nfs_bio.c revision 230605
1/*- 2 * Copyright (c) 1989, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * Rick Macklem at The University of Guelph. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 4. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 33 */ 34 35#include <sys/cdefs.h> 36__FBSDID("$FreeBSD: head/sys/nfsclient/nfs_bio.c 230605 2012-01-27 02:46:12Z rmacklem $"); 37 38#include "opt_kdtrace.h" 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/mbuf.h> 46#include <sys/mount.h> 47#include <sys/proc.h> 48#include <sys/vmmeter.h> 49#include <sys/vnode.h> 50 51#include <vm/vm.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 <nfs/nfsproto.h> 59#include <nfsclient/nfs.h> 60#include <nfsclient/nfsmount.h> 61#include <nfsclient/nfsnode.h> 62#include <nfs/nfs_kdtrace.h> 63 64static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 65 struct thread *td); 66static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 67 struct ucred *cred, int ioflag); 68 69extern int nfs_directio_enable; 70extern int nfs_directio_allow_mmap; 71 72/* 73 * Vnode op for VM getpages. 74 */ 75int 76nfs_getpages(struct vop_getpages_args *ap) 77{ 78 int i, error, nextoff, size, toff, count, npages; 79 struct uio uio; 80 struct iovec iov; 81 vm_offset_t kva; 82 struct buf *bp; 83 struct vnode *vp; 84 struct thread *td; 85 struct ucred *cred; 86 struct nfsmount *nmp; 87 vm_object_t object; 88 vm_page_t *pages; 89 struct nfsnode *np; 90 91 vp = ap->a_vp; 92 np = VTONFS(vp); 93 td = curthread; /* XXX */ 94 cred = curthread->td_ucred; /* XXX */ 95 nmp = VFSTONFS(vp->v_mount); 96 pages = ap->a_m; 97 count = ap->a_count; 98 99 if ((object = vp->v_object) == NULL) { 100 nfs_printf("nfs_getpages: called with non-merged cache vnode??\n"); 101 return (VM_PAGER_ERROR); 102 } 103 104 if (nfs_directio_enable && !nfs_directio_allow_mmap) { 105 mtx_lock(&np->n_mtx); 106 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 107 mtx_unlock(&np->n_mtx); 108 nfs_printf("nfs_getpages: called on non-cacheable vnode??\n"); 109 return (VM_PAGER_ERROR); 110 } else 111 mtx_unlock(&np->n_mtx); 112 } 113 114 mtx_lock(&nmp->nm_mtx); 115 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 116 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 117 mtx_unlock(&nmp->nm_mtx); 118 /* We'll never get here for v4, because we always have fsinfo */ 119 (void)nfs_fsinfo(nmp, vp, cred, td); 120 } else 121 mtx_unlock(&nmp->nm_mtx); 122 123 npages = btoc(count); 124 125 /* 126 * If the requested page is partially valid, just return it and 127 * allow the pager to zero-out the blanks. Partially valid pages 128 * can only occur at the file EOF. 129 */ 130 VM_OBJECT_LOCK(object); 131 if (pages[ap->a_reqpage]->valid != 0) { 132 for (i = 0; i < npages; ++i) { 133 if (i != ap->a_reqpage) { 134 vm_page_lock(pages[i]); 135 vm_page_free(pages[i]); 136 vm_page_unlock(pages[i]); 137 } 138 } 139 VM_OBJECT_UNLOCK(object); 140 return (0); 141 } 142 VM_OBJECT_UNLOCK(object); 143 144 /* 145 * We use only the kva address for the buffer, but this is extremely 146 * convienient and fast. 147 */ 148 bp = getpbuf(&nfs_pbuf_freecnt); 149 150 kva = (vm_offset_t) bp->b_data; 151 pmap_qenter(kva, pages, npages); 152 PCPU_INC(cnt.v_vnodein); 153 PCPU_ADD(cnt.v_vnodepgsin, npages); 154 155 iov.iov_base = (caddr_t) kva; 156 iov.iov_len = count; 157 uio.uio_iov = &iov; 158 uio.uio_iovcnt = 1; 159 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 160 uio.uio_resid = count; 161 uio.uio_segflg = UIO_SYSSPACE; 162 uio.uio_rw = UIO_READ; 163 uio.uio_td = td; 164 165 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred); 166 pmap_qremove(kva, npages); 167 168 relpbuf(bp, &nfs_pbuf_freecnt); 169 170 if (error && (uio.uio_resid == count)) { 171 nfs_printf("nfs_getpages: error %d\n", error); 172 VM_OBJECT_LOCK(object); 173 for (i = 0; i < npages; ++i) { 174 if (i != ap->a_reqpage) { 175 vm_page_lock(pages[i]); 176 vm_page_free(pages[i]); 177 vm_page_unlock(pages[i]); 178 } 179 } 180 VM_OBJECT_UNLOCK(object); 181 return (VM_PAGER_ERROR); 182 } 183 184 /* 185 * Calculate the number of bytes read and validate only that number 186 * of bytes. Note that due to pending writes, size may be 0. This 187 * does not mean that the remaining data is invalid! 188 */ 189 190 size = count - uio.uio_resid; 191 VM_OBJECT_LOCK(object); 192 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 193 vm_page_t m; 194 nextoff = toff + PAGE_SIZE; 195 m = pages[i]; 196 197 if (nextoff <= size) { 198 /* 199 * Read operation filled an entire page 200 */ 201 m->valid = VM_PAGE_BITS_ALL; 202 KASSERT(m->dirty == 0, 203 ("nfs_getpages: page %p is dirty", m)); 204 } else if (size > toff) { 205 /* 206 * Read operation filled a partial page. 207 */ 208 m->valid = 0; 209 vm_page_set_valid_range(m, 0, size - toff); 210 KASSERT(m->dirty == 0, 211 ("nfs_getpages: page %p is dirty", m)); 212 } else { 213 /* 214 * Read operation was short. If no error occured 215 * we may have hit a zero-fill section. We simply 216 * leave valid set to 0. 217 */ 218 ; 219 } 220 if (i != ap->a_reqpage) { 221 /* 222 * Whether or not to leave the page activated is up in 223 * the air, but we should put the page on a page queue 224 * somewhere (it already is in the object). Result: 225 * It appears that emperical results show that 226 * deactivating pages is best. 227 */ 228 229 /* 230 * Just in case someone was asking for this page we 231 * now tell them that it is ok to use. 232 */ 233 if (!error) { 234 if (m->oflags & VPO_WANTED) { 235 vm_page_lock(m); 236 vm_page_activate(m); 237 vm_page_unlock(m); 238 } else { 239 vm_page_lock(m); 240 vm_page_deactivate(m); 241 vm_page_unlock(m); 242 } 243 vm_page_wakeup(m); 244 } else { 245 vm_page_lock(m); 246 vm_page_free(m); 247 vm_page_unlock(m); 248 } 249 } 250 } 251 VM_OBJECT_UNLOCK(object); 252 return (0); 253} 254 255/* 256 * Vnode op for VM putpages. 257 */ 258int 259nfs_putpages(struct vop_putpages_args *ap) 260{ 261 struct uio uio; 262 struct iovec iov; 263 vm_offset_t kva; 264 struct buf *bp; 265 int iomode, must_commit, i, error, npages, count; 266 off_t offset; 267 int *rtvals; 268 struct vnode *vp; 269 struct thread *td; 270 struct ucred *cred; 271 struct nfsmount *nmp; 272 struct nfsnode *np; 273 vm_page_t *pages; 274 275 vp = ap->a_vp; 276 np = VTONFS(vp); 277 td = curthread; /* XXX */ 278 cred = curthread->td_ucred; /* XXX */ 279 nmp = VFSTONFS(vp->v_mount); 280 pages = ap->a_m; 281 count = ap->a_count; 282 rtvals = ap->a_rtvals; 283 npages = btoc(count); 284 offset = IDX_TO_OFF(pages[0]->pindex); 285 286 mtx_lock(&nmp->nm_mtx); 287 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 288 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 289 mtx_unlock(&nmp->nm_mtx); 290 (void)nfs_fsinfo(nmp, vp, cred, td); 291 } else 292 mtx_unlock(&nmp->nm_mtx); 293 294 mtx_lock(&np->n_mtx); 295 if (nfs_directio_enable && !nfs_directio_allow_mmap && 296 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 297 mtx_unlock(&np->n_mtx); 298 nfs_printf("nfs_putpages: called on noncache-able vnode??\n"); 299 mtx_lock(&np->n_mtx); 300 } 301 302 for (i = 0; i < npages; i++) 303 rtvals[i] = VM_PAGER_ERROR; 304 305 /* 306 * When putting pages, do not extend file past EOF. 307 */ 308 if (offset + count > np->n_size) { 309 count = np->n_size - offset; 310 if (count < 0) 311 count = 0; 312 } 313 mtx_unlock(&np->n_mtx); 314 315 /* 316 * We use only the kva address for the buffer, but this is extremely 317 * convienient and fast. 318 */ 319 bp = getpbuf(&nfs_pbuf_freecnt); 320 321 kva = (vm_offset_t) bp->b_data; 322 pmap_qenter(kva, pages, npages); 323 PCPU_INC(cnt.v_vnodeout); 324 PCPU_ADD(cnt.v_vnodepgsout, count); 325 326 iov.iov_base = (caddr_t) kva; 327 iov.iov_len = count; 328 uio.uio_iov = &iov; 329 uio.uio_iovcnt = 1; 330 uio.uio_offset = offset; 331 uio.uio_resid = count; 332 uio.uio_segflg = UIO_SYSSPACE; 333 uio.uio_rw = UIO_WRITE; 334 uio.uio_td = td; 335 336 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 337 iomode = NFSV3WRITE_UNSTABLE; 338 else 339 iomode = NFSV3WRITE_FILESYNC; 340 341 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit); 342 343 pmap_qremove(kva, npages); 344 relpbuf(bp, &nfs_pbuf_freecnt); 345 346 if (!error) { 347 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid); 348 if (must_commit) { 349 nfs_clearcommit(vp->v_mount); 350 } 351 } 352 return rtvals[0]; 353} 354 355/* 356 * For nfs, cache consistency can only be maintained approximately. 357 * Although RFC1094 does not specify the criteria, the following is 358 * believed to be compatible with the reference port. 359 * For nfs: 360 * If the file's modify time on the server has changed since the 361 * last read rpc or you have written to the file, 362 * you may have lost data cache consistency with the 363 * server, so flush all of the file's data out of the cache. 364 * Then force a getattr rpc to ensure that you have up to date 365 * attributes. 366 * NB: This implies that cache data can be read when up to 367 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 368 * attributes this could be forced by setting n_attrstamp to 0 before 369 * the VOP_GETATTR() call. 370 */ 371static inline int 372nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 373{ 374 int error = 0; 375 struct vattr vattr; 376 struct nfsnode *np = VTONFS(vp); 377 int old_lock; 378 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 379 380 /* 381 * Grab the exclusive lock before checking whether the cache is 382 * consistent. 383 * XXX - We can make this cheaper later (by acquiring cheaper locks). 384 * But for now, this suffices. 385 */ 386 old_lock = nfs_upgrade_vnlock(vp); 387 if (vp->v_iflag & VI_DOOMED) { 388 nfs_downgrade_vnlock(vp, old_lock); 389 return (EBADF); 390 } 391 392 mtx_lock(&np->n_mtx); 393 if (np->n_flag & NMODIFIED) { 394 mtx_unlock(&np->n_mtx); 395 if (vp->v_type != VREG) { 396 if (vp->v_type != VDIR) 397 panic("nfs: bioread, not dir"); 398 (nmp->nm_rpcops->nr_invaldir)(vp); 399 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 400 if (error) 401 goto out; 402 } 403 np->n_attrstamp = 0; 404 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 405 error = VOP_GETATTR(vp, &vattr, cred); 406 if (error) 407 goto out; 408 mtx_lock(&np->n_mtx); 409 np->n_mtime = vattr.va_mtime; 410 mtx_unlock(&np->n_mtx); 411 } else { 412 mtx_unlock(&np->n_mtx); 413 error = VOP_GETATTR(vp, &vattr, cred); 414 if (error) 415 return (error); 416 mtx_lock(&np->n_mtx); 417 if ((np->n_flag & NSIZECHANGED) 418 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 419 mtx_unlock(&np->n_mtx); 420 if (vp->v_type == VDIR) 421 (nmp->nm_rpcops->nr_invaldir)(vp); 422 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 423 if (error) 424 goto out; 425 mtx_lock(&np->n_mtx); 426 np->n_mtime = vattr.va_mtime; 427 np->n_flag &= ~NSIZECHANGED; 428 } 429 mtx_unlock(&np->n_mtx); 430 } 431out: 432 nfs_downgrade_vnlock(vp, old_lock); 433 return error; 434} 435 436/* 437 * Vnode op for read using bio 438 */ 439int 440nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 441{ 442 struct nfsnode *np = VTONFS(vp); 443 int biosize, i; 444 struct buf *bp, *rabp; 445 struct thread *td; 446 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 447 daddr_t lbn, rabn; 448 off_t end; 449 int bcount; 450 int seqcount; 451 int nra, error = 0, n = 0, on = 0; 452 453 KASSERT(uio->uio_rw == UIO_READ, ("nfs_read mode")); 454 if (uio->uio_resid == 0) 455 return (0); 456 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 457 return (EINVAL); 458 td = uio->uio_td; 459 460 mtx_lock(&nmp->nm_mtx); 461 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 462 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 463 mtx_unlock(&nmp->nm_mtx); 464 (void)nfs_fsinfo(nmp, vp, cred, td); 465 } else 466 mtx_unlock(&nmp->nm_mtx); 467 468 end = uio->uio_offset + uio->uio_resid; 469 if (vp->v_type != VDIR && 470 (end > nmp->nm_maxfilesize || end < uio->uio_offset)) 471 return (EFBIG); 472 473 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 474 /* No caching/ no readaheads. Just read data into the user buffer */ 475 return nfs_readrpc(vp, uio, cred); 476 477 biosize = vp->v_bufobj.bo_bsize; 478 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 479 480 error = nfs_bioread_check_cons(vp, td, cred); 481 if (error) 482 return error; 483 484 do { 485 u_quad_t nsize; 486 487 mtx_lock(&np->n_mtx); 488 nsize = np->n_size; 489 mtx_unlock(&np->n_mtx); 490 491 switch (vp->v_type) { 492 case VREG: 493 nfsstats.biocache_reads++; 494 lbn = uio->uio_offset / biosize; 495 on = uio->uio_offset & (biosize - 1); 496 497 /* 498 * Start the read ahead(s), as required. 499 */ 500 if (nmp->nm_readahead > 0) { 501 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 502 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 503 rabn = lbn + 1 + nra; 504 if (incore(&vp->v_bufobj, rabn) == NULL) { 505 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 506 if (!rabp) { 507 error = nfs_sigintr(nmp, td); 508 return (error ? error : EINTR); 509 } 510 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 511 rabp->b_flags |= B_ASYNC; 512 rabp->b_iocmd = BIO_READ; 513 vfs_busy_pages(rabp, 0); 514 if (nfs_asyncio(nmp, rabp, cred, td)) { 515 rabp->b_flags |= B_INVAL; 516 rabp->b_ioflags |= BIO_ERROR; 517 vfs_unbusy_pages(rabp); 518 brelse(rabp); 519 break; 520 } 521 } else { 522 brelse(rabp); 523 } 524 } 525 } 526 } 527 528 /* Note that bcount is *not* DEV_BSIZE aligned. */ 529 bcount = biosize; 530 if ((off_t)lbn * biosize >= nsize) { 531 bcount = 0; 532 } else if ((off_t)(lbn + 1) * biosize > nsize) { 533 bcount = nsize - (off_t)lbn * biosize; 534 } 535 bp = nfs_getcacheblk(vp, lbn, bcount, td); 536 537 if (!bp) { 538 error = nfs_sigintr(nmp, td); 539 return (error ? error : EINTR); 540 } 541 542 /* 543 * If B_CACHE is not set, we must issue the read. If this 544 * fails, we return an error. 545 */ 546 547 if ((bp->b_flags & B_CACHE) == 0) { 548 bp->b_iocmd = BIO_READ; 549 vfs_busy_pages(bp, 0); 550 error = nfs_doio(vp, bp, cred, td); 551 if (error) { 552 brelse(bp); 553 return (error); 554 } 555 } 556 557 /* 558 * on is the offset into the current bp. Figure out how many 559 * bytes we can copy out of the bp. Note that bcount is 560 * NOT DEV_BSIZE aligned. 561 * 562 * Then figure out how many bytes we can copy into the uio. 563 */ 564 565 n = 0; 566 if (on < bcount) 567 n = min((unsigned)(bcount - on), uio->uio_resid); 568 break; 569 case VLNK: 570 nfsstats.biocache_readlinks++; 571 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 572 if (!bp) { 573 error = nfs_sigintr(nmp, td); 574 return (error ? error : EINTR); 575 } 576 if ((bp->b_flags & B_CACHE) == 0) { 577 bp->b_iocmd = BIO_READ; 578 vfs_busy_pages(bp, 0); 579 error = nfs_doio(vp, bp, cred, td); 580 if (error) { 581 bp->b_ioflags |= BIO_ERROR; 582 brelse(bp); 583 return (error); 584 } 585 } 586 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 587 on = 0; 588 break; 589 case VDIR: 590 nfsstats.biocache_readdirs++; 591 if (np->n_direofoffset 592 && uio->uio_offset >= np->n_direofoffset) { 593 return (0); 594 } 595 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 596 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 597 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 598 if (!bp) { 599 error = nfs_sigintr(nmp, td); 600 return (error ? error : EINTR); 601 } 602 if ((bp->b_flags & B_CACHE) == 0) { 603 bp->b_iocmd = BIO_READ; 604 vfs_busy_pages(bp, 0); 605 error = nfs_doio(vp, bp, cred, td); 606 if (error) { 607 brelse(bp); 608 } 609 while (error == NFSERR_BAD_COOKIE) { 610 (nmp->nm_rpcops->nr_invaldir)(vp); 611 error = nfs_vinvalbuf(vp, 0, td, 1); 612 /* 613 * Yuck! The directory has been modified on the 614 * server. The only way to get the block is by 615 * reading from the beginning to get all the 616 * offset cookies. 617 * 618 * Leave the last bp intact unless there is an error. 619 * Loop back up to the while if the error is another 620 * NFSERR_BAD_COOKIE (double yuch!). 621 */ 622 for (i = 0; i <= lbn && !error; i++) { 623 if (np->n_direofoffset 624 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 625 return (0); 626 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 627 if (!bp) { 628 error = nfs_sigintr(nmp, td); 629 return (error ? error : EINTR); 630 } 631 if ((bp->b_flags & B_CACHE) == 0) { 632 bp->b_iocmd = BIO_READ; 633 vfs_busy_pages(bp, 0); 634 error = nfs_doio(vp, bp, cred, td); 635 /* 636 * no error + B_INVAL == directory EOF, 637 * use the block. 638 */ 639 if (error == 0 && (bp->b_flags & B_INVAL)) 640 break; 641 } 642 /* 643 * An error will throw away the block and the 644 * for loop will break out. If no error and this 645 * is not the block we want, we throw away the 646 * block and go for the next one via the for loop. 647 */ 648 if (error || i < lbn) 649 brelse(bp); 650 } 651 } 652 /* 653 * The above while is repeated if we hit another cookie 654 * error. If we hit an error and it wasn't a cookie error, 655 * we give up. 656 */ 657 if (error) 658 return (error); 659 } 660 661 /* 662 * If not eof and read aheads are enabled, start one. 663 * (You need the current block first, so that you have the 664 * directory offset cookie of the next block.) 665 */ 666 if (nmp->nm_readahead > 0 && 667 (bp->b_flags & B_INVAL) == 0 && 668 (np->n_direofoffset == 0 || 669 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 670 incore(&vp->v_bufobj, lbn + 1) == NULL) { 671 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 672 if (rabp) { 673 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 674 rabp->b_flags |= B_ASYNC; 675 rabp->b_iocmd = BIO_READ; 676 vfs_busy_pages(rabp, 0); 677 if (nfs_asyncio(nmp, rabp, cred, td)) { 678 rabp->b_flags |= B_INVAL; 679 rabp->b_ioflags |= BIO_ERROR; 680 vfs_unbusy_pages(rabp); 681 brelse(rabp); 682 } 683 } else { 684 brelse(rabp); 685 } 686 } 687 } 688 /* 689 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 690 * chopped for the EOF condition, we cannot tell how large 691 * NFS directories are going to be until we hit EOF. So 692 * an NFS directory buffer is *not* chopped to its EOF. Now, 693 * it just so happens that b_resid will effectively chop it 694 * to EOF. *BUT* this information is lost if the buffer goes 695 * away and is reconstituted into a B_CACHE state ( due to 696 * being VMIO ) later. So we keep track of the directory eof 697 * in np->n_direofoffset and chop it off as an extra step 698 * right here. 699 */ 700 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 701 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 702 n = np->n_direofoffset - uio->uio_offset; 703 break; 704 default: 705 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type); 706 bp = NULL; 707 break; 708 }; 709 710 if (n > 0) { 711 error = uiomove(bp->b_data + on, (int)n, uio); 712 } 713 if (vp->v_type == VLNK) 714 n = 0; 715 if (bp != NULL) 716 brelse(bp); 717 } while (error == 0 && uio->uio_resid > 0 && n > 0); 718 return (error); 719} 720 721/* 722 * The NFS write path cannot handle iovecs with len > 1. So we need to 723 * break up iovecs accordingly (restricting them to wsize). 724 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 725 * For the ASYNC case, 2 copies are needed. The first a copy from the 726 * user buffer to a staging buffer and then a second copy from the staging 727 * buffer to mbufs. This can be optimized by copying from the user buffer 728 * directly into mbufs and passing the chain down, but that requires a 729 * fair amount of re-working of the relevant codepaths (and can be done 730 * later). 731 */ 732static int 733nfs_directio_write(vp, uiop, cred, ioflag) 734 struct vnode *vp; 735 struct uio *uiop; 736 struct ucred *cred; 737 int ioflag; 738{ 739 int error; 740 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 741 struct thread *td = uiop->uio_td; 742 int size; 743 int wsize; 744 745 mtx_lock(&nmp->nm_mtx); 746 wsize = nmp->nm_wsize; 747 mtx_unlock(&nmp->nm_mtx); 748 if (ioflag & IO_SYNC) { 749 int iomode, must_commit; 750 struct uio uio; 751 struct iovec iov; 752do_sync: 753 while (uiop->uio_resid > 0) { 754 size = min(uiop->uio_resid, wsize); 755 size = min(uiop->uio_iov->iov_len, size); 756 iov.iov_base = uiop->uio_iov->iov_base; 757 iov.iov_len = size; 758 uio.uio_iov = &iov; 759 uio.uio_iovcnt = 1; 760 uio.uio_offset = uiop->uio_offset; 761 uio.uio_resid = size; 762 uio.uio_segflg = UIO_USERSPACE; 763 uio.uio_rw = UIO_WRITE; 764 uio.uio_td = td; 765 iomode = NFSV3WRITE_FILESYNC; 766 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, 767 &iomode, &must_commit); 768 KASSERT((must_commit == 0), 769 ("nfs_directio_write: Did not commit write")); 770 if (error) 771 return (error); 772 uiop->uio_offset += size; 773 uiop->uio_resid -= size; 774 if (uiop->uio_iov->iov_len <= size) { 775 uiop->uio_iovcnt--; 776 uiop->uio_iov++; 777 } else { 778 uiop->uio_iov->iov_base = 779 (char *)uiop->uio_iov->iov_base + size; 780 uiop->uio_iov->iov_len -= size; 781 } 782 } 783 } else { 784 struct uio *t_uio; 785 struct iovec *t_iov; 786 struct buf *bp; 787 788 /* 789 * Break up the write into blocksize chunks and hand these 790 * over to nfsiod's for write back. 791 * Unfortunately, this incurs a copy of the data. Since 792 * the user could modify the buffer before the write is 793 * initiated. 794 * 795 * The obvious optimization here is that one of the 2 copies 796 * in the async write path can be eliminated by copying the 797 * data here directly into mbufs and passing the mbuf chain 798 * down. But that will require a fair amount of re-working 799 * of the code and can be done if there's enough interest 800 * in NFS directio access. 801 */ 802 while (uiop->uio_resid > 0) { 803 size = min(uiop->uio_resid, wsize); 804 size = min(uiop->uio_iov->iov_len, size); 805 bp = getpbuf(&nfs_pbuf_freecnt); 806 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 807 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 808 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 809 t_iov->iov_len = size; 810 t_uio->uio_iov = t_iov; 811 t_uio->uio_iovcnt = 1; 812 t_uio->uio_offset = uiop->uio_offset; 813 t_uio->uio_resid = size; 814 t_uio->uio_segflg = UIO_SYSSPACE; 815 t_uio->uio_rw = UIO_WRITE; 816 t_uio->uio_td = td; 817 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); 818 bp->b_flags |= B_DIRECT; 819 bp->b_iocmd = BIO_WRITE; 820 if (cred != NOCRED) { 821 crhold(cred); 822 bp->b_wcred = cred; 823 } else 824 bp->b_wcred = NOCRED; 825 bp->b_caller1 = (void *)t_uio; 826 bp->b_vp = vp; 827 error = nfs_asyncio(nmp, bp, NOCRED, td); 828 if (error) { 829 free(t_iov->iov_base, M_NFSDIRECTIO); 830 free(t_iov, M_NFSDIRECTIO); 831 free(t_uio, M_NFSDIRECTIO); 832 bp->b_vp = NULL; 833 relpbuf(bp, &nfs_pbuf_freecnt); 834 if (error == EINTR) 835 return (error); 836 goto do_sync; 837 } 838 uiop->uio_offset += size; 839 uiop->uio_resid -= size; 840 if (uiop->uio_iov->iov_len <= size) { 841 uiop->uio_iovcnt--; 842 uiop->uio_iov++; 843 } else { 844 uiop->uio_iov->iov_base = 845 (char *)uiop->uio_iov->iov_base + size; 846 uiop->uio_iov->iov_len -= size; 847 } 848 } 849 } 850 return (0); 851} 852 853/* 854 * Vnode op for write using bio 855 */ 856int 857nfs_write(struct vop_write_args *ap) 858{ 859 int biosize; 860 struct uio *uio = ap->a_uio; 861 struct thread *td = uio->uio_td; 862 struct vnode *vp = ap->a_vp; 863 struct nfsnode *np = VTONFS(vp); 864 struct ucred *cred = ap->a_cred; 865 int ioflag = ap->a_ioflag; 866 struct buf *bp; 867 struct vattr vattr; 868 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 869 daddr_t lbn; 870 off_t end; 871 int bcount; 872 int n, on, error = 0; 873 874 KASSERT(uio->uio_rw == UIO_WRITE, ("nfs_write mode")); 875 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, 876 ("nfs_write proc")); 877 if (vp->v_type != VREG) 878 return (EIO); 879 mtx_lock(&np->n_mtx); 880 if (np->n_flag & NWRITEERR) { 881 np->n_flag &= ~NWRITEERR; 882 mtx_unlock(&np->n_mtx); 883 return (np->n_error); 884 } else 885 mtx_unlock(&np->n_mtx); 886 mtx_lock(&nmp->nm_mtx); 887 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 888 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 889 mtx_unlock(&nmp->nm_mtx); 890 (void)nfs_fsinfo(nmp, vp, cred, td); 891 } else 892 mtx_unlock(&nmp->nm_mtx); 893 894 /* 895 * Synchronously flush pending buffers if we are in synchronous 896 * mode or if we are appending. 897 */ 898 if (ioflag & (IO_APPEND | IO_SYNC)) { 899 mtx_lock(&np->n_mtx); 900 if (np->n_flag & NMODIFIED) { 901 mtx_unlock(&np->n_mtx); 902#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 903 /* 904 * Require non-blocking, synchronous writes to 905 * dirty files to inform the program it needs 906 * to fsync(2) explicitly. 907 */ 908 if (ioflag & IO_NDELAY) 909 return (EAGAIN); 910#endif 911flush_and_restart: 912 np->n_attrstamp = 0; 913 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 914 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 915 if (error) 916 return (error); 917 } else 918 mtx_unlock(&np->n_mtx); 919 } 920 921 /* 922 * If IO_APPEND then load uio_offset. We restart here if we cannot 923 * get the append lock. 924 */ 925 if (ioflag & IO_APPEND) { 926 np->n_attrstamp = 0; 927 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 928 error = VOP_GETATTR(vp, &vattr, cred); 929 if (error) 930 return (error); 931 mtx_lock(&np->n_mtx); 932 uio->uio_offset = np->n_size; 933 mtx_unlock(&np->n_mtx); 934 } 935 936 if (uio->uio_offset < 0) 937 return (EINVAL); 938 end = uio->uio_offset + uio->uio_resid; 939 if (end > nmp->nm_maxfilesize || end < uio->uio_offset) 940 return (EFBIG); 941 if (uio->uio_resid == 0) 942 return (0); 943 944 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 945 return nfs_directio_write(vp, uio, cred, ioflag); 946 947 /* 948 * Maybe this should be above the vnode op call, but so long as 949 * file servers have no limits, i don't think it matters 950 */ 951 if (vn_rlimit_fsize(vp, uio, td)) 952 return (EFBIG); 953 954 biosize = vp->v_bufobj.bo_bsize; 955 /* 956 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 957 * would exceed the local maximum per-file write commit size when 958 * combined with those, we must decide whether to flush, 959 * go synchronous, or return error. We don't bother checking 960 * IO_UNIT -- we just make all writes atomic anyway, as there's 961 * no point optimizing for something that really won't ever happen. 962 */ 963 if (!(ioflag & IO_SYNC)) { 964 int nflag; 965 966 mtx_lock(&np->n_mtx); 967 nflag = np->n_flag; 968 mtx_unlock(&np->n_mtx); 969 int needrestart = 0; 970 if (nmp->nm_wcommitsize < uio->uio_resid) { 971 /* 972 * If this request could not possibly be completed 973 * without exceeding the maximum outstanding write 974 * commit size, see if we can convert it into a 975 * synchronous write operation. 976 */ 977 if (ioflag & IO_NDELAY) 978 return (EAGAIN); 979 ioflag |= IO_SYNC; 980 if (nflag & NMODIFIED) 981 needrestart = 1; 982 } else if (nflag & NMODIFIED) { 983 int wouldcommit = 0; 984 BO_LOCK(&vp->v_bufobj); 985 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 986 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 987 b_bobufs) { 988 if (bp->b_flags & B_NEEDCOMMIT) 989 wouldcommit += bp->b_bcount; 990 } 991 } 992 BO_UNLOCK(&vp->v_bufobj); 993 /* 994 * Since we're not operating synchronously and 995 * bypassing the buffer cache, we are in a commit 996 * and holding all of these buffers whether 997 * transmitted or not. If not limited, this 998 * will lead to the buffer cache deadlocking, 999 * as no one else can flush our uncommitted buffers. 1000 */ 1001 wouldcommit += uio->uio_resid; 1002 /* 1003 * If we would initially exceed the maximum 1004 * outstanding write commit size, flush and restart. 1005 */ 1006 if (wouldcommit > nmp->nm_wcommitsize) 1007 needrestart = 1; 1008 } 1009 if (needrestart) 1010 goto flush_and_restart; 1011 } 1012 1013 do { 1014 nfsstats.biocache_writes++; 1015 lbn = uio->uio_offset / biosize; 1016 on = uio->uio_offset & (biosize-1); 1017 n = min((unsigned)(biosize - on), uio->uio_resid); 1018again: 1019 /* 1020 * Handle direct append and file extension cases, calculate 1021 * unaligned buffer size. 1022 */ 1023 mtx_lock(&np->n_mtx); 1024 if (uio->uio_offset == np->n_size && n) { 1025 mtx_unlock(&np->n_mtx); 1026 /* 1027 * Get the buffer (in its pre-append state to maintain 1028 * B_CACHE if it was previously set). Resize the 1029 * nfsnode after we have locked the buffer to prevent 1030 * readers from reading garbage. 1031 */ 1032 bcount = on; 1033 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1034 1035 if (bp != NULL) { 1036 long save; 1037 1038 mtx_lock(&np->n_mtx); 1039 np->n_size = uio->uio_offset + n; 1040 np->n_flag |= NMODIFIED; 1041 vnode_pager_setsize(vp, np->n_size); 1042 mtx_unlock(&np->n_mtx); 1043 1044 save = bp->b_flags & B_CACHE; 1045 bcount += n; 1046 allocbuf(bp, bcount); 1047 bp->b_flags |= save; 1048 } 1049 } else { 1050 /* 1051 * Obtain the locked cache block first, and then 1052 * adjust the file's size as appropriate. 1053 */ 1054 bcount = on + n; 1055 if ((off_t)lbn * biosize + bcount < np->n_size) { 1056 if ((off_t)(lbn + 1) * biosize < np->n_size) 1057 bcount = biosize; 1058 else 1059 bcount = np->n_size - (off_t)lbn * biosize; 1060 } 1061 mtx_unlock(&np->n_mtx); 1062 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1063 mtx_lock(&np->n_mtx); 1064 if (uio->uio_offset + n > np->n_size) { 1065 np->n_size = uio->uio_offset + n; 1066 np->n_flag |= NMODIFIED; 1067 vnode_pager_setsize(vp, np->n_size); 1068 } 1069 mtx_unlock(&np->n_mtx); 1070 } 1071 1072 if (!bp) { 1073 error = nfs_sigintr(nmp, td); 1074 if (!error) 1075 error = EINTR; 1076 break; 1077 } 1078 1079 /* 1080 * Issue a READ if B_CACHE is not set. In special-append 1081 * mode, B_CACHE is based on the buffer prior to the write 1082 * op and is typically set, avoiding the read. If a read 1083 * is required in special append mode, the server will 1084 * probably send us a short-read since we extended the file 1085 * on our end, resulting in b_resid == 0 and, thusly, 1086 * B_CACHE getting set. 1087 * 1088 * We can also avoid issuing the read if the write covers 1089 * the entire buffer. We have to make sure the buffer state 1090 * is reasonable in this case since we will not be initiating 1091 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1092 * more information. 1093 * 1094 * B_CACHE may also be set due to the buffer being cached 1095 * normally. 1096 */ 1097 1098 if (on == 0 && n == bcount) { 1099 bp->b_flags |= B_CACHE; 1100 bp->b_flags &= ~B_INVAL; 1101 bp->b_ioflags &= ~BIO_ERROR; 1102 } 1103 1104 if ((bp->b_flags & B_CACHE) == 0) { 1105 bp->b_iocmd = BIO_READ; 1106 vfs_busy_pages(bp, 0); 1107 error = nfs_doio(vp, bp, cred, td); 1108 if (error) { 1109 brelse(bp); 1110 break; 1111 } 1112 } 1113 if (bp->b_wcred == NOCRED) 1114 bp->b_wcred = crhold(cred); 1115 mtx_lock(&np->n_mtx); 1116 np->n_flag |= NMODIFIED; 1117 mtx_unlock(&np->n_mtx); 1118 1119 /* 1120 * If dirtyend exceeds file size, chop it down. This should 1121 * not normally occur but there is an append race where it 1122 * might occur XXX, so we log it. 1123 * 1124 * If the chopping creates a reverse-indexed or degenerate 1125 * situation with dirtyoff/end, we 0 both of them. 1126 */ 1127 1128 if (bp->b_dirtyend > bcount) { 1129 nfs_printf("NFS append race @%lx:%d\n", 1130 (long)bp->b_blkno * DEV_BSIZE, 1131 bp->b_dirtyend - bcount); 1132 bp->b_dirtyend = bcount; 1133 } 1134 1135 if (bp->b_dirtyoff >= bp->b_dirtyend) 1136 bp->b_dirtyoff = bp->b_dirtyend = 0; 1137 1138 /* 1139 * If the new write will leave a contiguous dirty 1140 * area, just update the b_dirtyoff and b_dirtyend, 1141 * otherwise force a write rpc of the old dirty area. 1142 * 1143 * While it is possible to merge discontiguous writes due to 1144 * our having a B_CACHE buffer ( and thus valid read data 1145 * for the hole), we don't because it could lead to 1146 * significant cache coherency problems with multiple clients, 1147 * especially if locking is implemented later on. 1148 * 1149 * as an optimization we could theoretically maintain 1150 * a linked list of discontinuous areas, but we would still 1151 * have to commit them separately so there isn't much 1152 * advantage to it except perhaps a bit of asynchronization. 1153 */ 1154 1155 if (bp->b_dirtyend > 0 && 1156 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1157 if (bwrite(bp) == EINTR) { 1158 error = EINTR; 1159 break; 1160 } 1161 goto again; 1162 } 1163 1164 error = uiomove((char *)bp->b_data + on, n, uio); 1165 1166 /* 1167 * Since this block is being modified, it must be written 1168 * again and not just committed. Since write clustering does 1169 * not work for the stage 1 data write, only the stage 2 1170 * commit rpc, we have to clear B_CLUSTEROK as well. 1171 */ 1172 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1173 1174 if (error) { 1175 bp->b_ioflags |= BIO_ERROR; 1176 brelse(bp); 1177 break; 1178 } 1179 1180 /* 1181 * Only update dirtyoff/dirtyend if not a degenerate 1182 * condition. 1183 */ 1184 if (n) { 1185 if (bp->b_dirtyend > 0) { 1186 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1187 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1188 } else { 1189 bp->b_dirtyoff = on; 1190 bp->b_dirtyend = on + n; 1191 } 1192 vfs_bio_set_valid(bp, on, n); 1193 } 1194 1195 /* 1196 * If IO_SYNC do bwrite(). 1197 * 1198 * IO_INVAL appears to be unused. The idea appears to be 1199 * to turn off caching in this case. Very odd. XXX 1200 */ 1201 if ((ioflag & IO_SYNC)) { 1202 if (ioflag & IO_INVAL) 1203 bp->b_flags |= B_NOCACHE; 1204 error = bwrite(bp); 1205 if (error) 1206 break; 1207 } else if ((n + on) == biosize) { 1208 bp->b_flags |= B_ASYNC; 1209 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL); 1210 } else { 1211 bdwrite(bp); 1212 } 1213 } while (uio->uio_resid > 0 && n > 0); 1214 1215 return (error); 1216} 1217 1218/* 1219 * Get an nfs cache block. 1220 * 1221 * Allocate a new one if the block isn't currently in the cache 1222 * and return the block marked busy. If the calling process is 1223 * interrupted by a signal for an interruptible mount point, return 1224 * NULL. 1225 * 1226 * The caller must carefully deal with the possible B_INVAL state of 1227 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1228 * indirectly), so synchronous reads can be issued without worrying about 1229 * the B_INVAL state. We have to be a little more careful when dealing 1230 * with writes (see comments in nfs_write()) when extending a file past 1231 * its EOF. 1232 */ 1233static struct buf * 1234nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1235{ 1236 struct buf *bp; 1237 struct mount *mp; 1238 struct nfsmount *nmp; 1239 1240 mp = vp->v_mount; 1241 nmp = VFSTONFS(mp); 1242 1243 if (nmp->nm_flag & NFSMNT_INT) { 1244 sigset_t oldset; 1245 1246 nfs_set_sigmask(td, &oldset); 1247 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0); 1248 nfs_restore_sigmask(td, &oldset); 1249 while (bp == NULL) { 1250 if (nfs_sigintr(nmp, td)) 1251 return (NULL); 1252 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1253 } 1254 } else { 1255 bp = getblk(vp, bn, size, 0, 0, 0); 1256 } 1257 1258 if (vp->v_type == VREG) 1259 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE); 1260 return (bp); 1261} 1262 1263/* 1264 * Flush and invalidate all dirty buffers. If another process is already 1265 * doing the flush, just wait for completion. 1266 */ 1267int 1268nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1269{ 1270 struct nfsnode *np = VTONFS(vp); 1271 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1272 int error = 0, slpflag, slptimeo; 1273 int old_lock = 0; 1274 1275 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf"); 1276 1277 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1278 intrflg = 0; 1279 if (intrflg) { 1280 slpflag = NFS_PCATCH; 1281 slptimeo = 2 * hz; 1282 } else { 1283 slpflag = 0; 1284 slptimeo = 0; 1285 } 1286 1287 old_lock = nfs_upgrade_vnlock(vp); 1288 if (vp->v_iflag & VI_DOOMED) { 1289 /* 1290 * Since vgonel() uses the generic vinvalbuf() to flush 1291 * dirty buffers and it does not call this function, it 1292 * is safe to just return OK when VI_DOOMED is set. 1293 */ 1294 nfs_downgrade_vnlock(vp, old_lock); 1295 return (0); 1296 } 1297 1298 /* 1299 * Now, flush as required. 1300 */ 1301 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1302 VM_OBJECT_LOCK(vp->v_bufobj.bo_object); 1303 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1304 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); 1305 /* 1306 * If the page clean was interrupted, fail the invalidation. 1307 * Not doing so, we run the risk of losing dirty pages in the 1308 * vinvalbuf() call below. 1309 */ 1310 if (intrflg && (error = nfs_sigintr(nmp, td))) 1311 goto out; 1312 } 1313 1314 error = vinvalbuf(vp, flags, slpflag, 0); 1315 while (error) { 1316 if (intrflg && (error = nfs_sigintr(nmp, td))) 1317 goto out; 1318 error = vinvalbuf(vp, flags, 0, slptimeo); 1319 } 1320 mtx_lock(&np->n_mtx); 1321 if (np->n_directio_asyncwr == 0) 1322 np->n_flag &= ~NMODIFIED; 1323 mtx_unlock(&np->n_mtx); 1324out: 1325 nfs_downgrade_vnlock(vp, old_lock); 1326 return error; 1327} 1328 1329/* 1330 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1331 * This is mainly to avoid queueing async I/O requests when the nfsiods 1332 * are all hung on a dead server. 1333 * 1334 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1335 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1336 */ 1337int 1338nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1339{ 1340 int iod; 1341 int gotiod; 1342 int slpflag = 0; 1343 int slptimeo = 0; 1344 int error, error2; 1345 1346 /* 1347 * Commits are usually short and sweet so lets save some cpu and 1348 * leave the async daemons for more important rpc's (such as reads 1349 * and writes). 1350 */ 1351 mtx_lock(&nfs_iod_mtx); 1352 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1353 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1354 mtx_unlock(&nfs_iod_mtx); 1355 return(EIO); 1356 } 1357again: 1358 if (nmp->nm_flag & NFSMNT_INT) 1359 slpflag = NFS_PCATCH; 1360 gotiod = FALSE; 1361 1362 /* 1363 * Find a free iod to process this request. 1364 */ 1365 for (iod = 0; iod < nfs_numasync; iod++) 1366 if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) { 1367 gotiod = TRUE; 1368 break; 1369 } 1370 1371 /* 1372 * Try to create one if none are free. 1373 */ 1374 if (!gotiod) 1375 nfs_nfsiodnew(); 1376 else { 1377 /* 1378 * Found one, so wake it up and tell it which 1379 * mount to process. 1380 */ 1381 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n", 1382 iod, nmp)); 1383 nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE; 1384 nfs_iodmount[iod] = nmp; 1385 nmp->nm_bufqiods++; 1386 wakeup(&nfs_iodwant[iod]); 1387 } 1388 1389 /* 1390 * If none are free, we may already have an iod working on this mount 1391 * point. If so, it will process our request. 1392 */ 1393 if (!gotiod) { 1394 if (nmp->nm_bufqiods > 0) { 1395 NFS_DPF(ASYNCIO, 1396 ("nfs_asyncio: %d iods are already processing mount %p\n", 1397 nmp->nm_bufqiods, nmp)); 1398 gotiod = TRUE; 1399 } 1400 } 1401 1402 /* 1403 * If we have an iod which can process the request, then queue 1404 * the buffer. 1405 */ 1406 if (gotiod) { 1407 /* 1408 * Ensure that the queue never grows too large. We still want 1409 * to asynchronize so we block rather then return EIO. 1410 */ 1411 while (nmp->nm_bufqlen >= 2 * nfs_numasync) { 1412 NFS_DPF(ASYNCIO, 1413 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1414 nmp->nm_bufqwant = TRUE; 1415 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx, 1416 slpflag | PRIBIO, 1417 "nfsaio", slptimeo); 1418 if (error) { 1419 error2 = nfs_sigintr(nmp, td); 1420 if (error2) { 1421 mtx_unlock(&nfs_iod_mtx); 1422 return (error2); 1423 } 1424 if (slpflag == NFS_PCATCH) { 1425 slpflag = 0; 1426 slptimeo = 2 * hz; 1427 } 1428 } 1429 /* 1430 * We might have lost our iod while sleeping, 1431 * so check and loop if nescessary. 1432 */ 1433 goto again; 1434 } 1435 1436 /* We might have lost our nfsiod */ 1437 if (nmp->nm_bufqiods == 0) { 1438 NFS_DPF(ASYNCIO, 1439("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1440 goto again; 1441 } 1442 1443 if (bp->b_iocmd == BIO_READ) { 1444 if (bp->b_rcred == NOCRED && cred != NOCRED) 1445 bp->b_rcred = crhold(cred); 1446 } else { 1447 if (bp->b_wcred == NOCRED && cred != NOCRED) 1448 bp->b_wcred = crhold(cred); 1449 } 1450 1451 if (bp->b_flags & B_REMFREE) 1452 bremfreef(bp); 1453 BUF_KERNPROC(bp); 1454 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1455 nmp->nm_bufqlen++; 1456 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1457 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1458 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1459 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1460 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1461 } 1462 mtx_unlock(&nfs_iod_mtx); 1463 return (0); 1464 } 1465 1466 mtx_unlock(&nfs_iod_mtx); 1467 1468 /* 1469 * All the iods are busy on other mounts, so return EIO to 1470 * force the caller to process the i/o synchronously. 1471 */ 1472 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1473 return (EIO); 1474} 1475 1476void 1477nfs_doio_directwrite(struct buf *bp) 1478{ 1479 int iomode, must_commit; 1480 struct uio *uiop = (struct uio *)bp->b_caller1; 1481 char *iov_base = uiop->uio_iov->iov_base; 1482 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount); 1483 1484 iomode = NFSV3WRITE_FILESYNC; 1485 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1486 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit); 1487 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write")); 1488 free(iov_base, M_NFSDIRECTIO); 1489 free(uiop->uio_iov, M_NFSDIRECTIO); 1490 free(uiop, M_NFSDIRECTIO); 1491 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1492 struct nfsnode *np = VTONFS(bp->b_vp); 1493 mtx_lock(&np->n_mtx); 1494 np->n_directio_asyncwr--; 1495 if (np->n_directio_asyncwr == 0) { 1496 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED; 1497 if ((np->n_flag & NFSYNCWAIT)) { 1498 np->n_flag &= ~NFSYNCWAIT; 1499 wakeup((caddr_t)&np->n_directio_asyncwr); 1500 } 1501 } 1502 mtx_unlock(&np->n_mtx); 1503 } 1504 bp->b_vp = NULL; 1505 relpbuf(bp, &nfs_pbuf_freecnt); 1506} 1507 1508/* 1509 * Do an I/O operation to/from a cache block. This may be called 1510 * synchronously or from an nfsiod. 1511 */ 1512int 1513nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) 1514{ 1515 struct uio *uiop; 1516 struct nfsnode *np; 1517 struct nfsmount *nmp; 1518 int error = 0, iomode, must_commit = 0; 1519 struct uio uio; 1520 struct iovec io; 1521 struct proc *p = td ? td->td_proc : NULL; 1522 uint8_t iocmd; 1523 1524 np = VTONFS(vp); 1525 nmp = VFSTONFS(vp->v_mount); 1526 uiop = &uio; 1527 uiop->uio_iov = &io; 1528 uiop->uio_iovcnt = 1; 1529 uiop->uio_segflg = UIO_SYSSPACE; 1530 uiop->uio_td = td; 1531 1532 /* 1533 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1534 * do this here so we do not have to do it in all the code that 1535 * calls us. 1536 */ 1537 bp->b_flags &= ~B_INVAL; 1538 bp->b_ioflags &= ~BIO_ERROR; 1539 1540 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1541 iocmd = bp->b_iocmd; 1542 if (iocmd == BIO_READ) { 1543 io.iov_len = uiop->uio_resid = bp->b_bcount; 1544 io.iov_base = bp->b_data; 1545 uiop->uio_rw = UIO_READ; 1546 1547 switch (vp->v_type) { 1548 case VREG: 1549 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1550 nfsstats.read_bios++; 1551 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr); 1552 1553 if (!error) { 1554 if (uiop->uio_resid) { 1555 /* 1556 * If we had a short read with no error, we must have 1557 * hit a file hole. We should zero-fill the remainder. 1558 * This can also occur if the server hits the file EOF. 1559 * 1560 * Holes used to be able to occur due to pending 1561 * writes, but that is not possible any longer. 1562 */ 1563 int nread = bp->b_bcount - uiop->uio_resid; 1564 int left = uiop->uio_resid; 1565 1566 if (left > 0) 1567 bzero((char *)bp->b_data + nread, left); 1568 uiop->uio_resid = 0; 1569 } 1570 } 1571 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */ 1572 if (p && (vp->v_vflag & VV_TEXT)) { 1573 mtx_lock(&np->n_mtx); 1574 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) { 1575 mtx_unlock(&np->n_mtx); 1576 PROC_LOCK(p); 1577 killproc(p, "text file modification"); 1578 PROC_UNLOCK(p); 1579 } else 1580 mtx_unlock(&np->n_mtx); 1581 } 1582 break; 1583 case VLNK: 1584 uiop->uio_offset = (off_t)0; 1585 nfsstats.readlink_bios++; 1586 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr); 1587 break; 1588 case VDIR: 1589 nfsstats.readdir_bios++; 1590 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1591 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1592 error = nfs_readdirplusrpc(vp, uiop, cr); 1593 if (error == NFSERR_NOTSUPP) 1594 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1595 } 1596 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1597 error = nfs_readdirrpc(vp, uiop, cr); 1598 /* 1599 * end-of-directory sets B_INVAL but does not generate an 1600 * error. 1601 */ 1602 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1603 bp->b_flags |= B_INVAL; 1604 break; 1605 default: 1606 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type); 1607 break; 1608 }; 1609 if (error) { 1610 bp->b_ioflags |= BIO_ERROR; 1611 bp->b_error = error; 1612 } 1613 } else { 1614 /* 1615 * If we only need to commit, try to commit 1616 */ 1617 if (bp->b_flags & B_NEEDCOMMIT) { 1618 int retv; 1619 off_t off; 1620 1621 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1622 retv = (nmp->nm_rpcops->nr_commit)( 1623 vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1624 bp->b_wcred, td); 1625 if (retv == 0) { 1626 bp->b_dirtyoff = bp->b_dirtyend = 0; 1627 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1628 bp->b_resid = 0; 1629 bufdone(bp); 1630 return (0); 1631 } 1632 if (retv == NFSERR_STALEWRITEVERF) { 1633 nfs_clearcommit(vp->v_mount); 1634 } 1635 } 1636 1637 /* 1638 * Setup for actual write 1639 */ 1640 mtx_lock(&np->n_mtx); 1641 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1642 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1643 mtx_unlock(&np->n_mtx); 1644 1645 if (bp->b_dirtyend > bp->b_dirtyoff) { 1646 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1647 - bp->b_dirtyoff; 1648 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1649 + bp->b_dirtyoff; 1650 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1651 uiop->uio_rw = UIO_WRITE; 1652 nfsstats.write_bios++; 1653 1654 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1655 iomode = NFSV3WRITE_UNSTABLE; 1656 else 1657 iomode = NFSV3WRITE_FILESYNC; 1658 1659 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit); 1660 1661 /* 1662 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1663 * to cluster the buffers needing commit. This will allow 1664 * the system to submit a single commit rpc for the whole 1665 * cluster. We can do this even if the buffer is not 100% 1666 * dirty (relative to the NFS blocksize), so we optimize the 1667 * append-to-file-case. 1668 * 1669 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1670 * cleared because write clustering only works for commit 1671 * rpc's, not for the data portion of the write). 1672 */ 1673 1674 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1675 bp->b_flags |= B_NEEDCOMMIT; 1676 if (bp->b_dirtyoff == 0 1677 && bp->b_dirtyend == bp->b_bcount) 1678 bp->b_flags |= B_CLUSTEROK; 1679 } else { 1680 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1681 } 1682 1683 /* 1684 * For an interrupted write, the buffer is still valid 1685 * and the write hasn't been pushed to the server yet, 1686 * so we can't set BIO_ERROR and report the interruption 1687 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1688 * is not relevant, so the rpc attempt is essentially 1689 * a noop. For the case of a V3 write rpc not being 1690 * committed to stable storage, the block is still 1691 * dirty and requires either a commit rpc or another 1692 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1693 * the block is reused. This is indicated by setting 1694 * the B_DELWRI and B_NEEDCOMMIT flags. 1695 * 1696 * If the buffer is marked B_PAGING, it does not reside on 1697 * the vp's paging queues so we cannot call bdirty(). The 1698 * bp in this case is not an NFS cache block so we should 1699 * be safe. XXX 1700 * 1701 * The logic below breaks up errors into recoverable and 1702 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1703 * and keep the buffer around for potential write retries. 1704 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1705 * and save the error in the nfsnode. This is less than ideal 1706 * but necessary. Keeping such buffers around could potentially 1707 * cause buffer exhaustion eventually (they can never be written 1708 * out, so will get constantly be re-dirtied). It also causes 1709 * all sorts of vfs panics. For non-recoverable write errors, 1710 * also invalidate the attrcache, so we'll be forced to go over 1711 * the wire for this object, returning an error to user on next 1712 * call (most of the time). 1713 */ 1714 if (error == EINTR || error == EIO || error == ETIMEDOUT 1715 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1716 int s; 1717 1718 s = splbio(); 1719 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1720 if ((bp->b_flags & B_PAGING) == 0) { 1721 bdirty(bp); 1722 bp->b_flags &= ~B_DONE; 1723 } 1724 if (error && (bp->b_flags & B_ASYNC) == 0) 1725 bp->b_flags |= B_EINTR; 1726 splx(s); 1727 } else { 1728 if (error) { 1729 bp->b_ioflags |= BIO_ERROR; 1730 bp->b_flags |= B_INVAL; 1731 bp->b_error = np->n_error = error; 1732 mtx_lock(&np->n_mtx); 1733 np->n_flag |= NWRITEERR; 1734 np->n_attrstamp = 0; 1735 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1736 mtx_unlock(&np->n_mtx); 1737 } 1738 bp->b_dirtyoff = bp->b_dirtyend = 0; 1739 } 1740 } else { 1741 bp->b_resid = 0; 1742 bufdone(bp); 1743 return (0); 1744 } 1745 } 1746 bp->b_resid = uiop->uio_resid; 1747 if (must_commit) 1748 nfs_clearcommit(vp->v_mount); 1749 bufdone(bp); 1750 return (error); 1751} 1752 1753/* 1754 * Used to aid in handling ftruncate() operations on the NFS client side. 1755 * Truncation creates a number of special problems for NFS. We have to 1756 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1757 * we have to properly handle VM pages or (potentially dirty) buffers 1758 * that straddle the truncation point. 1759 */ 1760 1761int 1762nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1763{ 1764 struct nfsnode *np = VTONFS(vp); 1765 u_quad_t tsize; 1766 int biosize = vp->v_bufobj.bo_bsize; 1767 int error = 0; 1768 1769 mtx_lock(&np->n_mtx); 1770 tsize = np->n_size; 1771 np->n_size = nsize; 1772 mtx_unlock(&np->n_mtx); 1773 1774 if (nsize < tsize) { 1775 struct buf *bp; 1776 daddr_t lbn; 1777 int bufsize; 1778 1779 /* 1780 * vtruncbuf() doesn't get the buffer overlapping the 1781 * truncation point. We may have a B_DELWRI and/or B_CACHE 1782 * buffer that now needs to be truncated. 1783 */ 1784 error = vtruncbuf(vp, cred, td, nsize, biosize); 1785 lbn = nsize / biosize; 1786 bufsize = nsize & (biosize - 1); 1787 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1788 if (!bp) 1789 return EINTR; 1790 if (bp->b_dirtyoff > bp->b_bcount) 1791 bp->b_dirtyoff = bp->b_bcount; 1792 if (bp->b_dirtyend > bp->b_bcount) 1793 bp->b_dirtyend = bp->b_bcount; 1794 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1795 brelse(bp); 1796 } else { 1797 vnode_pager_setsize(vp, nsize); 1798 } 1799 return(error); 1800} 1801 1802