nfs_bio.c revision 207728
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 207728 2010-05-06 18:58:32Z alc $"); 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 vm_page_lock(m); 198 vm_page_lock_queues(); 199 200 if (nextoff <= size) { 201 /* 202 * Read operation filled an entire page 203 */ 204 m->valid = VM_PAGE_BITS_ALL; 205 KASSERT(m->dirty == 0, 206 ("nfs_getpages: page %p is dirty", m)); 207 } else if (size > toff) { 208 /* 209 * Read operation filled a partial page. 210 */ 211 m->valid = 0; 212 vm_page_set_valid(m, 0, size - toff); 213 KASSERT(m->dirty == 0, 214 ("nfs_getpages: page %p is dirty", m)); 215 } else { 216 /* 217 * Read operation was short. If no error occured 218 * we may have hit a zero-fill section. We simply 219 * leave valid set to 0. 220 */ 221 ; 222 } 223 if (i != ap->a_reqpage) { 224 /* 225 * Whether or not to leave the page activated is up in 226 * the air, but we should put the page on a page queue 227 * somewhere (it already is in the object). Result: 228 * It appears that emperical results show that 229 * deactivating pages is best. 230 */ 231 232 /* 233 * Just in case someone was asking for this page we 234 * now tell them that it is ok to use. 235 */ 236 if (!error) { 237 if (m->oflags & VPO_WANTED) 238 vm_page_activate(m); 239 else 240 vm_page_deactivate(m); 241 vm_page_wakeup(m); 242 } else { 243 vm_page_free(m); 244 } 245 } 246 247 vm_page_unlock_queues(); 248 vm_page_unlock(m); 249 } 250 VM_OBJECT_UNLOCK(object); 251 return (0); 252} 253 254/* 255 * Vnode op for VM putpages. 256 */ 257int 258nfs_putpages(struct vop_putpages_args *ap) 259{ 260 struct uio uio; 261 struct iovec iov; 262 vm_offset_t kva; 263 struct buf *bp; 264 int iomode, must_commit, i, error, npages, count; 265 off_t offset; 266 int *rtvals; 267 struct vnode *vp; 268 struct thread *td; 269 struct ucred *cred; 270 struct nfsmount *nmp; 271 struct nfsnode *np; 272 vm_page_t *pages; 273 274 vp = ap->a_vp; 275 np = VTONFS(vp); 276 td = curthread; /* XXX */ 277 cred = curthread->td_ucred; /* XXX */ 278 nmp = VFSTONFS(vp->v_mount); 279 pages = ap->a_m; 280 count = ap->a_count; 281 rtvals = ap->a_rtvals; 282 npages = btoc(count); 283 offset = IDX_TO_OFF(pages[0]->pindex); 284 285 mtx_lock(&nmp->nm_mtx); 286 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 287 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 288 mtx_unlock(&nmp->nm_mtx); 289 (void)nfs_fsinfo(nmp, vp, cred, td); 290 } else 291 mtx_unlock(&nmp->nm_mtx); 292 293 mtx_lock(&np->n_mtx); 294 if (nfs_directio_enable && !nfs_directio_allow_mmap && 295 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 296 mtx_unlock(&np->n_mtx); 297 nfs_printf("nfs_putpages: called on noncache-able vnode??\n"); 298 mtx_lock(&np->n_mtx); 299 } 300 301 for (i = 0; i < npages; i++) 302 rtvals[i] = VM_PAGER_AGAIN; 303 304 /* 305 * When putting pages, do not extend file past EOF. 306 */ 307 if (offset + count > np->n_size) { 308 count = np->n_size - offset; 309 if (count < 0) 310 count = 0; 311 } 312 mtx_unlock(&np->n_mtx); 313 314 /* 315 * We use only the kva address for the buffer, but this is extremely 316 * convienient and fast. 317 */ 318 bp = getpbuf(&nfs_pbuf_freecnt); 319 320 kva = (vm_offset_t) bp->b_data; 321 pmap_qenter(kva, pages, npages); 322 PCPU_INC(cnt.v_vnodeout); 323 PCPU_ADD(cnt.v_vnodepgsout, count); 324 325 iov.iov_base = (caddr_t) kva; 326 iov.iov_len = count; 327 uio.uio_iov = &iov; 328 uio.uio_iovcnt = 1; 329 uio.uio_offset = offset; 330 uio.uio_resid = count; 331 uio.uio_segflg = UIO_SYSSPACE; 332 uio.uio_rw = UIO_WRITE; 333 uio.uio_td = td; 334 335 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 336 iomode = NFSV3WRITE_UNSTABLE; 337 else 338 iomode = NFSV3WRITE_FILESYNC; 339 340 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit); 341 342 pmap_qremove(kva, npages); 343 relpbuf(bp, &nfs_pbuf_freecnt); 344 345 if (!error) { 346 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 347 for (i = 0; i < nwritten; i++) { 348 rtvals[i] = VM_PAGER_OK; 349 vm_page_undirty(pages[i]); 350 } 351 if (must_commit) { 352 nfs_clearcommit(vp->v_mount); 353 } 354 } 355 return rtvals[0]; 356} 357 358/* 359 * For nfs, cache consistency can only be maintained approximately. 360 * Although RFC1094 does not specify the criteria, the following is 361 * believed to be compatible with the reference port. 362 * For nfs: 363 * If the file's modify time on the server has changed since the 364 * last read rpc or you have written to the file, 365 * you may have lost data cache consistency with the 366 * server, so flush all of the file's data out of the cache. 367 * Then force a getattr rpc to ensure that you have up to date 368 * attributes. 369 * NB: This implies that cache data can be read when up to 370 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 371 * attributes this could be forced by setting n_attrstamp to 0 before 372 * the VOP_GETATTR() call. 373 */ 374static inline int 375nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 376{ 377 int error = 0; 378 struct vattr vattr; 379 struct nfsnode *np = VTONFS(vp); 380 int old_lock; 381 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 382 383 /* 384 * Grab the exclusive lock before checking whether the cache is 385 * consistent. 386 * XXX - We can make this cheaper later (by acquiring cheaper locks). 387 * But for now, this suffices. 388 */ 389 old_lock = nfs_upgrade_vnlock(vp); 390 if (vp->v_iflag & VI_DOOMED) { 391 nfs_downgrade_vnlock(vp, old_lock); 392 return (EBADF); 393 } 394 395 mtx_lock(&np->n_mtx); 396 if (np->n_flag & NMODIFIED) { 397 mtx_unlock(&np->n_mtx); 398 if (vp->v_type != VREG) { 399 if (vp->v_type != VDIR) 400 panic("nfs: bioread, not dir"); 401 (nmp->nm_rpcops->nr_invaldir)(vp); 402 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 403 if (error) 404 goto out; 405 } 406 np->n_attrstamp = 0; 407 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 408 error = VOP_GETATTR(vp, &vattr, cred); 409 if (error) 410 goto out; 411 mtx_lock(&np->n_mtx); 412 np->n_mtime = vattr.va_mtime; 413 mtx_unlock(&np->n_mtx); 414 } else { 415 mtx_unlock(&np->n_mtx); 416 error = VOP_GETATTR(vp, &vattr, cred); 417 if (error) 418 return (error); 419 mtx_lock(&np->n_mtx); 420 if ((np->n_flag & NSIZECHANGED) 421 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 422 mtx_unlock(&np->n_mtx); 423 if (vp->v_type == VDIR) 424 (nmp->nm_rpcops->nr_invaldir)(vp); 425 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 426 if (error) 427 goto out; 428 mtx_lock(&np->n_mtx); 429 np->n_mtime = vattr.va_mtime; 430 np->n_flag &= ~NSIZECHANGED; 431 } 432 mtx_unlock(&np->n_mtx); 433 } 434out: 435 nfs_downgrade_vnlock(vp, old_lock); 436 return error; 437} 438 439/* 440 * Vnode op for read using bio 441 */ 442int 443nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 444{ 445 struct nfsnode *np = VTONFS(vp); 446 int biosize, i; 447 struct buf *bp, *rabp; 448 struct thread *td; 449 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 450 daddr_t lbn, rabn; 451 int bcount; 452 int seqcount; 453 int nra, error = 0, n = 0, on = 0; 454 455#ifdef DIAGNOSTIC 456 if (uio->uio_rw != UIO_READ) 457 panic("nfs_read mode"); 458#endif 459 if (uio->uio_resid == 0) 460 return (0); 461 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 462 return (EINVAL); 463 td = uio->uio_td; 464 465 mtx_lock(&nmp->nm_mtx); 466 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 467 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 468 mtx_unlock(&nmp->nm_mtx); 469 (void)nfs_fsinfo(nmp, vp, cred, td); 470 } else 471 mtx_unlock(&nmp->nm_mtx); 472 473 if (vp->v_type != VDIR && 474 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 475 return (EFBIG); 476 477 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 478 /* No caching/ no readaheads. Just read data into the user buffer */ 479 return nfs_readrpc(vp, uio, cred); 480 481 biosize = vp->v_mount->mnt_stat.f_iosize; 482 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 483 484 error = nfs_bioread_check_cons(vp, td, cred); 485 if (error) 486 return error; 487 488 do { 489 u_quad_t nsize; 490 491 mtx_lock(&np->n_mtx); 492 nsize = np->n_size; 493 mtx_unlock(&np->n_mtx); 494 495 switch (vp->v_type) { 496 case VREG: 497 nfsstats.biocache_reads++; 498 lbn = uio->uio_offset / biosize; 499 on = uio->uio_offset & (biosize - 1); 500 501 /* 502 * Start the read ahead(s), as required. 503 */ 504 if (nmp->nm_readahead > 0) { 505 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 506 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 507 rabn = lbn + 1 + nra; 508 if (incore(&vp->v_bufobj, rabn) == NULL) { 509 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 510 if (!rabp) { 511 error = nfs_sigintr(nmp, td); 512 return (error ? error : EINTR); 513 } 514 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 515 rabp->b_flags |= B_ASYNC; 516 rabp->b_iocmd = BIO_READ; 517 vfs_busy_pages(rabp, 0); 518 if (nfs_asyncio(nmp, rabp, cred, td)) { 519 rabp->b_flags |= B_INVAL; 520 rabp->b_ioflags |= BIO_ERROR; 521 vfs_unbusy_pages(rabp); 522 brelse(rabp); 523 break; 524 } 525 } else { 526 brelse(rabp); 527 } 528 } 529 } 530 } 531 532 /* Note that bcount is *not* DEV_BSIZE aligned. */ 533 bcount = biosize; 534 if ((off_t)lbn * biosize >= nsize) { 535 bcount = 0; 536 } else if ((off_t)(lbn + 1) * biosize > nsize) { 537 bcount = nsize - (off_t)lbn * biosize; 538 } 539 bp = nfs_getcacheblk(vp, lbn, bcount, td); 540 541 if (!bp) { 542 error = nfs_sigintr(nmp, td); 543 return (error ? error : EINTR); 544 } 545 546 /* 547 * If B_CACHE is not set, we must issue the read. If this 548 * fails, we return an error. 549 */ 550 551 if ((bp->b_flags & B_CACHE) == 0) { 552 bp->b_iocmd = BIO_READ; 553 vfs_busy_pages(bp, 0); 554 error = nfs_doio(vp, bp, cred, td); 555 if (error) { 556 brelse(bp); 557 return (error); 558 } 559 } 560 561 /* 562 * on is the offset into the current bp. Figure out how many 563 * bytes we can copy out of the bp. Note that bcount is 564 * NOT DEV_BSIZE aligned. 565 * 566 * Then figure out how many bytes we can copy into the uio. 567 */ 568 569 n = 0; 570 if (on < bcount) 571 n = min((unsigned)(bcount - on), uio->uio_resid); 572 break; 573 case VLNK: 574 nfsstats.biocache_readlinks++; 575 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 576 if (!bp) { 577 error = nfs_sigintr(nmp, td); 578 return (error ? error : EINTR); 579 } 580 if ((bp->b_flags & B_CACHE) == 0) { 581 bp->b_iocmd = BIO_READ; 582 vfs_busy_pages(bp, 0); 583 error = nfs_doio(vp, bp, cred, td); 584 if (error) { 585 bp->b_ioflags |= BIO_ERROR; 586 brelse(bp); 587 return (error); 588 } 589 } 590 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 591 on = 0; 592 break; 593 case VDIR: 594 nfsstats.biocache_readdirs++; 595 if (np->n_direofoffset 596 && uio->uio_offset >= np->n_direofoffset) { 597 return (0); 598 } 599 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 600 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 601 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 602 if (!bp) { 603 error = nfs_sigintr(nmp, td); 604 return (error ? error : EINTR); 605 } 606 if ((bp->b_flags & B_CACHE) == 0) { 607 bp->b_iocmd = BIO_READ; 608 vfs_busy_pages(bp, 0); 609 error = nfs_doio(vp, bp, cred, td); 610 if (error) { 611 brelse(bp); 612 } 613 while (error == NFSERR_BAD_COOKIE) { 614 (nmp->nm_rpcops->nr_invaldir)(vp); 615 error = nfs_vinvalbuf(vp, 0, td, 1); 616 /* 617 * Yuck! The directory has been modified on the 618 * server. The only way to get the block is by 619 * reading from the beginning to get all the 620 * offset cookies. 621 * 622 * Leave the last bp intact unless there is an error. 623 * Loop back up to the while if the error is another 624 * NFSERR_BAD_COOKIE (double yuch!). 625 */ 626 for (i = 0; i <= lbn && !error; i++) { 627 if (np->n_direofoffset 628 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 629 return (0); 630 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 631 if (!bp) { 632 error = nfs_sigintr(nmp, td); 633 return (error ? error : EINTR); 634 } 635 if ((bp->b_flags & B_CACHE) == 0) { 636 bp->b_iocmd = BIO_READ; 637 vfs_busy_pages(bp, 0); 638 error = nfs_doio(vp, bp, cred, td); 639 /* 640 * no error + B_INVAL == directory EOF, 641 * use the block. 642 */ 643 if (error == 0 && (bp->b_flags & B_INVAL)) 644 break; 645 } 646 /* 647 * An error will throw away the block and the 648 * for loop will break out. If no error and this 649 * is not the block we want, we throw away the 650 * block and go for the next one via the for loop. 651 */ 652 if (error || i < lbn) 653 brelse(bp); 654 } 655 } 656 /* 657 * The above while is repeated if we hit another cookie 658 * error. If we hit an error and it wasn't a cookie error, 659 * we give up. 660 */ 661 if (error) 662 return (error); 663 } 664 665 /* 666 * If not eof and read aheads are enabled, start one. 667 * (You need the current block first, so that you have the 668 * directory offset cookie of the next block.) 669 */ 670 if (nmp->nm_readahead > 0 && 671 (bp->b_flags & B_INVAL) == 0 && 672 (np->n_direofoffset == 0 || 673 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 674 incore(&vp->v_bufobj, lbn + 1) == NULL) { 675 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 676 if (rabp) { 677 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 678 rabp->b_flags |= B_ASYNC; 679 rabp->b_iocmd = BIO_READ; 680 vfs_busy_pages(rabp, 0); 681 if (nfs_asyncio(nmp, rabp, cred, td)) { 682 rabp->b_flags |= B_INVAL; 683 rabp->b_ioflags |= BIO_ERROR; 684 vfs_unbusy_pages(rabp); 685 brelse(rabp); 686 } 687 } else { 688 brelse(rabp); 689 } 690 } 691 } 692 /* 693 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 694 * chopped for the EOF condition, we cannot tell how large 695 * NFS directories are going to be until we hit EOF. So 696 * an NFS directory buffer is *not* chopped to its EOF. Now, 697 * it just so happens that b_resid will effectively chop it 698 * to EOF. *BUT* this information is lost if the buffer goes 699 * away and is reconstituted into a B_CACHE state ( due to 700 * being VMIO ) later. So we keep track of the directory eof 701 * in np->n_direofoffset and chop it off as an extra step 702 * right here. 703 */ 704 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 705 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 706 n = np->n_direofoffset - uio->uio_offset; 707 break; 708 default: 709 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type); 710 bp = NULL; 711 break; 712 }; 713 714 if (n > 0) { 715 error = uiomove(bp->b_data + on, (int)n, uio); 716 } 717 if (vp->v_type == VLNK) 718 n = 0; 719 if (bp != NULL) 720 brelse(bp); 721 } while (error == 0 && uio->uio_resid > 0 && n > 0); 722 return (error); 723} 724 725/* 726 * The NFS write path cannot handle iovecs with len > 1. So we need to 727 * break up iovecs accordingly (restricting them to wsize). 728 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 729 * For the ASYNC case, 2 copies are needed. The first a copy from the 730 * user buffer to a staging buffer and then a second copy from the staging 731 * buffer to mbufs. This can be optimized by copying from the user buffer 732 * directly into mbufs and passing the chain down, but that requires a 733 * fair amount of re-working of the relevant codepaths (and can be done 734 * later). 735 */ 736static int 737nfs_directio_write(vp, uiop, cred, ioflag) 738 struct vnode *vp; 739 struct uio *uiop; 740 struct ucred *cred; 741 int ioflag; 742{ 743 int error; 744 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 745 struct thread *td = uiop->uio_td; 746 int size; 747 int wsize; 748 749 mtx_lock(&nmp->nm_mtx); 750 wsize = nmp->nm_wsize; 751 mtx_unlock(&nmp->nm_mtx); 752 if (ioflag & IO_SYNC) { 753 int iomode, must_commit; 754 struct uio uio; 755 struct iovec iov; 756do_sync: 757 while (uiop->uio_resid > 0) { 758 size = min(uiop->uio_resid, wsize); 759 size = min(uiop->uio_iov->iov_len, size); 760 iov.iov_base = uiop->uio_iov->iov_base; 761 iov.iov_len = size; 762 uio.uio_iov = &iov; 763 uio.uio_iovcnt = 1; 764 uio.uio_offset = uiop->uio_offset; 765 uio.uio_resid = size; 766 uio.uio_segflg = UIO_USERSPACE; 767 uio.uio_rw = UIO_WRITE; 768 uio.uio_td = td; 769 iomode = NFSV3WRITE_FILESYNC; 770 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, 771 &iomode, &must_commit); 772 KASSERT((must_commit == 0), 773 ("nfs_directio_write: Did not commit write")); 774 if (error) 775 return (error); 776 uiop->uio_offset += size; 777 uiop->uio_resid -= size; 778 if (uiop->uio_iov->iov_len <= size) { 779 uiop->uio_iovcnt--; 780 uiop->uio_iov++; 781 } else { 782 uiop->uio_iov->iov_base = 783 (char *)uiop->uio_iov->iov_base + size; 784 uiop->uio_iov->iov_len -= size; 785 } 786 } 787 } else { 788 struct uio *t_uio; 789 struct iovec *t_iov; 790 struct buf *bp; 791 792 /* 793 * Break up the write into blocksize chunks and hand these 794 * over to nfsiod's for write back. 795 * Unfortunately, this incurs a copy of the data. Since 796 * the user could modify the buffer before the write is 797 * initiated. 798 * 799 * The obvious optimization here is that one of the 2 copies 800 * in the async write path can be eliminated by copying the 801 * data here directly into mbufs and passing the mbuf chain 802 * down. But that will require a fair amount of re-working 803 * of the code and can be done if there's enough interest 804 * in NFS directio access. 805 */ 806 while (uiop->uio_resid > 0) { 807 size = min(uiop->uio_resid, wsize); 808 size = min(uiop->uio_iov->iov_len, size); 809 bp = getpbuf(&nfs_pbuf_freecnt); 810 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 811 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 812 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 813 t_iov->iov_len = size; 814 t_uio->uio_iov = t_iov; 815 t_uio->uio_iovcnt = 1; 816 t_uio->uio_offset = uiop->uio_offset; 817 t_uio->uio_resid = size; 818 t_uio->uio_segflg = UIO_SYSSPACE; 819 t_uio->uio_rw = UIO_WRITE; 820 t_uio->uio_td = td; 821 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); 822 bp->b_flags |= B_DIRECT; 823 bp->b_iocmd = BIO_WRITE; 824 if (cred != NOCRED) { 825 crhold(cred); 826 bp->b_wcred = cred; 827 } else 828 bp->b_wcred = NOCRED; 829 bp->b_caller1 = (void *)t_uio; 830 bp->b_vp = vp; 831 error = nfs_asyncio(nmp, bp, NOCRED, td); 832 if (error) { 833 free(t_iov->iov_base, M_NFSDIRECTIO); 834 free(t_iov, M_NFSDIRECTIO); 835 free(t_uio, M_NFSDIRECTIO); 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; 876 877#ifdef DIAGNOSTIC 878 if (uio->uio_rw != UIO_WRITE) 879 panic("nfs_write mode"); 880 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 881 panic("nfs_write proc"); 882#endif 883 if (vp->v_type != VREG) 884 return (EIO); 885 mtx_lock(&np->n_mtx); 886 if (np->n_flag & NWRITEERR) { 887 np->n_flag &= ~NWRITEERR; 888 mtx_unlock(&np->n_mtx); 889 return (np->n_error); 890 } else 891 mtx_unlock(&np->n_mtx); 892 mtx_lock(&nmp->nm_mtx); 893 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 894 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 895 mtx_unlock(&nmp->nm_mtx); 896 (void)nfs_fsinfo(nmp, vp, cred, td); 897 } else 898 mtx_unlock(&nmp->nm_mtx); 899 900 /* 901 * Synchronously flush pending buffers if we are in synchronous 902 * mode or if we are appending. 903 */ 904 if (ioflag & (IO_APPEND | IO_SYNC)) { 905 mtx_lock(&np->n_mtx); 906 if (np->n_flag & NMODIFIED) { 907 mtx_unlock(&np->n_mtx); 908#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 909 /* 910 * Require non-blocking, synchronous writes to 911 * dirty files to inform the program it needs 912 * to fsync(2) explicitly. 913 */ 914 if (ioflag & IO_NDELAY) 915 return (EAGAIN); 916#endif 917flush_and_restart: 918 np->n_attrstamp = 0; 919 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 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 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 934 error = VOP_GETATTR(vp, &vattr, cred); 935 if (error) 936 return (error); 937 mtx_lock(&np->n_mtx); 938 uio->uio_offset = np->n_size; 939 mtx_unlock(&np->n_mtx); 940 } 941 942 if (uio->uio_offset < 0) 943 return (EINVAL); 944 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 945 return (EFBIG); 946 if (uio->uio_resid == 0) 947 return (0); 948 949 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 950 return nfs_directio_write(vp, uio, cred, ioflag); 951 952 /* 953 * Maybe this should be above the vnode op call, but so long as 954 * file servers have no limits, i don't think it matters 955 */ 956 if (vn_rlimit_fsize(vp, uio, td)) 957 return (EFBIG); 958 959 biosize = vp->v_mount->mnt_stat.f_iosize; 960 /* 961 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 962 * would exceed the local maximum per-file write commit size when 963 * combined with those, we must decide whether to flush, 964 * go synchronous, or return error. We don't bother checking 965 * IO_UNIT -- we just make all writes atomic anyway, as there's 966 * no point optimizing for something that really won't ever happen. 967 */ 968 if (!(ioflag & IO_SYNC)) { 969 int nflag; 970 971 mtx_lock(&np->n_mtx); 972 nflag = np->n_flag; 973 mtx_unlock(&np->n_mtx); 974 int needrestart = 0; 975 if (nmp->nm_wcommitsize < uio->uio_resid) { 976 /* 977 * If this request could not possibly be completed 978 * without exceeding the maximum outstanding write 979 * commit size, see if we can convert it into a 980 * synchronous write operation. 981 */ 982 if (ioflag & IO_NDELAY) 983 return (EAGAIN); 984 ioflag |= IO_SYNC; 985 if (nflag & NMODIFIED) 986 needrestart = 1; 987 } else if (nflag & NMODIFIED) { 988 int wouldcommit = 0; 989 BO_LOCK(&vp->v_bufobj); 990 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 991 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 992 b_bobufs) { 993 if (bp->b_flags & B_NEEDCOMMIT) 994 wouldcommit += bp->b_bcount; 995 } 996 } 997 BO_UNLOCK(&vp->v_bufobj); 998 /* 999 * Since we're not operating synchronously and 1000 * bypassing the buffer cache, we are in a commit 1001 * and holding all of these buffers whether 1002 * transmitted or not. If not limited, this 1003 * will lead to the buffer cache deadlocking, 1004 * as no one else can flush our uncommitted buffers. 1005 */ 1006 wouldcommit += uio->uio_resid; 1007 /* 1008 * If we would initially exceed the maximum 1009 * outstanding write commit size, flush and restart. 1010 */ 1011 if (wouldcommit > nmp->nm_wcommitsize) 1012 needrestart = 1; 1013 } 1014 if (needrestart) 1015 goto flush_and_restart; 1016 } 1017 1018 do { 1019 nfsstats.biocache_writes++; 1020 lbn = uio->uio_offset / biosize; 1021 on = uio->uio_offset & (biosize-1); 1022 n = min((unsigned)(biosize - on), uio->uio_resid); 1023again: 1024 /* 1025 * Handle direct append and file extension cases, calculate 1026 * unaligned buffer size. 1027 */ 1028 mtx_lock(&np->n_mtx); 1029 if (uio->uio_offset == np->n_size && n) { 1030 mtx_unlock(&np->n_mtx); 1031 /* 1032 * Get the buffer (in its pre-append state to maintain 1033 * B_CACHE if it was previously set). Resize the 1034 * nfsnode after we have locked the buffer to prevent 1035 * readers from reading garbage. 1036 */ 1037 bcount = on; 1038 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1039 1040 if (bp != NULL) { 1041 long save; 1042 1043 mtx_lock(&np->n_mtx); 1044 np->n_size = uio->uio_offset + n; 1045 np->n_flag |= NMODIFIED; 1046 vnode_pager_setsize(vp, np->n_size); 1047 mtx_unlock(&np->n_mtx); 1048 1049 save = bp->b_flags & B_CACHE; 1050 bcount += n; 1051 allocbuf(bp, bcount); 1052 bp->b_flags |= save; 1053 } 1054 } else { 1055 /* 1056 * Obtain the locked cache block first, and then 1057 * adjust the file's size as appropriate. 1058 */ 1059 bcount = on + n; 1060 if ((off_t)lbn * biosize + bcount < np->n_size) { 1061 if ((off_t)(lbn + 1) * biosize < np->n_size) 1062 bcount = biosize; 1063 else 1064 bcount = np->n_size - (off_t)lbn * biosize; 1065 } 1066 mtx_unlock(&np->n_mtx); 1067 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1068 mtx_lock(&np->n_mtx); 1069 if (uio->uio_offset + n > np->n_size) { 1070 np->n_size = uio->uio_offset + n; 1071 np->n_flag |= NMODIFIED; 1072 vnode_pager_setsize(vp, np->n_size); 1073 } 1074 mtx_unlock(&np->n_mtx); 1075 } 1076 1077 if (!bp) { 1078 error = nfs_sigintr(nmp, td); 1079 if (!error) 1080 error = EINTR; 1081 break; 1082 } 1083 1084 /* 1085 * Issue a READ if B_CACHE is not set. In special-append 1086 * mode, B_CACHE is based on the buffer prior to the write 1087 * op and is typically set, avoiding the read. If a read 1088 * is required in special append mode, the server will 1089 * probably send us a short-read since we extended the file 1090 * on our end, resulting in b_resid == 0 and, thusly, 1091 * B_CACHE getting set. 1092 * 1093 * We can also avoid issuing the read if the write covers 1094 * the entire buffer. We have to make sure the buffer state 1095 * is reasonable in this case since we will not be initiating 1096 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1097 * more information. 1098 * 1099 * B_CACHE may also be set due to the buffer being cached 1100 * normally. 1101 */ 1102 1103 if (on == 0 && n == bcount) { 1104 bp->b_flags |= B_CACHE; 1105 bp->b_flags &= ~B_INVAL; 1106 bp->b_ioflags &= ~BIO_ERROR; 1107 } 1108 1109 if ((bp->b_flags & B_CACHE) == 0) { 1110 bp->b_iocmd = BIO_READ; 1111 vfs_busy_pages(bp, 0); 1112 error = nfs_doio(vp, bp, cred, td); 1113 if (error) { 1114 brelse(bp); 1115 break; 1116 } 1117 } 1118 if (bp->b_wcred == NOCRED) 1119 bp->b_wcred = crhold(cred); 1120 mtx_lock(&np->n_mtx); 1121 np->n_flag |= NMODIFIED; 1122 mtx_unlock(&np->n_mtx); 1123 1124 /* 1125 * If dirtyend exceeds file size, chop it down. This should 1126 * not normally occur but there is an append race where it 1127 * might occur XXX, so we log it. 1128 * 1129 * If the chopping creates a reverse-indexed or degenerate 1130 * situation with dirtyoff/end, we 0 both of them. 1131 */ 1132 1133 if (bp->b_dirtyend > bcount) { 1134 nfs_printf("NFS append race @%lx:%d\n", 1135 (long)bp->b_blkno * DEV_BSIZE, 1136 bp->b_dirtyend - bcount); 1137 bp->b_dirtyend = bcount; 1138 } 1139 1140 if (bp->b_dirtyoff >= bp->b_dirtyend) 1141 bp->b_dirtyoff = bp->b_dirtyend = 0; 1142 1143 /* 1144 * If the new write will leave a contiguous dirty 1145 * area, just update the b_dirtyoff and b_dirtyend, 1146 * otherwise force a write rpc of the old dirty area. 1147 * 1148 * While it is possible to merge discontiguous writes due to 1149 * our having a B_CACHE buffer ( and thus valid read data 1150 * for the hole), we don't because it could lead to 1151 * significant cache coherency problems with multiple clients, 1152 * especially if locking is implemented later on. 1153 * 1154 * as an optimization we could theoretically maintain 1155 * a linked list of discontinuous areas, but we would still 1156 * have to commit them separately so there isn't much 1157 * advantage to it except perhaps a bit of asynchronization. 1158 */ 1159 1160 if (bp->b_dirtyend > 0 && 1161 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1162 if (bwrite(bp) == EINTR) { 1163 error = EINTR; 1164 break; 1165 } 1166 goto again; 1167 } 1168 1169 error = uiomove((char *)bp->b_data + on, n, uio); 1170 1171 /* 1172 * Since this block is being modified, it must be written 1173 * again and not just committed. Since write clustering does 1174 * not work for the stage 1 data write, only the stage 2 1175 * commit rpc, we have to clear B_CLUSTEROK as well. 1176 */ 1177 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1178 1179 if (error) { 1180 bp->b_ioflags |= BIO_ERROR; 1181 brelse(bp); 1182 break; 1183 } 1184 1185 /* 1186 * Only update dirtyoff/dirtyend if not a degenerate 1187 * condition. 1188 */ 1189 if (n) { 1190 if (bp->b_dirtyend > 0) { 1191 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1192 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1193 } else { 1194 bp->b_dirtyoff = on; 1195 bp->b_dirtyend = on + n; 1196 } 1197 vfs_bio_set_valid(bp, on, n); 1198 } 1199 1200 /* 1201 * If IO_SYNC do bwrite(). 1202 * 1203 * IO_INVAL appears to be unused. The idea appears to be 1204 * to turn off caching in this case. Very odd. XXX 1205 */ 1206 if ((ioflag & IO_SYNC)) { 1207 if (ioflag & IO_INVAL) 1208 bp->b_flags |= B_NOCACHE; 1209 error = bwrite(bp); 1210 if (error) 1211 break; 1212 } else if ((n + on) == biosize) { 1213 bp->b_flags |= B_ASYNC; 1214 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL); 1215 } else { 1216 bdwrite(bp); 1217 } 1218 } while (uio->uio_resid > 0 && n > 0); 1219 1220 return (error); 1221} 1222 1223/* 1224 * Get an nfs cache block. 1225 * 1226 * Allocate a new one if the block isn't currently in the cache 1227 * and return the block marked busy. If the calling process is 1228 * interrupted by a signal for an interruptible mount point, return 1229 * NULL. 1230 * 1231 * The caller must carefully deal with the possible B_INVAL state of 1232 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1233 * indirectly), so synchronous reads can be issued without worrying about 1234 * the B_INVAL state. We have to be a little more careful when dealing 1235 * with writes (see comments in nfs_write()) when extending a file past 1236 * its EOF. 1237 */ 1238static struct buf * 1239nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1240{ 1241 struct buf *bp; 1242 struct mount *mp; 1243 struct nfsmount *nmp; 1244 1245 mp = vp->v_mount; 1246 nmp = VFSTONFS(mp); 1247 1248 if (nmp->nm_flag & NFSMNT_INT) { 1249 sigset_t oldset; 1250 1251 nfs_set_sigmask(td, &oldset); 1252 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0); 1253 nfs_restore_sigmask(td, &oldset); 1254 while (bp == NULL) { 1255 if (nfs_sigintr(nmp, td)) 1256 return (NULL); 1257 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1258 } 1259 } else { 1260 bp = getblk(vp, bn, size, 0, 0, 0); 1261 } 1262 1263 if (vp->v_type == VREG) { 1264 int biosize; 1265 1266 biosize = mp->mnt_stat.f_iosize; 1267 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1268 } 1269 return (bp); 1270} 1271 1272/* 1273 * Flush and invalidate all dirty buffers. If another process is already 1274 * doing the flush, just wait for completion. 1275 */ 1276int 1277nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1278{ 1279 struct nfsnode *np = VTONFS(vp); 1280 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1281 int error = 0, slpflag, slptimeo; 1282 int old_lock = 0; 1283 1284 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf"); 1285 1286 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1287 intrflg = 0; 1288 if (intrflg) { 1289 slpflag = NFS_PCATCH; 1290 slptimeo = 2 * hz; 1291 } else { 1292 slpflag = 0; 1293 slptimeo = 0; 1294 } 1295 1296 old_lock = nfs_upgrade_vnlock(vp); 1297 if (vp->v_iflag & VI_DOOMED) { 1298 /* 1299 * Since vgonel() uses the generic vinvalbuf() to flush 1300 * dirty buffers and it does not call this function, it 1301 * is safe to just return OK when VI_DOOMED is set. 1302 */ 1303 nfs_downgrade_vnlock(vp, old_lock); 1304 return (0); 1305 } 1306 1307 /* 1308 * Now, flush as required. 1309 */ 1310 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1311 VM_OBJECT_LOCK(vp->v_bufobj.bo_object); 1312 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1313 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); 1314 /* 1315 * If the page clean was interrupted, fail the invalidation. 1316 * Not doing so, we run the risk of losing dirty pages in the 1317 * vinvalbuf() call below. 1318 */ 1319 if (intrflg && (error = nfs_sigintr(nmp, td))) 1320 goto out; 1321 } 1322 1323 error = vinvalbuf(vp, flags, slpflag, 0); 1324 while (error) { 1325 if (intrflg && (error = nfs_sigintr(nmp, td))) 1326 goto out; 1327 error = vinvalbuf(vp, flags, 0, slptimeo); 1328 } 1329 mtx_lock(&np->n_mtx); 1330 if (np->n_directio_asyncwr == 0) 1331 np->n_flag &= ~NMODIFIED; 1332 mtx_unlock(&np->n_mtx); 1333out: 1334 nfs_downgrade_vnlock(vp, old_lock); 1335 return error; 1336} 1337 1338/* 1339 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1340 * This is mainly to avoid queueing async I/O requests when the nfsiods 1341 * are all hung on a dead server. 1342 * 1343 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1344 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1345 */ 1346int 1347nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1348{ 1349 int iod; 1350 int gotiod; 1351 int slpflag = 0; 1352 int slptimeo = 0; 1353 int error, error2; 1354 1355 /* 1356 * Commits are usually short and sweet so lets save some cpu and 1357 * leave the async daemons for more important rpc's (such as reads 1358 * and writes). 1359 */ 1360 mtx_lock(&nfs_iod_mtx); 1361 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1362 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1363 mtx_unlock(&nfs_iod_mtx); 1364 return(EIO); 1365 } 1366again: 1367 if (nmp->nm_flag & NFSMNT_INT) 1368 slpflag = NFS_PCATCH; 1369 gotiod = FALSE; 1370 1371 /* 1372 * Find a free iod to process this request. 1373 */ 1374 for (iod = 0; iod < nfs_numasync; iod++) 1375 if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) { 1376 gotiod = TRUE; 1377 break; 1378 } 1379 1380 /* 1381 * Try to create one if none are free. 1382 */ 1383 if (!gotiod) { 1384 iod = nfs_nfsiodnew(1); 1385 if (iod != -1) 1386 gotiod = TRUE; 1387 } 1388 1389 if (gotiod) { 1390 /* 1391 * Found one, so wake it up and tell it which 1392 * mount to process. 1393 */ 1394 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n", 1395 iod, nmp)); 1396 nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE; 1397 nfs_iodmount[iod] = nmp; 1398 nmp->nm_bufqiods++; 1399 wakeup(&nfs_iodwant[iod]); 1400 } 1401 1402 /* 1403 * If none are free, we may already have an iod working on this mount 1404 * point. If so, it will process our request. 1405 */ 1406 if (!gotiod) { 1407 if (nmp->nm_bufqiods > 0) { 1408 NFS_DPF(ASYNCIO, 1409 ("nfs_asyncio: %d iods are already processing mount %p\n", 1410 nmp->nm_bufqiods, nmp)); 1411 gotiod = TRUE; 1412 } 1413 } 1414 1415 /* 1416 * If we have an iod which can process the request, then queue 1417 * the buffer. 1418 */ 1419 if (gotiod) { 1420 /* 1421 * Ensure that the queue never grows too large. We still want 1422 * to asynchronize so we block rather then return EIO. 1423 */ 1424 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1425 NFS_DPF(ASYNCIO, 1426 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1427 nmp->nm_bufqwant = TRUE; 1428 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx, 1429 slpflag | PRIBIO, 1430 "nfsaio", slptimeo); 1431 if (error) { 1432 error2 = nfs_sigintr(nmp, td); 1433 if (error2) { 1434 mtx_unlock(&nfs_iod_mtx); 1435 return (error2); 1436 } 1437 if (slpflag == NFS_PCATCH) { 1438 slpflag = 0; 1439 slptimeo = 2 * hz; 1440 } 1441 } 1442 /* 1443 * We might have lost our iod while sleeping, 1444 * so check and loop if nescessary. 1445 */ 1446 if (nmp->nm_bufqiods == 0) { 1447 NFS_DPF(ASYNCIO, 1448 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1449 goto again; 1450 } 1451 } 1452 1453 /* We might have lost our nfsiod */ 1454 if (nmp->nm_bufqiods == 0) { 1455 NFS_DPF(ASYNCIO, 1456 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1457 goto again; 1458 } 1459 1460 if (bp->b_iocmd == BIO_READ) { 1461 if (bp->b_rcred == NOCRED && cred != NOCRED) 1462 bp->b_rcred = crhold(cred); 1463 } else { 1464 if (bp->b_wcred == NOCRED && cred != NOCRED) 1465 bp->b_wcred = crhold(cred); 1466 } 1467 1468 if (bp->b_flags & B_REMFREE) 1469 bremfreef(bp); 1470 BUF_KERNPROC(bp); 1471 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1472 nmp->nm_bufqlen++; 1473 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1474 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1475 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1476 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1477 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1478 } 1479 mtx_unlock(&nfs_iod_mtx); 1480 return (0); 1481 } 1482 1483 mtx_unlock(&nfs_iod_mtx); 1484 1485 /* 1486 * All the iods are busy on other mounts, so return EIO to 1487 * force the caller to process the i/o synchronously. 1488 */ 1489 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1490 return (EIO); 1491} 1492 1493void 1494nfs_doio_directwrite(struct buf *bp) 1495{ 1496 int iomode, must_commit; 1497 struct uio *uiop = (struct uio *)bp->b_caller1; 1498 char *iov_base = uiop->uio_iov->iov_base; 1499 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount); 1500 1501 iomode = NFSV3WRITE_FILESYNC; 1502 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1503 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit); 1504 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write")); 1505 free(iov_base, M_NFSDIRECTIO); 1506 free(uiop->uio_iov, M_NFSDIRECTIO); 1507 free(uiop, M_NFSDIRECTIO); 1508 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1509 struct nfsnode *np = VTONFS(bp->b_vp); 1510 mtx_lock(&np->n_mtx); 1511 np->n_directio_asyncwr--; 1512 if (np->n_directio_asyncwr == 0) { 1513 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED; 1514 if ((np->n_flag & NFSYNCWAIT)) { 1515 np->n_flag &= ~NFSYNCWAIT; 1516 wakeup((caddr_t)&np->n_directio_asyncwr); 1517 } 1518 } 1519 mtx_unlock(&np->n_mtx); 1520 } 1521 bp->b_vp = NULL; 1522 relpbuf(bp, &nfs_pbuf_freecnt); 1523} 1524 1525/* 1526 * Do an I/O operation to/from a cache block. This may be called 1527 * synchronously or from an nfsiod. 1528 */ 1529int 1530nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) 1531{ 1532 struct uio *uiop; 1533 struct nfsnode *np; 1534 struct nfsmount *nmp; 1535 int error = 0, iomode, must_commit = 0; 1536 struct uio uio; 1537 struct iovec io; 1538 struct proc *p = td ? td->td_proc : NULL; 1539 uint8_t iocmd; 1540 1541 np = VTONFS(vp); 1542 nmp = VFSTONFS(vp->v_mount); 1543 uiop = &uio; 1544 uiop->uio_iov = &io; 1545 uiop->uio_iovcnt = 1; 1546 uiop->uio_segflg = UIO_SYSSPACE; 1547 uiop->uio_td = td; 1548 1549 /* 1550 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1551 * do this here so we do not have to do it in all the code that 1552 * calls us. 1553 */ 1554 bp->b_flags &= ~B_INVAL; 1555 bp->b_ioflags &= ~BIO_ERROR; 1556 1557 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1558 iocmd = bp->b_iocmd; 1559 if (iocmd == BIO_READ) { 1560 io.iov_len = uiop->uio_resid = bp->b_bcount; 1561 io.iov_base = bp->b_data; 1562 uiop->uio_rw = UIO_READ; 1563 1564 switch (vp->v_type) { 1565 case VREG: 1566 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1567 nfsstats.read_bios++; 1568 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr); 1569 1570 if (!error) { 1571 if (uiop->uio_resid) { 1572 /* 1573 * If we had a short read with no error, we must have 1574 * hit a file hole. We should zero-fill the remainder. 1575 * This can also occur if the server hits the file EOF. 1576 * 1577 * Holes used to be able to occur due to pending 1578 * writes, but that is not possible any longer. 1579 */ 1580 int nread = bp->b_bcount - uiop->uio_resid; 1581 int left = uiop->uio_resid; 1582 1583 if (left > 0) 1584 bzero((char *)bp->b_data + nread, left); 1585 uiop->uio_resid = 0; 1586 } 1587 } 1588 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */ 1589 if (p && (vp->v_vflag & VV_TEXT)) { 1590 mtx_lock(&np->n_mtx); 1591 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) { 1592 mtx_unlock(&np->n_mtx); 1593 PROC_LOCK(p); 1594 killproc(p, "text file modification"); 1595 PROC_UNLOCK(p); 1596 } else 1597 mtx_unlock(&np->n_mtx); 1598 } 1599 break; 1600 case VLNK: 1601 uiop->uio_offset = (off_t)0; 1602 nfsstats.readlink_bios++; 1603 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr); 1604 break; 1605 case VDIR: 1606 nfsstats.readdir_bios++; 1607 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1608 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1609 error = nfs_readdirplusrpc(vp, uiop, cr); 1610 if (error == NFSERR_NOTSUPP) 1611 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1612 } 1613 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1614 error = nfs_readdirrpc(vp, uiop, cr); 1615 /* 1616 * end-of-directory sets B_INVAL but does not generate an 1617 * error. 1618 */ 1619 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1620 bp->b_flags |= B_INVAL; 1621 break; 1622 default: 1623 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type); 1624 break; 1625 }; 1626 if (error) { 1627 bp->b_ioflags |= BIO_ERROR; 1628 bp->b_error = error; 1629 } 1630 } else { 1631 /* 1632 * If we only need to commit, try to commit 1633 */ 1634 if (bp->b_flags & B_NEEDCOMMIT) { 1635 int retv; 1636 off_t off; 1637 1638 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1639 retv = (nmp->nm_rpcops->nr_commit)( 1640 vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1641 bp->b_wcred, td); 1642 if (retv == 0) { 1643 bp->b_dirtyoff = bp->b_dirtyend = 0; 1644 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1645 bp->b_resid = 0; 1646 bufdone(bp); 1647 return (0); 1648 } 1649 if (retv == NFSERR_STALEWRITEVERF) { 1650 nfs_clearcommit(vp->v_mount); 1651 } 1652 } 1653 1654 /* 1655 * Setup for actual write 1656 */ 1657 mtx_lock(&np->n_mtx); 1658 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1659 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1660 mtx_unlock(&np->n_mtx); 1661 1662 if (bp->b_dirtyend > bp->b_dirtyoff) { 1663 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1664 - bp->b_dirtyoff; 1665 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1666 + bp->b_dirtyoff; 1667 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1668 uiop->uio_rw = UIO_WRITE; 1669 nfsstats.write_bios++; 1670 1671 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1672 iomode = NFSV3WRITE_UNSTABLE; 1673 else 1674 iomode = NFSV3WRITE_FILESYNC; 1675 1676 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit); 1677 1678 /* 1679 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1680 * to cluster the buffers needing commit. This will allow 1681 * the system to submit a single commit rpc for the whole 1682 * cluster. We can do this even if the buffer is not 100% 1683 * dirty (relative to the NFS blocksize), so we optimize the 1684 * append-to-file-case. 1685 * 1686 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1687 * cleared because write clustering only works for commit 1688 * rpc's, not for the data portion of the write). 1689 */ 1690 1691 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1692 bp->b_flags |= B_NEEDCOMMIT; 1693 if (bp->b_dirtyoff == 0 1694 && bp->b_dirtyend == bp->b_bcount) 1695 bp->b_flags |= B_CLUSTEROK; 1696 } else { 1697 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1698 } 1699 1700 /* 1701 * For an interrupted write, the buffer is still valid 1702 * and the write hasn't been pushed to the server yet, 1703 * so we can't set BIO_ERROR and report the interruption 1704 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1705 * is not relevant, so the rpc attempt is essentially 1706 * a noop. For the case of a V3 write rpc not being 1707 * committed to stable storage, the block is still 1708 * dirty and requires either a commit rpc or another 1709 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1710 * the block is reused. This is indicated by setting 1711 * the B_DELWRI and B_NEEDCOMMIT flags. 1712 * 1713 * If the buffer is marked B_PAGING, it does not reside on 1714 * the vp's paging queues so we cannot call bdirty(). The 1715 * bp in this case is not an NFS cache block so we should 1716 * be safe. XXX 1717 * 1718 * The logic below breaks up errors into recoverable and 1719 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1720 * and keep the buffer around for potential write retries. 1721 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1722 * and save the error in the nfsnode. This is less than ideal 1723 * but necessary. Keeping such buffers around could potentially 1724 * cause buffer exhaustion eventually (they can never be written 1725 * out, so will get constantly be re-dirtied). It also causes 1726 * all sorts of vfs panics. For non-recoverable write errors, 1727 * also invalidate the attrcache, so we'll be forced to go over 1728 * the wire for this object, returning an error to user on next 1729 * call (most of the time). 1730 */ 1731 if (error == EINTR || error == EIO || error == ETIMEDOUT 1732 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1733 int s; 1734 1735 s = splbio(); 1736 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1737 if ((bp->b_flags & B_PAGING) == 0) { 1738 bdirty(bp); 1739 bp->b_flags &= ~B_DONE; 1740 } 1741 if (error && (bp->b_flags & B_ASYNC) == 0) 1742 bp->b_flags |= B_EINTR; 1743 splx(s); 1744 } else { 1745 if (error) { 1746 bp->b_ioflags |= BIO_ERROR; 1747 bp->b_flags |= B_INVAL; 1748 bp->b_error = np->n_error = error; 1749 mtx_lock(&np->n_mtx); 1750 np->n_flag |= NWRITEERR; 1751 np->n_attrstamp = 0; 1752 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1753 mtx_unlock(&np->n_mtx); 1754 } 1755 bp->b_dirtyoff = bp->b_dirtyend = 0; 1756 } 1757 } else { 1758 bp->b_resid = 0; 1759 bufdone(bp); 1760 return (0); 1761 } 1762 } 1763 bp->b_resid = uiop->uio_resid; 1764 if (must_commit) 1765 nfs_clearcommit(vp->v_mount); 1766 bufdone(bp); 1767 return (error); 1768} 1769 1770/* 1771 * Used to aid in handling ftruncate() operations on the NFS client side. 1772 * Truncation creates a number of special problems for NFS. We have to 1773 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1774 * we have to properly handle VM pages or (potentially dirty) buffers 1775 * that straddle the truncation point. 1776 */ 1777 1778int 1779nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1780{ 1781 struct nfsnode *np = VTONFS(vp); 1782 u_quad_t tsize; 1783 int biosize = vp->v_mount->mnt_stat.f_iosize; 1784 int error = 0; 1785 1786 mtx_lock(&np->n_mtx); 1787 tsize = np->n_size; 1788 np->n_size = nsize; 1789 mtx_unlock(&np->n_mtx); 1790 1791 if (nsize < tsize) { 1792 struct buf *bp; 1793 daddr_t lbn; 1794 int bufsize; 1795 1796 /* 1797 * vtruncbuf() doesn't get the buffer overlapping the 1798 * truncation point. We may have a B_DELWRI and/or B_CACHE 1799 * buffer that now needs to be truncated. 1800 */ 1801 error = vtruncbuf(vp, cred, td, nsize, biosize); 1802 lbn = nsize / biosize; 1803 bufsize = nsize & (biosize - 1); 1804 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1805 if (!bp) 1806 return EINTR; 1807 if (bp->b_dirtyoff > bp->b_bcount) 1808 bp->b_dirtyoff = bp->b_bcount; 1809 if (bp->b_dirtyend > bp->b_bcount) 1810 bp->b_dirtyend = bp->b_bcount; 1811 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1812 brelse(bp); 1813 } else { 1814 vnode_pager_setsize(vp, nsize); 1815 } 1816 return(error); 1817} 1818 1819