742 /* 1743 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1744 * to cluster the buffers needing commit. This will allow 1745 * the system to submit a single commit rpc for the whole 1746 * cluster. We can do this even if the buffer is not 100% 1747 * dirty (relative to the NFS blocksize), so we optimize the 1748 * append-to-file-case. 1749 * 1750 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1751 * cleared because write clustering only works for commit 1752 * rpc's, not for the data portion of the write). 1753 */ 1754 1755 if (!error && iomode == NFSWRITE_UNSTABLE) { 1756 bp->b_flags |= B_NEEDCOMMIT; 1757 if (bp->b_dirtyoff == 0 1758 && bp->b_dirtyend == bp->b_bcount) 1759 bp->b_flags |= B_CLUSTEROK; 1760 } else { 1761 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1762 } 1763 1764 /* 1765 * For an interrupted write, the buffer is still valid 1766 * and the write hasn't been pushed to the server yet, 1767 * so we can't set BIO_ERROR and report the interruption 1768 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1769 * is not relevant, so the rpc attempt is essentially 1770 * a noop. For the case of a V3 write rpc not being 1771 * committed to stable storage, the block is still 1772 * dirty and requires either a commit rpc or another 1773 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1774 * the block is reused. This is indicated by setting 1775 * the B_DELWRI and B_NEEDCOMMIT flags. 1776 * 1777 * EIO is returned by ncl_writerpc() to indicate a recoverable 1778 * write error and is handled as above, except that 1779 * B_EINTR isn't set. One cause of this is a stale stateid 1780 * error for the RPC that indicates recovery is required, 1781 * when called with called_from_strategy != 0. 1782 * 1783 * If the buffer is marked B_PAGING, it does not reside on 1784 * the vp's paging queues so we cannot call bdirty(). The 1785 * bp in this case is not an NFS cache block so we should 1786 * be safe. XXX 1787 * 1788 * The logic below breaks up errors into recoverable and 1789 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1790 * and keep the buffer around for potential write retries. 1791 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1792 * and save the error in the nfsnode. This is less than ideal 1793 * but necessary. Keeping such buffers around could potentially 1794 * cause buffer exhaustion eventually (they can never be written 1795 * out, so will get constantly be re-dirtied). It also causes 1796 * all sorts of vfs panics. For non-recoverable write errors, 1797 * also invalidate the attrcache, so we'll be forced to go over 1798 * the wire for this object, returning an error to user on next 1799 * call (most of the time). 1800 */ 1801 if (error == EINTR || error == EIO || error == ETIMEDOUT 1802 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1803 int s; 1804 1805 s = splbio(); 1806 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1807 if ((bp->b_flags & B_PAGING) == 0) { 1808 bdirty(bp); 1809 bp->b_flags &= ~B_DONE; 1810 } 1811 if ((error == EINTR || error == ETIMEDOUT) && 1812 (bp->b_flags & B_ASYNC) == 0) 1813 bp->b_flags |= B_EINTR; 1814 splx(s); 1815 } else { 1816 if (error) { 1817 bp->b_ioflags |= BIO_ERROR; 1818 bp->b_flags |= B_INVAL; 1819 bp->b_error = np->n_error = error; 1820 mtx_lock(&np->n_mtx); 1821 np->n_flag |= NWRITEERR; 1822 np->n_attrstamp = 0; 1823 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1824 mtx_unlock(&np->n_mtx); 1825 } 1826 bp->b_dirtyoff = bp->b_dirtyend = 0; 1827 } 1828 } else { 1829 bp->b_resid = 0; 1830 bufdone(bp); 1831 return (0); 1832 } 1833 } 1834 bp->b_resid = uiop->uio_resid; 1835 if (must_commit) 1836 ncl_clearcommit(vp->v_mount); 1837 bufdone(bp); 1838 return (error); 1839 } 1840 1841 /* 1842 * Used to aid in handling ftruncate() operations on the NFS client side. 1843 * Truncation creates a number of special problems for NFS. We have to 1844 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1845 * we have to properly handle VM pages or (potentially dirty) buffers 1846 * that straddle the truncation point. 1847 */ 1848 1849 int 1850 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1851 { 1852 struct nfsnode *np = VTONFS(vp); 1853 u_quad_t tsize; 1854 int biosize = vp->v_bufobj.bo_bsize; 1855 int error = 0; 1856 1857 mtx_lock(&np->n_mtx); 1858 tsize = np->n_size; 1859 np->n_size = nsize; 1860 mtx_unlock(&np->n_mtx); 1861 1862 if (nsize < tsize) { 1863 struct buf *bp; 1864 daddr_t lbn; 1865 int bufsize; 1866 1867 /* 1868 * vtruncbuf() doesn't get the buffer overlapping the 1869 * truncation point. We may have a B_DELWRI and/or B_CACHE 1870 * buffer that now needs to be truncated. 1871 */ 1872 error = vtruncbuf(vp, cred, nsize, biosize); 1873 lbn = nsize / biosize; 1874 bufsize = nsize - (lbn * biosize); 1875 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1876 if (!bp) 1877 return EINTR; 1878 if (bp->b_dirtyoff > bp->b_bcount) 1879 bp->b_dirtyoff = bp->b_bcount; 1880 if (bp->b_dirtyend > bp->b_bcount) 1881 bp->b_dirtyend = bp->b_bcount; 1882 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1883 brelse(bp); 1884 } else { 1885 vnode_pager_setsize(vp, nsize); 1886 } 1887 return(error); 1888 } 1889 1890