nfs_bio.c revision 214026
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 214026 2010-10-18 19:06:46Z kib $"); 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 <nfsclient/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(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_AGAIN; 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 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 348 for (i = 0; i < nwritten; i++) { 349 rtvals[i] = VM_PAGER_OK; 350 vm_page_undirty(pages[i]); 351 } 352 if (must_commit) { 353 nfs_clearcommit(vp->v_mount); 354 } 355 } 356 return rtvals[0]; 357} 358 359/* 360 * For nfs, cache consistency can only be maintained approximately. 361 * Although RFC1094 does not specify the criteria, the following is 362 * believed to be compatible with the reference port. 363 * For nfs: 364 * If the file's modify time on the server has changed since the 365 * last read rpc or you have written to the file, 366 * you may have lost data cache consistency with the 367 * server, so flush all of the file's data out of the cache. 368 * Then force a getattr rpc to ensure that you have up to date 369 * attributes. 370 * NB: This implies that cache data can be read when up to 371 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 372 * attributes this could be forced by setting n_attrstamp to 0 before 373 * the VOP_GETATTR() call. 374 */ 375static inline int 376nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 377{ 378 int error = 0; 379 struct vattr vattr; 380 struct nfsnode *np = VTONFS(vp); 381 int old_lock; 382 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 383 384 /* 385 * Grab the exclusive lock before checking whether the cache is 386 * consistent. 387 * XXX - We can make this cheaper later (by acquiring cheaper locks). 388 * But for now, this suffices. 389 */ 390 old_lock = nfs_upgrade_vnlock(vp); 391 if (vp->v_iflag & VI_DOOMED) { 392 nfs_downgrade_vnlock(vp, old_lock); 393 return (EBADF); 394 } 395 396 mtx_lock(&np->n_mtx); 397 if (np->n_flag & NMODIFIED) { 398 mtx_unlock(&np->n_mtx); 399 if (vp->v_type != VREG) { 400 if (vp->v_type != VDIR) 401 panic("nfs: bioread, not dir"); 402 (nmp->nm_rpcops->nr_invaldir)(vp); 403 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 404 if (error) 405 goto out; 406 } 407 np->n_attrstamp = 0; 408 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 409 error = VOP_GETATTR(vp, &vattr, cred); 410 if (error) 411 goto out; 412 mtx_lock(&np->n_mtx); 413 np->n_mtime = vattr.va_mtime; 414 mtx_unlock(&np->n_mtx); 415 } else { 416 mtx_unlock(&np->n_mtx); 417 error = VOP_GETATTR(vp, &vattr, cred); 418 if (error) 419 return (error); 420 mtx_lock(&np->n_mtx); 421 if ((np->n_flag & NSIZECHANGED) 422 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 423 mtx_unlock(&np->n_mtx); 424 if (vp->v_type == VDIR) 425 (nmp->nm_rpcops->nr_invaldir)(vp); 426 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 427 if (error) 428 goto out; 429 mtx_lock(&np->n_mtx); 430 np->n_mtime = vattr.va_mtime; 431 np->n_flag &= ~NSIZECHANGED; 432 } 433 mtx_unlock(&np->n_mtx); 434 } 435out: 436 nfs_downgrade_vnlock(vp, old_lock); 437 return error; 438} 439 440/* 441 * Vnode op for read using bio 442 */ 443int 444nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 445{ 446 struct nfsnode *np = VTONFS(vp); 447 int biosize, i; 448 struct buf *bp, *rabp; 449 struct thread *td; 450 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 451 daddr_t lbn, rabn; 452 int bcount; 453 int seqcount; 454 int nra, error = 0, n = 0, on = 0; 455 456 KASSERT(uio->uio_rw == UIO_READ, ("nfs_read mode")); 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, 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, 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, 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, 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, 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 error = nfs_asyncio(nmp, bp, NOCRED, td); 830 if (error) { 831 free(t_iov->iov_base, M_NFSDIRECTIO); 832 free(t_iov, M_NFSDIRECTIO); 833 free(t_uio, M_NFSDIRECTIO); 834 bp->b_vp = NULL; 835 relpbuf(bp, &nfs_pbuf_freecnt); 836 if (error == EINTR) 837 return (error); 838 goto do_sync; 839 } 840 uiop->uio_offset += size; 841 uiop->uio_resid -= size; 842 if (uiop->uio_iov->iov_len <= size) { 843 uiop->uio_iovcnt--; 844 uiop->uio_iov++; 845 } else { 846 uiop->uio_iov->iov_base = 847 (char *)uiop->uio_iov->iov_base + size; 848 uiop->uio_iov->iov_len -= size; 849 } 850 } 851 } 852 return (0); 853} 854 855/* 856 * Vnode op for write using bio 857 */ 858int 859nfs_write(struct vop_write_args *ap) 860{ 861 int biosize; 862 struct uio *uio = ap->a_uio; 863 struct thread *td = uio->uio_td; 864 struct vnode *vp = ap->a_vp; 865 struct nfsnode *np = VTONFS(vp); 866 struct ucred *cred = ap->a_cred; 867 int ioflag = ap->a_ioflag; 868 struct buf *bp; 869 struct vattr vattr; 870 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 871 daddr_t lbn; 872 int bcount; 873 int n, on, error = 0; 874 875 KASSERT(uio->uio_rw == UIO_WRITE, ("nfs_write mode")); 876 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, 877 ("nfs_write proc")); 878 if (vp->v_type != VREG) 879 return (EIO); 880 mtx_lock(&np->n_mtx); 881 if (np->n_flag & NWRITEERR) { 882 np->n_flag &= ~NWRITEERR; 883 mtx_unlock(&np->n_mtx); 884 return (np->n_error); 885 } else 886 mtx_unlock(&np->n_mtx); 887 mtx_lock(&nmp->nm_mtx); 888 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 889 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 890 mtx_unlock(&nmp->nm_mtx); 891 (void)nfs_fsinfo(nmp, vp, cred, td); 892 } else 893 mtx_unlock(&nmp->nm_mtx); 894 895 /* 896 * Synchronously flush pending buffers if we are in synchronous 897 * mode or if we are appending. 898 */ 899 if (ioflag & (IO_APPEND | IO_SYNC)) { 900 mtx_lock(&np->n_mtx); 901 if (np->n_flag & NMODIFIED) { 902 mtx_unlock(&np->n_mtx); 903#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 904 /* 905 * Require non-blocking, synchronous writes to 906 * dirty files to inform the program it needs 907 * to fsync(2) explicitly. 908 */ 909 if (ioflag & IO_NDELAY) 910 return (EAGAIN); 911#endif 912flush_and_restart: 913 np->n_attrstamp = 0; 914 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 915 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 916 if (error) 917 return (error); 918 } else 919 mtx_unlock(&np->n_mtx); 920 } 921 922 /* 923 * If IO_APPEND then load uio_offset. We restart here if we cannot 924 * get the append lock. 925 */ 926 if (ioflag & IO_APPEND) { 927 np->n_attrstamp = 0; 928 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 929 error = VOP_GETATTR(vp, &vattr, cred); 930 if (error) 931 return (error); 932 mtx_lock(&np->n_mtx); 933 uio->uio_offset = np->n_size; 934 mtx_unlock(&np->n_mtx); 935 } 936 937 if (uio->uio_offset < 0) 938 return (EINVAL); 939 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 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_mount->mnt_stat.f_iosize; 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 int biosize; 1260 1261 biosize = mp->mnt_stat.f_iosize; 1262 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1263 } 1264 return (bp); 1265} 1266 1267/* 1268 * Flush and invalidate all dirty buffers. If another process is already 1269 * doing the flush, just wait for completion. 1270 */ 1271int 1272nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1273{ 1274 struct nfsnode *np = VTONFS(vp); 1275 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1276 int error = 0, slpflag, slptimeo; 1277 int old_lock = 0; 1278 1279 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf"); 1280 1281 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1282 intrflg = 0; 1283 if (intrflg) { 1284 slpflag = NFS_PCATCH; 1285 slptimeo = 2 * hz; 1286 } else { 1287 slpflag = 0; 1288 slptimeo = 0; 1289 } 1290 1291 old_lock = nfs_upgrade_vnlock(vp); 1292 if (vp->v_iflag & VI_DOOMED) { 1293 /* 1294 * Since vgonel() uses the generic vinvalbuf() to flush 1295 * dirty buffers and it does not call this function, it 1296 * is safe to just return OK when VI_DOOMED is set. 1297 */ 1298 nfs_downgrade_vnlock(vp, old_lock); 1299 return (0); 1300 } 1301 1302 /* 1303 * Now, flush as required. 1304 */ 1305 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1306 VM_OBJECT_LOCK(vp->v_bufobj.bo_object); 1307 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1308 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); 1309 /* 1310 * If the page clean was interrupted, fail the invalidation. 1311 * Not doing so, we run the risk of losing dirty pages in the 1312 * vinvalbuf() call below. 1313 */ 1314 if (intrflg && (error = nfs_sigintr(nmp, td))) 1315 goto out; 1316 } 1317 1318 error = vinvalbuf(vp, flags, slpflag, 0); 1319 while (error) { 1320 if (intrflg && (error = nfs_sigintr(nmp, td))) 1321 goto out; 1322 error = vinvalbuf(vp, flags, 0, slptimeo); 1323 } 1324 mtx_lock(&np->n_mtx); 1325 if (np->n_directio_asyncwr == 0) 1326 np->n_flag &= ~NMODIFIED; 1327 mtx_unlock(&np->n_mtx); 1328out: 1329 nfs_downgrade_vnlock(vp, old_lock); 1330 return error; 1331} 1332 1333/* 1334 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1335 * This is mainly to avoid queueing async I/O requests when the nfsiods 1336 * are all hung on a dead server. 1337 * 1338 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1339 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1340 */ 1341int 1342nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1343{ 1344 int iod; 1345 int gotiod; 1346 int slpflag = 0; 1347 int slptimeo = 0; 1348 int error, error2; 1349 1350 /* 1351 * Commits are usually short and sweet so lets save some cpu and 1352 * leave the async daemons for more important rpc's (such as reads 1353 * and writes). 1354 */ 1355 mtx_lock(&nfs_iod_mtx); 1356 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1357 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1358 mtx_unlock(&nfs_iod_mtx); 1359 return(EIO); 1360 } 1361again: 1362 if (nmp->nm_flag & NFSMNT_INT) 1363 slpflag = NFS_PCATCH; 1364 gotiod = FALSE; 1365 1366 /* 1367 * Find a free iod to process this request. 1368 */ 1369 for (iod = 0; iod < nfs_numasync; iod++) 1370 if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) { 1371 gotiod = TRUE; 1372 break; 1373 } 1374 1375 /* 1376 * Try to create one if none are free. 1377 */ 1378 if (!gotiod) 1379 nfs_nfsiodnew(); 1380 else { 1381 /* 1382 * Found one, so wake it up and tell it which 1383 * mount to process. 1384 */ 1385 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n", 1386 iod, nmp)); 1387 nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE; 1388 nfs_iodmount[iod] = nmp; 1389 nmp->nm_bufqiods++; 1390 wakeup(&nfs_iodwant[iod]); 1391 } 1392 1393 /* 1394 * If none are free, we may already have an iod working on this mount 1395 * point. If so, it will process our request. 1396 */ 1397 if (!gotiod) { 1398 if (nmp->nm_bufqiods > 0) { 1399 NFS_DPF(ASYNCIO, 1400 ("nfs_asyncio: %d iods are already processing mount %p\n", 1401 nmp->nm_bufqiods, nmp)); 1402 gotiod = TRUE; 1403 } 1404 } 1405 1406 /* 1407 * If we have an iod which can process the request, then queue 1408 * the buffer. 1409 */ 1410 if (gotiod) { 1411 /* 1412 * Ensure that the queue never grows too large. We still want 1413 * to asynchronize so we block rather then return EIO. 1414 */ 1415 while (nmp->nm_bufqlen >= 2 * nfs_numasync) { 1416 NFS_DPF(ASYNCIO, 1417 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1418 nmp->nm_bufqwant = TRUE; 1419 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx, 1420 slpflag | PRIBIO, 1421 "nfsaio", slptimeo); 1422 if (error) { 1423 error2 = nfs_sigintr(nmp, td); 1424 if (error2) { 1425 mtx_unlock(&nfs_iod_mtx); 1426 return (error2); 1427 } 1428 if (slpflag == NFS_PCATCH) { 1429 slpflag = 0; 1430 slptimeo = 2 * hz; 1431 } 1432 } 1433 /* 1434 * We might have lost our iod while sleeping, 1435 * so check and loop if nescessary. 1436 */ 1437 goto again; 1438 } 1439 1440 /* We might have lost our nfsiod */ 1441 if (nmp->nm_bufqiods == 0) { 1442 NFS_DPF(ASYNCIO, 1443("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1444 goto again; 1445 } 1446 1447 if (bp->b_iocmd == BIO_READ) { 1448 if (bp->b_rcred == NOCRED && cred != NOCRED) 1449 bp->b_rcred = crhold(cred); 1450 } else { 1451 if (bp->b_wcred == NOCRED && cred != NOCRED) 1452 bp->b_wcred = crhold(cred); 1453 } 1454 1455 if (bp->b_flags & B_REMFREE) 1456 bremfreef(bp); 1457 BUF_KERNPROC(bp); 1458 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1459 nmp->nm_bufqlen++; 1460 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1461 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1462 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1463 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1464 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1465 } 1466 mtx_unlock(&nfs_iod_mtx); 1467 return (0); 1468 } 1469 1470 mtx_unlock(&nfs_iod_mtx); 1471 1472 /* 1473 * All the iods are busy on other mounts, so return EIO to 1474 * force the caller to process the i/o synchronously. 1475 */ 1476 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1477 return (EIO); 1478} 1479 1480void 1481nfs_doio_directwrite(struct buf *bp) 1482{ 1483 int iomode, must_commit; 1484 struct uio *uiop = (struct uio *)bp->b_caller1; 1485 char *iov_base = uiop->uio_iov->iov_base; 1486 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount); 1487 1488 iomode = NFSV3WRITE_FILESYNC; 1489 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1490 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit); 1491 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write")); 1492 free(iov_base, M_NFSDIRECTIO); 1493 free(uiop->uio_iov, M_NFSDIRECTIO); 1494 free(uiop, M_NFSDIRECTIO); 1495 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1496 struct nfsnode *np = VTONFS(bp->b_vp); 1497 mtx_lock(&np->n_mtx); 1498 np->n_directio_asyncwr--; 1499 if (np->n_directio_asyncwr == 0) { 1500 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED; 1501 if ((np->n_flag & NFSYNCWAIT)) { 1502 np->n_flag &= ~NFSYNCWAIT; 1503 wakeup((caddr_t)&np->n_directio_asyncwr); 1504 } 1505 } 1506 mtx_unlock(&np->n_mtx); 1507 } 1508 bp->b_vp = NULL; 1509 relpbuf(bp, &nfs_pbuf_freecnt); 1510} 1511 1512/* 1513 * Do an I/O operation to/from a cache block. This may be called 1514 * synchronously or from an nfsiod. 1515 */ 1516int 1517nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) 1518{ 1519 struct uio *uiop; 1520 struct nfsnode *np; 1521 struct nfsmount *nmp; 1522 int error = 0, iomode, must_commit = 0; 1523 struct uio uio; 1524 struct iovec io; 1525 struct proc *p = td ? td->td_proc : NULL; 1526 uint8_t iocmd; 1527 1528 np = VTONFS(vp); 1529 nmp = VFSTONFS(vp->v_mount); 1530 uiop = &uio; 1531 uiop->uio_iov = &io; 1532 uiop->uio_iovcnt = 1; 1533 uiop->uio_segflg = UIO_SYSSPACE; 1534 uiop->uio_td = td; 1535 1536 /* 1537 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1538 * do this here so we do not have to do it in all the code that 1539 * calls us. 1540 */ 1541 bp->b_flags &= ~B_INVAL; 1542 bp->b_ioflags &= ~BIO_ERROR; 1543 1544 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1545 iocmd = bp->b_iocmd; 1546 if (iocmd == BIO_READ) { 1547 io.iov_len = uiop->uio_resid = bp->b_bcount; 1548 io.iov_base = bp->b_data; 1549 uiop->uio_rw = UIO_READ; 1550 1551 switch (vp->v_type) { 1552 case VREG: 1553 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1554 nfsstats.read_bios++; 1555 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr); 1556 1557 if (!error) { 1558 if (uiop->uio_resid) { 1559 /* 1560 * If we had a short read with no error, we must have 1561 * hit a file hole. We should zero-fill the remainder. 1562 * This can also occur if the server hits the file EOF. 1563 * 1564 * Holes used to be able to occur due to pending 1565 * writes, but that is not possible any longer. 1566 */ 1567 int nread = bp->b_bcount - uiop->uio_resid; 1568 int left = uiop->uio_resid; 1569 1570 if (left > 0) 1571 bzero((char *)bp->b_data + nread, left); 1572 uiop->uio_resid = 0; 1573 } 1574 } 1575 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */ 1576 if (p && (vp->v_vflag & VV_TEXT)) { 1577 mtx_lock(&np->n_mtx); 1578 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) { 1579 mtx_unlock(&np->n_mtx); 1580 PROC_LOCK(p); 1581 killproc(p, "text file modification"); 1582 PROC_UNLOCK(p); 1583 } else 1584 mtx_unlock(&np->n_mtx); 1585 } 1586 break; 1587 case VLNK: 1588 uiop->uio_offset = (off_t)0; 1589 nfsstats.readlink_bios++; 1590 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr); 1591 break; 1592 case VDIR: 1593 nfsstats.readdir_bios++; 1594 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1595 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1596 error = nfs_readdirplusrpc(vp, uiop, cr); 1597 if (error == NFSERR_NOTSUPP) 1598 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1599 } 1600 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1601 error = nfs_readdirrpc(vp, uiop, cr); 1602 /* 1603 * end-of-directory sets B_INVAL but does not generate an 1604 * error. 1605 */ 1606 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1607 bp->b_flags |= B_INVAL; 1608 break; 1609 default: 1610 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type); 1611 break; 1612 }; 1613 if (error) { 1614 bp->b_ioflags |= BIO_ERROR; 1615 bp->b_error = error; 1616 } 1617 } else { 1618 /* 1619 * If we only need to commit, try to commit 1620 */ 1621 if (bp->b_flags & B_NEEDCOMMIT) { 1622 int retv; 1623 off_t off; 1624 1625 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1626 retv = (nmp->nm_rpcops->nr_commit)( 1627 vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1628 bp->b_wcred, td); 1629 if (retv == 0) { 1630 bp->b_dirtyoff = bp->b_dirtyend = 0; 1631 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1632 bp->b_resid = 0; 1633 bufdone(bp); 1634 return (0); 1635 } 1636 if (retv == NFSERR_STALEWRITEVERF) { 1637 nfs_clearcommit(vp->v_mount); 1638 } 1639 } 1640 1641 /* 1642 * Setup for actual write 1643 */ 1644 mtx_lock(&np->n_mtx); 1645 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1646 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1647 mtx_unlock(&np->n_mtx); 1648 1649 if (bp->b_dirtyend > bp->b_dirtyoff) { 1650 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1651 - bp->b_dirtyoff; 1652 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1653 + bp->b_dirtyoff; 1654 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1655 uiop->uio_rw = UIO_WRITE; 1656 nfsstats.write_bios++; 1657 1658 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1659 iomode = NFSV3WRITE_UNSTABLE; 1660 else 1661 iomode = NFSV3WRITE_FILESYNC; 1662 1663 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit); 1664 1665 /* 1666 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1667 * to cluster the buffers needing commit. This will allow 1668 * the system to submit a single commit rpc for the whole 1669 * cluster. We can do this even if the buffer is not 100% 1670 * dirty (relative to the NFS blocksize), so we optimize the 1671 * append-to-file-case. 1672 * 1673 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1674 * cleared because write clustering only works for commit 1675 * rpc's, not for the data portion of the write). 1676 */ 1677 1678 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1679 bp->b_flags |= B_NEEDCOMMIT; 1680 if (bp->b_dirtyoff == 0 1681 && bp->b_dirtyend == bp->b_bcount) 1682 bp->b_flags |= B_CLUSTEROK; 1683 } else { 1684 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1685 } 1686 1687 /* 1688 * For an interrupted write, the buffer is still valid 1689 * and the write hasn't been pushed to the server yet, 1690 * so we can't set BIO_ERROR and report the interruption 1691 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1692 * is not relevant, so the rpc attempt is essentially 1693 * a noop. For the case of a V3 write rpc not being 1694 * committed to stable storage, the block is still 1695 * dirty and requires either a commit rpc or another 1696 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1697 * the block is reused. This is indicated by setting 1698 * the B_DELWRI and B_NEEDCOMMIT flags. 1699 * 1700 * If the buffer is marked B_PAGING, it does not reside on 1701 * the vp's paging queues so we cannot call bdirty(). The 1702 * bp in this case is not an NFS cache block so we should 1703 * be safe. XXX 1704 * 1705 * The logic below breaks up errors into recoverable and 1706 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1707 * and keep the buffer around for potential write retries. 1708 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1709 * and save the error in the nfsnode. This is less than ideal 1710 * but necessary. Keeping such buffers around could potentially 1711 * cause buffer exhaustion eventually (they can never be written 1712 * out, so will get constantly be re-dirtied). It also causes 1713 * all sorts of vfs panics. For non-recoverable write errors, 1714 * also invalidate the attrcache, so we'll be forced to go over 1715 * the wire for this object, returning an error to user on next 1716 * call (most of the time). 1717 */ 1718 if (error == EINTR || error == EIO || error == ETIMEDOUT 1719 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1720 int s; 1721 1722 s = splbio(); 1723 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1724 if ((bp->b_flags & B_PAGING) == 0) { 1725 bdirty(bp); 1726 bp->b_flags &= ~B_DONE; 1727 } 1728 if (error && (bp->b_flags & B_ASYNC) == 0) 1729 bp->b_flags |= B_EINTR; 1730 splx(s); 1731 } else { 1732 if (error) { 1733 bp->b_ioflags |= BIO_ERROR; 1734 bp->b_flags |= B_INVAL; 1735 bp->b_error = np->n_error = error; 1736 mtx_lock(&np->n_mtx); 1737 np->n_flag |= NWRITEERR; 1738 np->n_attrstamp = 0; 1739 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1740 mtx_unlock(&np->n_mtx); 1741 } 1742 bp->b_dirtyoff = bp->b_dirtyend = 0; 1743 } 1744 } else { 1745 bp->b_resid = 0; 1746 bufdone(bp); 1747 return (0); 1748 } 1749 } 1750 bp->b_resid = uiop->uio_resid; 1751 if (must_commit) 1752 nfs_clearcommit(vp->v_mount); 1753 bufdone(bp); 1754 return (error); 1755} 1756 1757/* 1758 * Used to aid in handling ftruncate() operations on the NFS client side. 1759 * Truncation creates a number of special problems for NFS. We have to 1760 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1761 * we have to properly handle VM pages or (potentially dirty) buffers 1762 * that straddle the truncation point. 1763 */ 1764 1765int 1766nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1767{ 1768 struct nfsnode *np = VTONFS(vp); 1769 u_quad_t tsize; 1770 int biosize = vp->v_mount->mnt_stat.f_iosize; 1771 int error = 0; 1772 1773 mtx_lock(&np->n_mtx); 1774 tsize = np->n_size; 1775 np->n_size = nsize; 1776 mtx_unlock(&np->n_mtx); 1777 1778 if (nsize < tsize) { 1779 struct buf *bp; 1780 daddr_t lbn; 1781 int bufsize; 1782 1783 /* 1784 * vtruncbuf() doesn't get the buffer overlapping the 1785 * truncation point. We may have a B_DELWRI and/or B_CACHE 1786 * buffer that now needs to be truncated. 1787 */ 1788 error = vtruncbuf(vp, cred, td, nsize, biosize); 1789 lbn = nsize / biosize; 1790 bufsize = nsize & (biosize - 1); 1791 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1792 if (!bp) 1793 return EINTR; 1794 if (bp->b_dirtyoff > bp->b_bcount) 1795 bp->b_dirtyoff = bp->b_bcount; 1796 if (bp->b_dirtyend > bp->b_bcount) 1797 bp->b_dirtyend = bp->b_bcount; 1798 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1799 brelse(bp); 1800 } else { 1801 vnode_pager_setsize(vp, nsize); 1802 } 1803 return(error); 1804} 1805 1806