nfs_bio.c revision 194425
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 194425 2009-06-18 05:56:24Z 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/resourcevar.h> 49#include <sys/signalvar.h> 50#include <sys/vmmeter.h> 51#include <sys/vnode.h> 52 53#include <vm/vm.h> 54#include <vm/vm_extern.h> 55#include <vm/vm_page.h> 56#include <vm/vm_object.h> 57#include <vm/vm_pager.h> 58#include <vm/vnode_pager.h> 59 60#include <nfs/rpcv2.h> 61#include <nfs/nfsproto.h> 62#include <nfsclient/nfs.h> 63#include <nfsclient/nfsmount.h> 64#include <nfsclient/nfsnode.h> 65#include <nfsclient/nfs_kdtrace.h> 66 67static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 68 struct thread *td); 69static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 70 struct ucred *cred, int ioflag); 71 72extern int nfs_directio_enable; 73extern int nfs_directio_allow_mmap; 74 75/* 76 * Vnode op for VM getpages. 77 */ 78int 79nfs_getpages(struct vop_getpages_args *ap) 80{ 81 int i, error, nextoff, size, toff, count, npages; 82 struct uio uio; 83 struct iovec iov; 84 vm_offset_t kva; 85 struct buf *bp; 86 struct vnode *vp; 87 struct thread *td; 88 struct ucred *cred; 89 struct nfsmount *nmp; 90 vm_object_t object; 91 vm_page_t *pages; 92 struct nfsnode *np; 93 94 vp = ap->a_vp; 95 np = VTONFS(vp); 96 td = curthread; /* XXX */ 97 cred = curthread->td_ucred; /* XXX */ 98 nmp = VFSTONFS(vp->v_mount); 99 pages = ap->a_m; 100 count = ap->a_count; 101 102 if ((object = vp->v_object) == NULL) { 103 nfs_printf("nfs_getpages: called with non-merged cache vnode??\n"); 104 return (VM_PAGER_ERROR); 105 } 106 107 if (nfs_directio_enable && !nfs_directio_allow_mmap) { 108 mtx_lock(&np->n_mtx); 109 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 110 mtx_unlock(&np->n_mtx); 111 nfs_printf("nfs_getpages: called on non-cacheable vnode??\n"); 112 return (VM_PAGER_ERROR); 113 } else 114 mtx_unlock(&np->n_mtx); 115 } 116 117 mtx_lock(&nmp->nm_mtx); 118 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 119 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 120 mtx_unlock(&nmp->nm_mtx); 121 /* We'll never get here for v4, because we always have fsinfo */ 122 (void)nfs_fsinfo(nmp, vp, cred, td); 123 } else 124 mtx_unlock(&nmp->nm_mtx); 125 126 npages = btoc(count); 127 128 /* 129 * If the requested page is partially valid, just return it and 130 * allow the pager to zero-out the blanks. Partially valid pages 131 * can only occur at the file EOF. 132 */ 133 VM_OBJECT_LOCK(object); 134 if (pages[ap->a_reqpage]->valid != 0) { 135 vm_page_lock_queues(); 136 for (i = 0; i < npages; ++i) { 137 if (i != ap->a_reqpage) 138 vm_page_free(pages[i]); 139 } 140 vm_page_unlock_queues(); 141 VM_OBJECT_UNLOCK(object); 142 return (0); 143 } 144 VM_OBJECT_UNLOCK(object); 145 146 /* 147 * We use only the kva address for the buffer, but this is extremely 148 * convienient and fast. 149 */ 150 bp = getpbuf(&nfs_pbuf_freecnt); 151 152 kva = (vm_offset_t) bp->b_data; 153 pmap_qenter(kva, pages, npages); 154 PCPU_INC(cnt.v_vnodein); 155 PCPU_ADD(cnt.v_vnodepgsin, npages); 156 157 iov.iov_base = (caddr_t) kva; 158 iov.iov_len = count; 159 uio.uio_iov = &iov; 160 uio.uio_iovcnt = 1; 161 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 162 uio.uio_resid = count; 163 uio.uio_segflg = UIO_SYSSPACE; 164 uio.uio_rw = UIO_READ; 165 uio.uio_td = td; 166 167 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred); 168 pmap_qremove(kva, npages); 169 170 relpbuf(bp, &nfs_pbuf_freecnt); 171 172 if (error && (uio.uio_resid == count)) { 173 nfs_printf("nfs_getpages: error %d\n", error); 174 VM_OBJECT_LOCK(object); 175 vm_page_lock_queues(); 176 for (i = 0; i < npages; ++i) { 177 if (i != ap->a_reqpage) 178 vm_page_free(pages[i]); 179 } 180 vm_page_unlock_queues(); 181 VM_OBJECT_UNLOCK(object); 182 return (VM_PAGER_ERROR); 183 } 184 185 /* 186 * Calculate the number of bytes read and validate only that number 187 * of bytes. Note that due to pending writes, size may be 0. This 188 * does not mean that the remaining data is invalid! 189 */ 190 191 size = count - uio.uio_resid; 192 VM_OBJECT_LOCK(object); 193 vm_page_lock_queues(); 194 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 195 vm_page_t m; 196 nextoff = toff + PAGE_SIZE; 197 m = pages[i]; 198 199 if (nextoff <= size) { 200 /* 201 * Read operation filled an entire page 202 */ 203 m->valid = VM_PAGE_BITS_ALL; 204 KASSERT(m->dirty == 0, 205 ("nfs_getpages: page %p is dirty", m)); 206 } else if (size > toff) { 207 /* 208 * Read operation filled a partial page. 209 */ 210 m->valid = 0; 211 vm_page_set_valid(m, 0, size - toff); 212 KASSERT(m->dirty == 0, 213 ("nfs_getpages: page %p is dirty", m)); 214 } else { 215 /* 216 * Read operation was short. If no error occured 217 * we may have hit a zero-fill section. We simply 218 * leave valid set to 0. 219 */ 220 ; 221 } 222 if (i != ap->a_reqpage) { 223 /* 224 * Whether or not to leave the page activated is up in 225 * the air, but we should put the page on a page queue 226 * somewhere (it already is in the object). Result: 227 * It appears that emperical results show that 228 * deactivating pages is best. 229 */ 230 231 /* 232 * Just in case someone was asking for this page we 233 * now tell them that it is ok to use. 234 */ 235 if (!error) { 236 if (m->oflags & VPO_WANTED) 237 vm_page_activate(m); 238 else 239 vm_page_deactivate(m); 240 vm_page_wakeup(m); 241 } else { 242 vm_page_free(m); 243 } 244 } 245 } 246 vm_page_unlock_queues(); 247 VM_OBJECT_UNLOCK(object); 248 return (0); 249} 250 251/* 252 * Vnode op for VM putpages. 253 */ 254int 255nfs_putpages(struct vop_putpages_args *ap) 256{ 257 struct uio uio; 258 struct iovec iov; 259 vm_offset_t kva; 260 struct buf *bp; 261 int iomode, must_commit, i, error, npages, count; 262 off_t offset; 263 int *rtvals; 264 struct vnode *vp; 265 struct thread *td; 266 struct ucred *cred; 267 struct nfsmount *nmp; 268 struct nfsnode *np; 269 vm_page_t *pages; 270 271 vp = ap->a_vp; 272 np = VTONFS(vp); 273 td = curthread; /* XXX */ 274 cred = curthread->td_ucred; /* XXX */ 275 nmp = VFSTONFS(vp->v_mount); 276 pages = ap->a_m; 277 count = ap->a_count; 278 rtvals = ap->a_rtvals; 279 npages = btoc(count); 280 offset = IDX_TO_OFF(pages[0]->pindex); 281 282 mtx_lock(&nmp->nm_mtx); 283 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 284 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 285 mtx_unlock(&nmp->nm_mtx); 286 (void)nfs_fsinfo(nmp, vp, cred, td); 287 } else 288 mtx_unlock(&nmp->nm_mtx); 289 290 mtx_lock(&np->n_mtx); 291 if (nfs_directio_enable && !nfs_directio_allow_mmap && 292 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 293 mtx_unlock(&np->n_mtx); 294 nfs_printf("nfs_putpages: called on noncache-able vnode??\n"); 295 mtx_lock(&np->n_mtx); 296 } 297 298 for (i = 0; i < npages; i++) 299 rtvals[i] = VM_PAGER_AGAIN; 300 301 /* 302 * When putting pages, do not extend file past EOF. 303 */ 304 if (offset + count > np->n_size) { 305 count = np->n_size - offset; 306 if (count < 0) 307 count = 0; 308 } 309 mtx_unlock(&np->n_mtx); 310 311 /* 312 * We use only the kva address for the buffer, but this is extremely 313 * convienient and fast. 314 */ 315 bp = getpbuf(&nfs_pbuf_freecnt); 316 317 kva = (vm_offset_t) bp->b_data; 318 pmap_qenter(kva, pages, npages); 319 PCPU_INC(cnt.v_vnodeout); 320 PCPU_ADD(cnt.v_vnodepgsout, count); 321 322 iov.iov_base = (caddr_t) kva; 323 iov.iov_len = count; 324 uio.uio_iov = &iov; 325 uio.uio_iovcnt = 1; 326 uio.uio_offset = offset; 327 uio.uio_resid = count; 328 uio.uio_segflg = UIO_SYSSPACE; 329 uio.uio_rw = UIO_WRITE; 330 uio.uio_td = td; 331 332 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 333 iomode = NFSV3WRITE_UNSTABLE; 334 else 335 iomode = NFSV3WRITE_FILESYNC; 336 337 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit); 338 339 pmap_qremove(kva, npages); 340 relpbuf(bp, &nfs_pbuf_freecnt); 341 342 if (!error) { 343 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 344 for (i = 0; i < nwritten; i++) { 345 rtvals[i] = VM_PAGER_OK; 346 vm_page_undirty(pages[i]); 347 } 348 if (must_commit) { 349 nfs_clearcommit(vp->v_mount); 350 } 351 } 352 return rtvals[0]; 353} 354 355/* 356 * For nfs, cache consistency can only be maintained approximately. 357 * Although RFC1094 does not specify the criteria, the following is 358 * believed to be compatible with the reference port. 359 * For nfs: 360 * If the file's modify time on the server has changed since the 361 * last read rpc or you have written to the file, 362 * you may have lost data cache consistency with the 363 * server, so flush all of the file's data out of the cache. 364 * Then force a getattr rpc to ensure that you have up to date 365 * attributes. 366 * NB: This implies that cache data can be read when up to 367 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 368 * attributes this could be forced by setting n_attrstamp to 0 before 369 * the VOP_GETATTR() call. 370 */ 371static inline int 372nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 373{ 374 int error = 0; 375 struct vattr vattr; 376 struct nfsnode *np = VTONFS(vp); 377 int old_lock; 378 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 379 380 /* 381 * Grab the exclusive lock before checking whether the cache is 382 * consistent. 383 * XXX - We can make this cheaper later (by acquiring cheaper locks). 384 * But for now, this suffices. 385 */ 386 old_lock = nfs_upgrade_vnlock(vp); 387 if (vp->v_iflag & VI_DOOMED) { 388 nfs_downgrade_vnlock(vp, old_lock); 389 return (EBADF); 390 } 391 392 mtx_lock(&np->n_mtx); 393 if (np->n_flag & NMODIFIED) { 394 mtx_unlock(&np->n_mtx); 395 if (vp->v_type != VREG) { 396 if (vp->v_type != VDIR) 397 panic("nfs: bioread, not dir"); 398 (nmp->nm_rpcops->nr_invaldir)(vp); 399 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 400 if (error) 401 goto out; 402 } 403 np->n_attrstamp = 0; 404 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 405 error = VOP_GETATTR(vp, &vattr, cred); 406 if (error) 407 goto out; 408 mtx_lock(&np->n_mtx); 409 np->n_mtime = vattr.va_mtime; 410 mtx_unlock(&np->n_mtx); 411 } else { 412 mtx_unlock(&np->n_mtx); 413 error = VOP_GETATTR(vp, &vattr, cred); 414 if (error) 415 return (error); 416 mtx_lock(&np->n_mtx); 417 if ((np->n_flag & NSIZECHANGED) 418 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 419 mtx_unlock(&np->n_mtx); 420 if (vp->v_type == VDIR) 421 (nmp->nm_rpcops->nr_invaldir)(vp); 422 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 423 if (error) 424 goto out; 425 mtx_lock(&np->n_mtx); 426 np->n_mtime = vattr.va_mtime; 427 np->n_flag &= ~NSIZECHANGED; 428 } 429 mtx_unlock(&np->n_mtx); 430 } 431out: 432 nfs_downgrade_vnlock(vp, old_lock); 433 return error; 434} 435 436/* 437 * Vnode op for read using bio 438 */ 439int 440nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 441{ 442 struct nfsnode *np = VTONFS(vp); 443 int biosize, i; 444 struct buf *bp, *rabp; 445 struct thread *td; 446 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 447 daddr_t lbn, rabn; 448 int bcount; 449 int seqcount; 450 int nra, error = 0, n = 0, on = 0; 451 452#ifdef DIAGNOSTIC 453 if (uio->uio_rw != UIO_READ) 454 panic("nfs_read mode"); 455#endif 456 if (uio->uio_resid == 0) 457 return (0); 458 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 459 return (EINVAL); 460 td = uio->uio_td; 461 462 mtx_lock(&nmp->nm_mtx); 463 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 464 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 465 mtx_unlock(&nmp->nm_mtx); 466 (void)nfs_fsinfo(nmp, vp, cred, td); 467 } else 468 mtx_unlock(&nmp->nm_mtx); 469 470 if (vp->v_type != VDIR && 471 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 472 return (EFBIG); 473 474 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 475 /* No caching/ no readaheads. Just read data into the user buffer */ 476 return nfs_readrpc(vp, uio, cred); 477 478 biosize = vp->v_mount->mnt_stat.f_iosize; 479 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 480 481 error = nfs_bioread_check_cons(vp, td, cred); 482 if (error) 483 return error; 484 485 do { 486 u_quad_t nsize; 487 488 mtx_lock(&np->n_mtx); 489 nsize = np->n_size; 490 mtx_unlock(&np->n_mtx); 491 492 switch (vp->v_type) { 493 case VREG: 494 nfsstats.biocache_reads++; 495 lbn = uio->uio_offset / biosize; 496 on = uio->uio_offset & (biosize - 1); 497 498 /* 499 * Start the read ahead(s), as required. 500 */ 501 if (nmp->nm_readahead > 0) { 502 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 503 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 504 rabn = lbn + 1 + nra; 505 if (incore(&vp->v_bufobj, rabn) == NULL) { 506 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 507 if (!rabp) { 508 error = nfs_sigintr(nmp, NULL, td); 509 return (error ? error : EINTR); 510 } 511 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 512 rabp->b_flags |= B_ASYNC; 513 rabp->b_iocmd = BIO_READ; 514 vfs_busy_pages(rabp, 0); 515 if (nfs_asyncio(nmp, rabp, cred, td)) { 516 rabp->b_flags |= B_INVAL; 517 rabp->b_ioflags |= BIO_ERROR; 518 vfs_unbusy_pages(rabp); 519 brelse(rabp); 520 break; 521 } 522 } else { 523 brelse(rabp); 524 } 525 } 526 } 527 } 528 529 /* Note that bcount is *not* DEV_BSIZE aligned. */ 530 bcount = biosize; 531 if ((off_t)lbn * biosize >= nsize) { 532 bcount = 0; 533 } else if ((off_t)(lbn + 1) * biosize > nsize) { 534 bcount = nsize - (off_t)lbn * biosize; 535 } 536 bp = nfs_getcacheblk(vp, lbn, bcount, td); 537 538 if (!bp) { 539 error = nfs_sigintr(nmp, NULL, td); 540 return (error ? error : EINTR); 541 } 542 543 /* 544 * If B_CACHE is not set, we must issue the read. If this 545 * fails, we return an error. 546 */ 547 548 if ((bp->b_flags & B_CACHE) == 0) { 549 bp->b_iocmd = BIO_READ; 550 vfs_busy_pages(bp, 0); 551 error = nfs_doio(vp, bp, cred, td); 552 if (error) { 553 brelse(bp); 554 return (error); 555 } 556 } 557 558 /* 559 * on is the offset into the current bp. Figure out how many 560 * bytes we can copy out of the bp. Note that bcount is 561 * NOT DEV_BSIZE aligned. 562 * 563 * Then figure out how many bytes we can copy into the uio. 564 */ 565 566 n = 0; 567 if (on < bcount) 568 n = min((unsigned)(bcount - on), uio->uio_resid); 569 break; 570 case VLNK: 571 nfsstats.biocache_readlinks++; 572 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 573 if (!bp) { 574 error = nfs_sigintr(nmp, NULL, td); 575 return (error ? error : EINTR); 576 } 577 if ((bp->b_flags & B_CACHE) == 0) { 578 bp->b_iocmd = BIO_READ; 579 vfs_busy_pages(bp, 0); 580 error = nfs_doio(vp, bp, cred, td); 581 if (error) { 582 bp->b_ioflags |= BIO_ERROR; 583 brelse(bp); 584 return (error); 585 } 586 } 587 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 588 on = 0; 589 break; 590 case VDIR: 591 nfsstats.biocache_readdirs++; 592 if (np->n_direofoffset 593 && uio->uio_offset >= np->n_direofoffset) { 594 return (0); 595 } 596 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 597 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 598 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 599 if (!bp) { 600 error = nfs_sigintr(nmp, NULL, td); 601 return (error ? error : EINTR); 602 } 603 if ((bp->b_flags & B_CACHE) == 0) { 604 bp->b_iocmd = BIO_READ; 605 vfs_busy_pages(bp, 0); 606 error = nfs_doio(vp, bp, cred, td); 607 if (error) { 608 brelse(bp); 609 } 610 while (error == NFSERR_BAD_COOKIE) { 611 (nmp->nm_rpcops->nr_invaldir)(vp); 612 error = nfs_vinvalbuf(vp, 0, td, 1); 613 /* 614 * Yuck! The directory has been modified on the 615 * server. The only way to get the block is by 616 * reading from the beginning to get all the 617 * offset cookies. 618 * 619 * Leave the last bp intact unless there is an error. 620 * Loop back up to the while if the error is another 621 * NFSERR_BAD_COOKIE (double yuch!). 622 */ 623 for (i = 0; i <= lbn && !error; i++) { 624 if (np->n_direofoffset 625 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 626 return (0); 627 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 628 if (!bp) { 629 error = nfs_sigintr(nmp, NULL, td); 630 return (error ? error : EINTR); 631 } 632 if ((bp->b_flags & B_CACHE) == 0) { 633 bp->b_iocmd = BIO_READ; 634 vfs_busy_pages(bp, 0); 635 error = nfs_doio(vp, bp, cred, td); 636 /* 637 * no error + B_INVAL == directory EOF, 638 * use the block. 639 */ 640 if (error == 0 && (bp->b_flags & B_INVAL)) 641 break; 642 } 643 /* 644 * An error will throw away the block and the 645 * for loop will break out. If no error and this 646 * is not the block we want, we throw away the 647 * block and go for the next one via the for loop. 648 */ 649 if (error || i < lbn) 650 brelse(bp); 651 } 652 } 653 /* 654 * The above while is repeated if we hit another cookie 655 * error. If we hit an error and it wasn't a cookie error, 656 * we give up. 657 */ 658 if (error) 659 return (error); 660 } 661 662 /* 663 * If not eof and read aheads are enabled, start one. 664 * (You need the current block first, so that you have the 665 * directory offset cookie of the next block.) 666 */ 667 if (nmp->nm_readahead > 0 && 668 (bp->b_flags & B_INVAL) == 0 && 669 (np->n_direofoffset == 0 || 670 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 671 incore(&vp->v_bufobj, lbn + 1) == NULL) { 672 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 673 if (rabp) { 674 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 675 rabp->b_flags |= B_ASYNC; 676 rabp->b_iocmd = BIO_READ; 677 vfs_busy_pages(rabp, 0); 678 if (nfs_asyncio(nmp, rabp, cred, td)) { 679 rabp->b_flags |= B_INVAL; 680 rabp->b_ioflags |= BIO_ERROR; 681 vfs_unbusy_pages(rabp); 682 brelse(rabp); 683 } 684 } else { 685 brelse(rabp); 686 } 687 } 688 } 689 /* 690 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 691 * chopped for the EOF condition, we cannot tell how large 692 * NFS directories are going to be until we hit EOF. So 693 * an NFS directory buffer is *not* chopped to its EOF. Now, 694 * it just so happens that b_resid will effectively chop it 695 * to EOF. *BUT* this information is lost if the buffer goes 696 * away and is reconstituted into a B_CACHE state ( due to 697 * being VMIO ) later. So we keep track of the directory eof 698 * in np->n_direofoffset and chop it off as an extra step 699 * right here. 700 */ 701 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 702 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 703 n = np->n_direofoffset - uio->uio_offset; 704 break; 705 default: 706 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type); 707 bp = NULL; 708 break; 709 }; 710 711 if (n > 0) { 712 error = uiomove(bp->b_data + on, (int)n, uio); 713 } 714 if (vp->v_type == VLNK) 715 n = 0; 716 if (bp != NULL) 717 brelse(bp); 718 } while (error == 0 && uio->uio_resid > 0 && n > 0); 719 return (error); 720} 721 722/* 723 * The NFS write path cannot handle iovecs with len > 1. So we need to 724 * break up iovecs accordingly (restricting them to wsize). 725 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 726 * For the ASYNC case, 2 copies are needed. The first a copy from the 727 * user buffer to a staging buffer and then a second copy from the staging 728 * buffer to mbufs. This can be optimized by copying from the user buffer 729 * directly into mbufs and passing the chain down, but that requires a 730 * fair amount of re-working of the relevant codepaths (and can be done 731 * later). 732 */ 733static int 734nfs_directio_write(vp, uiop, cred, ioflag) 735 struct vnode *vp; 736 struct uio *uiop; 737 struct ucred *cred; 738 int ioflag; 739{ 740 int error; 741 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 742 struct thread *td = uiop->uio_td; 743 int size; 744 int wsize; 745 746 mtx_lock(&nmp->nm_mtx); 747 wsize = nmp->nm_wsize; 748 mtx_unlock(&nmp->nm_mtx); 749 if (ioflag & IO_SYNC) { 750 int iomode, must_commit; 751 struct uio uio; 752 struct iovec iov; 753do_sync: 754 while (uiop->uio_resid > 0) { 755 size = min(uiop->uio_resid, wsize); 756 size = min(uiop->uio_iov->iov_len, size); 757 iov.iov_base = uiop->uio_iov->iov_base; 758 iov.iov_len = size; 759 uio.uio_iov = &iov; 760 uio.uio_iovcnt = 1; 761 uio.uio_offset = uiop->uio_offset; 762 uio.uio_resid = size; 763 uio.uio_segflg = UIO_USERSPACE; 764 uio.uio_rw = UIO_WRITE; 765 uio.uio_td = td; 766 iomode = NFSV3WRITE_FILESYNC; 767 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, 768 &iomode, &must_commit); 769 KASSERT((must_commit == 0), 770 ("nfs_directio_write: Did not commit write")); 771 if (error) 772 return (error); 773 uiop->uio_offset += size; 774 uiop->uio_resid -= size; 775 if (uiop->uio_iov->iov_len <= size) { 776 uiop->uio_iovcnt--; 777 uiop->uio_iov++; 778 } else { 779 uiop->uio_iov->iov_base = 780 (char *)uiop->uio_iov->iov_base + size; 781 uiop->uio_iov->iov_len -= size; 782 } 783 } 784 } else { 785 struct uio *t_uio; 786 struct iovec *t_iov; 787 struct buf *bp; 788 789 /* 790 * Break up the write into blocksize chunks and hand these 791 * over to nfsiod's for write back. 792 * Unfortunately, this incurs a copy of the data. Since 793 * the user could modify the buffer before the write is 794 * initiated. 795 * 796 * The obvious optimization here is that one of the 2 copies 797 * in the async write path can be eliminated by copying the 798 * data here directly into mbufs and passing the mbuf chain 799 * down. But that will require a fair amount of re-working 800 * of the code and can be done if there's enough interest 801 * in NFS directio access. 802 */ 803 while (uiop->uio_resid > 0) { 804 size = min(uiop->uio_resid, wsize); 805 size = min(uiop->uio_iov->iov_len, size); 806 bp = getpbuf(&nfs_pbuf_freecnt); 807 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 808 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 809 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 810 t_iov->iov_len = size; 811 t_uio->uio_iov = t_iov; 812 t_uio->uio_iovcnt = 1; 813 t_uio->uio_offset = uiop->uio_offset; 814 t_uio->uio_resid = size; 815 t_uio->uio_segflg = UIO_SYSSPACE; 816 t_uio->uio_rw = UIO_WRITE; 817 t_uio->uio_td = td; 818 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); 819 bp->b_flags |= B_DIRECT; 820 bp->b_iocmd = BIO_WRITE; 821 if (cred != NOCRED) { 822 crhold(cred); 823 bp->b_wcred = cred; 824 } else 825 bp->b_wcred = NOCRED; 826 bp->b_caller1 = (void *)t_uio; 827 bp->b_vp = vp; 828 error = nfs_asyncio(nmp, bp, NOCRED, td); 829 if (error) { 830 free(t_iov->iov_base, M_NFSDIRECTIO); 831 free(t_iov, M_NFSDIRECTIO); 832 free(t_uio, M_NFSDIRECTIO); 833 bp->b_vp = NULL; 834 relpbuf(bp, &nfs_pbuf_freecnt); 835 if (error == EINTR) 836 return (error); 837 goto do_sync; 838 } 839 uiop->uio_offset += size; 840 uiop->uio_resid -= size; 841 if (uiop->uio_iov->iov_len <= size) { 842 uiop->uio_iovcnt--; 843 uiop->uio_iov++; 844 } else { 845 uiop->uio_iov->iov_base = 846 (char *)uiop->uio_iov->iov_base + size; 847 uiop->uio_iov->iov_len -= size; 848 } 849 } 850 } 851 return (0); 852} 853 854/* 855 * Vnode op for write using bio 856 */ 857int 858nfs_write(struct vop_write_args *ap) 859{ 860 int biosize; 861 struct uio *uio = ap->a_uio; 862 struct thread *td = uio->uio_td; 863 struct vnode *vp = ap->a_vp; 864 struct nfsnode *np = VTONFS(vp); 865 struct ucred *cred = ap->a_cred; 866 int ioflag = ap->a_ioflag; 867 struct buf *bp; 868 struct vattr vattr; 869 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 870 daddr_t lbn; 871 int bcount; 872 int n, on, error = 0; 873 struct proc *p = td?td->td_proc:NULL; 874 875#ifdef DIAGNOSTIC 876 if (uio->uio_rw != UIO_WRITE) 877 panic("nfs_write mode"); 878 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 879 panic("nfs_write proc"); 880#endif 881 if (vp->v_type != VREG) 882 return (EIO); 883 mtx_lock(&np->n_mtx); 884 if (np->n_flag & NWRITEERR) { 885 np->n_flag &= ~NWRITEERR; 886 mtx_unlock(&np->n_mtx); 887 return (np->n_error); 888 } else 889 mtx_unlock(&np->n_mtx); 890 mtx_lock(&nmp->nm_mtx); 891 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 892 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 893 mtx_unlock(&nmp->nm_mtx); 894 (void)nfs_fsinfo(nmp, vp, cred, td); 895 } else 896 mtx_unlock(&nmp->nm_mtx); 897 898 /* 899 * Synchronously flush pending buffers if we are in synchronous 900 * mode or if we are appending. 901 */ 902 if (ioflag & (IO_APPEND | IO_SYNC)) { 903 mtx_lock(&np->n_mtx); 904 if (np->n_flag & NMODIFIED) { 905 mtx_unlock(&np->n_mtx); 906#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 907 /* 908 * Require non-blocking, synchronous writes to 909 * dirty files to inform the program it needs 910 * to fsync(2) explicitly. 911 */ 912 if (ioflag & IO_NDELAY) 913 return (EAGAIN); 914#endif 915flush_and_restart: 916 np->n_attrstamp = 0; 917 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 918 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 919 if (error) 920 return (error); 921 } else 922 mtx_unlock(&np->n_mtx); 923 } 924 925 /* 926 * If IO_APPEND then load uio_offset. We restart here if we cannot 927 * get the append lock. 928 */ 929 if (ioflag & IO_APPEND) { 930 np->n_attrstamp = 0; 931 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 932 error = VOP_GETATTR(vp, &vattr, cred); 933 if (error) 934 return (error); 935 mtx_lock(&np->n_mtx); 936 uio->uio_offset = np->n_size; 937 mtx_unlock(&np->n_mtx); 938 } 939 940 if (uio->uio_offset < 0) 941 return (EINVAL); 942 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 943 return (EFBIG); 944 if (uio->uio_resid == 0) 945 return (0); 946 947 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 948 return nfs_directio_write(vp, uio, cred, ioflag); 949 950 /* 951 * Maybe this should be above the vnode op call, but so long as 952 * file servers have no limits, i don't think it matters 953 */ 954 if (p != NULL) { 955 PROC_LOCK(p); 956 if (uio->uio_offset + uio->uio_resid > 957 lim_cur(p, RLIMIT_FSIZE)) { 958 psignal(p, SIGXFSZ); 959 PROC_UNLOCK(p); 960 return (EFBIG); 961 } 962 PROC_UNLOCK(p); 963 } 964 965 biosize = vp->v_mount->mnt_stat.f_iosize; 966 /* 967 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 968 * would exceed the local maximum per-file write commit size when 969 * combined with those, we must decide whether to flush, 970 * go synchronous, or return error. We don't bother checking 971 * IO_UNIT -- we just make all writes atomic anyway, as there's 972 * no point optimizing for something that really won't ever happen. 973 */ 974 if (!(ioflag & IO_SYNC)) { 975 int nflag; 976 977 mtx_lock(&np->n_mtx); 978 nflag = np->n_flag; 979 mtx_unlock(&np->n_mtx); 980 int needrestart = 0; 981 if (nmp->nm_wcommitsize < uio->uio_resid) { 982 /* 983 * If this request could not possibly be completed 984 * without exceeding the maximum outstanding write 985 * commit size, see if we can convert it into a 986 * synchronous write operation. 987 */ 988 if (ioflag & IO_NDELAY) 989 return (EAGAIN); 990 ioflag |= IO_SYNC; 991 if (nflag & NMODIFIED) 992 needrestart = 1; 993 } else if (nflag & NMODIFIED) { 994 int wouldcommit = 0; 995 BO_LOCK(&vp->v_bufobj); 996 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 997 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 998 b_bobufs) { 999 if (bp->b_flags & B_NEEDCOMMIT) 1000 wouldcommit += bp->b_bcount; 1001 } 1002 } 1003 BO_UNLOCK(&vp->v_bufobj); 1004 /* 1005 * Since we're not operating synchronously and 1006 * bypassing the buffer cache, we are in a commit 1007 * and holding all of these buffers whether 1008 * transmitted or not. If not limited, this 1009 * will lead to the buffer cache deadlocking, 1010 * as no one else can flush our uncommitted buffers. 1011 */ 1012 wouldcommit += uio->uio_resid; 1013 /* 1014 * If we would initially exceed the maximum 1015 * outstanding write commit size, flush and restart. 1016 */ 1017 if (wouldcommit > nmp->nm_wcommitsize) 1018 needrestart = 1; 1019 } 1020 if (needrestart) 1021 goto flush_and_restart; 1022 } 1023 1024 do { 1025 nfsstats.biocache_writes++; 1026 lbn = uio->uio_offset / biosize; 1027 on = uio->uio_offset & (biosize-1); 1028 n = min((unsigned)(biosize - on), uio->uio_resid); 1029again: 1030 /* 1031 * Handle direct append and file extension cases, calculate 1032 * unaligned buffer size. 1033 */ 1034 mtx_lock(&np->n_mtx); 1035 if (uio->uio_offset == np->n_size && n) { 1036 mtx_unlock(&np->n_mtx); 1037 /* 1038 * Get the buffer (in its pre-append state to maintain 1039 * B_CACHE if it was previously set). Resize the 1040 * nfsnode after we have locked the buffer to prevent 1041 * readers from reading garbage. 1042 */ 1043 bcount = on; 1044 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1045 1046 if (bp != NULL) { 1047 long save; 1048 1049 mtx_lock(&np->n_mtx); 1050 np->n_size = uio->uio_offset + n; 1051 np->n_flag |= NMODIFIED; 1052 vnode_pager_setsize(vp, np->n_size); 1053 mtx_unlock(&np->n_mtx); 1054 1055 save = bp->b_flags & B_CACHE; 1056 bcount += n; 1057 allocbuf(bp, bcount); 1058 bp->b_flags |= save; 1059 } 1060 } else { 1061 /* 1062 * Obtain the locked cache block first, and then 1063 * adjust the file's size as appropriate. 1064 */ 1065 bcount = on + n; 1066 if ((off_t)lbn * biosize + bcount < np->n_size) { 1067 if ((off_t)(lbn + 1) * biosize < np->n_size) 1068 bcount = biosize; 1069 else 1070 bcount = np->n_size - (off_t)lbn * biosize; 1071 } 1072 mtx_unlock(&np->n_mtx); 1073 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1074 mtx_lock(&np->n_mtx); 1075 if (uio->uio_offset + n > np->n_size) { 1076 np->n_size = uio->uio_offset + n; 1077 np->n_flag |= NMODIFIED; 1078 vnode_pager_setsize(vp, np->n_size); 1079 } 1080 mtx_unlock(&np->n_mtx); 1081 } 1082 1083 if (!bp) { 1084 error = nfs_sigintr(nmp, NULL, td); 1085 if (!error) 1086 error = EINTR; 1087 break; 1088 } 1089 1090 /* 1091 * Issue a READ if B_CACHE is not set. In special-append 1092 * mode, B_CACHE is based on the buffer prior to the write 1093 * op and is typically set, avoiding the read. If a read 1094 * is required in special append mode, the server will 1095 * probably send us a short-read since we extended the file 1096 * on our end, resulting in b_resid == 0 and, thusly, 1097 * B_CACHE getting set. 1098 * 1099 * We can also avoid issuing the read if the write covers 1100 * the entire buffer. We have to make sure the buffer state 1101 * is reasonable in this case since we will not be initiating 1102 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1103 * more information. 1104 * 1105 * B_CACHE may also be set due to the buffer being cached 1106 * normally. 1107 */ 1108 1109 if (on == 0 && n == bcount) { 1110 bp->b_flags |= B_CACHE; 1111 bp->b_flags &= ~B_INVAL; 1112 bp->b_ioflags &= ~BIO_ERROR; 1113 } 1114 1115 if ((bp->b_flags & B_CACHE) == 0) { 1116 bp->b_iocmd = BIO_READ; 1117 vfs_busy_pages(bp, 0); 1118 error = nfs_doio(vp, bp, cred, td); 1119 if (error) { 1120 brelse(bp); 1121 break; 1122 } 1123 } 1124 if (bp->b_wcred == NOCRED) 1125 bp->b_wcred = crhold(cred); 1126 mtx_lock(&np->n_mtx); 1127 np->n_flag |= NMODIFIED; 1128 mtx_unlock(&np->n_mtx); 1129 1130 /* 1131 * If dirtyend exceeds file size, chop it down. This should 1132 * not normally occur but there is an append race where it 1133 * might occur XXX, so we log it. 1134 * 1135 * If the chopping creates a reverse-indexed or degenerate 1136 * situation with dirtyoff/end, we 0 both of them. 1137 */ 1138 1139 if (bp->b_dirtyend > bcount) { 1140 nfs_printf("NFS append race @%lx:%d\n", 1141 (long)bp->b_blkno * DEV_BSIZE, 1142 bp->b_dirtyend - bcount); 1143 bp->b_dirtyend = bcount; 1144 } 1145 1146 if (bp->b_dirtyoff >= bp->b_dirtyend) 1147 bp->b_dirtyoff = bp->b_dirtyend = 0; 1148 1149 /* 1150 * If the new write will leave a contiguous dirty 1151 * area, just update the b_dirtyoff and b_dirtyend, 1152 * otherwise force a write rpc of the old dirty area. 1153 * 1154 * While it is possible to merge discontiguous writes due to 1155 * our having a B_CACHE buffer ( and thus valid read data 1156 * for the hole), we don't because it could lead to 1157 * significant cache coherency problems with multiple clients, 1158 * especially if locking is implemented later on. 1159 * 1160 * as an optimization we could theoretically maintain 1161 * a linked list of discontinuous areas, but we would still 1162 * have to commit them separately so there isn't much 1163 * advantage to it except perhaps a bit of asynchronization. 1164 */ 1165 1166 if (bp->b_dirtyend > 0 && 1167 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1168 if (bwrite(bp) == EINTR) { 1169 error = EINTR; 1170 break; 1171 } 1172 goto again; 1173 } 1174 1175 error = uiomove((char *)bp->b_data + on, n, uio); 1176 1177 /* 1178 * Since this block is being modified, it must be written 1179 * again and not just committed. Since write clustering does 1180 * not work for the stage 1 data write, only the stage 2 1181 * commit rpc, we have to clear B_CLUSTEROK as well. 1182 */ 1183 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1184 1185 if (error) { 1186 bp->b_ioflags |= BIO_ERROR; 1187 brelse(bp); 1188 break; 1189 } 1190 1191 /* 1192 * Only update dirtyoff/dirtyend if not a degenerate 1193 * condition. 1194 */ 1195 if (n) { 1196 if (bp->b_dirtyend > 0) { 1197 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1198 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1199 } else { 1200 bp->b_dirtyoff = on; 1201 bp->b_dirtyend = on + n; 1202 } 1203 vfs_bio_set_valid(bp, on, n); 1204 } 1205 1206 /* 1207 * If IO_SYNC do bwrite(). 1208 * 1209 * IO_INVAL appears to be unused. The idea appears to be 1210 * to turn off caching in this case. Very odd. XXX 1211 */ 1212 if ((ioflag & IO_SYNC)) { 1213 if (ioflag & IO_INVAL) 1214 bp->b_flags |= B_NOCACHE; 1215 error = bwrite(bp); 1216 if (error) 1217 break; 1218 } else if ((n + on) == biosize) { 1219 bp->b_flags |= B_ASYNC; 1220 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL); 1221 } else { 1222 bdwrite(bp); 1223 } 1224 } while (uio->uio_resid > 0 && n > 0); 1225 1226 return (error); 1227} 1228 1229/* 1230 * Get an nfs cache block. 1231 * 1232 * Allocate a new one if the block isn't currently in the cache 1233 * and return the block marked busy. If the calling process is 1234 * interrupted by a signal for an interruptible mount point, return 1235 * NULL. 1236 * 1237 * The caller must carefully deal with the possible B_INVAL state of 1238 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1239 * indirectly), so synchronous reads can be issued without worrying about 1240 * the B_INVAL state. We have to be a little more careful when dealing 1241 * with writes (see comments in nfs_write()) when extending a file past 1242 * its EOF. 1243 */ 1244static struct buf * 1245nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1246{ 1247 struct buf *bp; 1248 struct mount *mp; 1249 struct nfsmount *nmp; 1250 1251 mp = vp->v_mount; 1252 nmp = VFSTONFS(mp); 1253 1254 if (nmp->nm_flag & NFSMNT_INT) { 1255 sigset_t oldset; 1256 1257 nfs_set_sigmask(td, &oldset); 1258 bp = getblk(vp, bn, size, PCATCH, 0, 0); 1259 nfs_restore_sigmask(td, &oldset); 1260 while (bp == NULL) { 1261 if (nfs_sigintr(nmp, NULL, td)) 1262 return (NULL); 1263 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1264 } 1265 } else { 1266 bp = getblk(vp, bn, size, 0, 0, 0); 1267 } 1268 1269 if (vp->v_type == VREG) { 1270 int biosize; 1271 1272 biosize = mp->mnt_stat.f_iosize; 1273 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1274 } 1275 return (bp); 1276} 1277 1278/* 1279 * Flush and invalidate all dirty buffers. If another process is already 1280 * doing the flush, just wait for completion. 1281 */ 1282int 1283nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1284{ 1285 struct nfsnode *np = VTONFS(vp); 1286 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1287 int error = 0, slpflag, slptimeo; 1288 int old_lock = 0; 1289 1290 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf"); 1291 1292 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1293 intrflg = 0; 1294 if (intrflg) { 1295 slpflag = PCATCH; 1296 slptimeo = 2 * hz; 1297 } else { 1298 slpflag = 0; 1299 slptimeo = 0; 1300 } 1301 1302 old_lock = nfs_upgrade_vnlock(vp); 1303 if (vp->v_iflag & VI_DOOMED) { 1304 /* 1305 * Since vgonel() uses the generic vinvalbuf() to flush 1306 * dirty buffers and it does not call this function, it 1307 * is safe to just return OK when VI_DOOMED is set. 1308 */ 1309 nfs_downgrade_vnlock(vp, old_lock); 1310 return (0); 1311 } 1312 1313 /* 1314 * Now, flush as required. 1315 */ 1316 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1317 VM_OBJECT_LOCK(vp->v_bufobj.bo_object); 1318 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1319 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); 1320 /* 1321 * If the page clean was interrupted, fail the invalidation. 1322 * Not doing so, we run the risk of losing dirty pages in the 1323 * vinvalbuf() call below. 1324 */ 1325 if (intrflg && (error = nfs_sigintr(nmp, NULL, td))) 1326 goto out; 1327 } 1328 1329 error = vinvalbuf(vp, flags, slpflag, 0); 1330 while (error) { 1331 if (intrflg && (error = nfs_sigintr(nmp, NULL, td))) 1332 goto out; 1333 error = vinvalbuf(vp, flags, 0, slptimeo); 1334 } 1335 mtx_lock(&np->n_mtx); 1336 if (np->n_directio_asyncwr == 0) 1337 np->n_flag &= ~NMODIFIED; 1338 mtx_unlock(&np->n_mtx); 1339out: 1340 nfs_downgrade_vnlock(vp, old_lock); 1341 return error; 1342} 1343 1344/* 1345 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1346 * This is mainly to avoid queueing async I/O requests when the nfsiods 1347 * are all hung on a dead server. 1348 * 1349 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1350 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1351 */ 1352int 1353nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1354{ 1355 int iod; 1356 int gotiod; 1357 int slpflag = 0; 1358 int slptimeo = 0; 1359 int error, error2; 1360 1361 /* 1362 * Commits are usually short and sweet so lets save some cpu and 1363 * leave the async daemons for more important rpc's (such as reads 1364 * and writes). 1365 */ 1366 mtx_lock(&nfs_iod_mtx); 1367 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1368 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1369 mtx_unlock(&nfs_iod_mtx); 1370 return(EIO); 1371 } 1372again: 1373 if (nmp->nm_flag & NFSMNT_INT) 1374 slpflag = PCATCH; 1375 gotiod = FALSE; 1376 1377 /* 1378 * Find a free iod to process this request. 1379 */ 1380 for (iod = 0; iod < nfs_numasync; iod++) 1381 if (nfs_iodwant[iod]) { 1382 gotiod = TRUE; 1383 break; 1384 } 1385 1386 /* 1387 * Try to create one if none are free. 1388 */ 1389 if (!gotiod) { 1390 iod = nfs_nfsiodnew(); 1391 if (iod != -1) 1392 gotiod = TRUE; 1393 } 1394 1395 if (gotiod) { 1396 /* 1397 * Found one, so wake it up and tell it which 1398 * mount to process. 1399 */ 1400 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n", 1401 iod, nmp)); 1402 nfs_iodwant[iod] = NULL; 1403 nfs_iodmount[iod] = nmp; 1404 nmp->nm_bufqiods++; 1405 wakeup(&nfs_iodwant[iod]); 1406 } 1407 1408 /* 1409 * If none are free, we may already have an iod working on this mount 1410 * point. If so, it will process our request. 1411 */ 1412 if (!gotiod) { 1413 if (nmp->nm_bufqiods > 0) { 1414 NFS_DPF(ASYNCIO, 1415 ("nfs_asyncio: %d iods are already processing mount %p\n", 1416 nmp->nm_bufqiods, nmp)); 1417 gotiod = TRUE; 1418 } 1419 } 1420 1421 /* 1422 * If we have an iod which can process the request, then queue 1423 * the buffer. 1424 */ 1425 if (gotiod) { 1426 /* 1427 * Ensure that the queue never grows too large. We still want 1428 * to asynchronize so we block rather then return EIO. 1429 */ 1430 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1431 NFS_DPF(ASYNCIO, 1432 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1433 nmp->nm_bufqwant = TRUE; 1434 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx, 1435 slpflag | PRIBIO, 1436 "nfsaio", slptimeo); 1437 if (error) { 1438 error2 = nfs_sigintr(nmp, NULL, td); 1439 if (error2) { 1440 mtx_unlock(&nfs_iod_mtx); 1441 return (error2); 1442 } 1443 if (slpflag == PCATCH) { 1444 slpflag = 0; 1445 slptimeo = 2 * hz; 1446 } 1447 } 1448 /* 1449 * We might have lost our iod while sleeping, 1450 * so check and loop if nescessary. 1451 */ 1452 if (nmp->nm_bufqiods == 0) { 1453 NFS_DPF(ASYNCIO, 1454 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1455 goto again; 1456 } 1457 } 1458 1459 /* We might have lost our nfsiod */ 1460 if (nmp->nm_bufqiods == 0) { 1461 NFS_DPF(ASYNCIO, 1462 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1463 goto again; 1464 } 1465 1466 if (bp->b_iocmd == BIO_READ) { 1467 if (bp->b_rcred == NOCRED && cred != NOCRED) 1468 bp->b_rcred = crhold(cred); 1469 } else { 1470 if (bp->b_wcred == NOCRED && cred != NOCRED) 1471 bp->b_wcred = crhold(cred); 1472 } 1473 1474 if (bp->b_flags & B_REMFREE) 1475 bremfreef(bp); 1476 BUF_KERNPROC(bp); 1477 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1478 nmp->nm_bufqlen++; 1479 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1480 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1481 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1482 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1483 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1484 } 1485 mtx_unlock(&nfs_iod_mtx); 1486 return (0); 1487 } 1488 1489 mtx_unlock(&nfs_iod_mtx); 1490 1491 /* 1492 * All the iods are busy on other mounts, so return EIO to 1493 * force the caller to process the i/o synchronously. 1494 */ 1495 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1496 return (EIO); 1497} 1498 1499void 1500nfs_doio_directwrite(struct buf *bp) 1501{ 1502 int iomode, must_commit; 1503 struct uio *uiop = (struct uio *)bp->b_caller1; 1504 char *iov_base = uiop->uio_iov->iov_base; 1505 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount); 1506 1507 iomode = NFSV3WRITE_FILESYNC; 1508 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1509 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit); 1510 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write")); 1511 free(iov_base, M_NFSDIRECTIO); 1512 free(uiop->uio_iov, M_NFSDIRECTIO); 1513 free(uiop, M_NFSDIRECTIO); 1514 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1515 struct nfsnode *np = VTONFS(bp->b_vp); 1516 mtx_lock(&np->n_mtx); 1517 np->n_directio_asyncwr--; 1518 if (np->n_directio_asyncwr == 0) { 1519 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED; 1520 if ((np->n_flag & NFSYNCWAIT)) { 1521 np->n_flag &= ~NFSYNCWAIT; 1522 wakeup((caddr_t)&np->n_directio_asyncwr); 1523 } 1524 } 1525 mtx_unlock(&np->n_mtx); 1526 } 1527 bp->b_vp = NULL; 1528 relpbuf(bp, &nfs_pbuf_freecnt); 1529} 1530 1531/* 1532 * Do an I/O operation to/from a cache block. This may be called 1533 * synchronously or from an nfsiod. 1534 */ 1535int 1536nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) 1537{ 1538 struct uio *uiop; 1539 struct nfsnode *np; 1540 struct nfsmount *nmp; 1541 int error = 0, iomode, must_commit = 0; 1542 struct uio uio; 1543 struct iovec io; 1544 struct proc *p = td ? td->td_proc : NULL; 1545 uint8_t iocmd; 1546 1547 np = VTONFS(vp); 1548 nmp = VFSTONFS(vp->v_mount); 1549 uiop = &uio; 1550 uiop->uio_iov = &io; 1551 uiop->uio_iovcnt = 1; 1552 uiop->uio_segflg = UIO_SYSSPACE; 1553 uiop->uio_td = td; 1554 1555 /* 1556 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1557 * do this here so we do not have to do it in all the code that 1558 * calls us. 1559 */ 1560 bp->b_flags &= ~B_INVAL; 1561 bp->b_ioflags &= ~BIO_ERROR; 1562 1563 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1564 iocmd = bp->b_iocmd; 1565 if (iocmd == BIO_READ) { 1566 io.iov_len = uiop->uio_resid = bp->b_bcount; 1567 io.iov_base = bp->b_data; 1568 uiop->uio_rw = UIO_READ; 1569 1570 switch (vp->v_type) { 1571 case VREG: 1572 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1573 nfsstats.read_bios++; 1574 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr); 1575 1576 if (!error) { 1577 if (uiop->uio_resid) { 1578 /* 1579 * If we had a short read with no error, we must have 1580 * hit a file hole. We should zero-fill the remainder. 1581 * This can also occur if the server hits the file EOF. 1582 * 1583 * Holes used to be able to occur due to pending 1584 * writes, but that is not possible any longer. 1585 */ 1586 int nread = bp->b_bcount - uiop->uio_resid; 1587 int left = uiop->uio_resid; 1588 1589 if (left > 0) 1590 bzero((char *)bp->b_data + nread, left); 1591 uiop->uio_resid = 0; 1592 } 1593 } 1594 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */ 1595 if (p && (vp->v_vflag & VV_TEXT)) { 1596 mtx_lock(&np->n_mtx); 1597 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) { 1598 mtx_unlock(&np->n_mtx); 1599 PROC_LOCK(p); 1600 killproc(p, "text file modification"); 1601 PROC_UNLOCK(p); 1602 } else 1603 mtx_unlock(&np->n_mtx); 1604 } 1605 break; 1606 case VLNK: 1607 uiop->uio_offset = (off_t)0; 1608 nfsstats.readlink_bios++; 1609 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr); 1610 break; 1611 case VDIR: 1612 nfsstats.readdir_bios++; 1613 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1614 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1615 error = nfs_readdirplusrpc(vp, uiop, cr); 1616 if (error == NFSERR_NOTSUPP) 1617 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1618 } 1619 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1620 error = nfs_readdirrpc(vp, uiop, cr); 1621 /* 1622 * end-of-directory sets B_INVAL but does not generate an 1623 * error. 1624 */ 1625 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1626 bp->b_flags |= B_INVAL; 1627 break; 1628 default: 1629 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type); 1630 break; 1631 }; 1632 if (error) { 1633 bp->b_ioflags |= BIO_ERROR; 1634 bp->b_error = error; 1635 } 1636 } else { 1637 /* 1638 * If we only need to commit, try to commit 1639 */ 1640 if (bp->b_flags & B_NEEDCOMMIT) { 1641 int retv; 1642 off_t off; 1643 1644 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1645 retv = (nmp->nm_rpcops->nr_commit)( 1646 vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1647 bp->b_wcred, td); 1648 if (retv == 0) { 1649 bp->b_dirtyoff = bp->b_dirtyend = 0; 1650 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1651 bp->b_resid = 0; 1652 bufdone(bp); 1653 return (0); 1654 } 1655 if (retv == NFSERR_STALEWRITEVERF) { 1656 nfs_clearcommit(vp->v_mount); 1657 } 1658 } 1659 1660 /* 1661 * Setup for actual write 1662 */ 1663 mtx_lock(&np->n_mtx); 1664 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1665 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1666 mtx_unlock(&np->n_mtx); 1667 1668 if (bp->b_dirtyend > bp->b_dirtyoff) { 1669 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1670 - bp->b_dirtyoff; 1671 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1672 + bp->b_dirtyoff; 1673 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1674 uiop->uio_rw = UIO_WRITE; 1675 nfsstats.write_bios++; 1676 1677 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1678 iomode = NFSV3WRITE_UNSTABLE; 1679 else 1680 iomode = NFSV3WRITE_FILESYNC; 1681 1682 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit); 1683 1684 /* 1685 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1686 * to cluster the buffers needing commit. This will allow 1687 * the system to submit a single commit rpc for the whole 1688 * cluster. We can do this even if the buffer is not 100% 1689 * dirty (relative to the NFS blocksize), so we optimize the 1690 * append-to-file-case. 1691 * 1692 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1693 * cleared because write clustering only works for commit 1694 * rpc's, not for the data portion of the write). 1695 */ 1696 1697 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1698 bp->b_flags |= B_NEEDCOMMIT; 1699 if (bp->b_dirtyoff == 0 1700 && bp->b_dirtyend == bp->b_bcount) 1701 bp->b_flags |= B_CLUSTEROK; 1702 } else { 1703 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1704 } 1705 1706 /* 1707 * For an interrupted write, the buffer is still valid 1708 * and the write hasn't been pushed to the server yet, 1709 * so we can't set BIO_ERROR and report the interruption 1710 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1711 * is not relevant, so the rpc attempt is essentially 1712 * a noop. For the case of a V3 write rpc not being 1713 * committed to stable storage, the block is still 1714 * dirty and requires either a commit rpc or another 1715 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1716 * the block is reused. This is indicated by setting 1717 * the B_DELWRI and B_NEEDCOMMIT flags. 1718 * 1719 * If the buffer is marked B_PAGING, it does not reside on 1720 * the vp's paging queues so we cannot call bdirty(). The 1721 * bp in this case is not an NFS cache block so we should 1722 * be safe. XXX 1723 * 1724 * The logic below breaks up errors into recoverable and 1725 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1726 * and keep the buffer around for potential write retries. 1727 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1728 * and save the error in the nfsnode. This is less than ideal 1729 * but necessary. Keeping such buffers around could potentially 1730 * cause buffer exhaustion eventually (they can never be written 1731 * out, so will get constantly be re-dirtied). It also causes 1732 * all sorts of vfs panics. For non-recoverable write errors, 1733 * also invalidate the attrcache, so we'll be forced to go over 1734 * the wire for this object, returning an error to user on next 1735 * call (most of the time). 1736 */ 1737 if (error == EINTR || error == EIO || error == ETIMEDOUT 1738 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1739 int s; 1740 1741 s = splbio(); 1742 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1743 if ((bp->b_flags & B_PAGING) == 0) { 1744 bdirty(bp); 1745 bp->b_flags &= ~B_DONE; 1746 } 1747 if (error && (bp->b_flags & B_ASYNC) == 0) 1748 bp->b_flags |= B_EINTR; 1749 splx(s); 1750 } else { 1751 if (error) { 1752 bp->b_ioflags |= BIO_ERROR; 1753 bp->b_flags |= B_INVAL; 1754 bp->b_error = np->n_error = error; 1755 mtx_lock(&np->n_mtx); 1756 np->n_flag |= NWRITEERR; 1757 np->n_attrstamp = 0; 1758 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1759 mtx_unlock(&np->n_mtx); 1760 } 1761 bp->b_dirtyoff = bp->b_dirtyend = 0; 1762 } 1763 } else { 1764 bp->b_resid = 0; 1765 bufdone(bp); 1766 return (0); 1767 } 1768 } 1769 bp->b_resid = uiop->uio_resid; 1770 if (must_commit) 1771 nfs_clearcommit(vp->v_mount); 1772 bufdone(bp); 1773 return (error); 1774} 1775 1776/* 1777 * Used to aid in handling ftruncate() operations on the NFS client side. 1778 * Truncation creates a number of special problems for NFS. We have to 1779 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1780 * we have to properly handle VM pages or (potentially dirty) buffers 1781 * that straddle the truncation point. 1782 */ 1783 1784int 1785nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1786{ 1787 struct nfsnode *np = VTONFS(vp); 1788 u_quad_t tsize; 1789 int biosize = vp->v_mount->mnt_stat.f_iosize; 1790 int error = 0; 1791 1792 mtx_lock(&np->n_mtx); 1793 tsize = np->n_size; 1794 np->n_size = nsize; 1795 mtx_unlock(&np->n_mtx); 1796 1797 if (nsize < tsize) { 1798 struct buf *bp; 1799 daddr_t lbn; 1800 int bufsize; 1801 1802 /* 1803 * vtruncbuf() doesn't get the buffer overlapping the 1804 * truncation point. We may have a B_DELWRI and/or B_CACHE 1805 * buffer that now needs to be truncated. 1806 */ 1807 error = vtruncbuf(vp, cred, td, nsize, biosize); 1808 lbn = nsize / biosize; 1809 bufsize = nsize & (biosize - 1); 1810 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1811 if (!bp) 1812 return EINTR; 1813 if (bp->b_dirtyoff > bp->b_bcount) 1814 bp->b_dirtyoff = bp->b_bcount; 1815 if (bp->b_dirtyend > bp->b_bcount) 1816 bp->b_dirtyend = bp->b_bcount; 1817 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1818 brelse(bp); 1819 } else { 1820 vnode_pager_setsize(vp, nsize); 1821 } 1822 return(error); 1823} 1824 1825