nfs_bio.c revision 158739
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 158739 2006-05-19 00:04:24Z mohans $"); 37 38#include <sys/param.h> 39#include <sys/systm.h> 40#include <sys/bio.h> 41#include <sys/buf.h> 42#include <sys/kernel.h> 43#include <sys/mount.h> 44#include <sys/proc.h> 45#include <sys/resourcevar.h> 46#include <sys/signalvar.h> 47#include <sys/vmmeter.h> 48#include <sys/vnode.h> 49 50#include <vm/vm.h> 51#include <vm/vm_extern.h> 52#include <vm/vm_page.h> 53#include <vm/vm_object.h> 54#include <vm/vm_pager.h> 55#include <vm/vnode_pager.h> 56 57#include <rpc/rpcclnt.h> 58 59#include <nfs/rpcv2.h> 60#include <nfs/nfsproto.h> 61#include <nfsclient/nfs.h> 62#include <nfsclient/nfsmount.h> 63#include <nfsclient/nfsnode.h> 64 65#include <nfs4client/nfs4.h> 66 67static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 68 struct thread *td); 69static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 70 struct ucred *cred, int ioflag); 71 72extern int nfs_directio_enable; 73extern int nfs_directio_allow_mmap; 74 75/* 76 * Vnode op for VM getpages. 77 */ 78int 79nfs_getpages(struct vop_getpages_args *ap) 80{ 81 int i, error, nextoff, size, toff, count, npages; 82 struct uio uio; 83 struct iovec iov; 84 vm_offset_t kva; 85 struct buf *bp; 86 struct vnode *vp; 87 struct thread *td; 88 struct ucred *cred; 89 struct nfsmount *nmp; 90 vm_object_t object; 91 vm_page_t *pages; 92 struct nfsnode *np; 93 94 vp = ap->a_vp; 95 np = VTONFS(vp); 96 td = curthread; /* XXX */ 97 cred = curthread->td_ucred; /* XXX */ 98 nmp = VFSTONFS(vp->v_mount); 99 pages = ap->a_m; 100 count = ap->a_count; 101 102 if ((object = vp->v_object) == NULL) { 103 nfs_printf("nfs_getpages: called with non-merged cache vnode??\n"); 104 return VM_PAGER_ERROR; 105 } 106 107 if (nfs_directio_enable && !nfs_directio_allow_mmap) { 108 mtx_lock(&np->n_mtx); 109 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 110 mtx_unlock(&np->n_mtx); 111 nfs_printf("nfs_getpages: called on non-cacheable vnode??\n"); 112 return VM_PAGER_ERROR; 113 } else 114 mtx_unlock(&np->n_mtx); 115 } 116 117 mtx_lock(&nmp->nm_mtx); 118 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 119 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 120 mtx_unlock(&nmp->nm_mtx); 121 /* We'll never get here for v4, because we always have fsinfo */ 122 (void)nfs_fsinfo(nmp, vp, cred, td); 123 } else 124 mtx_unlock(&nmp->nm_mtx); 125 126 npages = btoc(count); 127 128 /* 129 * If the requested page is partially valid, just return it and 130 * allow the pager to zero-out the blanks. Partially valid pages 131 * can only occur at the file EOF. 132 */ 133 134 { 135 vm_page_t m = pages[ap->a_reqpage]; 136 137 VM_OBJECT_LOCK(object); 138 vm_page_lock_queues(); 139 if (m->valid != 0) { 140 /* handled by vm_fault now */ 141 /* vm_page_zero_invalid(m, TRUE); */ 142 for (i = 0; i < npages; ++i) { 143 if (i != ap->a_reqpage) 144 vm_page_free(pages[i]); 145 } 146 vm_page_unlock_queues(); 147 VM_OBJECT_UNLOCK(object); 148 return(0); 149 } 150 vm_page_unlock_queues(); 151 VM_OBJECT_UNLOCK(object); 152 } 153 154 /* 155 * We use only the kva address for the buffer, but this is extremely 156 * convienient and fast. 157 */ 158 bp = getpbuf(&nfs_pbuf_freecnt); 159 160 kva = (vm_offset_t) bp->b_data; 161 pmap_qenter(kva, pages, npages); 162 cnt.v_vnodein++; 163 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 = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred); 176 pmap_qremove(kva, npages); 177 178 relpbuf(bp, &nfs_pbuf_freecnt); 179 180 if (error && (uio.uio_resid == count)) { 181 nfs_printf("nfs_getpages: error %d\n", error); 182 VM_OBJECT_LOCK(object); 183 vm_page_lock_queues(); 184 for (i = 0; i < npages; ++i) { 185 if (i != ap->a_reqpage) 186 vm_page_free(pages[i]); 187 } 188 vm_page_unlock_queues(); 189 VM_OBJECT_UNLOCK(object); 190 return VM_PAGER_ERROR; 191 } 192 193 /* 194 * Calculate the number of bytes read and validate only that number 195 * of bytes. Note that due to pending writes, size may be 0. This 196 * does not mean that the remaining data is invalid! 197 */ 198 199 size = count - uio.uio_resid; 200 VM_OBJECT_LOCK(object); 201 vm_page_lock_queues(); 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 vm_page_undirty(m); 213 } else if (size > toff) { 214 /* 215 * Read operation filled a partial page. 216 */ 217 m->valid = 0; 218 vm_page_set_validclean(m, 0, size - toff); 219 /* handled by vm_fault now */ 220 /* vm_page_zero_invalid(m, TRUE); */ 221 } else { 222 /* 223 * Read operation was short. If no error occured 224 * we may have hit a zero-fill section. We simply 225 * leave valid set to 0. 226 */ 227 ; 228 } 229 if (i != ap->a_reqpage) { 230 /* 231 * Whether or not to leave the page activated is up in 232 * the air, but we should put the page on a page queue 233 * somewhere (it already is in the object). Result: 234 * It appears that emperical results show that 235 * deactivating pages is best. 236 */ 237 238 /* 239 * Just in case someone was asking for this page we 240 * now tell them that it is ok to use. 241 */ 242 if (!error) { 243 if (m->flags & PG_WANTED) 244 vm_page_activate(m); 245 else 246 vm_page_deactivate(m); 247 vm_page_wakeup(m); 248 } else { 249 vm_page_free(m); 250 } 251 } 252 } 253 vm_page_unlock_queues(); 254 VM_OBJECT_UNLOCK(object); 255 return 0; 256} 257 258/* 259 * Vnode op for VM putpages. 260 */ 261int 262nfs_putpages(struct vop_putpages_args *ap) 263{ 264 struct uio uio; 265 struct iovec iov; 266 vm_offset_t kva; 267 struct buf *bp; 268 int iomode, must_commit, i, error, npages, count; 269 off_t offset; 270 int *rtvals; 271 struct vnode *vp; 272 struct thread *td; 273 struct ucred *cred; 274 struct nfsmount *nmp; 275 struct nfsnode *np; 276 vm_page_t *pages; 277 278 vp = ap->a_vp; 279 np = VTONFS(vp); 280 td = curthread; /* XXX */ 281 cred = curthread->td_ucred; /* XXX */ 282 nmp = VFSTONFS(vp->v_mount); 283 pages = ap->a_m; 284 count = ap->a_count; 285 rtvals = ap->a_rtvals; 286 npages = btoc(count); 287 offset = IDX_TO_OFF(pages[0]->pindex); 288 289 mtx_lock(&nmp->nm_mtx); 290 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 291 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 292 mtx_unlock(&nmp->nm_mtx); 293 (void)nfs_fsinfo(nmp, vp, cred, td); 294 } else 295 mtx_unlock(&nmp->nm_mtx); 296 297 mtx_lock(&np->n_mtx); 298 if (nfs_directio_enable && !nfs_directio_allow_mmap && 299 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 300 mtx_unlock(&np->n_mtx); 301 nfs_printf("nfs_putpages: called on noncache-able vnode??\n"); 302 mtx_lock(&np->n_mtx); 303 } 304 305 for (i = 0; i < npages; i++) 306 rtvals[i] = VM_PAGER_AGAIN; 307 308 /* 309 * When putting pages, do not extend file past EOF. 310 */ 311 if (offset + count > np->n_size) { 312 count = np->n_size - offset; 313 if (count < 0) 314 count = 0; 315 } 316 mtx_unlock(&np->n_mtx); 317 318 /* 319 * We use only the kva address for the buffer, but this is extremely 320 * convienient and fast. 321 */ 322 bp = getpbuf(&nfs_pbuf_freecnt); 323 324 kva = (vm_offset_t) bp->b_data; 325 pmap_qenter(kva, pages, npages); 326 cnt.v_vnodeout++; 327 cnt.v_vnodepgsout += count; 328 329 iov.iov_base = (caddr_t) kva; 330 iov.iov_len = count; 331 uio.uio_iov = &iov; 332 uio.uio_iovcnt = 1; 333 uio.uio_offset = offset; 334 uio.uio_resid = count; 335 uio.uio_segflg = UIO_SYSSPACE; 336 uio.uio_rw = UIO_WRITE; 337 uio.uio_td = td; 338 339 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 340 iomode = NFSV3WRITE_UNSTABLE; 341 else 342 iomode = NFSV3WRITE_FILESYNC; 343 344 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit); 345 346 pmap_qremove(kva, npages); 347 relpbuf(bp, &nfs_pbuf_freecnt); 348 349 if (!error) { 350 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 351 for (i = 0; i < nwritten; i++) { 352 rtvals[i] = VM_PAGER_OK; 353 vm_page_undirty(pages[i]); 354 } 355 if (must_commit) { 356 nfs_clearcommit(vp->v_mount); 357 } 358 } 359 return rtvals[0]; 360} 361 362/* 363 * For nfs, cache consistency can only be maintained approximately. 364 * Although RFC1094 does not specify the criteria, the following is 365 * believed to be compatible with the reference port. 366 * For nfs: 367 * If the file's modify time on the server has changed since the 368 * last read rpc or you have written to the file, 369 * you may have lost data cache consistency with the 370 * server, so flush all of the file's data out of the cache. 371 * Then force a getattr rpc to ensure that you have up to date 372 * attributes. 373 * NB: This implies that cache data can be read when up to 374 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 375 * attributes this could be forced by setting n_attrstamp to 0 before 376 * the VOP_GETATTR() call. 377 */ 378static inline int 379nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 380{ 381 int error = 0; 382 struct vattr vattr; 383 struct nfsnode *np = VTONFS(vp); 384 int old_lock; 385 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 386 387 /* 388 * Grab the exclusive lock before checking whether the cache is 389 * consistent. 390 * XXX - We can make this cheaper later (by acquiring cheaper locks). 391 * But for now, this suffices. 392 */ 393 old_lock = nfs_upgrade_vnlock(vp, td); 394 mtx_lock(&np->n_mtx); 395 if (np->n_flag & NMODIFIED) { 396 mtx_unlock(&np->n_mtx); 397 if (vp->v_type != VREG) { 398 if (vp->v_type != VDIR) 399 panic("nfs: bioread, not dir"); 400 (nmp->nm_rpcops->nr_invaldir)(vp); 401 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 402 if (error) 403 goto out; 404 } 405 np->n_attrstamp = 0; 406 error = VOP_GETATTR(vp, &vattr, cred, td); 407 if (error) 408 goto out; 409 mtx_lock(&np->n_mtx); 410 np->n_mtime = vattr.va_mtime; 411 mtx_unlock(&np->n_mtx); 412 } else { 413 mtx_unlock(&np->n_mtx); 414 error = VOP_GETATTR(vp, &vattr, cred, td); 415 if (error) 416 return (error); 417 mtx_lock(&np->n_mtx); 418 if ((np->n_flag & NSIZECHANGED) 419 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 420 mtx_unlock(&np->n_mtx); 421 if (vp->v_type == VDIR) 422 (nmp->nm_rpcops->nr_invaldir)(vp); 423 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 424 if (error) 425 goto out; 426 mtx_lock(&np->n_mtx); 427 np->n_mtime = vattr.va_mtime; 428 np->n_flag &= ~NSIZECHANGED; 429 } 430 mtx_unlock(&np->n_mtx); 431 } 432out: 433 nfs_downgrade_vnlock(vp, td, old_lock); 434 return error; 435} 436 437/* 438 * Vnode op for read using bio 439 */ 440int 441nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 442{ 443 struct nfsnode *np = VTONFS(vp); 444 int biosize, i; 445 struct buf *bp, *rabp; 446 struct thread *td; 447 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 448 daddr_t lbn, rabn; 449 int bcount; 450 int seqcount; 451 int nra, error = 0, n = 0, on = 0; 452 453#ifdef DIAGNOSTIC 454 if (uio->uio_rw != UIO_READ) 455 panic("nfs_read mode"); 456#endif 457 if (uio->uio_resid == 0) 458 return (0); 459 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 460 return (EINVAL); 461 td = uio->uio_td; 462 463 mtx_lock(&nmp->nm_mtx); 464 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 465 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 466 mtx_unlock(&nmp->nm_mtx); 467 (void)nfs_fsinfo(nmp, vp, cred, td); 468 } else 469 mtx_unlock(&nmp->nm_mtx); 470 471 if (vp->v_type != VDIR && 472 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 473 return (EFBIG); 474 475 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 476 /* No caching/ no readaheads. Just read data into the user buffer */ 477 return nfs_readrpc(vp, uio, cred); 478 479 biosize = vp->v_mount->mnt_stat.f_iosize; 480 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 481 482 error = nfs_bioread_check_cons(vp, td, cred); 483 if (error) 484 return error; 485 486 do { 487 u_quad_t nsize; 488 489 mtx_lock(&np->n_mtx); 490 nsize = np->n_size; 491 mtx_unlock(&np->n_mtx); 492 493 switch (vp->v_type) { 494 case VREG: 495 nfsstats.biocache_reads++; 496 lbn = uio->uio_offset / biosize; 497 on = uio->uio_offset & (biosize - 1); 498 499 /* 500 * Start the read ahead(s), as required. 501 */ 502 if (nmp->nm_readahead > 0) { 503 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 504 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 505 rabn = lbn + 1 + nra; 506 if (incore(&vp->v_bufobj, rabn) == NULL) { 507 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 508 if (!rabp) { 509 error = nfs_sigintr(nmp, NULL, td); 510 return (error ? error : EINTR); 511 } 512 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 513 rabp->b_flags |= B_ASYNC; 514 rabp->b_iocmd = BIO_READ; 515 vfs_busy_pages(rabp, 0); 516 if (nfs_asyncio(nmp, rabp, cred, td)) { 517 rabp->b_flags |= B_INVAL; 518 rabp->b_ioflags |= BIO_ERROR; 519 vfs_unbusy_pages(rabp); 520 brelse(rabp); 521 break; 522 } 523 } else { 524 brelse(rabp); 525 } 526 } 527 } 528 } 529 530 /* Note that bcount is *not* DEV_BSIZE aligned. */ 531 bcount = biosize; 532 if ((off_t)lbn * biosize >= nsize) { 533 bcount = 0; 534 } else if ((off_t)(lbn + 1) * biosize > nsize) { 535 bcount = nsize - (off_t)lbn * biosize; 536 } 537 bp = nfs_getcacheblk(vp, lbn, bcount, td); 538 539 if (!bp) { 540 error = nfs_sigintr(nmp, NULL, td); 541 return (error ? error : EINTR); 542 } 543 544 /* 545 * If B_CACHE is not set, we must issue the read. If this 546 * fails, we return an error. 547 */ 548 549 if ((bp->b_flags & B_CACHE) == 0) { 550 bp->b_iocmd = BIO_READ; 551 vfs_busy_pages(bp, 0); 552 error = nfs_doio(vp, bp, cred, td); 553 if (error) { 554 brelse(bp); 555 return (error); 556 } 557 } 558 559 /* 560 * on is the offset into the current bp. Figure out how many 561 * bytes we can copy out of the bp. Note that bcount is 562 * NOT DEV_BSIZE aligned. 563 * 564 * Then figure out how many bytes we can copy into the uio. 565 */ 566 567 n = 0; 568 if (on < bcount) 569 n = min((unsigned)(bcount - on), uio->uio_resid); 570 break; 571 case VLNK: 572 nfsstats.biocache_readlinks++; 573 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 574 if (!bp) { 575 error = nfs_sigintr(nmp, NULL, td); 576 return (error ? error : EINTR); 577 } 578 if ((bp->b_flags & B_CACHE) == 0) { 579 bp->b_iocmd = BIO_READ; 580 vfs_busy_pages(bp, 0); 581 error = nfs_doio(vp, bp, cred, td); 582 if (error) { 583 bp->b_ioflags |= BIO_ERROR; 584 brelse(bp); 585 return (error); 586 } 587 } 588 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 589 on = 0; 590 break; 591 case VDIR: 592 nfsstats.biocache_readdirs++; 593 if (np->n_direofoffset 594 && uio->uio_offset >= np->n_direofoffset) { 595 return (0); 596 } 597 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 598 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 599 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 600 if (!bp) { 601 error = nfs_sigintr(nmp, NULL, td); 602 return (error ? error : EINTR); 603 } 604 if ((bp->b_flags & B_CACHE) == 0) { 605 bp->b_iocmd = BIO_READ; 606 vfs_busy_pages(bp, 0); 607 error = nfs_doio(vp, bp, cred, td); 608 if (error) { 609 brelse(bp); 610 } 611 while (error == NFSERR_BAD_COOKIE) { 612 (nmp->nm_rpcops->nr_invaldir)(vp); 613 error = nfs_vinvalbuf(vp, 0, td, 1); 614 /* 615 * Yuck! The directory has been modified on the 616 * server. The only way to get the block is by 617 * reading from the beginning to get all the 618 * offset cookies. 619 * 620 * Leave the last bp intact unless there is an error. 621 * Loop back up to the while if the error is another 622 * NFSERR_BAD_COOKIE (double yuch!). 623 */ 624 for (i = 0; i <= lbn && !error; i++) { 625 if (np->n_direofoffset 626 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 627 return (0); 628 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 629 if (!bp) { 630 error = nfs_sigintr(nmp, NULL, td); 631 return (error ? error : EINTR); 632 } 633 if ((bp->b_flags & B_CACHE) == 0) { 634 bp->b_iocmd = BIO_READ; 635 vfs_busy_pages(bp, 0); 636 error = nfs_doio(vp, bp, cred, td); 637 /* 638 * no error + B_INVAL == directory EOF, 639 * use the block. 640 */ 641 if (error == 0 && (bp->b_flags & B_INVAL)) 642 break; 643 } 644 /* 645 * An error will throw away the block and the 646 * for loop will break out. If no error and this 647 * is not the block we want, we throw away the 648 * block and go for the next one via the for loop. 649 */ 650 if (error || i < lbn) 651 brelse(bp); 652 } 653 } 654 /* 655 * The above while is repeated if we hit another cookie 656 * error. If we hit an error and it wasn't a cookie error, 657 * we give up. 658 */ 659 if (error) 660 return (error); 661 } 662 663 /* 664 * If not eof and read aheads are enabled, start one. 665 * (You need the current block first, so that you have the 666 * directory offset cookie of the next block.) 667 */ 668 if (nmp->nm_readahead > 0 && 669 (bp->b_flags & B_INVAL) == 0 && 670 (np->n_direofoffset == 0 || 671 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 672 incore(&vp->v_bufobj, lbn + 1) == NULL) { 673 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 674 if (rabp) { 675 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 676 rabp->b_flags |= B_ASYNC; 677 rabp->b_iocmd = BIO_READ; 678 vfs_busy_pages(rabp, 0); 679 if (nfs_asyncio(nmp, rabp, cred, td)) { 680 rabp->b_flags |= B_INVAL; 681 rabp->b_ioflags |= BIO_ERROR; 682 vfs_unbusy_pages(rabp); 683 brelse(rabp); 684 } 685 } else { 686 brelse(rabp); 687 } 688 } 689 } 690 /* 691 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 692 * chopped for the EOF condition, we cannot tell how large 693 * NFS directories are going to be until we hit EOF. So 694 * an NFS directory buffer is *not* chopped to its EOF. Now, 695 * it just so happens that b_resid will effectively chop it 696 * to EOF. *BUT* this information is lost if the buffer goes 697 * away and is reconstituted into a B_CACHE state ( due to 698 * being VMIO ) later. So we keep track of the directory eof 699 * in np->n_direofoffset and chop it off as an extra step 700 * right here. 701 */ 702 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 703 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 704 n = np->n_direofoffset - uio->uio_offset; 705 break; 706 default: 707 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type); 708 bp = NULL; 709 break; 710 }; 711 712 if (n > 0) { 713 error = uiomove(bp->b_data + on, (int)n, uio); 714 } 715 if (vp->v_type == VLNK) 716 n = 0; 717 if (bp != NULL) 718 brelse(bp); 719 } while (error == 0 && uio->uio_resid > 0 && n > 0); 720 return (error); 721} 722 723/* 724 * The NFS write path cannot handle iovecs with len > 1. So we need to 725 * break up iovecs accordingly (restricting them to wsize). 726 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 727 * For the ASYNC case, 2 copies are needed. The first a copy from the 728 * user buffer to a staging buffer and then a second copy from the staging 729 * buffer to mbufs. This can be optimized by copying from the user buffer 730 * directly into mbufs and passing the chain down, but that requires a 731 * fair amount of re-working of the relevant codepaths (and can be done 732 * later). 733 */ 734static int 735nfs_directio_write(vp, uiop, cred, ioflag) 736 struct vnode *vp; 737 struct uio *uiop; 738 struct ucred *cred; 739 int ioflag; 740{ 741 int error; 742 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 743 struct thread *td = uiop->uio_td; 744 int size; 745 int wsize; 746 747 mtx_lock(&nmp->nm_mtx); 748 wsize = nmp->nm_wsize; 749 mtx_unlock(&nmp->nm_mtx); 750 if (ioflag & IO_SYNC) { 751 int iomode, must_commit; 752 struct uio uio; 753 struct iovec iov; 754do_sync: 755 while (uiop->uio_resid > 0) { 756 size = min(uiop->uio_resid, wsize); 757 size = min(uiop->uio_iov->iov_len, size); 758 iov.iov_base = uiop->uio_iov->iov_base; 759 iov.iov_len = size; 760 uio.uio_iov = &iov; 761 uio.uio_iovcnt = 1; 762 uio.uio_offset = uiop->uio_offset; 763 uio.uio_resid = size; 764 uio.uio_segflg = UIO_USERSPACE; 765 uio.uio_rw = UIO_WRITE; 766 uio.uio_td = td; 767 iomode = NFSV3WRITE_FILESYNC; 768 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, 769 &iomode, &must_commit); 770 KASSERT((must_commit == 0), 771 ("nfs_directio_write: Did not commit write")); 772 if (error) 773 return (error); 774 uiop->uio_offset += size; 775 uiop->uio_resid -= size; 776 if (uiop->uio_iov->iov_len <= size) { 777 uiop->uio_iovcnt--; 778 uiop->uio_iov++; 779 } else { 780 uiop->uio_iov->iov_base = 781 (char *)uiop->uio_iov->iov_base + size; 782 uiop->uio_iov->iov_len -= size; 783 } 784 } 785 } else { 786 struct uio *t_uio; 787 struct iovec *t_iov; 788 struct buf *bp; 789 790 /* 791 * Break up the write into blocksize chunks and hand these 792 * over to nfsiod's for write back. 793 * Unfortunately, this incurs a copy of the data. Since 794 * the user could modify the buffer before the write is 795 * initiated. 796 * 797 * The obvious optimization here is that one of the 2 copies 798 * in the async write path can be eliminated by copying the 799 * data here directly into mbufs and passing the mbuf chain 800 * down. But that will require a fair amount of re-working 801 * of the code and can be done if there's enough interest 802 * in NFS directio access. 803 */ 804 while (uiop->uio_resid > 0) { 805 size = min(uiop->uio_resid, wsize); 806 size = min(uiop->uio_iov->iov_len, size); 807 bp = getpbuf(&nfs_pbuf_freecnt); 808 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 809 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 810 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 811 t_iov->iov_len = size; 812 t_uio->uio_iov = t_iov; 813 t_uio->uio_iovcnt = 1; 814 t_uio->uio_offset = uiop->uio_offset; 815 t_uio->uio_resid = size; 816 t_uio->uio_segflg = UIO_SYSSPACE; 817 t_uio->uio_rw = UIO_WRITE; 818 t_uio->uio_td = td; 819 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); 820 bp->b_flags |= B_DIRECT; 821 bp->b_iocmd = BIO_WRITE; 822 if (cred != NOCRED) { 823 crhold(cred); 824 bp->b_wcred = cred; 825 } else 826 bp->b_wcred = NOCRED; 827 bp->b_caller1 = (void *)t_uio; 828 bp->b_vp = vp; 829 vhold(vp); 830 error = nfs_asyncio(nmp, bp, NOCRED, td); 831 if (error) { 832 free(t_iov->iov_base, M_NFSDIRECTIO); 833 free(t_iov, M_NFSDIRECTIO); 834 free(t_uio, M_NFSDIRECTIO); 835 vdrop(bp->b_vp); 836 bp->b_vp = NULL; 837 relpbuf(bp, &nfs_pbuf_freecnt); 838 if (error == EINTR) 839 return (error); 840 goto do_sync; 841 } 842 uiop->uio_offset += size; 843 uiop->uio_resid -= size; 844 if (uiop->uio_iov->iov_len <= size) { 845 uiop->uio_iovcnt--; 846 uiop->uio_iov++; 847 } else { 848 uiop->uio_iov->iov_base = 849 (char *)uiop->uio_iov->iov_base + size; 850 uiop->uio_iov->iov_len -= size; 851 } 852 } 853 } 854 return (0); 855} 856 857/* 858 * Vnode op for write using bio 859 */ 860int 861nfs_write(struct vop_write_args *ap) 862{ 863 int biosize; 864 struct uio *uio = ap->a_uio; 865 struct thread *td = uio->uio_td; 866 struct vnode *vp = ap->a_vp; 867 struct nfsnode *np = VTONFS(vp); 868 struct ucred *cred = ap->a_cred; 869 int ioflag = ap->a_ioflag; 870 struct buf *bp; 871 struct vattr vattr; 872 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 873 daddr_t lbn; 874 int bcount; 875 int n, on, error = 0; 876 struct proc *p = td?td->td_proc:NULL; 877 878#ifdef DIAGNOSTIC 879 if (uio->uio_rw != UIO_WRITE) 880 panic("nfs_write mode"); 881 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 882 panic("nfs_write proc"); 883#endif 884 if (vp->v_type != VREG) 885 return (EIO); 886 mtx_lock(&np->n_mtx); 887 if (np->n_flag & NWRITEERR) { 888 np->n_flag &= ~NWRITEERR; 889 mtx_unlock(&np->n_mtx); 890 return (np->n_error); 891 } else 892 mtx_unlock(&np->n_mtx); 893 mtx_lock(&nmp->nm_mtx); 894 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 895 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 896 mtx_unlock(&nmp->nm_mtx); 897 (void)nfs_fsinfo(nmp, vp, cred, td); 898 } else 899 mtx_unlock(&nmp->nm_mtx); 900 901 /* 902 * Synchronously flush pending buffers if we are in synchronous 903 * mode or if we are appending. 904 */ 905 if (ioflag & (IO_APPEND | IO_SYNC)) { 906 mtx_lock(&np->n_mtx); 907 if (np->n_flag & NMODIFIED) { 908 mtx_unlock(&np->n_mtx); 909#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 910 /* 911 * Require non-blocking, synchronous writes to 912 * dirty files to inform the program it needs 913 * to fsync(2) explicitly. 914 */ 915 if (ioflag & IO_NDELAY) 916 return (EAGAIN); 917#endif 918flush_and_restart: 919 np->n_attrstamp = 0; 920 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 921 if (error) 922 return (error); 923 } else 924 mtx_unlock(&np->n_mtx); 925 } 926 927 /* 928 * If IO_APPEND then load uio_offset. We restart here if we cannot 929 * get the append lock. 930 */ 931 if (ioflag & IO_APPEND) { 932 np->n_attrstamp = 0; 933 error = VOP_GETATTR(vp, &vattr, cred, td); 934 if (error) 935 return (error); 936 mtx_lock(&np->n_mtx); 937 uio->uio_offset = np->n_size; 938 mtx_unlock(&np->n_mtx); 939 } 940 941 if (uio->uio_offset < 0) 942 return (EINVAL); 943 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 944 return (EFBIG); 945 if (uio->uio_resid == 0) 946 return (0); 947 948 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 949 return nfs_directio_write(vp, uio, cred, ioflag); 950 951 /* 952 * Maybe this should be above the vnode op call, but so long as 953 * file servers have no limits, i don't think it matters 954 */ 955 if (p != NULL) { 956 PROC_LOCK(p); 957 if (uio->uio_offset + uio->uio_resid > 958 lim_cur(p, RLIMIT_FSIZE)) { 959 psignal(p, SIGXFSZ); 960 PROC_UNLOCK(p); 961 return (EFBIG); 962 } 963 PROC_UNLOCK(p); 964 } 965 966 biosize = vp->v_mount->mnt_stat.f_iosize; 967 /* 968 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 969 * would exceed the local maximum per-file write commit size when 970 * combined with those, we must decide whether to flush, 971 * go synchronous, or return error. We don't bother checking 972 * IO_UNIT -- we just make all writes atomic anyway, as there's 973 * no point optimizing for something that really won't ever happen. 974 */ 975 if (!(ioflag & IO_SYNC)) { 976 int nflag; 977 978 mtx_lock(&np->n_mtx); 979 nflag = np->n_flag; 980 mtx_unlock(&np->n_mtx); 981 int needrestart = 0; 982 if (nmp->nm_wcommitsize < uio->uio_resid) { 983 /* 984 * If this request could not possibly be completed 985 * without exceeding the maximum outstanding write 986 * commit size, see if we can convert it into a 987 * synchronous write operation. 988 */ 989 if (ioflag & IO_NDELAY) 990 return (EAGAIN); 991 ioflag |= IO_SYNC; 992 if (nflag & NMODIFIED) 993 needrestart = 1; 994 } else if (nflag & NMODIFIED) { 995 int wouldcommit = 0; 996 BO_LOCK(&vp->v_bufobj); 997 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 998 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 999 b_bobufs) { 1000 if (bp->b_flags & B_NEEDCOMMIT) 1001 wouldcommit += bp->b_bcount; 1002 } 1003 } 1004 BO_UNLOCK(&vp->v_bufobj); 1005 /* 1006 * Since we're not operating synchronously and 1007 * bypassing the buffer cache, we are in a commit 1008 * and holding all of these buffers whether 1009 * transmitted or not. If not limited, this 1010 * will lead to the buffer cache deadlocking, 1011 * as no one else can flush our uncommitted buffers. 1012 */ 1013 wouldcommit += uio->uio_resid; 1014 /* 1015 * If we would initially exceed the maximum 1016 * outstanding write commit size, flush and restart. 1017 */ 1018 if (wouldcommit > nmp->nm_wcommitsize) 1019 needrestart = 1; 1020 } 1021 if (needrestart) 1022 goto flush_and_restart; 1023 } 1024 1025 do { 1026 nfsstats.biocache_writes++; 1027 lbn = uio->uio_offset / biosize; 1028 on = uio->uio_offset & (biosize-1); 1029 n = min((unsigned)(biosize - on), uio->uio_resid); 1030again: 1031 /* 1032 * Handle direct append and file extension cases, calculate 1033 * unaligned buffer size. 1034 */ 1035 mtx_lock(&np->n_mtx); 1036 if (uio->uio_offset == np->n_size && n) { 1037 mtx_unlock(&np->n_mtx); 1038 /* 1039 * Get the buffer (in its pre-append state to maintain 1040 * B_CACHE if it was previously set). Resize the 1041 * nfsnode after we have locked the buffer to prevent 1042 * readers from reading garbage. 1043 */ 1044 bcount = on; 1045 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1046 1047 if (bp != NULL) { 1048 long save; 1049 1050 mtx_lock(&np->n_mtx); 1051 np->n_size = uio->uio_offset + n; 1052 np->n_flag |= NMODIFIED; 1053 vnode_pager_setsize(vp, np->n_size); 1054 mtx_unlock(&np->n_mtx); 1055 1056 save = bp->b_flags & B_CACHE; 1057 bcount += n; 1058 allocbuf(bp, bcount); 1059 bp->b_flags |= save; 1060 } 1061 } else { 1062 /* 1063 * Obtain the locked cache block first, and then 1064 * adjust the file's size as appropriate. 1065 */ 1066 bcount = on + n; 1067 if ((off_t)lbn * biosize + bcount < np->n_size) { 1068 if ((off_t)(lbn + 1) * biosize < np->n_size) 1069 bcount = biosize; 1070 else 1071 bcount = np->n_size - (off_t)lbn * biosize; 1072 } 1073 mtx_unlock(&np->n_mtx); 1074 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1075 mtx_lock(&np->n_mtx); 1076 if (uio->uio_offset + n > np->n_size) { 1077 np->n_size = uio->uio_offset + n; 1078 np->n_flag |= NMODIFIED; 1079 vnode_pager_setsize(vp, np->n_size); 1080 } 1081 mtx_unlock(&np->n_mtx); 1082 } 1083 1084 if (!bp) { 1085 error = nfs_sigintr(nmp, NULL, td); 1086 if (!error) 1087 error = EINTR; 1088 break; 1089 } 1090 1091 /* 1092 * Issue a READ if B_CACHE is not set. In special-append 1093 * mode, B_CACHE is based on the buffer prior to the write 1094 * op and is typically set, avoiding the read. If a read 1095 * is required in special append mode, the server will 1096 * probably send us a short-read since we extended the file 1097 * on our end, resulting in b_resid == 0 and, thusly, 1098 * B_CACHE getting set. 1099 * 1100 * We can also avoid issuing the read if the write covers 1101 * the entire buffer. We have to make sure the buffer state 1102 * is reasonable in this case since we will not be initiating 1103 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1104 * more information. 1105 * 1106 * B_CACHE may also be set due to the buffer being cached 1107 * normally. 1108 */ 1109 1110 if (on == 0 && n == bcount) { 1111 bp->b_flags |= B_CACHE; 1112 bp->b_flags &= ~B_INVAL; 1113 bp->b_ioflags &= ~BIO_ERROR; 1114 } 1115 1116 if ((bp->b_flags & B_CACHE) == 0) { 1117 bp->b_iocmd = BIO_READ; 1118 vfs_busy_pages(bp, 0); 1119 error = nfs_doio(vp, bp, cred, td); 1120 if (error) { 1121 brelse(bp); 1122 break; 1123 } 1124 } 1125 if (bp->b_wcred == NOCRED) 1126 bp->b_wcred = crhold(cred); 1127 mtx_lock(&np->n_mtx); 1128 np->n_flag |= NMODIFIED; 1129 mtx_unlock(&np->n_mtx); 1130 1131 /* 1132 * If dirtyend exceeds file size, chop it down. This should 1133 * not normally occur but there is an append race where it 1134 * might occur XXX, so we log it. 1135 * 1136 * If the chopping creates a reverse-indexed or degenerate 1137 * situation with dirtyoff/end, we 0 both of them. 1138 */ 1139 1140 if (bp->b_dirtyend > bcount) { 1141 nfs_printf("NFS append race @%lx:%d\n", 1142 (long)bp->b_blkno * DEV_BSIZE, 1143 bp->b_dirtyend - bcount); 1144 bp->b_dirtyend = bcount; 1145 } 1146 1147 if (bp->b_dirtyoff >= bp->b_dirtyend) 1148 bp->b_dirtyoff = bp->b_dirtyend = 0; 1149 1150 /* 1151 * If the new write will leave a contiguous dirty 1152 * area, just update the b_dirtyoff and b_dirtyend, 1153 * otherwise force a write rpc of the old dirty area. 1154 * 1155 * While it is possible to merge discontiguous writes due to 1156 * our having a B_CACHE buffer ( and thus valid read data 1157 * for the hole), we don't because it could lead to 1158 * significant cache coherency problems with multiple clients, 1159 * especially if locking is implemented later on. 1160 * 1161 * as an optimization we could theoretically maintain 1162 * a linked list of discontinuous areas, but we would still 1163 * have to commit them separately so there isn't much 1164 * advantage to it except perhaps a bit of asynchronization. 1165 */ 1166 1167 if (bp->b_dirtyend > 0 && 1168 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1169 if (bwrite(bp) == EINTR) { 1170 error = EINTR; 1171 break; 1172 } 1173 goto again; 1174 } 1175 1176 error = uiomove((char *)bp->b_data + on, n, uio); 1177 1178 /* 1179 * Since this block is being modified, it must be written 1180 * again and not just committed. Since write clustering does 1181 * not work for the stage 1 data write, only the stage 2 1182 * commit rpc, we have to clear B_CLUSTEROK as well. 1183 */ 1184 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1185 1186 if (error) { 1187 bp->b_ioflags |= BIO_ERROR; 1188 brelse(bp); 1189 break; 1190 } 1191 1192 /* 1193 * Only update dirtyoff/dirtyend if not a degenerate 1194 * condition. 1195 */ 1196 if (n) { 1197 if (bp->b_dirtyend > 0) { 1198 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1199 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1200 } else { 1201 bp->b_dirtyoff = on; 1202 bp->b_dirtyend = on + n; 1203 } 1204 vfs_bio_set_validclean(bp, on, n); 1205 } 1206 1207 /* 1208 * If IO_SYNC do bwrite(). 1209 * 1210 * IO_INVAL appears to be unused. The idea appears to be 1211 * to turn off caching in this case. Very odd. XXX 1212 */ 1213 if ((ioflag & IO_SYNC)) { 1214 if (ioflag & IO_INVAL) 1215 bp->b_flags |= B_NOCACHE; 1216 error = bwrite(bp); 1217 if (error) 1218 break; 1219 } else if ((n + on) == biosize) { 1220 bp->b_flags |= B_ASYNC; 1221 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL); 1222 } else { 1223 bdwrite(bp); 1224 } 1225 } while (uio->uio_resid > 0 && n > 0); 1226 1227 return (error); 1228} 1229 1230/* 1231 * Get an nfs cache block. 1232 * 1233 * Allocate a new one if the block isn't currently in the cache 1234 * and return the block marked busy. If the calling process is 1235 * interrupted by a signal for an interruptible mount point, return 1236 * NULL. 1237 * 1238 * The caller must carefully deal with the possible B_INVAL state of 1239 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1240 * indirectly), so synchronous reads can be issued without worrying about 1241 * the B_INVAL state. We have to be a little more careful when dealing 1242 * with writes (see comments in nfs_write()) when extending a file past 1243 * its EOF. 1244 */ 1245static struct buf * 1246nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1247{ 1248 struct buf *bp; 1249 struct mount *mp; 1250 struct nfsmount *nmp; 1251 1252 mp = vp->v_mount; 1253 nmp = VFSTONFS(mp); 1254 1255 if (nmp->nm_flag & NFSMNT_INT) { 1256 sigset_t oldset; 1257 1258 nfs_set_sigmask(td, &oldset); 1259 bp = getblk(vp, bn, size, PCATCH, 0, 0); 1260 nfs_restore_sigmask(td, &oldset); 1261 while (bp == NULL) { 1262 if (nfs_sigintr(nmp, NULL, td)) 1263 return (NULL); 1264 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1265 } 1266 } else { 1267 bp = getblk(vp, bn, size, 0, 0, 0); 1268 } 1269 1270 if (vp->v_type == VREG) { 1271 int biosize; 1272 1273 biosize = mp->mnt_stat.f_iosize; 1274 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1275 } 1276 return (bp); 1277} 1278 1279/* 1280 * Flush and invalidate all dirty buffers. If another process is already 1281 * doing the flush, just wait for completion. 1282 */ 1283int 1284nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1285{ 1286 struct nfsnode *np = VTONFS(vp); 1287 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1288 int error = 0, slpflag, slptimeo; 1289 int old_lock = 0; 1290 1291 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf"); 1292 1293 /* 1294 * XXX This check stops us from needlessly doing a vinvalbuf when 1295 * being called through vclean(). It is not clear that this is 1296 * unsafe. 1297 */ 1298 if (vp->v_iflag & VI_DOOMED) 1299 return (0); 1300 1301 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1302 intrflg = 0; 1303 if (intrflg) { 1304 slpflag = PCATCH; 1305 slptimeo = 2 * hz; 1306 } else { 1307 slpflag = 0; 1308 slptimeo = 0; 1309 } 1310 1311 old_lock = nfs_upgrade_vnlock(vp, td); 1312 /* 1313 * Now, flush as required. 1314 */ 1315 error = vinvalbuf(vp, flags, td, slpflag, 0); 1316 while (error) { 1317 if (intrflg && (error = nfs_sigintr(nmp, NULL, td))) 1318 goto out; 1319 error = vinvalbuf(vp, flags, td, 0, slptimeo); 1320 } 1321 mtx_lock(&np->n_mtx); 1322 if (np->n_directio_asyncwr == 0) 1323 np->n_flag &= ~NMODIFIED; 1324 mtx_unlock(&np->n_mtx); 1325out: 1326 nfs_downgrade_vnlock(vp, td, old_lock); 1327 return error; 1328} 1329 1330/* 1331 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1332 * This is mainly to avoid queueing async I/O requests when the nfsiods 1333 * are all hung on a dead server. 1334 * 1335 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1336 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1337 */ 1338int 1339nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1340{ 1341 int iod; 1342 int gotiod; 1343 int slpflag = 0; 1344 int slptimeo = 0; 1345 int error, error2; 1346 1347 /* 1348 * Commits are usually short and sweet so lets save some cpu and 1349 * leave the async daemons for more important rpc's (such as reads 1350 * and writes). 1351 */ 1352 mtx_lock(&nfs_iod_mtx); 1353 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1354 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1355 mtx_unlock(&nfs_iod_mtx); 1356 return(EIO); 1357 } 1358again: 1359 if (nmp->nm_flag & NFSMNT_INT) 1360 slpflag = PCATCH; 1361 gotiod = FALSE; 1362 1363 /* 1364 * Find a free iod to process this request. 1365 */ 1366 for (iod = 0; iod < nfs_numasync; iod++) 1367 if (nfs_iodwant[iod]) { 1368 gotiod = TRUE; 1369 break; 1370 } 1371 1372 /* 1373 * Try to create one if none are free. 1374 */ 1375 if (!gotiod) { 1376 iod = nfs_nfsiodnew(); 1377 if (iod != -1) 1378 gotiod = TRUE; 1379 } 1380 1381 if (gotiod) { 1382 /* 1383 * Found one, so wake it up and tell it which 1384 * mount to process. 1385 */ 1386 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n", 1387 iod, nmp)); 1388 nfs_iodwant[iod] = NULL; 1389 nfs_iodmount[iod] = nmp; 1390 nmp->nm_bufqiods++; 1391 wakeup(&nfs_iodwant[iod]); 1392 } 1393 1394 /* 1395 * If none are free, we may already have an iod working on this mount 1396 * point. If so, it will process our request. 1397 */ 1398 if (!gotiod) { 1399 if (nmp->nm_bufqiods > 0) { 1400 NFS_DPF(ASYNCIO, 1401 ("nfs_asyncio: %d iods are already processing mount %p\n", 1402 nmp->nm_bufqiods, nmp)); 1403 gotiod = TRUE; 1404 } 1405 } 1406 1407 /* 1408 * If we have an iod which can process the request, then queue 1409 * the buffer. 1410 */ 1411 if (gotiod) { 1412 /* 1413 * Ensure that the queue never grows too large. We still want 1414 * to asynchronize so we block rather then return EIO. 1415 */ 1416 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1417 NFS_DPF(ASYNCIO, 1418 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1419 nmp->nm_bufqwant = TRUE; 1420 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx, 1421 slpflag | PRIBIO, 1422 "nfsaio", slptimeo); 1423 if (error) { 1424 error2 = nfs_sigintr(nmp, NULL, td); 1425 if (error2) { 1426 mtx_unlock(&nfs_iod_mtx); 1427 return (error2); 1428 } 1429 if (slpflag == PCATCH) { 1430 slpflag = 0; 1431 slptimeo = 2 * hz; 1432 } 1433 } 1434 /* 1435 * We might have lost our iod while sleeping, 1436 * so check and loop if nescessary. 1437 */ 1438 if (nmp->nm_bufqiods == 0) { 1439 NFS_DPF(ASYNCIO, 1440 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1441 goto again; 1442 } 1443 } 1444 1445 if (bp->b_iocmd == BIO_READ) { 1446 if (bp->b_rcred == NOCRED && cred != NOCRED) 1447 bp->b_rcred = crhold(cred); 1448 } else { 1449 if (bp->b_wcred == NOCRED && cred != NOCRED) 1450 bp->b_wcred = crhold(cred); 1451 } 1452 1453 if (bp->b_flags & B_REMFREE) 1454 bremfreef(bp); 1455 BUF_KERNPROC(bp); 1456 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1457 nmp->nm_bufqlen++; 1458 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1459 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1460 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1461 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1462 } 1463 mtx_unlock(&nfs_iod_mtx); 1464 return (0); 1465 } 1466 1467 mtx_unlock(&nfs_iod_mtx); 1468 1469 /* 1470 * All the iods are busy on other mounts, so return EIO to 1471 * force the caller to process the i/o synchronously. 1472 */ 1473 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1474 return (EIO); 1475} 1476 1477void 1478nfs_doio_directwrite(struct buf *bp) 1479{ 1480 int iomode, must_commit; 1481 struct uio *uiop = (struct uio *)bp->b_caller1; 1482 char *iov_base = uiop->uio_iov->iov_base; 1483 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount); 1484 1485 iomode = NFSV3WRITE_FILESYNC; 1486 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1487 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit); 1488 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write")); 1489 free(iov_base, M_NFSDIRECTIO); 1490 free(uiop->uio_iov, M_NFSDIRECTIO); 1491 free(uiop, M_NFSDIRECTIO); 1492 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1493 struct nfsnode *np = VTONFS(bp->b_vp); 1494 mtx_lock(&np->n_mtx); 1495 np->n_directio_asyncwr--; 1496 if ((np->n_flag & NFSYNCWAIT) && np->n_directio_asyncwr == 0) { 1497 np->n_flag &= ~NFSYNCWAIT; 1498 wakeup((caddr_t)&np->n_directio_asyncwr); 1499 } 1500 mtx_unlock(&np->n_mtx); 1501 } 1502 vdrop(bp->b_vp); 1503 bp->b_vp = NULL; 1504 relpbuf(bp, &nfs_pbuf_freecnt); 1505} 1506 1507/* 1508 * Do an I/O operation to/from a cache block. This may be called 1509 * synchronously or from an nfsiod. 1510 */ 1511int 1512nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) 1513{ 1514 struct uio *uiop; 1515 struct nfsnode *np; 1516 struct nfsmount *nmp; 1517 int error = 0, iomode, must_commit = 0; 1518 struct uio uio; 1519 struct iovec io; 1520 struct proc *p = td ? td->td_proc : NULL; 1521 uint8_t iocmd; 1522 1523 np = VTONFS(vp); 1524 nmp = VFSTONFS(vp->v_mount); 1525 uiop = &uio; 1526 uiop->uio_iov = &io; 1527 uiop->uio_iovcnt = 1; 1528 uiop->uio_segflg = UIO_SYSSPACE; 1529 uiop->uio_td = td; 1530 1531 /* 1532 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1533 * do this here so we do not have to do it in all the code that 1534 * calls us. 1535 */ 1536 bp->b_flags &= ~B_INVAL; 1537 bp->b_ioflags &= ~BIO_ERROR; 1538 1539 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1540 iocmd = bp->b_iocmd; 1541 if (iocmd == BIO_READ) { 1542 io.iov_len = uiop->uio_resid = bp->b_bcount; 1543 io.iov_base = bp->b_data; 1544 uiop->uio_rw = UIO_READ; 1545 1546 switch (vp->v_type) { 1547 case VREG: 1548 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1549 nfsstats.read_bios++; 1550 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr); 1551 1552 if (!error) { 1553 if (uiop->uio_resid) { 1554 /* 1555 * If we had a short read with no error, we must have 1556 * hit a file hole. We should zero-fill the remainder. 1557 * This can also occur if the server hits the file EOF. 1558 * 1559 * Holes used to be able to occur due to pending 1560 * writes, but that is not possible any longer. 1561 */ 1562 int nread = bp->b_bcount - uiop->uio_resid; 1563 int left = uiop->uio_resid; 1564 1565 if (left > 0) 1566 bzero((char *)bp->b_data + nread, left); 1567 uiop->uio_resid = 0; 1568 } 1569 } 1570 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */ 1571 if (p && (vp->v_vflag & VV_TEXT)) { 1572 mtx_lock(&np->n_mtx); 1573 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) { 1574 mtx_unlock(&np->n_mtx); 1575 PROC_LOCK(p); 1576 killproc(p, "text file modification"); 1577 PROC_UNLOCK(p); 1578 } else 1579 mtx_unlock(&np->n_mtx); 1580 } 1581 break; 1582 case VLNK: 1583 uiop->uio_offset = (off_t)0; 1584 nfsstats.readlink_bios++; 1585 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr); 1586 break; 1587 case VDIR: 1588 nfsstats.readdir_bios++; 1589 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1590 if ((nmp->nm_flag & NFSMNT_NFSV4) != 0) 1591 error = nfs4_readdirrpc(vp, uiop, cr); 1592 else { 1593 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1594 error = nfs_readdirplusrpc(vp, uiop, cr); 1595 if (error == NFSERR_NOTSUPP) 1596 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1597 } 1598 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1599 error = nfs_readdirrpc(vp, uiop, cr); 1600 } 1601 /* 1602 * end-of-directory sets B_INVAL but does not generate an 1603 * error. 1604 */ 1605 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1606 bp->b_flags |= B_INVAL; 1607 break; 1608 default: 1609 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type); 1610 break; 1611 }; 1612 if (error) { 1613 bp->b_ioflags |= BIO_ERROR; 1614 bp->b_error = error; 1615 } 1616 } else { 1617 /* 1618 * If we only need to commit, try to commit 1619 */ 1620 if (bp->b_flags & B_NEEDCOMMIT) { 1621 int retv; 1622 off_t off; 1623 1624 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1625 retv = (nmp->nm_rpcops->nr_commit)( 1626 vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1627 bp->b_wcred, td); 1628 if (retv == 0) { 1629 bp->b_dirtyoff = bp->b_dirtyend = 0; 1630 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1631 bp->b_resid = 0; 1632 bufdone(bp); 1633 return (0); 1634 } 1635 if (retv == NFSERR_STALEWRITEVERF) { 1636 nfs_clearcommit(vp->v_mount); 1637 } 1638 } 1639 1640 /* 1641 * Setup for actual write 1642 */ 1643 mtx_lock(&np->n_mtx); 1644 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1645 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1646 mtx_unlock(&np->n_mtx); 1647 1648 if (bp->b_dirtyend > bp->b_dirtyoff) { 1649 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1650 - bp->b_dirtyoff; 1651 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1652 + bp->b_dirtyoff; 1653 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1654 uiop->uio_rw = UIO_WRITE; 1655 nfsstats.write_bios++; 1656 1657 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1658 iomode = NFSV3WRITE_UNSTABLE; 1659 else 1660 iomode = NFSV3WRITE_FILESYNC; 1661 1662 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit); 1663 1664 /* 1665 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1666 * to cluster the buffers needing commit. This will allow 1667 * the system to submit a single commit rpc for the whole 1668 * cluster. We can do this even if the buffer is not 100% 1669 * dirty (relative to the NFS blocksize), so we optimize the 1670 * append-to-file-case. 1671 * 1672 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1673 * cleared because write clustering only works for commit 1674 * rpc's, not for the data portion of the write). 1675 */ 1676 1677 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1678 bp->b_flags |= B_NEEDCOMMIT; 1679 if (bp->b_dirtyoff == 0 1680 && bp->b_dirtyend == bp->b_bcount) 1681 bp->b_flags |= B_CLUSTEROK; 1682 } else { 1683 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1684 } 1685 1686 /* 1687 * For an interrupted write, the buffer is still valid 1688 * and the write hasn't been pushed to the server yet, 1689 * so we can't set BIO_ERROR and report the interruption 1690 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1691 * is not relevant, so the rpc attempt is essentially 1692 * a noop. For the case of a V3 write rpc not being 1693 * committed to stable storage, the block is still 1694 * dirty and requires either a commit rpc or another 1695 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1696 * the block is reused. This is indicated by setting 1697 * the B_DELWRI and B_NEEDCOMMIT flags. 1698 * 1699 * If the buffer is marked B_PAGING, it does not reside on 1700 * the vp's paging queues so we cannot call bdirty(). The 1701 * bp in this case is not an NFS cache block so we should 1702 * be safe. XXX 1703 */ 1704 if (error == EINTR || error == EIO || error == ETIMEDOUT 1705 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1706 int s; 1707 1708 s = splbio(); 1709 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1710 if ((bp->b_flags & B_PAGING) == 0) { 1711 bdirty(bp); 1712 bp->b_flags &= ~B_DONE; 1713 } 1714 if (error && (bp->b_flags & B_ASYNC) == 0) 1715 bp->b_flags |= B_EINTR; 1716 splx(s); 1717 } else { 1718 if (error) { 1719 bp->b_ioflags |= BIO_ERROR; 1720 bp->b_error = np->n_error = error; 1721 mtx_lock(&np->n_mtx); 1722 np->n_flag |= NWRITEERR; 1723 mtx_unlock(&np->n_mtx); 1724 } 1725 bp->b_dirtyoff = bp->b_dirtyend = 0; 1726 } 1727 } else { 1728 bp->b_resid = 0; 1729 bufdone(bp); 1730 return (0); 1731 } 1732 } 1733 bp->b_resid = uiop->uio_resid; 1734 if (must_commit) 1735 nfs_clearcommit(vp->v_mount); 1736 bufdone(bp); 1737 return (error); 1738} 1739 1740/* 1741 * Used to aid in handling ftruncate() operations on the NFS client side. 1742 * Truncation creates a number of special problems for NFS. We have to 1743 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1744 * we have to properly handle VM pages or (potentially dirty) buffers 1745 * that straddle the truncation point. 1746 */ 1747 1748int 1749nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1750{ 1751 struct nfsnode *np = VTONFS(vp); 1752 u_quad_t tsize; 1753 int biosize = vp->v_mount->mnt_stat.f_iosize; 1754 int error = 0; 1755 1756 mtx_lock(&np->n_mtx); 1757 tsize = np->n_size; 1758 np->n_size = nsize; 1759 mtx_unlock(&np->n_mtx); 1760 1761 if (nsize < tsize) { 1762 struct buf *bp; 1763 daddr_t lbn; 1764 int bufsize; 1765 1766 /* 1767 * vtruncbuf() doesn't get the buffer overlapping the 1768 * truncation point. We may have a B_DELWRI and/or B_CACHE 1769 * buffer that now needs to be truncated. 1770 */ 1771 error = vtruncbuf(vp, cred, td, nsize, biosize); 1772 lbn = nsize / biosize; 1773 bufsize = nsize & (biosize - 1); 1774 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1775 if (!bp) 1776 return EINTR; 1777 if (bp->b_dirtyoff > bp->b_bcount) 1778 bp->b_dirtyoff = bp->b_bcount; 1779 if (bp->b_dirtyend > bp->b_bcount) 1780 bp->b_dirtyend = bp->b_bcount; 1781 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1782 brelse(bp); 1783 } else { 1784 vnode_pager_setsize(vp, nsize); 1785 } 1786 return(error); 1787} 1788 1789