nfs_bio.c revision 195203
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 195203 2009-06-30 19:10:17Z dfr $"); 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/nfsproto.h> 61#include <nfsclient/nfs.h> 62#include <nfsclient/nfsmount.h> 63#include <nfsclient/nfsnode.h> 64#include <nfsclient/nfs_kdtrace.h> 65 66static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 67 struct thread *td); 68static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 69 struct ucred *cred, int ioflag); 70 71extern int nfs_directio_enable; 72extern int nfs_directio_allow_mmap; 73 74/* 75 * Vnode op for VM getpages. 76 */ 77int 78nfs_getpages(struct vop_getpages_args *ap) 79{ 80 int i, error, nextoff, size, toff, count, npages; 81 struct uio uio; 82 struct iovec iov; 83 vm_offset_t kva; 84 struct buf *bp; 85 struct vnode *vp; 86 struct thread *td; 87 struct ucred *cred; 88 struct nfsmount *nmp; 89 vm_object_t object; 90 vm_page_t *pages; 91 struct nfsnode *np; 92 93 vp = ap->a_vp; 94 np = VTONFS(vp); 95 td = curthread; /* XXX */ 96 cred = curthread->td_ucred; /* XXX */ 97 nmp = VFSTONFS(vp->v_mount); 98 pages = ap->a_m; 99 count = ap->a_count; 100 101 if ((object = vp->v_object) == NULL) { 102 nfs_printf("nfs_getpages: called with non-merged cache vnode??\n"); 103 return (VM_PAGER_ERROR); 104 } 105 106 if (nfs_directio_enable && !nfs_directio_allow_mmap) { 107 mtx_lock(&np->n_mtx); 108 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 109 mtx_unlock(&np->n_mtx); 110 nfs_printf("nfs_getpages: called on non-cacheable vnode??\n"); 111 return (VM_PAGER_ERROR); 112 } else 113 mtx_unlock(&np->n_mtx); 114 } 115 116 mtx_lock(&nmp->nm_mtx); 117 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 118 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 119 mtx_unlock(&nmp->nm_mtx); 120 /* We'll never get here for v4, because we always have fsinfo */ 121 (void)nfs_fsinfo(nmp, vp, cred, td); 122 } else 123 mtx_unlock(&nmp->nm_mtx); 124 125 npages = btoc(count); 126 127 /* 128 * If the requested page is partially valid, just return it and 129 * allow the pager to zero-out the blanks. Partially valid pages 130 * can only occur at the file EOF. 131 */ 132 VM_OBJECT_LOCK(object); 133 if (pages[ap->a_reqpage]->valid != 0) { 134 vm_page_lock_queues(); 135 for (i = 0; i < npages; ++i) { 136 if (i != ap->a_reqpage) 137 vm_page_free(pages[i]); 138 } 139 vm_page_unlock_queues(); 140 VM_OBJECT_UNLOCK(object); 141 return (0); 142 } 143 VM_OBJECT_UNLOCK(object); 144 145 /* 146 * We use only the kva address for the buffer, but this is extremely 147 * convienient and fast. 148 */ 149 bp = getpbuf(&nfs_pbuf_freecnt); 150 151 kva = (vm_offset_t) bp->b_data; 152 pmap_qenter(kva, pages, npages); 153 PCPU_INC(cnt.v_vnodein); 154 PCPU_ADD(cnt.v_vnodepgsin, npages); 155 156 iov.iov_base = (caddr_t) kva; 157 iov.iov_len = count; 158 uio.uio_iov = &iov; 159 uio.uio_iovcnt = 1; 160 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 161 uio.uio_resid = count; 162 uio.uio_segflg = UIO_SYSSPACE; 163 uio.uio_rw = UIO_READ; 164 uio.uio_td = td; 165 166 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred); 167 pmap_qremove(kva, npages); 168 169 relpbuf(bp, &nfs_pbuf_freecnt); 170 171 if (error && (uio.uio_resid == count)) { 172 nfs_printf("nfs_getpages: error %d\n", error); 173 VM_OBJECT_LOCK(object); 174 vm_page_lock_queues(); 175 for (i = 0; i < npages; ++i) { 176 if (i != ap->a_reqpage) 177 vm_page_free(pages[i]); 178 } 179 vm_page_unlock_queues(); 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 vm_page_lock_queues(); 193 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 194 vm_page_t m; 195 nextoff = toff + PAGE_SIZE; 196 m = pages[i]; 197 198 if (nextoff <= size) { 199 /* 200 * Read operation filled an entire page 201 */ 202 m->valid = VM_PAGE_BITS_ALL; 203 KASSERT(m->dirty == 0, 204 ("nfs_getpages: page %p is dirty", m)); 205 } else if (size > toff) { 206 /* 207 * Read operation filled a partial page. 208 */ 209 m->valid = 0; 210 vm_page_set_valid(m, 0, size - toff); 211 KASSERT(m->dirty == 0, 212 ("nfs_getpages: page %p is dirty", m)); 213 } else { 214 /* 215 * Read operation was short. If no error occured 216 * we may have hit a zero-fill section. We simply 217 * leave valid set to 0. 218 */ 219 ; 220 } 221 if (i != ap->a_reqpage) { 222 /* 223 * Whether or not to leave the page activated is up in 224 * the air, but we should put the page on a page queue 225 * somewhere (it already is in the object). Result: 226 * It appears that emperical results show that 227 * deactivating pages is best. 228 */ 229 230 /* 231 * Just in case someone was asking for this page we 232 * now tell them that it is ok to use. 233 */ 234 if (!error) { 235 if (m->oflags & VPO_WANTED) 236 vm_page_activate(m); 237 else 238 vm_page_deactivate(m); 239 vm_page_wakeup(m); 240 } else { 241 vm_page_free(m); 242 } 243 } 244 } 245 vm_page_unlock_queues(); 246 VM_OBJECT_UNLOCK(object); 247 return (0); 248} 249 250/* 251 * Vnode op for VM putpages. 252 */ 253int 254nfs_putpages(struct vop_putpages_args *ap) 255{ 256 struct uio uio; 257 struct iovec iov; 258 vm_offset_t kva; 259 struct buf *bp; 260 int iomode, must_commit, i, error, npages, count; 261 off_t offset; 262 int *rtvals; 263 struct vnode *vp; 264 struct thread *td; 265 struct ucred *cred; 266 struct nfsmount *nmp; 267 struct nfsnode *np; 268 vm_page_t *pages; 269 270 vp = ap->a_vp; 271 np = VTONFS(vp); 272 td = curthread; /* XXX */ 273 cred = curthread->td_ucred; /* XXX */ 274 nmp = VFSTONFS(vp->v_mount); 275 pages = ap->a_m; 276 count = ap->a_count; 277 rtvals = ap->a_rtvals; 278 npages = btoc(count); 279 offset = IDX_TO_OFF(pages[0]->pindex); 280 281 mtx_lock(&nmp->nm_mtx); 282 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 283 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 284 mtx_unlock(&nmp->nm_mtx); 285 (void)nfs_fsinfo(nmp, vp, cred, td); 286 } else 287 mtx_unlock(&nmp->nm_mtx); 288 289 mtx_lock(&np->n_mtx); 290 if (nfs_directio_enable && !nfs_directio_allow_mmap && 291 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 292 mtx_unlock(&np->n_mtx); 293 nfs_printf("nfs_putpages: called on noncache-able vnode??\n"); 294 mtx_lock(&np->n_mtx); 295 } 296 297 for (i = 0; i < npages; i++) 298 rtvals[i] = VM_PAGER_AGAIN; 299 300 /* 301 * When putting pages, do not extend file past EOF. 302 */ 303 if (offset + count > np->n_size) { 304 count = np->n_size - offset; 305 if (count < 0) 306 count = 0; 307 } 308 mtx_unlock(&np->n_mtx); 309 310 /* 311 * We use only the kva address for the buffer, but this is extremely 312 * convienient and fast. 313 */ 314 bp = getpbuf(&nfs_pbuf_freecnt); 315 316 kva = (vm_offset_t) bp->b_data; 317 pmap_qenter(kva, pages, npages); 318 PCPU_INC(cnt.v_vnodeout); 319 PCPU_ADD(cnt.v_vnodepgsout, count); 320 321 iov.iov_base = (caddr_t) kva; 322 iov.iov_len = count; 323 uio.uio_iov = &iov; 324 uio.uio_iovcnt = 1; 325 uio.uio_offset = offset; 326 uio.uio_resid = count; 327 uio.uio_segflg = UIO_SYSSPACE; 328 uio.uio_rw = UIO_WRITE; 329 uio.uio_td = td; 330 331 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 332 iomode = NFSV3WRITE_UNSTABLE; 333 else 334 iomode = NFSV3WRITE_FILESYNC; 335 336 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit); 337 338 pmap_qremove(kva, npages); 339 relpbuf(bp, &nfs_pbuf_freecnt); 340 341 if (!error) { 342 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 343 for (i = 0; i < nwritten; i++) { 344 rtvals[i] = VM_PAGER_OK; 345 vm_page_undirty(pages[i]); 346 } 347 if (must_commit) { 348 nfs_clearcommit(vp->v_mount); 349 } 350 } 351 return rtvals[0]; 352} 353 354/* 355 * For nfs, cache consistency can only be maintained approximately. 356 * Although RFC1094 does not specify the criteria, the following is 357 * believed to be compatible with the reference port. 358 * For nfs: 359 * If the file's modify time on the server has changed since the 360 * last read rpc or you have written to the file, 361 * you may have lost data cache consistency with the 362 * server, so flush all of the file's data out of the cache. 363 * Then force a getattr rpc to ensure that you have up to date 364 * attributes. 365 * NB: This implies that cache data can be read when up to 366 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 367 * attributes this could be forced by setting n_attrstamp to 0 before 368 * the VOP_GETATTR() call. 369 */ 370static inline int 371nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 372{ 373 int error = 0; 374 struct vattr vattr; 375 struct nfsnode *np = VTONFS(vp); 376 int old_lock; 377 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 378 379 /* 380 * Grab the exclusive lock before checking whether the cache is 381 * consistent. 382 * XXX - We can make this cheaper later (by acquiring cheaper locks). 383 * But for now, this suffices. 384 */ 385 old_lock = nfs_upgrade_vnlock(vp); 386 if (vp->v_iflag & VI_DOOMED) { 387 nfs_downgrade_vnlock(vp, old_lock); 388 return (EBADF); 389 } 390 391 mtx_lock(&np->n_mtx); 392 if (np->n_flag & NMODIFIED) { 393 mtx_unlock(&np->n_mtx); 394 if (vp->v_type != VREG) { 395 if (vp->v_type != VDIR) 396 panic("nfs: bioread, not dir"); 397 (nmp->nm_rpcops->nr_invaldir)(vp); 398 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 399 if (error) 400 goto out; 401 } 402 np->n_attrstamp = 0; 403 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 404 error = VOP_GETATTR(vp, &vattr, cred); 405 if (error) 406 goto out; 407 mtx_lock(&np->n_mtx); 408 np->n_mtime = vattr.va_mtime; 409 mtx_unlock(&np->n_mtx); 410 } else { 411 mtx_unlock(&np->n_mtx); 412 error = VOP_GETATTR(vp, &vattr, cred); 413 if (error) 414 return (error); 415 mtx_lock(&np->n_mtx); 416 if ((np->n_flag & NSIZECHANGED) 417 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 418 mtx_unlock(&np->n_mtx); 419 if (vp->v_type == VDIR) 420 (nmp->nm_rpcops->nr_invaldir)(vp); 421 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 422 if (error) 423 goto out; 424 mtx_lock(&np->n_mtx); 425 np->n_mtime = vattr.va_mtime; 426 np->n_flag &= ~NSIZECHANGED; 427 } 428 mtx_unlock(&np->n_mtx); 429 } 430out: 431 nfs_downgrade_vnlock(vp, old_lock); 432 return error; 433} 434 435/* 436 * Vnode op for read using bio 437 */ 438int 439nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 440{ 441 struct nfsnode *np = VTONFS(vp); 442 int biosize, i; 443 struct buf *bp, *rabp; 444 struct thread *td; 445 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 446 daddr_t lbn, rabn; 447 int bcount; 448 int seqcount; 449 int nra, error = 0, n = 0, on = 0; 450 451#ifdef DIAGNOSTIC 452 if (uio->uio_rw != UIO_READ) 453 panic("nfs_read mode"); 454#endif 455 if (uio->uio_resid == 0) 456 return (0); 457 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 458 return (EINVAL); 459 td = uio->uio_td; 460 461 mtx_lock(&nmp->nm_mtx); 462 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 463 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 464 mtx_unlock(&nmp->nm_mtx); 465 (void)nfs_fsinfo(nmp, vp, cred, td); 466 } else 467 mtx_unlock(&nmp->nm_mtx); 468 469 if (vp->v_type != VDIR && 470 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 471 return (EFBIG); 472 473 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 474 /* No caching/ no readaheads. Just read data into the user buffer */ 475 return nfs_readrpc(vp, uio, cred); 476 477 biosize = vp->v_mount->mnt_stat.f_iosize; 478 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 479 480 error = nfs_bioread_check_cons(vp, td, cred); 481 if (error) 482 return error; 483 484 do { 485 u_quad_t nsize; 486 487 mtx_lock(&np->n_mtx); 488 nsize = np->n_size; 489 mtx_unlock(&np->n_mtx); 490 491 switch (vp->v_type) { 492 case VREG: 493 nfsstats.biocache_reads++; 494 lbn = uio->uio_offset / biosize; 495 on = uio->uio_offset & (biosize - 1); 496 497 /* 498 * Start the read ahead(s), as required. 499 */ 500 if (nmp->nm_readahead > 0) { 501 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 502 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 503 rabn = lbn + 1 + nra; 504 if (incore(&vp->v_bufobj, rabn) == NULL) { 505 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 506 if (!rabp) { 507 error = nfs_sigintr(nmp, td); 508 return (error ? error : EINTR); 509 } 510 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 511 rabp->b_flags |= B_ASYNC; 512 rabp->b_iocmd = BIO_READ; 513 vfs_busy_pages(rabp, 0); 514 if (nfs_asyncio(nmp, rabp, cred, td)) { 515 rabp->b_flags |= B_INVAL; 516 rabp->b_ioflags |= BIO_ERROR; 517 vfs_unbusy_pages(rabp); 518 brelse(rabp); 519 break; 520 } 521 } else { 522 brelse(rabp); 523 } 524 } 525 } 526 } 527 528 /* Note that bcount is *not* DEV_BSIZE aligned. */ 529 bcount = biosize; 530 if ((off_t)lbn * biosize >= nsize) { 531 bcount = 0; 532 } else if ((off_t)(lbn + 1) * biosize > nsize) { 533 bcount = nsize - (off_t)lbn * biosize; 534 } 535 bp = nfs_getcacheblk(vp, lbn, bcount, td); 536 537 if (!bp) { 538 error = nfs_sigintr(nmp, td); 539 return (error ? error : EINTR); 540 } 541 542 /* 543 * If B_CACHE is not set, we must issue the read. If this 544 * fails, we return an error. 545 */ 546 547 if ((bp->b_flags & B_CACHE) == 0) { 548 bp->b_iocmd = BIO_READ; 549 vfs_busy_pages(bp, 0); 550 error = nfs_doio(vp, bp, cred, td); 551 if (error) { 552 brelse(bp); 553 return (error); 554 } 555 } 556 557 /* 558 * on is the offset into the current bp. Figure out how many 559 * bytes we can copy out of the bp. Note that bcount is 560 * NOT DEV_BSIZE aligned. 561 * 562 * Then figure out how many bytes we can copy into the uio. 563 */ 564 565 n = 0; 566 if (on < bcount) 567 n = min((unsigned)(bcount - on), uio->uio_resid); 568 break; 569 case VLNK: 570 nfsstats.biocache_readlinks++; 571 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 572 if (!bp) { 573 error = nfs_sigintr(nmp, td); 574 return (error ? error : EINTR); 575 } 576 if ((bp->b_flags & B_CACHE) == 0) { 577 bp->b_iocmd = BIO_READ; 578 vfs_busy_pages(bp, 0); 579 error = nfs_doio(vp, bp, cred, td); 580 if (error) { 581 bp->b_ioflags |= BIO_ERROR; 582 brelse(bp); 583 return (error); 584 } 585 } 586 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 587 on = 0; 588 break; 589 case VDIR: 590 nfsstats.biocache_readdirs++; 591 if (np->n_direofoffset 592 && uio->uio_offset >= np->n_direofoffset) { 593 return (0); 594 } 595 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 596 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 597 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 598 if (!bp) { 599 error = nfs_sigintr(nmp, td); 600 return (error ? error : EINTR); 601 } 602 if ((bp->b_flags & B_CACHE) == 0) { 603 bp->b_iocmd = BIO_READ; 604 vfs_busy_pages(bp, 0); 605 error = nfs_doio(vp, bp, cred, td); 606 if (error) { 607 brelse(bp); 608 } 609 while (error == NFSERR_BAD_COOKIE) { 610 (nmp->nm_rpcops->nr_invaldir)(vp); 611 error = nfs_vinvalbuf(vp, 0, td, 1); 612 /* 613 * Yuck! The directory has been modified on the 614 * server. The only way to get the block is by 615 * reading from the beginning to get all the 616 * offset cookies. 617 * 618 * Leave the last bp intact unless there is an error. 619 * Loop back up to the while if the error is another 620 * NFSERR_BAD_COOKIE (double yuch!). 621 */ 622 for (i = 0; i <= lbn && !error; i++) { 623 if (np->n_direofoffset 624 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 625 return (0); 626 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 627 if (!bp) { 628 error = nfs_sigintr(nmp, td); 629 return (error ? error : EINTR); 630 } 631 if ((bp->b_flags & B_CACHE) == 0) { 632 bp->b_iocmd = BIO_READ; 633 vfs_busy_pages(bp, 0); 634 error = nfs_doio(vp, bp, cred, td); 635 /* 636 * no error + B_INVAL == directory EOF, 637 * use the block. 638 */ 639 if (error == 0 && (bp->b_flags & B_INVAL)) 640 break; 641 } 642 /* 643 * An error will throw away the block and the 644 * for loop will break out. If no error and this 645 * is not the block we want, we throw away the 646 * block and go for the next one via the for loop. 647 */ 648 if (error || i < lbn) 649 brelse(bp); 650 } 651 } 652 /* 653 * The above while is repeated if we hit another cookie 654 * error. If we hit an error and it wasn't a cookie error, 655 * we give up. 656 */ 657 if (error) 658 return (error); 659 } 660 661 /* 662 * If not eof and read aheads are enabled, start one. 663 * (You need the current block first, so that you have the 664 * directory offset cookie of the next block.) 665 */ 666 if (nmp->nm_readahead > 0 && 667 (bp->b_flags & B_INVAL) == 0 && 668 (np->n_direofoffset == 0 || 669 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 670 incore(&vp->v_bufobj, lbn + 1) == NULL) { 671 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 672 if (rabp) { 673 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 674 rabp->b_flags |= B_ASYNC; 675 rabp->b_iocmd = BIO_READ; 676 vfs_busy_pages(rabp, 0); 677 if (nfs_asyncio(nmp, rabp, cred, td)) { 678 rabp->b_flags |= B_INVAL; 679 rabp->b_ioflags |= BIO_ERROR; 680 vfs_unbusy_pages(rabp); 681 brelse(rabp); 682 } 683 } else { 684 brelse(rabp); 685 } 686 } 687 } 688 /* 689 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 690 * chopped for the EOF condition, we cannot tell how large 691 * NFS directories are going to be until we hit EOF. So 692 * an NFS directory buffer is *not* chopped to its EOF. Now, 693 * it just so happens that b_resid will effectively chop it 694 * to EOF. *BUT* this information is lost if the buffer goes 695 * away and is reconstituted into a B_CACHE state ( due to 696 * being VMIO ) later. So we keep track of the directory eof 697 * in np->n_direofoffset and chop it off as an extra step 698 * right here. 699 */ 700 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 701 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 702 n = np->n_direofoffset - uio->uio_offset; 703 break; 704 default: 705 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type); 706 bp = NULL; 707 break; 708 }; 709 710 if (n > 0) { 711 error = uiomove(bp->b_data + on, (int)n, uio); 712 } 713 if (vp->v_type == VLNK) 714 n = 0; 715 if (bp != NULL) 716 brelse(bp); 717 } while (error == 0 && uio->uio_resid > 0 && n > 0); 718 return (error); 719} 720 721/* 722 * The NFS write path cannot handle iovecs with len > 1. So we need to 723 * break up iovecs accordingly (restricting them to wsize). 724 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 725 * For the ASYNC case, 2 copies are needed. The first a copy from the 726 * user buffer to a staging buffer and then a second copy from the staging 727 * buffer to mbufs. This can be optimized by copying from the user buffer 728 * directly into mbufs and passing the chain down, but that requires a 729 * fair amount of re-working of the relevant codepaths (and can be done 730 * later). 731 */ 732static int 733nfs_directio_write(vp, uiop, cred, ioflag) 734 struct vnode *vp; 735 struct uio *uiop; 736 struct ucred *cred; 737 int ioflag; 738{ 739 int error; 740 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 741 struct thread *td = uiop->uio_td; 742 int size; 743 int wsize; 744 745 mtx_lock(&nmp->nm_mtx); 746 wsize = nmp->nm_wsize; 747 mtx_unlock(&nmp->nm_mtx); 748 if (ioflag & IO_SYNC) { 749 int iomode, must_commit; 750 struct uio uio; 751 struct iovec iov; 752do_sync: 753 while (uiop->uio_resid > 0) { 754 size = min(uiop->uio_resid, wsize); 755 size = min(uiop->uio_iov->iov_len, size); 756 iov.iov_base = uiop->uio_iov->iov_base; 757 iov.iov_len = size; 758 uio.uio_iov = &iov; 759 uio.uio_iovcnt = 1; 760 uio.uio_offset = uiop->uio_offset; 761 uio.uio_resid = size; 762 uio.uio_segflg = UIO_USERSPACE; 763 uio.uio_rw = UIO_WRITE; 764 uio.uio_td = td; 765 iomode = NFSV3WRITE_FILESYNC; 766 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, 767 &iomode, &must_commit); 768 KASSERT((must_commit == 0), 769 ("nfs_directio_write: Did not commit write")); 770 if (error) 771 return (error); 772 uiop->uio_offset += size; 773 uiop->uio_resid -= size; 774 if (uiop->uio_iov->iov_len <= size) { 775 uiop->uio_iovcnt--; 776 uiop->uio_iov++; 777 } else { 778 uiop->uio_iov->iov_base = 779 (char *)uiop->uio_iov->iov_base + size; 780 uiop->uio_iov->iov_len -= size; 781 } 782 } 783 } else { 784 struct uio *t_uio; 785 struct iovec *t_iov; 786 struct buf *bp; 787 788 /* 789 * Break up the write into blocksize chunks and hand these 790 * over to nfsiod's for write back. 791 * Unfortunately, this incurs a copy of the data. Since 792 * the user could modify the buffer before the write is 793 * initiated. 794 * 795 * The obvious optimization here is that one of the 2 copies 796 * in the async write path can be eliminated by copying the 797 * data here directly into mbufs and passing the mbuf chain 798 * down. But that will require a fair amount of re-working 799 * of the code and can be done if there's enough interest 800 * in NFS directio access. 801 */ 802 while (uiop->uio_resid > 0) { 803 size = min(uiop->uio_resid, wsize); 804 size = min(uiop->uio_iov->iov_len, size); 805 bp = getpbuf(&nfs_pbuf_freecnt); 806 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 807 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 808 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 809 t_iov->iov_len = size; 810 t_uio->uio_iov = t_iov; 811 t_uio->uio_iovcnt = 1; 812 t_uio->uio_offset = uiop->uio_offset; 813 t_uio->uio_resid = size; 814 t_uio->uio_segflg = UIO_SYSSPACE; 815 t_uio->uio_rw = UIO_WRITE; 816 t_uio->uio_td = td; 817 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); 818 bp->b_flags |= B_DIRECT; 819 bp->b_iocmd = BIO_WRITE; 820 if (cred != NOCRED) { 821 crhold(cred); 822 bp->b_wcred = cred; 823 } else 824 bp->b_wcred = NOCRED; 825 bp->b_caller1 = (void *)t_uio; 826 bp->b_vp = vp; 827 error = nfs_asyncio(nmp, bp, NOCRED, td); 828 if (error) { 829 free(t_iov->iov_base, M_NFSDIRECTIO); 830 free(t_iov, M_NFSDIRECTIO); 831 free(t_uio, M_NFSDIRECTIO); 832 bp->b_vp = NULL; 833 relpbuf(bp, &nfs_pbuf_freecnt); 834 if (error == EINTR) 835 return (error); 836 goto do_sync; 837 } 838 uiop->uio_offset += size; 839 uiop->uio_resid -= size; 840 if (uiop->uio_iov->iov_len <= size) { 841 uiop->uio_iovcnt--; 842 uiop->uio_iov++; 843 } else { 844 uiop->uio_iov->iov_base = 845 (char *)uiop->uio_iov->iov_base + size; 846 uiop->uio_iov->iov_len -= size; 847 } 848 } 849 } 850 return (0); 851} 852 853/* 854 * Vnode op for write using bio 855 */ 856int 857nfs_write(struct vop_write_args *ap) 858{ 859 int biosize; 860 struct uio *uio = ap->a_uio; 861 struct thread *td = uio->uio_td; 862 struct vnode *vp = ap->a_vp; 863 struct nfsnode *np = VTONFS(vp); 864 struct ucred *cred = ap->a_cred; 865 int ioflag = ap->a_ioflag; 866 struct buf *bp; 867 struct vattr vattr; 868 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 869 daddr_t lbn; 870 int bcount; 871 int n, on, error = 0; 872 struct proc *p = td?td->td_proc:NULL; 873 874#ifdef DIAGNOSTIC 875 if (uio->uio_rw != UIO_WRITE) 876 panic("nfs_write mode"); 877 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 878 panic("nfs_write proc"); 879#endif 880 if (vp->v_type != VREG) 881 return (EIO); 882 mtx_lock(&np->n_mtx); 883 if (np->n_flag & NWRITEERR) { 884 np->n_flag &= ~NWRITEERR; 885 mtx_unlock(&np->n_mtx); 886 return (np->n_error); 887 } else 888 mtx_unlock(&np->n_mtx); 889 mtx_lock(&nmp->nm_mtx); 890 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 891 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 892 mtx_unlock(&nmp->nm_mtx); 893 (void)nfs_fsinfo(nmp, vp, cred, td); 894 } else 895 mtx_unlock(&nmp->nm_mtx); 896 897 /* 898 * Synchronously flush pending buffers if we are in synchronous 899 * mode or if we are appending. 900 */ 901 if (ioflag & (IO_APPEND | IO_SYNC)) { 902 mtx_lock(&np->n_mtx); 903 if (np->n_flag & NMODIFIED) { 904 mtx_unlock(&np->n_mtx); 905#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 906 /* 907 * Require non-blocking, synchronous writes to 908 * dirty files to inform the program it needs 909 * to fsync(2) explicitly. 910 */ 911 if (ioflag & IO_NDELAY) 912 return (EAGAIN); 913#endif 914flush_and_restart: 915 np->n_attrstamp = 0; 916 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 917 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 918 if (error) 919 return (error); 920 } else 921 mtx_unlock(&np->n_mtx); 922 } 923 924 /* 925 * If IO_APPEND then load uio_offset. We restart here if we cannot 926 * get the append lock. 927 */ 928 if (ioflag & IO_APPEND) { 929 np->n_attrstamp = 0; 930 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 931 error = VOP_GETATTR(vp, &vattr, cred); 932 if (error) 933 return (error); 934 mtx_lock(&np->n_mtx); 935 uio->uio_offset = np->n_size; 936 mtx_unlock(&np->n_mtx); 937 } 938 939 if (uio->uio_offset < 0) 940 return (EINVAL); 941 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 942 return (EFBIG); 943 if (uio->uio_resid == 0) 944 return (0); 945 946 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 947 return nfs_directio_write(vp, uio, cred, ioflag); 948 949 /* 950 * Maybe this should be above the vnode op call, but so long as 951 * file servers have no limits, i don't think it matters 952 */ 953 if (p != NULL) { 954 PROC_LOCK(p); 955 if (uio->uio_offset + uio->uio_resid > 956 lim_cur(p, RLIMIT_FSIZE)) { 957 psignal(p, SIGXFSZ); 958 PROC_UNLOCK(p); 959 return (EFBIG); 960 } 961 PROC_UNLOCK(p); 962 } 963 964 biosize = vp->v_mount->mnt_stat.f_iosize; 965 /* 966 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 967 * would exceed the local maximum per-file write commit size when 968 * combined with those, we must decide whether to flush, 969 * go synchronous, or return error. We don't bother checking 970 * IO_UNIT -- we just make all writes atomic anyway, as there's 971 * no point optimizing for something that really won't ever happen. 972 */ 973 if (!(ioflag & IO_SYNC)) { 974 int nflag; 975 976 mtx_lock(&np->n_mtx); 977 nflag = np->n_flag; 978 mtx_unlock(&np->n_mtx); 979 int needrestart = 0; 980 if (nmp->nm_wcommitsize < uio->uio_resid) { 981 /* 982 * If this request could not possibly be completed 983 * without exceeding the maximum outstanding write 984 * commit size, see if we can convert it into a 985 * synchronous write operation. 986 */ 987 if (ioflag & IO_NDELAY) 988 return (EAGAIN); 989 ioflag |= IO_SYNC; 990 if (nflag & NMODIFIED) 991 needrestart = 1; 992 } else if (nflag & NMODIFIED) { 993 int wouldcommit = 0; 994 BO_LOCK(&vp->v_bufobj); 995 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 996 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 997 b_bobufs) { 998 if (bp->b_flags & B_NEEDCOMMIT) 999 wouldcommit += bp->b_bcount; 1000 } 1001 } 1002 BO_UNLOCK(&vp->v_bufobj); 1003 /* 1004 * Since we're not operating synchronously and 1005 * bypassing the buffer cache, we are in a commit 1006 * and holding all of these buffers whether 1007 * transmitted or not. If not limited, this 1008 * will lead to the buffer cache deadlocking, 1009 * as no one else can flush our uncommitted buffers. 1010 */ 1011 wouldcommit += uio->uio_resid; 1012 /* 1013 * If we would initially exceed the maximum 1014 * outstanding write commit size, flush and restart. 1015 */ 1016 if (wouldcommit > nmp->nm_wcommitsize) 1017 needrestart = 1; 1018 } 1019 if (needrestart) 1020 goto flush_and_restart; 1021 } 1022 1023 do { 1024 nfsstats.biocache_writes++; 1025 lbn = uio->uio_offset / biosize; 1026 on = uio->uio_offset & (biosize-1); 1027 n = min((unsigned)(biosize - on), uio->uio_resid); 1028again: 1029 /* 1030 * Handle direct append and file extension cases, calculate 1031 * unaligned buffer size. 1032 */ 1033 mtx_lock(&np->n_mtx); 1034 if (uio->uio_offset == np->n_size && n) { 1035 mtx_unlock(&np->n_mtx); 1036 /* 1037 * Get the buffer (in its pre-append state to maintain 1038 * B_CACHE if it was previously set). Resize the 1039 * nfsnode after we have locked the buffer to prevent 1040 * readers from reading garbage. 1041 */ 1042 bcount = on; 1043 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1044 1045 if (bp != NULL) { 1046 long save; 1047 1048 mtx_lock(&np->n_mtx); 1049 np->n_size = uio->uio_offset + n; 1050 np->n_flag |= NMODIFIED; 1051 vnode_pager_setsize(vp, np->n_size); 1052 mtx_unlock(&np->n_mtx); 1053 1054 save = bp->b_flags & B_CACHE; 1055 bcount += n; 1056 allocbuf(bp, bcount); 1057 bp->b_flags |= save; 1058 } 1059 } else { 1060 /* 1061 * Obtain the locked cache block first, and then 1062 * adjust the file's size as appropriate. 1063 */ 1064 bcount = on + n; 1065 if ((off_t)lbn * biosize + bcount < np->n_size) { 1066 if ((off_t)(lbn + 1) * biosize < np->n_size) 1067 bcount = biosize; 1068 else 1069 bcount = np->n_size - (off_t)lbn * biosize; 1070 } 1071 mtx_unlock(&np->n_mtx); 1072 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1073 mtx_lock(&np->n_mtx); 1074 if (uio->uio_offset + n > np->n_size) { 1075 np->n_size = uio->uio_offset + n; 1076 np->n_flag |= NMODIFIED; 1077 vnode_pager_setsize(vp, np->n_size); 1078 } 1079 mtx_unlock(&np->n_mtx); 1080 } 1081 1082 if (!bp) { 1083 error = nfs_sigintr(nmp, td); 1084 if (!error) 1085 error = EINTR; 1086 break; 1087 } 1088 1089 /* 1090 * Issue a READ if B_CACHE is not set. In special-append 1091 * mode, B_CACHE is based on the buffer prior to the write 1092 * op and is typically set, avoiding the read. If a read 1093 * is required in special append mode, the server will 1094 * probably send us a short-read since we extended the file 1095 * on our end, resulting in b_resid == 0 and, thusly, 1096 * B_CACHE getting set. 1097 * 1098 * We can also avoid issuing the read if the write covers 1099 * the entire buffer. We have to make sure the buffer state 1100 * is reasonable in this case since we will not be initiating 1101 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1102 * more information. 1103 * 1104 * B_CACHE may also be set due to the buffer being cached 1105 * normally. 1106 */ 1107 1108 if (on == 0 && n == bcount) { 1109 bp->b_flags |= B_CACHE; 1110 bp->b_flags &= ~B_INVAL; 1111 bp->b_ioflags &= ~BIO_ERROR; 1112 } 1113 1114 if ((bp->b_flags & B_CACHE) == 0) { 1115 bp->b_iocmd = BIO_READ; 1116 vfs_busy_pages(bp, 0); 1117 error = nfs_doio(vp, bp, cred, td); 1118 if (error) { 1119 brelse(bp); 1120 break; 1121 } 1122 } 1123 if (bp->b_wcred == NOCRED) 1124 bp->b_wcred = crhold(cred); 1125 mtx_lock(&np->n_mtx); 1126 np->n_flag |= NMODIFIED; 1127 mtx_unlock(&np->n_mtx); 1128 1129 /* 1130 * If dirtyend exceeds file size, chop it down. This should 1131 * not normally occur but there is an append race where it 1132 * might occur XXX, so we log it. 1133 * 1134 * If the chopping creates a reverse-indexed or degenerate 1135 * situation with dirtyoff/end, we 0 both of them. 1136 */ 1137 1138 if (bp->b_dirtyend > bcount) { 1139 nfs_printf("NFS append race @%lx:%d\n", 1140 (long)bp->b_blkno * DEV_BSIZE, 1141 bp->b_dirtyend - bcount); 1142 bp->b_dirtyend = bcount; 1143 } 1144 1145 if (bp->b_dirtyoff >= bp->b_dirtyend) 1146 bp->b_dirtyoff = bp->b_dirtyend = 0; 1147 1148 /* 1149 * If the new write will leave a contiguous dirty 1150 * area, just update the b_dirtyoff and b_dirtyend, 1151 * otherwise force a write rpc of the old dirty area. 1152 * 1153 * While it is possible to merge discontiguous writes due to 1154 * our having a B_CACHE buffer ( and thus valid read data 1155 * for the hole), we don't because it could lead to 1156 * significant cache coherency problems with multiple clients, 1157 * especially if locking is implemented later on. 1158 * 1159 * as an optimization we could theoretically maintain 1160 * a linked list of discontinuous areas, but we would still 1161 * have to commit them separately so there isn't much 1162 * advantage to it except perhaps a bit of asynchronization. 1163 */ 1164 1165 if (bp->b_dirtyend > 0 && 1166 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1167 if (bwrite(bp) == EINTR) { 1168 error = EINTR; 1169 break; 1170 } 1171 goto again; 1172 } 1173 1174 error = uiomove((char *)bp->b_data + on, n, uio); 1175 1176 /* 1177 * Since this block is being modified, it must be written 1178 * again and not just committed. Since write clustering does 1179 * not work for the stage 1 data write, only the stage 2 1180 * commit rpc, we have to clear B_CLUSTEROK as well. 1181 */ 1182 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1183 1184 if (error) { 1185 bp->b_ioflags |= BIO_ERROR; 1186 brelse(bp); 1187 break; 1188 } 1189 1190 /* 1191 * Only update dirtyoff/dirtyend if not a degenerate 1192 * condition. 1193 */ 1194 if (n) { 1195 if (bp->b_dirtyend > 0) { 1196 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1197 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1198 } else { 1199 bp->b_dirtyoff = on; 1200 bp->b_dirtyend = on + n; 1201 } 1202 vfs_bio_set_valid(bp, on, n); 1203 } 1204 1205 /* 1206 * If IO_SYNC do bwrite(). 1207 * 1208 * IO_INVAL appears to be unused. The idea appears to be 1209 * to turn off caching in this case. Very odd. XXX 1210 */ 1211 if ((ioflag & IO_SYNC)) { 1212 if (ioflag & IO_INVAL) 1213 bp->b_flags |= B_NOCACHE; 1214 error = bwrite(bp); 1215 if (error) 1216 break; 1217 } else if ((n + on) == biosize) { 1218 bp->b_flags |= B_ASYNC; 1219 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL); 1220 } else { 1221 bdwrite(bp); 1222 } 1223 } while (uio->uio_resid > 0 && n > 0); 1224 1225 return (error); 1226} 1227 1228/* 1229 * Get an nfs cache block. 1230 * 1231 * Allocate a new one if the block isn't currently in the cache 1232 * and return the block marked busy. If the calling process is 1233 * interrupted by a signal for an interruptible mount point, return 1234 * NULL. 1235 * 1236 * The caller must carefully deal with the possible B_INVAL state of 1237 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1238 * indirectly), so synchronous reads can be issued without worrying about 1239 * the B_INVAL state. We have to be a little more careful when dealing 1240 * with writes (see comments in nfs_write()) when extending a file past 1241 * its EOF. 1242 */ 1243static struct buf * 1244nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1245{ 1246 struct buf *bp; 1247 struct mount *mp; 1248 struct nfsmount *nmp; 1249 1250 mp = vp->v_mount; 1251 nmp = VFSTONFS(mp); 1252 1253 if (nmp->nm_flag & NFSMNT_INT) { 1254 sigset_t oldset; 1255 1256 nfs_set_sigmask(td, &oldset); 1257 bp = getblk(vp, bn, size, PCATCH, 0, 0); 1258 nfs_restore_sigmask(td, &oldset); 1259 while (bp == NULL) { 1260 if (nfs_sigintr(nmp, td)) 1261 return (NULL); 1262 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1263 } 1264 } else { 1265 bp = getblk(vp, bn, size, 0, 0, 0); 1266 } 1267 1268 if (vp->v_type == VREG) { 1269 int biosize; 1270 1271 biosize = mp->mnt_stat.f_iosize; 1272 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1273 } 1274 return (bp); 1275} 1276 1277/* 1278 * Flush and invalidate all dirty buffers. If another process is already 1279 * doing the flush, just wait for completion. 1280 */ 1281int 1282nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1283{ 1284 struct nfsnode *np = VTONFS(vp); 1285 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1286 int error = 0, slpflag, slptimeo; 1287 int old_lock = 0; 1288 1289 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf"); 1290 1291 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1292 intrflg = 0; 1293 if (intrflg) { 1294 slpflag = PCATCH; 1295 slptimeo = 2 * hz; 1296 } else { 1297 slpflag = 0; 1298 slptimeo = 0; 1299 } 1300 1301 old_lock = nfs_upgrade_vnlock(vp); 1302 if (vp->v_iflag & VI_DOOMED) { 1303 /* 1304 * Since vgonel() uses the generic vinvalbuf() to flush 1305 * dirty buffers and it does not call this function, it 1306 * is safe to just return OK when VI_DOOMED is set. 1307 */ 1308 nfs_downgrade_vnlock(vp, old_lock); 1309 return (0); 1310 } 1311 1312 /* 1313 * Now, flush as required. 1314 */ 1315 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1316 VM_OBJECT_LOCK(vp->v_bufobj.bo_object); 1317 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1318 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); 1319 /* 1320 * If the page clean was interrupted, fail the invalidation. 1321 * Not doing so, we run the risk of losing dirty pages in the 1322 * vinvalbuf() call below. 1323 */ 1324 if (intrflg && (error = nfs_sigintr(nmp, td))) 1325 goto out; 1326 } 1327 1328 error = vinvalbuf(vp, flags, slpflag, 0); 1329 while (error) { 1330 if (intrflg && (error = nfs_sigintr(nmp, td))) 1331 goto out; 1332 error = vinvalbuf(vp, flags, 0, slptimeo); 1333 } 1334 mtx_lock(&np->n_mtx); 1335 if (np->n_directio_asyncwr == 0) 1336 np->n_flag &= ~NMODIFIED; 1337 mtx_unlock(&np->n_mtx); 1338out: 1339 nfs_downgrade_vnlock(vp, old_lock); 1340 return error; 1341} 1342 1343/* 1344 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1345 * This is mainly to avoid queueing async I/O requests when the nfsiods 1346 * are all hung on a dead server. 1347 * 1348 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1349 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1350 */ 1351int 1352nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1353{ 1354 int iod; 1355 int gotiod; 1356 int slpflag = 0; 1357 int slptimeo = 0; 1358 int error, error2; 1359 1360 /* 1361 * Commits are usually short and sweet so lets save some cpu and 1362 * leave the async daemons for more important rpc's (such as reads 1363 * and writes). 1364 */ 1365 mtx_lock(&nfs_iod_mtx); 1366 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1367 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1368 mtx_unlock(&nfs_iod_mtx); 1369 return(EIO); 1370 } 1371again: 1372 if (nmp->nm_flag & NFSMNT_INT) 1373 slpflag = PCATCH; 1374 gotiod = FALSE; 1375 1376 /* 1377 * Find a free iod to process this request. 1378 */ 1379 for (iod = 0; iod < nfs_numasync; iod++) 1380 if (nfs_iodwant[iod]) { 1381 gotiod = TRUE; 1382 break; 1383 } 1384 1385 /* 1386 * Try to create one if none are free. 1387 */ 1388 if (!gotiod) { 1389 iod = nfs_nfsiodnew(); 1390 if (iod != -1) 1391 gotiod = TRUE; 1392 } 1393 1394 if (gotiod) { 1395 /* 1396 * Found one, so wake it up and tell it which 1397 * mount to process. 1398 */ 1399 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n", 1400 iod, nmp)); 1401 nfs_iodwant[iod] = NULL; 1402 nfs_iodmount[iod] = nmp; 1403 nmp->nm_bufqiods++; 1404 wakeup(&nfs_iodwant[iod]); 1405 } 1406 1407 /* 1408 * If none are free, we may already have an iod working on this mount 1409 * point. If so, it will process our request. 1410 */ 1411 if (!gotiod) { 1412 if (nmp->nm_bufqiods > 0) { 1413 NFS_DPF(ASYNCIO, 1414 ("nfs_asyncio: %d iods are already processing mount %p\n", 1415 nmp->nm_bufqiods, nmp)); 1416 gotiod = TRUE; 1417 } 1418 } 1419 1420 /* 1421 * If we have an iod which can process the request, then queue 1422 * the buffer. 1423 */ 1424 if (gotiod) { 1425 /* 1426 * Ensure that the queue never grows too large. We still want 1427 * to asynchronize so we block rather then return EIO. 1428 */ 1429 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1430 NFS_DPF(ASYNCIO, 1431 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1432 nmp->nm_bufqwant = TRUE; 1433 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx, 1434 slpflag | PRIBIO, 1435 "nfsaio", slptimeo); 1436 if (error) { 1437 error2 = nfs_sigintr(nmp, td); 1438 if (error2) { 1439 mtx_unlock(&nfs_iod_mtx); 1440 return (error2); 1441 } 1442 if (slpflag == PCATCH) { 1443 slpflag = 0; 1444 slptimeo = 2 * hz; 1445 } 1446 } 1447 /* 1448 * We might have lost our iod while sleeping, 1449 * so check and loop if nescessary. 1450 */ 1451 if (nmp->nm_bufqiods == 0) { 1452 NFS_DPF(ASYNCIO, 1453 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1454 goto again; 1455 } 1456 } 1457 1458 /* We might have lost our nfsiod */ 1459 if (nmp->nm_bufqiods == 0) { 1460 NFS_DPF(ASYNCIO, 1461 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1462 goto again; 1463 } 1464 1465 if (bp->b_iocmd == BIO_READ) { 1466 if (bp->b_rcred == NOCRED && cred != NOCRED) 1467 bp->b_rcred = crhold(cred); 1468 } else { 1469 if (bp->b_wcred == NOCRED && cred != NOCRED) 1470 bp->b_wcred = crhold(cred); 1471 } 1472 1473 if (bp->b_flags & B_REMFREE) 1474 bremfreef(bp); 1475 BUF_KERNPROC(bp); 1476 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1477 nmp->nm_bufqlen++; 1478 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1479 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1480 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1481 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1482 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1483 } 1484 mtx_unlock(&nfs_iod_mtx); 1485 return (0); 1486 } 1487 1488 mtx_unlock(&nfs_iod_mtx); 1489 1490 /* 1491 * All the iods are busy on other mounts, so return EIO to 1492 * force the caller to process the i/o synchronously. 1493 */ 1494 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1495 return (EIO); 1496} 1497 1498void 1499nfs_doio_directwrite(struct buf *bp) 1500{ 1501 int iomode, must_commit; 1502 struct uio *uiop = (struct uio *)bp->b_caller1; 1503 char *iov_base = uiop->uio_iov->iov_base; 1504 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount); 1505 1506 iomode = NFSV3WRITE_FILESYNC; 1507 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1508 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit); 1509 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write")); 1510 free(iov_base, M_NFSDIRECTIO); 1511 free(uiop->uio_iov, M_NFSDIRECTIO); 1512 free(uiop, M_NFSDIRECTIO); 1513 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1514 struct nfsnode *np = VTONFS(bp->b_vp); 1515 mtx_lock(&np->n_mtx); 1516 np->n_directio_asyncwr--; 1517 if (np->n_directio_asyncwr == 0) { 1518 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED; 1519 if ((np->n_flag & NFSYNCWAIT)) { 1520 np->n_flag &= ~NFSYNCWAIT; 1521 wakeup((caddr_t)&np->n_directio_asyncwr); 1522 } 1523 } 1524 mtx_unlock(&np->n_mtx); 1525 } 1526 bp->b_vp = NULL; 1527 relpbuf(bp, &nfs_pbuf_freecnt); 1528} 1529 1530/* 1531 * Do an I/O operation to/from a cache block. This may be called 1532 * synchronously or from an nfsiod. 1533 */ 1534int 1535nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) 1536{ 1537 struct uio *uiop; 1538 struct nfsnode *np; 1539 struct nfsmount *nmp; 1540 int error = 0, iomode, must_commit = 0; 1541 struct uio uio; 1542 struct iovec io; 1543 struct proc *p = td ? td->td_proc : NULL; 1544 uint8_t iocmd; 1545 1546 np = VTONFS(vp); 1547 nmp = VFSTONFS(vp->v_mount); 1548 uiop = &uio; 1549 uiop->uio_iov = &io; 1550 uiop->uio_iovcnt = 1; 1551 uiop->uio_segflg = UIO_SYSSPACE; 1552 uiop->uio_td = td; 1553 1554 /* 1555 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1556 * do this here so we do not have to do it in all the code that 1557 * calls us. 1558 */ 1559 bp->b_flags &= ~B_INVAL; 1560 bp->b_ioflags &= ~BIO_ERROR; 1561 1562 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1563 iocmd = bp->b_iocmd; 1564 if (iocmd == BIO_READ) { 1565 io.iov_len = uiop->uio_resid = bp->b_bcount; 1566 io.iov_base = bp->b_data; 1567 uiop->uio_rw = UIO_READ; 1568 1569 switch (vp->v_type) { 1570 case VREG: 1571 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1572 nfsstats.read_bios++; 1573 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr); 1574 1575 if (!error) { 1576 if (uiop->uio_resid) { 1577 /* 1578 * If we had a short read with no error, we must have 1579 * hit a file hole. We should zero-fill the remainder. 1580 * This can also occur if the server hits the file EOF. 1581 * 1582 * Holes used to be able to occur due to pending 1583 * writes, but that is not possible any longer. 1584 */ 1585 int nread = bp->b_bcount - uiop->uio_resid; 1586 int left = uiop->uio_resid; 1587 1588 if (left > 0) 1589 bzero((char *)bp->b_data + nread, left); 1590 uiop->uio_resid = 0; 1591 } 1592 } 1593 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */ 1594 if (p && (vp->v_vflag & VV_TEXT)) { 1595 mtx_lock(&np->n_mtx); 1596 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) { 1597 mtx_unlock(&np->n_mtx); 1598 PROC_LOCK(p); 1599 killproc(p, "text file modification"); 1600 PROC_UNLOCK(p); 1601 } else 1602 mtx_unlock(&np->n_mtx); 1603 } 1604 break; 1605 case VLNK: 1606 uiop->uio_offset = (off_t)0; 1607 nfsstats.readlink_bios++; 1608 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr); 1609 break; 1610 case VDIR: 1611 nfsstats.readdir_bios++; 1612 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1613 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1614 error = nfs_readdirplusrpc(vp, uiop, cr); 1615 if (error == NFSERR_NOTSUPP) 1616 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1617 } 1618 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1619 error = nfs_readdirrpc(vp, uiop, cr); 1620 /* 1621 * end-of-directory sets B_INVAL but does not generate an 1622 * error. 1623 */ 1624 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1625 bp->b_flags |= B_INVAL; 1626 break; 1627 default: 1628 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type); 1629 break; 1630 }; 1631 if (error) { 1632 bp->b_ioflags |= BIO_ERROR; 1633 bp->b_error = error; 1634 } 1635 } else { 1636 /* 1637 * If we only need to commit, try to commit 1638 */ 1639 if (bp->b_flags & B_NEEDCOMMIT) { 1640 int retv; 1641 off_t off; 1642 1643 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1644 retv = (nmp->nm_rpcops->nr_commit)( 1645 vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1646 bp->b_wcred, td); 1647 if (retv == 0) { 1648 bp->b_dirtyoff = bp->b_dirtyend = 0; 1649 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1650 bp->b_resid = 0; 1651 bufdone(bp); 1652 return (0); 1653 } 1654 if (retv == NFSERR_STALEWRITEVERF) { 1655 nfs_clearcommit(vp->v_mount); 1656 } 1657 } 1658 1659 /* 1660 * Setup for actual write 1661 */ 1662 mtx_lock(&np->n_mtx); 1663 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1664 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1665 mtx_unlock(&np->n_mtx); 1666 1667 if (bp->b_dirtyend > bp->b_dirtyoff) { 1668 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1669 - bp->b_dirtyoff; 1670 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1671 + bp->b_dirtyoff; 1672 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1673 uiop->uio_rw = UIO_WRITE; 1674 nfsstats.write_bios++; 1675 1676 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1677 iomode = NFSV3WRITE_UNSTABLE; 1678 else 1679 iomode = NFSV3WRITE_FILESYNC; 1680 1681 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit); 1682 1683 /* 1684 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1685 * to cluster the buffers needing commit. This will allow 1686 * the system to submit a single commit rpc for the whole 1687 * cluster. We can do this even if the buffer is not 100% 1688 * dirty (relative to the NFS blocksize), so we optimize the 1689 * append-to-file-case. 1690 * 1691 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1692 * cleared because write clustering only works for commit 1693 * rpc's, not for the data portion of the write). 1694 */ 1695 1696 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1697 bp->b_flags |= B_NEEDCOMMIT; 1698 if (bp->b_dirtyoff == 0 1699 && bp->b_dirtyend == bp->b_bcount) 1700 bp->b_flags |= B_CLUSTEROK; 1701 } else { 1702 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1703 } 1704 1705 /* 1706 * For an interrupted write, the buffer is still valid 1707 * and the write hasn't been pushed to the server yet, 1708 * so we can't set BIO_ERROR and report the interruption 1709 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1710 * is not relevant, so the rpc attempt is essentially 1711 * a noop. For the case of a V3 write rpc not being 1712 * committed to stable storage, the block is still 1713 * dirty and requires either a commit rpc or another 1714 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1715 * the block is reused. This is indicated by setting 1716 * the B_DELWRI and B_NEEDCOMMIT flags. 1717 * 1718 * If the buffer is marked B_PAGING, it does not reside on 1719 * the vp's paging queues so we cannot call bdirty(). The 1720 * bp in this case is not an NFS cache block so we should 1721 * be safe. XXX 1722 * 1723 * The logic below breaks up errors into recoverable and 1724 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1725 * and keep the buffer around for potential write retries. 1726 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1727 * and save the error in the nfsnode. This is less than ideal 1728 * but necessary. Keeping such buffers around could potentially 1729 * cause buffer exhaustion eventually (they can never be written 1730 * out, so will get constantly be re-dirtied). It also causes 1731 * all sorts of vfs panics. For non-recoverable write errors, 1732 * also invalidate the attrcache, so we'll be forced to go over 1733 * the wire for this object, returning an error to user on next 1734 * call (most of the time). 1735 */ 1736 if (error == EINTR || error == EIO || error == ETIMEDOUT 1737 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1738 int s; 1739 1740 s = splbio(); 1741 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1742 if ((bp->b_flags & B_PAGING) == 0) { 1743 bdirty(bp); 1744 bp->b_flags &= ~B_DONE; 1745 } 1746 if (error && (bp->b_flags & B_ASYNC) == 0) 1747 bp->b_flags |= B_EINTR; 1748 splx(s); 1749 } else { 1750 if (error) { 1751 bp->b_ioflags |= BIO_ERROR; 1752 bp->b_flags |= B_INVAL; 1753 bp->b_error = np->n_error = error; 1754 mtx_lock(&np->n_mtx); 1755 np->n_flag |= NWRITEERR; 1756 np->n_attrstamp = 0; 1757 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1758 mtx_unlock(&np->n_mtx); 1759 } 1760 bp->b_dirtyoff = bp->b_dirtyend = 0; 1761 } 1762 } else { 1763 bp->b_resid = 0; 1764 bufdone(bp); 1765 return (0); 1766 } 1767 } 1768 bp->b_resid = uiop->uio_resid; 1769 if (must_commit) 1770 nfs_clearcommit(vp->v_mount); 1771 bufdone(bp); 1772 return (error); 1773} 1774 1775/* 1776 * Used to aid in handling ftruncate() operations on the NFS client side. 1777 * Truncation creates a number of special problems for NFS. We have to 1778 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1779 * we have to properly handle VM pages or (potentially dirty) buffers 1780 * that straddle the truncation point. 1781 */ 1782 1783int 1784nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1785{ 1786 struct nfsnode *np = VTONFS(vp); 1787 u_quad_t tsize; 1788 int biosize = vp->v_mount->mnt_stat.f_iosize; 1789 int error = 0; 1790 1791 mtx_lock(&np->n_mtx); 1792 tsize = np->n_size; 1793 np->n_size = nsize; 1794 mtx_unlock(&np->n_mtx); 1795 1796 if (nsize < tsize) { 1797 struct buf *bp; 1798 daddr_t lbn; 1799 int bufsize; 1800 1801 /* 1802 * vtruncbuf() doesn't get the buffer overlapping the 1803 * truncation point. We may have a B_DELWRI and/or B_CACHE 1804 * buffer that now needs to be truncated. 1805 */ 1806 error = vtruncbuf(vp, cred, td, nsize, biosize); 1807 lbn = nsize / biosize; 1808 bufsize = nsize & (biosize - 1); 1809 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1810 if (!bp) 1811 return EINTR; 1812 if (bp->b_dirtyoff > bp->b_bcount) 1813 bp->b_dirtyoff = bp->b_bcount; 1814 if (bp->b_dirtyend > bp->b_bcount) 1815 bp->b_dirtyend = bp->b_bcount; 1816 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1817 brelse(bp); 1818 } else { 1819 vnode_pager_setsize(vp, nsize); 1820 } 1821 return(error); 1822} 1823 1824