nfs_bio.c revision 190380
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 190380 2009-03-24 17:14:34Z rwatson $"); 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/mount.h> 46#include <sys/proc.h> 47#include <sys/resourcevar.h> 48#include <sys/signalvar.h> 49#include <sys/vmmeter.h> 50#include <sys/vnode.h> 51 52#include <vm/vm.h> 53#include <vm/vm_extern.h> 54#include <vm/vm_page.h> 55#include <vm/vm_object.h> 56#include <vm/vm_pager.h> 57#include <vm/vnode_pager.h> 58 59#include <rpc/rpcclnt.h> 60 61#include <nfs/rpcv2.h> 62#include <nfs/nfsproto.h> 63#include <nfsclient/nfs.h> 64#include <nfsclient/nfsmount.h> 65#include <nfsclient/nfsnode.h> 66#include <nfsclient/nfs_kdtrace.h> 67 68#include <nfs4client/nfs4.h> 69 70static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 71 struct thread *td); 72static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 73 struct ucred *cred, int ioflag); 74 75extern int nfs_directio_enable; 76extern int nfs_directio_allow_mmap; 77 78/* 79 * Vnode op for VM getpages. 80 */ 81int 82nfs_getpages(struct vop_getpages_args *ap) 83{ 84 int i, error, nextoff, size, toff, count, npages; 85 struct uio uio; 86 struct iovec iov; 87 vm_offset_t kva; 88 struct buf *bp; 89 struct vnode *vp; 90 struct thread *td; 91 struct ucred *cred; 92 struct nfsmount *nmp; 93 vm_object_t object; 94 vm_page_t *pages; 95 struct nfsnode *np; 96 97 vp = ap->a_vp; 98 np = VTONFS(vp); 99 td = curthread; /* XXX */ 100 cred = curthread->td_ucred; /* XXX */ 101 nmp = VFSTONFS(vp->v_mount); 102 pages = ap->a_m; 103 count = ap->a_count; 104 105 if ((object = vp->v_object) == NULL) { 106 nfs_printf("nfs_getpages: called with non-merged cache vnode??\n"); 107 return VM_PAGER_ERROR; 108 } 109 110 if (nfs_directio_enable && !nfs_directio_allow_mmap) { 111 mtx_lock(&np->n_mtx); 112 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 113 mtx_unlock(&np->n_mtx); 114 nfs_printf("nfs_getpages: called on non-cacheable vnode??\n"); 115 return VM_PAGER_ERROR; 116 } else 117 mtx_unlock(&np->n_mtx); 118 } 119 120 mtx_lock(&nmp->nm_mtx); 121 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 122 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 123 mtx_unlock(&nmp->nm_mtx); 124 /* We'll never get here for v4, because we always have fsinfo */ 125 (void)nfs_fsinfo(nmp, vp, cred, td); 126 } else 127 mtx_unlock(&nmp->nm_mtx); 128 129 npages = btoc(count); 130 131 /* 132 * If the requested page is partially valid, just return it and 133 * allow the pager to zero-out the blanks. Partially valid pages 134 * can only occur at the file EOF. 135 */ 136 137 { 138 vm_page_t m = pages[ap->a_reqpage]; 139 140 VM_OBJECT_LOCK(object); 141 vm_page_lock_queues(); 142 if (m->valid != 0) { 143 /* handled by vm_fault now */ 144 /* vm_page_zero_invalid(m, TRUE); */ 145 for (i = 0; i < npages; ++i) { 146 if (i != ap->a_reqpage) 147 vm_page_free(pages[i]); 148 } 149 vm_page_unlock_queues(); 150 VM_OBJECT_UNLOCK(object); 151 return(0); 152 } 153 vm_page_unlock_queues(); 154 VM_OBJECT_UNLOCK(object); 155 } 156 157 /* 158 * We use only the kva address for the buffer, but this is extremely 159 * convienient and fast. 160 */ 161 bp = getpbuf(&nfs_pbuf_freecnt); 162 163 kva = (vm_offset_t) bp->b_data; 164 pmap_qenter(kva, pages, npages); 165 PCPU_INC(cnt.v_vnodein); 166 PCPU_ADD(cnt.v_vnodepgsin, npages); 167 168 iov.iov_base = (caddr_t) kva; 169 iov.iov_len = count; 170 uio.uio_iov = &iov; 171 uio.uio_iovcnt = 1; 172 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 173 uio.uio_resid = count; 174 uio.uio_segflg = UIO_SYSSPACE; 175 uio.uio_rw = UIO_READ; 176 uio.uio_td = td; 177 178 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred); 179 pmap_qremove(kva, npages); 180 181 relpbuf(bp, &nfs_pbuf_freecnt); 182 183 if (error && (uio.uio_resid == count)) { 184 nfs_printf("nfs_getpages: error %d\n", error); 185 VM_OBJECT_LOCK(object); 186 vm_page_lock_queues(); 187 for (i = 0; i < npages; ++i) { 188 if (i != ap->a_reqpage) 189 vm_page_free(pages[i]); 190 } 191 vm_page_unlock_queues(); 192 VM_OBJECT_UNLOCK(object); 193 return VM_PAGER_ERROR; 194 } 195 196 /* 197 * Calculate the number of bytes read and validate only that number 198 * of bytes. Note that due to pending writes, size may be 0. This 199 * does not mean that the remaining data is invalid! 200 */ 201 202 size = count - uio.uio_resid; 203 VM_OBJECT_LOCK(object); 204 vm_page_lock_queues(); 205 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 206 vm_page_t m; 207 nextoff = toff + PAGE_SIZE; 208 m = pages[i]; 209 210 if (nextoff <= size) { 211 /* 212 * Read operation filled an entire page 213 */ 214 m->valid = VM_PAGE_BITS_ALL; 215 vm_page_undirty(m); 216 } else if (size > toff) { 217 /* 218 * Read operation filled a partial page. 219 */ 220 m->valid = 0; 221 vm_page_set_validclean(m, 0, size - toff); 222 /* handled by vm_fault now */ 223 /* vm_page_zero_invalid(m, TRUE); */ 224 } else { 225 /* 226 * Read operation was short. If no error occured 227 * we may have hit a zero-fill section. We simply 228 * leave valid set to 0. 229 */ 230 ; 231 } 232 if (i != ap->a_reqpage) { 233 /* 234 * Whether or not to leave the page activated is up in 235 * the air, but we should put the page on a page queue 236 * somewhere (it already is in the object). Result: 237 * It appears that emperical results show that 238 * deactivating pages is best. 239 */ 240 241 /* 242 * Just in case someone was asking for this page we 243 * now tell them that it is ok to use. 244 */ 245 if (!error) { 246 if (m->oflags & VPO_WANTED) 247 vm_page_activate(m); 248 else 249 vm_page_deactivate(m); 250 vm_page_wakeup(m); 251 } else { 252 vm_page_free(m); 253 } 254 } 255 } 256 vm_page_unlock_queues(); 257 VM_OBJECT_UNLOCK(object); 258 return 0; 259} 260 261/* 262 * Vnode op for VM putpages. 263 */ 264int 265nfs_putpages(struct vop_putpages_args *ap) 266{ 267 struct uio uio; 268 struct iovec iov; 269 vm_offset_t kva; 270 struct buf *bp; 271 int iomode, must_commit, i, error, npages, count; 272 off_t offset; 273 int *rtvals; 274 struct vnode *vp; 275 struct thread *td; 276 struct ucred *cred; 277 struct nfsmount *nmp; 278 struct nfsnode *np; 279 vm_page_t *pages; 280 281 vp = ap->a_vp; 282 np = VTONFS(vp); 283 td = curthread; /* XXX */ 284 cred = curthread->td_ucred; /* XXX */ 285 nmp = VFSTONFS(vp->v_mount); 286 pages = ap->a_m; 287 count = ap->a_count; 288 rtvals = ap->a_rtvals; 289 npages = btoc(count); 290 offset = IDX_TO_OFF(pages[0]->pindex); 291 292 mtx_lock(&nmp->nm_mtx); 293 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 294 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 295 mtx_unlock(&nmp->nm_mtx); 296 (void)nfs_fsinfo(nmp, vp, cred, td); 297 } else 298 mtx_unlock(&nmp->nm_mtx); 299 300 mtx_lock(&np->n_mtx); 301 if (nfs_directio_enable && !nfs_directio_allow_mmap && 302 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 303 mtx_unlock(&np->n_mtx); 304 nfs_printf("nfs_putpages: called on noncache-able vnode??\n"); 305 mtx_lock(&np->n_mtx); 306 } 307 308 for (i = 0; i < npages; i++) 309 rtvals[i] = VM_PAGER_AGAIN; 310 311 /* 312 * When putting pages, do not extend file past EOF. 313 */ 314 if (offset + count > np->n_size) { 315 count = np->n_size - offset; 316 if (count < 0) 317 count = 0; 318 } 319 mtx_unlock(&np->n_mtx); 320 321 /* 322 * We use only the kva address for the buffer, but this is extremely 323 * convienient and fast. 324 */ 325 bp = getpbuf(&nfs_pbuf_freecnt); 326 327 kva = (vm_offset_t) bp->b_data; 328 pmap_qenter(kva, pages, npages); 329 PCPU_INC(cnt.v_vnodeout); 330 PCPU_ADD(cnt.v_vnodepgsout, count); 331 332 iov.iov_base = (caddr_t) kva; 333 iov.iov_len = count; 334 uio.uio_iov = &iov; 335 uio.uio_iovcnt = 1; 336 uio.uio_offset = offset; 337 uio.uio_resid = count; 338 uio.uio_segflg = UIO_SYSSPACE; 339 uio.uio_rw = UIO_WRITE; 340 uio.uio_td = td; 341 342 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 343 iomode = NFSV3WRITE_UNSTABLE; 344 else 345 iomode = NFSV3WRITE_FILESYNC; 346 347 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit); 348 349 pmap_qremove(kva, npages); 350 relpbuf(bp, &nfs_pbuf_freecnt); 351 352 if (!error) { 353 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 354 for (i = 0; i < nwritten; i++) { 355 rtvals[i] = VM_PAGER_OK; 356 vm_page_undirty(pages[i]); 357 } 358 if (must_commit) { 359 nfs_clearcommit(vp->v_mount); 360 } 361 } 362 return rtvals[0]; 363} 364 365/* 366 * For nfs, cache consistency can only be maintained approximately. 367 * Although RFC1094 does not specify the criteria, the following is 368 * believed to be compatible with the reference port. 369 * For nfs: 370 * If the file's modify time on the server has changed since the 371 * last read rpc or you have written to the file, 372 * you may have lost data cache consistency with the 373 * server, so flush all of the file's data out of the cache. 374 * Then force a getattr rpc to ensure that you have up to date 375 * attributes. 376 * NB: This implies that cache data can be read when up to 377 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 378 * attributes this could be forced by setting n_attrstamp to 0 before 379 * the VOP_GETATTR() call. 380 */ 381static inline int 382nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 383{ 384 int error = 0; 385 struct vattr vattr; 386 struct nfsnode *np = VTONFS(vp); 387 int old_lock; 388 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 389 390 /* 391 * Grab the exclusive lock before checking whether the cache is 392 * consistent. 393 * XXX - We can make this cheaper later (by acquiring cheaper locks). 394 * But for now, this suffices. 395 */ 396 old_lock = nfs_upgrade_vnlock(vp); 397 mtx_lock(&np->n_mtx); 398 if (np->n_flag & NMODIFIED) { 399 mtx_unlock(&np->n_mtx); 400 if (vp->v_type != VREG) { 401 if (vp->v_type != VDIR) 402 panic("nfs: bioread, not dir"); 403 (nmp->nm_rpcops->nr_invaldir)(vp); 404 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 405 if (error) 406 goto out; 407 } 408 np->n_attrstamp = 0; 409 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 410 error = VOP_GETATTR(vp, &vattr, cred); 411 if (error) 412 goto out; 413 mtx_lock(&np->n_mtx); 414 np->n_mtime = vattr.va_mtime; 415 mtx_unlock(&np->n_mtx); 416 } else { 417 mtx_unlock(&np->n_mtx); 418 error = VOP_GETATTR(vp, &vattr, cred); 419 if (error) 420 return (error); 421 mtx_lock(&np->n_mtx); 422 if ((np->n_flag & NSIZECHANGED) 423 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 424 mtx_unlock(&np->n_mtx); 425 if (vp->v_type == VDIR) 426 (nmp->nm_rpcops->nr_invaldir)(vp); 427 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 428 if (error) 429 goto out; 430 mtx_lock(&np->n_mtx); 431 np->n_mtime = vattr.va_mtime; 432 np->n_flag &= ~NSIZECHANGED; 433 } 434 mtx_unlock(&np->n_mtx); 435 } 436out: 437 nfs_downgrade_vnlock(vp, old_lock); 438 return error; 439} 440 441/* 442 * Vnode op for read using bio 443 */ 444int 445nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 446{ 447 struct nfsnode *np = VTONFS(vp); 448 int biosize, i; 449 struct buf *bp, *rabp; 450 struct thread *td; 451 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 452 daddr_t lbn, rabn; 453 int bcount; 454 int seqcount; 455 int nra, error = 0, n = 0, on = 0; 456 457#ifdef DIAGNOSTIC 458 if (uio->uio_rw != UIO_READ) 459 panic("nfs_read mode"); 460#endif 461 if (uio->uio_resid == 0) 462 return (0); 463 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 464 return (EINVAL); 465 td = uio->uio_td; 466 467 mtx_lock(&nmp->nm_mtx); 468 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 469 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 470 mtx_unlock(&nmp->nm_mtx); 471 (void)nfs_fsinfo(nmp, vp, cred, td); 472 } else 473 mtx_unlock(&nmp->nm_mtx); 474 475 if (vp->v_type != VDIR && 476 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 477 return (EFBIG); 478 479 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 480 /* No caching/ no readaheads. Just read data into the user buffer */ 481 return nfs_readrpc(vp, uio, cred); 482 483 biosize = vp->v_mount->mnt_stat.f_iosize; 484 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 485 486 error = nfs_bioread_check_cons(vp, td, cred); 487 if (error) 488 return error; 489 490 do { 491 u_quad_t nsize; 492 493 mtx_lock(&np->n_mtx); 494 nsize = np->n_size; 495 mtx_unlock(&np->n_mtx); 496 497 switch (vp->v_type) { 498 case VREG: 499 nfsstats.biocache_reads++; 500 lbn = uio->uio_offset / biosize; 501 on = uio->uio_offset & (biosize - 1); 502 503 /* 504 * Start the read ahead(s), as required. 505 */ 506 if (nmp->nm_readahead > 0) { 507 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 508 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 509 rabn = lbn + 1 + nra; 510 if (incore(&vp->v_bufobj, rabn) == NULL) { 511 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 512 if (!rabp) { 513 error = nfs_sigintr(nmp, NULL, td); 514 return (error ? error : EINTR); 515 } 516 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 517 rabp->b_flags |= B_ASYNC; 518 rabp->b_iocmd = BIO_READ; 519 vfs_busy_pages(rabp, 0); 520 if (nfs_asyncio(nmp, rabp, cred, td)) { 521 rabp->b_flags |= B_INVAL; 522 rabp->b_ioflags |= BIO_ERROR; 523 vfs_unbusy_pages(rabp); 524 brelse(rabp); 525 break; 526 } 527 } else { 528 brelse(rabp); 529 } 530 } 531 } 532 } 533 534 /* Note that bcount is *not* DEV_BSIZE aligned. */ 535 bcount = biosize; 536 if ((off_t)lbn * biosize >= nsize) { 537 bcount = 0; 538 } else if ((off_t)(lbn + 1) * biosize > nsize) { 539 bcount = nsize - (off_t)lbn * biosize; 540 } 541 bp = nfs_getcacheblk(vp, lbn, bcount, td); 542 543 if (!bp) { 544 error = nfs_sigintr(nmp, NULL, td); 545 return (error ? error : EINTR); 546 } 547 548 /* 549 * If B_CACHE is not set, we must issue the read. If this 550 * fails, we return an error. 551 */ 552 553 if ((bp->b_flags & B_CACHE) == 0) { 554 bp->b_iocmd = BIO_READ; 555 vfs_busy_pages(bp, 0); 556 error = nfs_doio(vp, bp, cred, td); 557 if (error) { 558 brelse(bp); 559 return (error); 560 } 561 } 562 563 /* 564 * on is the offset into the current bp. Figure out how many 565 * bytes we can copy out of the bp. Note that bcount is 566 * NOT DEV_BSIZE aligned. 567 * 568 * Then figure out how many bytes we can copy into the uio. 569 */ 570 571 n = 0; 572 if (on < bcount) 573 n = min((unsigned)(bcount - on), uio->uio_resid); 574 break; 575 case VLNK: 576 nfsstats.biocache_readlinks++; 577 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 578 if (!bp) { 579 error = nfs_sigintr(nmp, NULL, td); 580 return (error ? error : EINTR); 581 } 582 if ((bp->b_flags & B_CACHE) == 0) { 583 bp->b_iocmd = BIO_READ; 584 vfs_busy_pages(bp, 0); 585 error = nfs_doio(vp, bp, cred, td); 586 if (error) { 587 bp->b_ioflags |= BIO_ERROR; 588 brelse(bp); 589 return (error); 590 } 591 } 592 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 593 on = 0; 594 break; 595 case VDIR: 596 nfsstats.biocache_readdirs++; 597 if (np->n_direofoffset 598 && uio->uio_offset >= np->n_direofoffset) { 599 return (0); 600 } 601 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 602 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 603 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 604 if (!bp) { 605 error = nfs_sigintr(nmp, NULL, td); 606 return (error ? error : EINTR); 607 } 608 if ((bp->b_flags & B_CACHE) == 0) { 609 bp->b_iocmd = BIO_READ; 610 vfs_busy_pages(bp, 0); 611 error = nfs_doio(vp, bp, cred, td); 612 if (error) { 613 brelse(bp); 614 } 615 while (error == NFSERR_BAD_COOKIE) { 616 (nmp->nm_rpcops->nr_invaldir)(vp); 617 error = nfs_vinvalbuf(vp, 0, td, 1); 618 /* 619 * Yuck! The directory has been modified on the 620 * server. The only way to get the block is by 621 * reading from the beginning to get all the 622 * offset cookies. 623 * 624 * Leave the last bp intact unless there is an error. 625 * Loop back up to the while if the error is another 626 * NFSERR_BAD_COOKIE (double yuch!). 627 */ 628 for (i = 0; i <= lbn && !error; i++) { 629 if (np->n_direofoffset 630 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 631 return (0); 632 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 633 if (!bp) { 634 error = nfs_sigintr(nmp, NULL, td); 635 return (error ? error : EINTR); 636 } 637 if ((bp->b_flags & B_CACHE) == 0) { 638 bp->b_iocmd = BIO_READ; 639 vfs_busy_pages(bp, 0); 640 error = nfs_doio(vp, bp, cred, td); 641 /* 642 * no error + B_INVAL == directory EOF, 643 * use the block. 644 */ 645 if (error == 0 && (bp->b_flags & B_INVAL)) 646 break; 647 } 648 /* 649 * An error will throw away the block and the 650 * for loop will break out. If no error and this 651 * is not the block we want, we throw away the 652 * block and go for the next one via the for loop. 653 */ 654 if (error || i < lbn) 655 brelse(bp); 656 } 657 } 658 /* 659 * The above while is repeated if we hit another cookie 660 * error. If we hit an error and it wasn't a cookie error, 661 * we give up. 662 */ 663 if (error) 664 return (error); 665 } 666 667 /* 668 * If not eof and read aheads are enabled, start one. 669 * (You need the current block first, so that you have the 670 * directory offset cookie of the next block.) 671 */ 672 if (nmp->nm_readahead > 0 && 673 (bp->b_flags & B_INVAL) == 0 && 674 (np->n_direofoffset == 0 || 675 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 676 incore(&vp->v_bufobj, lbn + 1) == NULL) { 677 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 678 if (rabp) { 679 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 680 rabp->b_flags |= B_ASYNC; 681 rabp->b_iocmd = BIO_READ; 682 vfs_busy_pages(rabp, 0); 683 if (nfs_asyncio(nmp, rabp, cred, td)) { 684 rabp->b_flags |= B_INVAL; 685 rabp->b_ioflags |= BIO_ERROR; 686 vfs_unbusy_pages(rabp); 687 brelse(rabp); 688 } 689 } else { 690 brelse(rabp); 691 } 692 } 693 } 694 /* 695 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 696 * chopped for the EOF condition, we cannot tell how large 697 * NFS directories are going to be until we hit EOF. So 698 * an NFS directory buffer is *not* chopped to its EOF. Now, 699 * it just so happens that b_resid will effectively chop it 700 * to EOF. *BUT* this information is lost if the buffer goes 701 * away and is reconstituted into a B_CACHE state ( due to 702 * being VMIO ) later. So we keep track of the directory eof 703 * in np->n_direofoffset and chop it off as an extra step 704 * right here. 705 */ 706 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 707 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 708 n = np->n_direofoffset - uio->uio_offset; 709 break; 710 default: 711 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type); 712 bp = NULL; 713 break; 714 }; 715 716 if (n > 0) { 717 error = uiomove(bp->b_data + on, (int)n, uio); 718 } 719 if (vp->v_type == VLNK) 720 n = 0; 721 if (bp != NULL) 722 brelse(bp); 723 } while (error == 0 && uio->uio_resid > 0 && n > 0); 724 return (error); 725} 726 727/* 728 * The NFS write path cannot handle iovecs with len > 1. So we need to 729 * break up iovecs accordingly (restricting them to wsize). 730 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 731 * For the ASYNC case, 2 copies are needed. The first a copy from the 732 * user buffer to a staging buffer and then a second copy from the staging 733 * buffer to mbufs. This can be optimized by copying from the user buffer 734 * directly into mbufs and passing the chain down, but that requires a 735 * fair amount of re-working of the relevant codepaths (and can be done 736 * later). 737 */ 738static int 739nfs_directio_write(vp, uiop, cred, ioflag) 740 struct vnode *vp; 741 struct uio *uiop; 742 struct ucred *cred; 743 int ioflag; 744{ 745 int error; 746 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 747 struct thread *td = uiop->uio_td; 748 int size; 749 int wsize; 750 751 mtx_lock(&nmp->nm_mtx); 752 wsize = nmp->nm_wsize; 753 mtx_unlock(&nmp->nm_mtx); 754 if (ioflag & IO_SYNC) { 755 int iomode, must_commit; 756 struct uio uio; 757 struct iovec iov; 758do_sync: 759 while (uiop->uio_resid > 0) { 760 size = min(uiop->uio_resid, wsize); 761 size = min(uiop->uio_iov->iov_len, size); 762 iov.iov_base = uiop->uio_iov->iov_base; 763 iov.iov_len = size; 764 uio.uio_iov = &iov; 765 uio.uio_iovcnt = 1; 766 uio.uio_offset = uiop->uio_offset; 767 uio.uio_resid = size; 768 uio.uio_segflg = UIO_USERSPACE; 769 uio.uio_rw = UIO_WRITE; 770 uio.uio_td = td; 771 iomode = NFSV3WRITE_FILESYNC; 772 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, 773 &iomode, &must_commit); 774 KASSERT((must_commit == 0), 775 ("nfs_directio_write: Did not commit write")); 776 if (error) 777 return (error); 778 uiop->uio_offset += size; 779 uiop->uio_resid -= size; 780 if (uiop->uio_iov->iov_len <= size) { 781 uiop->uio_iovcnt--; 782 uiop->uio_iov++; 783 } else { 784 uiop->uio_iov->iov_base = 785 (char *)uiop->uio_iov->iov_base + size; 786 uiop->uio_iov->iov_len -= size; 787 } 788 } 789 } else { 790 struct uio *t_uio; 791 struct iovec *t_iov; 792 struct buf *bp; 793 794 /* 795 * Break up the write into blocksize chunks and hand these 796 * over to nfsiod's for write back. 797 * Unfortunately, this incurs a copy of the data. Since 798 * the user could modify the buffer before the write is 799 * initiated. 800 * 801 * The obvious optimization here is that one of the 2 copies 802 * in the async write path can be eliminated by copying the 803 * data here directly into mbufs and passing the mbuf chain 804 * down. But that will require a fair amount of re-working 805 * of the code and can be done if there's enough interest 806 * in NFS directio access. 807 */ 808 while (uiop->uio_resid > 0) { 809 size = min(uiop->uio_resid, wsize); 810 size = min(uiop->uio_iov->iov_len, size); 811 bp = getpbuf(&nfs_pbuf_freecnt); 812 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 813 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 814 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 815 t_iov->iov_len = size; 816 t_uio->uio_iov = t_iov; 817 t_uio->uio_iovcnt = 1; 818 t_uio->uio_offset = uiop->uio_offset; 819 t_uio->uio_resid = size; 820 t_uio->uio_segflg = UIO_SYSSPACE; 821 t_uio->uio_rw = UIO_WRITE; 822 t_uio->uio_td = td; 823 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); 824 bp->b_flags |= B_DIRECT; 825 bp->b_iocmd = BIO_WRITE; 826 if (cred != NOCRED) { 827 crhold(cred); 828 bp->b_wcred = cred; 829 } else 830 bp->b_wcred = NOCRED; 831 bp->b_caller1 = (void *)t_uio; 832 bp->b_vp = vp; 833 error = nfs_asyncio(nmp, bp, NOCRED, td); 834 if (error) { 835 free(t_iov->iov_base, M_NFSDIRECTIO); 836 free(t_iov, M_NFSDIRECTIO); 837 free(t_uio, M_NFSDIRECTIO); 838 bp->b_vp = NULL; 839 relpbuf(bp, &nfs_pbuf_freecnt); 840 if (error == EINTR) 841 return (error); 842 goto do_sync; 843 } 844 uiop->uio_offset += size; 845 uiop->uio_resid -= size; 846 if (uiop->uio_iov->iov_len <= size) { 847 uiop->uio_iovcnt--; 848 uiop->uio_iov++; 849 } else { 850 uiop->uio_iov->iov_base = 851 (char *)uiop->uio_iov->iov_base + size; 852 uiop->uio_iov->iov_len -= size; 853 } 854 } 855 } 856 return (0); 857} 858 859/* 860 * Vnode op for write using bio 861 */ 862int 863nfs_write(struct vop_write_args *ap) 864{ 865 int biosize; 866 struct uio *uio = ap->a_uio; 867 struct thread *td = uio->uio_td; 868 struct vnode *vp = ap->a_vp; 869 struct nfsnode *np = VTONFS(vp); 870 struct ucred *cred = ap->a_cred; 871 int ioflag = ap->a_ioflag; 872 struct buf *bp; 873 struct vattr vattr; 874 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 875 daddr_t lbn; 876 int bcount; 877 int n, on, error = 0; 878 struct proc *p = td?td->td_proc:NULL; 879 880#ifdef DIAGNOSTIC 881 if (uio->uio_rw != UIO_WRITE) 882 panic("nfs_write mode"); 883 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 884 panic("nfs_write proc"); 885#endif 886 if (vp->v_type != VREG) 887 return (EIO); 888 mtx_lock(&np->n_mtx); 889 if (np->n_flag & NWRITEERR) { 890 np->n_flag &= ~NWRITEERR; 891 mtx_unlock(&np->n_mtx); 892 return (np->n_error); 893 } else 894 mtx_unlock(&np->n_mtx); 895 mtx_lock(&nmp->nm_mtx); 896 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 897 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 898 mtx_unlock(&nmp->nm_mtx); 899 (void)nfs_fsinfo(nmp, vp, cred, td); 900 } else 901 mtx_unlock(&nmp->nm_mtx); 902 903 /* 904 * Synchronously flush pending buffers if we are in synchronous 905 * mode or if we are appending. 906 */ 907 if (ioflag & (IO_APPEND | IO_SYNC)) { 908 mtx_lock(&np->n_mtx); 909 if (np->n_flag & NMODIFIED) { 910 mtx_unlock(&np->n_mtx); 911#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 912 /* 913 * Require non-blocking, synchronous writes to 914 * dirty files to inform the program it needs 915 * to fsync(2) explicitly. 916 */ 917 if (ioflag & IO_NDELAY) 918 return (EAGAIN); 919#endif 920flush_and_restart: 921 np->n_attrstamp = 0; 922 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 923 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 924 if (error) 925 return (error); 926 } else 927 mtx_unlock(&np->n_mtx); 928 } 929 930 /* 931 * If IO_APPEND then load uio_offset. We restart here if we cannot 932 * get the append lock. 933 */ 934 if (ioflag & IO_APPEND) { 935 np->n_attrstamp = 0; 936 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 937 error = VOP_GETATTR(vp, &vattr, cred); 938 if (error) 939 return (error); 940 mtx_lock(&np->n_mtx); 941 uio->uio_offset = np->n_size; 942 mtx_unlock(&np->n_mtx); 943 } 944 945 if (uio->uio_offset < 0) 946 return (EINVAL); 947 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 948 return (EFBIG); 949 if (uio->uio_resid == 0) 950 return (0); 951 952 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 953 return nfs_directio_write(vp, uio, cred, ioflag); 954 955 /* 956 * Maybe this should be above the vnode op call, but so long as 957 * file servers have no limits, i don't think it matters 958 */ 959 if (p != NULL) { 960 PROC_LOCK(p); 961 if (uio->uio_offset + uio->uio_resid > 962 lim_cur(p, RLIMIT_FSIZE)) { 963 psignal(p, SIGXFSZ); 964 PROC_UNLOCK(p); 965 return (EFBIG); 966 } 967 PROC_UNLOCK(p); 968 } 969 970 biosize = vp->v_mount->mnt_stat.f_iosize; 971 /* 972 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 973 * would exceed the local maximum per-file write commit size when 974 * combined with those, we must decide whether to flush, 975 * go synchronous, or return error. We don't bother checking 976 * IO_UNIT -- we just make all writes atomic anyway, as there's 977 * no point optimizing for something that really won't ever happen. 978 */ 979 if (!(ioflag & IO_SYNC)) { 980 int nflag; 981 982 mtx_lock(&np->n_mtx); 983 nflag = np->n_flag; 984 mtx_unlock(&np->n_mtx); 985 int needrestart = 0; 986 if (nmp->nm_wcommitsize < uio->uio_resid) { 987 /* 988 * If this request could not possibly be completed 989 * without exceeding the maximum outstanding write 990 * commit size, see if we can convert it into a 991 * synchronous write operation. 992 */ 993 if (ioflag & IO_NDELAY) 994 return (EAGAIN); 995 ioflag |= IO_SYNC; 996 if (nflag & NMODIFIED) 997 needrestart = 1; 998 } else if (nflag & NMODIFIED) { 999 int wouldcommit = 0; 1000 BO_LOCK(&vp->v_bufobj); 1001 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 1002 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 1003 b_bobufs) { 1004 if (bp->b_flags & B_NEEDCOMMIT) 1005 wouldcommit += bp->b_bcount; 1006 } 1007 } 1008 BO_UNLOCK(&vp->v_bufobj); 1009 /* 1010 * Since we're not operating synchronously and 1011 * bypassing the buffer cache, we are in a commit 1012 * and holding all of these buffers whether 1013 * transmitted or not. If not limited, this 1014 * will lead to the buffer cache deadlocking, 1015 * as no one else can flush our uncommitted buffers. 1016 */ 1017 wouldcommit += uio->uio_resid; 1018 /* 1019 * If we would initially exceed the maximum 1020 * outstanding write commit size, flush and restart. 1021 */ 1022 if (wouldcommit > nmp->nm_wcommitsize) 1023 needrestart = 1; 1024 } 1025 if (needrestart) 1026 goto flush_and_restart; 1027 } 1028 1029 do { 1030 nfsstats.biocache_writes++; 1031 lbn = uio->uio_offset / biosize; 1032 on = uio->uio_offset & (biosize-1); 1033 n = min((unsigned)(biosize - on), uio->uio_resid); 1034again: 1035 /* 1036 * Handle direct append and file extension cases, calculate 1037 * unaligned buffer size. 1038 */ 1039 mtx_lock(&np->n_mtx); 1040 if (uio->uio_offset == np->n_size && n) { 1041 mtx_unlock(&np->n_mtx); 1042 /* 1043 * Get the buffer (in its pre-append state to maintain 1044 * B_CACHE if it was previously set). Resize the 1045 * nfsnode after we have locked the buffer to prevent 1046 * readers from reading garbage. 1047 */ 1048 bcount = on; 1049 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1050 1051 if (bp != NULL) { 1052 long save; 1053 1054 mtx_lock(&np->n_mtx); 1055 np->n_size = uio->uio_offset + n; 1056 np->n_flag |= NMODIFIED; 1057 vnode_pager_setsize(vp, np->n_size); 1058 mtx_unlock(&np->n_mtx); 1059 1060 save = bp->b_flags & B_CACHE; 1061 bcount += n; 1062 allocbuf(bp, bcount); 1063 bp->b_flags |= save; 1064 } 1065 } else { 1066 /* 1067 * Obtain the locked cache block first, and then 1068 * adjust the file's size as appropriate. 1069 */ 1070 bcount = on + n; 1071 if ((off_t)lbn * biosize + bcount < np->n_size) { 1072 if ((off_t)(lbn + 1) * biosize < np->n_size) 1073 bcount = biosize; 1074 else 1075 bcount = np->n_size - (off_t)lbn * biosize; 1076 } 1077 mtx_unlock(&np->n_mtx); 1078 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1079 mtx_lock(&np->n_mtx); 1080 if (uio->uio_offset + n > np->n_size) { 1081 np->n_size = uio->uio_offset + n; 1082 np->n_flag |= NMODIFIED; 1083 vnode_pager_setsize(vp, np->n_size); 1084 } 1085 mtx_unlock(&np->n_mtx); 1086 } 1087 1088 if (!bp) { 1089 error = nfs_sigintr(nmp, NULL, td); 1090 if (!error) 1091 error = EINTR; 1092 break; 1093 } 1094 1095 /* 1096 * Issue a READ if B_CACHE is not set. In special-append 1097 * mode, B_CACHE is based on the buffer prior to the write 1098 * op and is typically set, avoiding the read. If a read 1099 * is required in special append mode, the server will 1100 * probably send us a short-read since we extended the file 1101 * on our end, resulting in b_resid == 0 and, thusly, 1102 * B_CACHE getting set. 1103 * 1104 * We can also avoid issuing the read if the write covers 1105 * the entire buffer. We have to make sure the buffer state 1106 * is reasonable in this case since we will not be initiating 1107 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1108 * more information. 1109 * 1110 * B_CACHE may also be set due to the buffer being cached 1111 * normally. 1112 */ 1113 1114 if (on == 0 && n == bcount) { 1115 bp->b_flags |= B_CACHE; 1116 bp->b_flags &= ~B_INVAL; 1117 bp->b_ioflags &= ~BIO_ERROR; 1118 } 1119 1120 if ((bp->b_flags & B_CACHE) == 0) { 1121 bp->b_iocmd = BIO_READ; 1122 vfs_busy_pages(bp, 0); 1123 error = nfs_doio(vp, bp, cred, td); 1124 if (error) { 1125 brelse(bp); 1126 break; 1127 } 1128 } 1129 if (bp->b_wcred == NOCRED) 1130 bp->b_wcred = crhold(cred); 1131 mtx_lock(&np->n_mtx); 1132 np->n_flag |= NMODIFIED; 1133 mtx_unlock(&np->n_mtx); 1134 1135 /* 1136 * If dirtyend exceeds file size, chop it down. This should 1137 * not normally occur but there is an append race where it 1138 * might occur XXX, so we log it. 1139 * 1140 * If the chopping creates a reverse-indexed or degenerate 1141 * situation with dirtyoff/end, we 0 both of them. 1142 */ 1143 1144 if (bp->b_dirtyend > bcount) { 1145 nfs_printf("NFS append race @%lx:%d\n", 1146 (long)bp->b_blkno * DEV_BSIZE, 1147 bp->b_dirtyend - bcount); 1148 bp->b_dirtyend = bcount; 1149 } 1150 1151 if (bp->b_dirtyoff >= bp->b_dirtyend) 1152 bp->b_dirtyoff = bp->b_dirtyend = 0; 1153 1154 /* 1155 * If the new write will leave a contiguous dirty 1156 * area, just update the b_dirtyoff and b_dirtyend, 1157 * otherwise force a write rpc of the old dirty area. 1158 * 1159 * While it is possible to merge discontiguous writes due to 1160 * our having a B_CACHE buffer ( and thus valid read data 1161 * for the hole), we don't because it could lead to 1162 * significant cache coherency problems with multiple clients, 1163 * especially if locking is implemented later on. 1164 * 1165 * as an optimization we could theoretically maintain 1166 * a linked list of discontinuous areas, but we would still 1167 * have to commit them separately so there isn't much 1168 * advantage to it except perhaps a bit of asynchronization. 1169 */ 1170 1171 if (bp->b_dirtyend > 0 && 1172 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1173 if (bwrite(bp) == EINTR) { 1174 error = EINTR; 1175 break; 1176 } 1177 goto again; 1178 } 1179 1180 error = uiomove((char *)bp->b_data + on, n, uio); 1181 1182 /* 1183 * Since this block is being modified, it must be written 1184 * again and not just committed. Since write clustering does 1185 * not work for the stage 1 data write, only the stage 2 1186 * commit rpc, we have to clear B_CLUSTEROK as well. 1187 */ 1188 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1189 1190 if (error) { 1191 bp->b_ioflags |= BIO_ERROR; 1192 brelse(bp); 1193 break; 1194 } 1195 1196 /* 1197 * Only update dirtyoff/dirtyend if not a degenerate 1198 * condition. 1199 */ 1200 if (n) { 1201 if (bp->b_dirtyend > 0) { 1202 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1203 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1204 } else { 1205 bp->b_dirtyoff = on; 1206 bp->b_dirtyend = on + n; 1207 } 1208 vfs_bio_set_validclean(bp, on, n); 1209 } 1210 1211 /* 1212 * If IO_SYNC do bwrite(). 1213 * 1214 * IO_INVAL appears to be unused. The idea appears to be 1215 * to turn off caching in this case. Very odd. XXX 1216 */ 1217 if ((ioflag & IO_SYNC)) { 1218 if (ioflag & IO_INVAL) 1219 bp->b_flags |= B_NOCACHE; 1220 error = bwrite(bp); 1221 if (error) 1222 break; 1223 } else if ((n + on) == biosize) { 1224 bp->b_flags |= B_ASYNC; 1225 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL); 1226 } else { 1227 bdwrite(bp); 1228 } 1229 } while (uio->uio_resid > 0 && n > 0); 1230 1231 return (error); 1232} 1233 1234/* 1235 * Get an nfs cache block. 1236 * 1237 * Allocate a new one if the block isn't currently in the cache 1238 * and return the block marked busy. If the calling process is 1239 * interrupted by a signal for an interruptible mount point, return 1240 * NULL. 1241 * 1242 * The caller must carefully deal with the possible B_INVAL state of 1243 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1244 * indirectly), so synchronous reads can be issued without worrying about 1245 * the B_INVAL state. We have to be a little more careful when dealing 1246 * with writes (see comments in nfs_write()) when extending a file past 1247 * its EOF. 1248 */ 1249static struct buf * 1250nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1251{ 1252 struct buf *bp; 1253 struct mount *mp; 1254 struct nfsmount *nmp; 1255 1256 mp = vp->v_mount; 1257 nmp = VFSTONFS(mp); 1258 1259 if (nmp->nm_flag & NFSMNT_INT) { 1260 sigset_t oldset; 1261 1262 nfs_set_sigmask(td, &oldset); 1263 bp = getblk(vp, bn, size, PCATCH, 0, 0); 1264 nfs_restore_sigmask(td, &oldset); 1265 while (bp == NULL) { 1266 if (nfs_sigintr(nmp, NULL, td)) 1267 return (NULL); 1268 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1269 } 1270 } else { 1271 bp = getblk(vp, bn, size, 0, 0, 0); 1272 } 1273 1274 if (vp->v_type == VREG) { 1275 int biosize; 1276 1277 biosize = mp->mnt_stat.f_iosize; 1278 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1279 } 1280 return (bp); 1281} 1282 1283/* 1284 * Flush and invalidate all dirty buffers. If another process is already 1285 * doing the flush, just wait for completion. 1286 */ 1287int 1288nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1289{ 1290 struct nfsnode *np = VTONFS(vp); 1291 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1292 int error = 0, slpflag, slptimeo; 1293 int old_lock = 0; 1294 1295 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf"); 1296 1297 /* 1298 * XXX This check stops us from needlessly doing a vinvalbuf when 1299 * being called through vclean(). It is not clear that this is 1300 * unsafe. 1301 */ 1302 if (vp->v_iflag & VI_DOOMED) 1303 return (0); 1304 1305 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1306 intrflg = 0; 1307 if (intrflg) { 1308 slpflag = PCATCH; 1309 slptimeo = 2 * hz; 1310 } else { 1311 slpflag = 0; 1312 slptimeo = 0; 1313 } 1314 1315 old_lock = nfs_upgrade_vnlock(vp); 1316 /* 1317 * Now, flush as required. 1318 */ 1319 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1320 VM_OBJECT_LOCK(vp->v_bufobj.bo_object); 1321 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1322 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); 1323 /* 1324 * If the page clean was interrupted, fail the invalidation. 1325 * Not doing so, we run the risk of losing dirty pages in the 1326 * vinvalbuf() call below. 1327 */ 1328 if (intrflg && (error = nfs_sigintr(nmp, NULL, td))) 1329 goto out; 1330 } 1331 1332 error = vinvalbuf(vp, flags, slpflag, 0); 1333 while (error) { 1334 if (intrflg && (error = nfs_sigintr(nmp, NULL, td))) 1335 goto out; 1336 error = vinvalbuf(vp, flags, 0, slptimeo); 1337 } 1338 mtx_lock(&np->n_mtx); 1339 if (np->n_directio_asyncwr == 0) 1340 np->n_flag &= ~NMODIFIED; 1341 mtx_unlock(&np->n_mtx); 1342out: 1343 nfs_downgrade_vnlock(vp, old_lock); 1344 return error; 1345} 1346 1347/* 1348 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1349 * This is mainly to avoid queueing async I/O requests when the nfsiods 1350 * are all hung on a dead server. 1351 * 1352 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1353 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1354 */ 1355int 1356nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1357{ 1358 int iod; 1359 int gotiod; 1360 int slpflag = 0; 1361 int slptimeo = 0; 1362 int error, error2; 1363 1364 /* 1365 * Commits are usually short and sweet so lets save some cpu and 1366 * leave the async daemons for more important rpc's (such as reads 1367 * and writes). 1368 */ 1369 mtx_lock(&nfs_iod_mtx); 1370 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1371 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1372 mtx_unlock(&nfs_iod_mtx); 1373 return(EIO); 1374 } 1375again: 1376 if (nmp->nm_flag & NFSMNT_INT) 1377 slpflag = PCATCH; 1378 gotiod = FALSE; 1379 1380 /* 1381 * Find a free iod to process this request. 1382 */ 1383 for (iod = 0; iod < nfs_numasync; iod++) 1384 if (nfs_iodwant[iod]) { 1385 gotiod = TRUE; 1386 break; 1387 } 1388 1389 /* 1390 * Try to create one if none are free. 1391 */ 1392 if (!gotiod) { 1393 iod = nfs_nfsiodnew(); 1394 if (iod != -1) 1395 gotiod = TRUE; 1396 } 1397 1398 if (gotiod) { 1399 /* 1400 * Found one, so wake it up and tell it which 1401 * mount to process. 1402 */ 1403 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n", 1404 iod, nmp)); 1405 nfs_iodwant[iod] = NULL; 1406 nfs_iodmount[iod] = nmp; 1407 nmp->nm_bufqiods++; 1408 wakeup(&nfs_iodwant[iod]); 1409 } 1410 1411 /* 1412 * If none are free, we may already have an iod working on this mount 1413 * point. If so, it will process our request. 1414 */ 1415 if (!gotiod) { 1416 if (nmp->nm_bufqiods > 0) { 1417 NFS_DPF(ASYNCIO, 1418 ("nfs_asyncio: %d iods are already processing mount %p\n", 1419 nmp->nm_bufqiods, nmp)); 1420 gotiod = TRUE; 1421 } 1422 } 1423 1424 /* 1425 * If we have an iod which can process the request, then queue 1426 * the buffer. 1427 */ 1428 if (gotiod) { 1429 /* 1430 * Ensure that the queue never grows too large. We still want 1431 * to asynchronize so we block rather then return EIO. 1432 */ 1433 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1434 NFS_DPF(ASYNCIO, 1435 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1436 nmp->nm_bufqwant = TRUE; 1437 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx, 1438 slpflag | PRIBIO, 1439 "nfsaio", slptimeo); 1440 if (error) { 1441 error2 = nfs_sigintr(nmp, NULL, td); 1442 if (error2) { 1443 mtx_unlock(&nfs_iod_mtx); 1444 return (error2); 1445 } 1446 if (slpflag == PCATCH) { 1447 slpflag = 0; 1448 slptimeo = 2 * hz; 1449 } 1450 } 1451 /* 1452 * We might have lost our iod while sleeping, 1453 * so check and loop if nescessary. 1454 */ 1455 if (nmp->nm_bufqiods == 0) { 1456 NFS_DPF(ASYNCIO, 1457 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1458 goto again; 1459 } 1460 } 1461 1462 /* We might have lost our nfsiod */ 1463 if (nmp->nm_bufqiods == 0) { 1464 NFS_DPF(ASYNCIO, 1465 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1466 goto again; 1467 } 1468 1469 if (bp->b_iocmd == BIO_READ) { 1470 if (bp->b_rcred == NOCRED && cred != NOCRED) 1471 bp->b_rcred = crhold(cred); 1472 } else { 1473 if (bp->b_wcred == NOCRED && cred != NOCRED) 1474 bp->b_wcred = crhold(cred); 1475 } 1476 1477 if (bp->b_flags & B_REMFREE) 1478 bremfreef(bp); 1479 BUF_KERNPROC(bp); 1480 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1481 nmp->nm_bufqlen++; 1482 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1483 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1484 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1485 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1486 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1487 } 1488 mtx_unlock(&nfs_iod_mtx); 1489 return (0); 1490 } 1491 1492 mtx_unlock(&nfs_iod_mtx); 1493 1494 /* 1495 * All the iods are busy on other mounts, so return EIO to 1496 * force the caller to process the i/o synchronously. 1497 */ 1498 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1499 return (EIO); 1500} 1501 1502void 1503nfs_doio_directwrite(struct buf *bp) 1504{ 1505 int iomode, must_commit; 1506 struct uio *uiop = (struct uio *)bp->b_caller1; 1507 char *iov_base = uiop->uio_iov->iov_base; 1508 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount); 1509 1510 iomode = NFSV3WRITE_FILESYNC; 1511 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1512 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit); 1513 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write")); 1514 free(iov_base, M_NFSDIRECTIO); 1515 free(uiop->uio_iov, M_NFSDIRECTIO); 1516 free(uiop, M_NFSDIRECTIO); 1517 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1518 struct nfsnode *np = VTONFS(bp->b_vp); 1519 mtx_lock(&np->n_mtx); 1520 np->n_directio_asyncwr--; 1521 if (np->n_directio_asyncwr == 0) { 1522 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED; 1523 if ((np->n_flag & NFSYNCWAIT)) { 1524 np->n_flag &= ~NFSYNCWAIT; 1525 wakeup((caddr_t)&np->n_directio_asyncwr); 1526 } 1527 } 1528 mtx_unlock(&np->n_mtx); 1529 } 1530 bp->b_vp = NULL; 1531 relpbuf(bp, &nfs_pbuf_freecnt); 1532} 1533 1534/* 1535 * Do an I/O operation to/from a cache block. This may be called 1536 * synchronously or from an nfsiod. 1537 */ 1538int 1539nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) 1540{ 1541 struct uio *uiop; 1542 struct nfsnode *np; 1543 struct nfsmount *nmp; 1544 int error = 0, iomode, must_commit = 0; 1545 struct uio uio; 1546 struct iovec io; 1547 struct proc *p = td ? td->td_proc : NULL; 1548 uint8_t iocmd; 1549 1550 np = VTONFS(vp); 1551 nmp = VFSTONFS(vp->v_mount); 1552 uiop = &uio; 1553 uiop->uio_iov = &io; 1554 uiop->uio_iovcnt = 1; 1555 uiop->uio_segflg = UIO_SYSSPACE; 1556 uiop->uio_td = td; 1557 1558 /* 1559 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1560 * do this here so we do not have to do it in all the code that 1561 * calls us. 1562 */ 1563 bp->b_flags &= ~B_INVAL; 1564 bp->b_ioflags &= ~BIO_ERROR; 1565 1566 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1567 iocmd = bp->b_iocmd; 1568 if (iocmd == BIO_READ) { 1569 io.iov_len = uiop->uio_resid = bp->b_bcount; 1570 io.iov_base = bp->b_data; 1571 uiop->uio_rw = UIO_READ; 1572 1573 switch (vp->v_type) { 1574 case VREG: 1575 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1576 nfsstats.read_bios++; 1577 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr); 1578 1579 if (!error) { 1580 if (uiop->uio_resid) { 1581 /* 1582 * If we had a short read with no error, we must have 1583 * hit a file hole. We should zero-fill the remainder. 1584 * This can also occur if the server hits the file EOF. 1585 * 1586 * Holes used to be able to occur due to pending 1587 * writes, but that is not possible any longer. 1588 */ 1589 int nread = bp->b_bcount - uiop->uio_resid; 1590 int left = uiop->uio_resid; 1591 1592 if (left > 0) 1593 bzero((char *)bp->b_data + nread, left); 1594 uiop->uio_resid = 0; 1595 } 1596 } 1597 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */ 1598 if (p && (vp->v_vflag & VV_TEXT)) { 1599 mtx_lock(&np->n_mtx); 1600 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) { 1601 mtx_unlock(&np->n_mtx); 1602 PROC_LOCK(p); 1603 killproc(p, "text file modification"); 1604 PROC_UNLOCK(p); 1605 } else 1606 mtx_unlock(&np->n_mtx); 1607 } 1608 break; 1609 case VLNK: 1610 uiop->uio_offset = (off_t)0; 1611 nfsstats.readlink_bios++; 1612 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr); 1613 break; 1614 case VDIR: 1615 nfsstats.readdir_bios++; 1616 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1617 if ((nmp->nm_flag & NFSMNT_NFSV4) != 0) 1618 error = nfs4_readdirrpc(vp, uiop, cr); 1619 else { 1620 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1621 error = nfs_readdirplusrpc(vp, uiop, cr); 1622 if (error == NFSERR_NOTSUPP) 1623 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1624 } 1625 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1626 error = nfs_readdirrpc(vp, uiop, cr); 1627 } 1628 /* 1629 * end-of-directory sets B_INVAL but does not generate an 1630 * error. 1631 */ 1632 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1633 bp->b_flags |= B_INVAL; 1634 break; 1635 default: 1636 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type); 1637 break; 1638 }; 1639 if (error) { 1640 bp->b_ioflags |= BIO_ERROR; 1641 bp->b_error = error; 1642 } 1643 } else { 1644 /* 1645 * If we only need to commit, try to commit 1646 */ 1647 if (bp->b_flags & B_NEEDCOMMIT) { 1648 int retv; 1649 off_t off; 1650 1651 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1652 retv = (nmp->nm_rpcops->nr_commit)( 1653 vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1654 bp->b_wcred, td); 1655 if (retv == 0) { 1656 bp->b_dirtyoff = bp->b_dirtyend = 0; 1657 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1658 bp->b_resid = 0; 1659 bufdone(bp); 1660 return (0); 1661 } 1662 if (retv == NFSERR_STALEWRITEVERF) { 1663 nfs_clearcommit(vp->v_mount); 1664 } 1665 } 1666 1667 /* 1668 * Setup for actual write 1669 */ 1670 mtx_lock(&np->n_mtx); 1671 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1672 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1673 mtx_unlock(&np->n_mtx); 1674 1675 if (bp->b_dirtyend > bp->b_dirtyoff) { 1676 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1677 - bp->b_dirtyoff; 1678 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1679 + bp->b_dirtyoff; 1680 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1681 uiop->uio_rw = UIO_WRITE; 1682 nfsstats.write_bios++; 1683 1684 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1685 iomode = NFSV3WRITE_UNSTABLE; 1686 else 1687 iomode = NFSV3WRITE_FILESYNC; 1688 1689 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit); 1690 1691 /* 1692 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1693 * to cluster the buffers needing commit. This will allow 1694 * the system to submit a single commit rpc for the whole 1695 * cluster. We can do this even if the buffer is not 100% 1696 * dirty (relative to the NFS blocksize), so we optimize the 1697 * append-to-file-case. 1698 * 1699 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1700 * cleared because write clustering only works for commit 1701 * rpc's, not for the data portion of the write). 1702 */ 1703 1704 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1705 bp->b_flags |= B_NEEDCOMMIT; 1706 if (bp->b_dirtyoff == 0 1707 && bp->b_dirtyend == bp->b_bcount) 1708 bp->b_flags |= B_CLUSTEROK; 1709 } else { 1710 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1711 } 1712 1713 /* 1714 * For an interrupted write, the buffer is still valid 1715 * and the write hasn't been pushed to the server yet, 1716 * so we can't set BIO_ERROR and report the interruption 1717 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1718 * is not relevant, so the rpc attempt is essentially 1719 * a noop. For the case of a V3 write rpc not being 1720 * committed to stable storage, the block is still 1721 * dirty and requires either a commit rpc or another 1722 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1723 * the block is reused. This is indicated by setting 1724 * the B_DELWRI and B_NEEDCOMMIT flags. 1725 * 1726 * If the buffer is marked B_PAGING, it does not reside on 1727 * the vp's paging queues so we cannot call bdirty(). The 1728 * bp in this case is not an NFS cache block so we should 1729 * be safe. XXX 1730 * 1731 * The logic below breaks up errors into recoverable and 1732 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1733 * and keep the buffer around for potential write retries. 1734 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1735 * and save the error in the nfsnode. This is less than ideal 1736 * but necessary. Keeping such buffers around could potentially 1737 * cause buffer exhaustion eventually (they can never be written 1738 * out, so will get constantly be re-dirtied). It also causes 1739 * all sorts of vfs panics. For non-recoverable write errors, 1740 * also invalidate the attrcache, so we'll be forced to go over 1741 * the wire for this object, returning an error to user on next 1742 * call (most of the time). 1743 */ 1744 if (error == EINTR || error == EIO || error == ETIMEDOUT 1745 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1746 int s; 1747 1748 s = splbio(); 1749 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1750 if ((bp->b_flags & B_PAGING) == 0) { 1751 bdirty(bp); 1752 bp->b_flags &= ~B_DONE; 1753 } 1754 if (error && (bp->b_flags & B_ASYNC) == 0) 1755 bp->b_flags |= B_EINTR; 1756 splx(s); 1757 } else { 1758 if (error) { 1759 bp->b_ioflags |= BIO_ERROR; 1760 bp->b_flags |= B_INVAL; 1761 bp->b_error = np->n_error = error; 1762 mtx_lock(&np->n_mtx); 1763 np->n_flag |= NWRITEERR; 1764 np->n_attrstamp = 0; 1765 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1766 mtx_unlock(&np->n_mtx); 1767 } 1768 bp->b_dirtyoff = bp->b_dirtyend = 0; 1769 } 1770 } else { 1771 bp->b_resid = 0; 1772 bufdone(bp); 1773 return (0); 1774 } 1775 } 1776 bp->b_resid = uiop->uio_resid; 1777 if (must_commit) 1778 nfs_clearcommit(vp->v_mount); 1779 bufdone(bp); 1780 return (error); 1781} 1782 1783/* 1784 * Used to aid in handling ftruncate() operations on the NFS client side. 1785 * Truncation creates a number of special problems for NFS. We have to 1786 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1787 * we have to properly handle VM pages or (potentially dirty) buffers 1788 * that straddle the truncation point. 1789 */ 1790 1791int 1792nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1793{ 1794 struct nfsnode *np = VTONFS(vp); 1795 u_quad_t tsize; 1796 int biosize = vp->v_mount->mnt_stat.f_iosize; 1797 int error = 0; 1798 1799 mtx_lock(&np->n_mtx); 1800 tsize = np->n_size; 1801 np->n_size = nsize; 1802 mtx_unlock(&np->n_mtx); 1803 1804 if (nsize < tsize) { 1805 struct buf *bp; 1806 daddr_t lbn; 1807 int bufsize; 1808 1809 /* 1810 * vtruncbuf() doesn't get the buffer overlapping the 1811 * truncation point. We may have a B_DELWRI and/or B_CACHE 1812 * buffer that now needs to be truncated. 1813 */ 1814 error = vtruncbuf(vp, cred, td, nsize, biosize); 1815 lbn = nsize / biosize; 1816 bufsize = nsize & (biosize - 1); 1817 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1818 if (!bp) 1819 return EINTR; 1820 if (bp->b_dirtyoff > bp->b_bcount) 1821 bp->b_dirtyoff = bp->b_bcount; 1822 if (bp->b_dirtyend > bp->b_bcount) 1823 bp->b_dirtyend = bp->b_bcount; 1824 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1825 brelse(bp); 1826 } else { 1827 vnode_pager_setsize(vp, nsize); 1828 } 1829 return(error); 1830} 1831 1832