nfs_bio.c revision 152656
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 152656 2005-11-21 19:23:46Z ps $"); 37 38#include <sys/param.h> 39#include <sys/systm.h> 40#include <sys/bio.h> 41#include <sys/buf.h> 42#include <sys/kernel.h> 43#include <sys/mount.h> 44#include <sys/proc.h> 45#include <sys/resourcevar.h> 46#include <sys/signalvar.h> 47#include <sys/vmmeter.h> 48#include <sys/vnode.h> 49 50#include <vm/vm.h> 51#include <vm/vm_extern.h> 52#include <vm/vm_page.h> 53#include <vm/vm_object.h> 54#include <vm/vm_pager.h> 55#include <vm/vnode_pager.h> 56 57#include <rpc/rpcclnt.h> 58 59#include <nfs/rpcv2.h> 60#include <nfs/nfsproto.h> 61#include <nfsclient/nfs.h> 62#include <nfsclient/nfsmount.h> 63#include <nfsclient/nfsnode.h> 64 65#include <nfs4client/nfs4.h> 66 67static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 68 struct thread *td); 69static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 70 struct ucred *cred, int ioflag); 71 72extern int nfs_directio_enable; 73extern int nfs_directio_allow_mmap; 74/* 75 * 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 GIANT_REQUIRED; 94 95 vp = ap->a_vp; 96 np = VTONFS(vp); 97 td = curthread; /* XXX */ 98 cred = curthread->td_ucred; /* XXX */ 99 nmp = VFSTONFS(vp->v_mount); 100 pages = ap->a_m; 101 count = ap->a_count; 102 103 if ((object = vp->v_object) == NULL) { 104 printf("nfs_getpages: called with non-merged cache vnode??\n"); 105 return VM_PAGER_ERROR; 106 } 107 108 if (!nfs_directio_allow_mmap && (np->n_flag & NNONCACHE) && 109 (vp->v_type == VREG)) { 110 printf("nfs_getpages: called on non-cacheable vnode??\n"); 111 return VM_PAGER_ERROR; 112 } 113 114 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 115 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 116 /* We'll never get here for v4, because we always have fsinfo */ 117 (void)nfs_fsinfo(nmp, vp, cred, td); 118 } 119 120 npages = btoc(count); 121 122 /* 123 * If the requested page is partially valid, just return it and 124 * allow the pager to zero-out the blanks. Partially valid pages 125 * can only occur at the file EOF. 126 */ 127 128 { 129 vm_page_t m = pages[ap->a_reqpage]; 130 131 VM_OBJECT_LOCK(object); 132 vm_page_lock_queues(); 133 if (m->valid != 0) { 134 /* handled by vm_fault now */ 135 /* vm_page_zero_invalid(m, TRUE); */ 136 for (i = 0; i < npages; ++i) { 137 if (i != ap->a_reqpage) 138 vm_page_free(pages[i]); 139 } 140 vm_page_unlock_queues(); 141 VM_OBJECT_UNLOCK(object); 142 return(0); 143 } 144 vm_page_unlock_queues(); 145 VM_OBJECT_UNLOCK(object); 146 } 147 148 /* 149 * We use only the kva address for the buffer, but this is extremely 150 * convienient and fast. 151 */ 152 bp = getpbuf(&nfs_pbuf_freecnt); 153 154 kva = (vm_offset_t) bp->b_data; 155 pmap_qenter(kva, pages, npages); 156 cnt.v_vnodein++; 157 cnt.v_vnodepgsin += npages; 158 159 iov.iov_base = (caddr_t) kva; 160 iov.iov_len = count; 161 uio.uio_iov = &iov; 162 uio.uio_iovcnt = 1; 163 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 164 uio.uio_resid = count; 165 uio.uio_segflg = UIO_SYSSPACE; 166 uio.uio_rw = UIO_READ; 167 uio.uio_td = td; 168 169 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred); 170 pmap_qremove(kva, npages); 171 172 relpbuf(bp, &nfs_pbuf_freecnt); 173 174 if (error && (uio.uio_resid == count)) { 175 printf("nfs_getpages: error %d\n", error); 176 VM_OBJECT_LOCK(object); 177 vm_page_lock_queues(); 178 for (i = 0; i < npages; ++i) { 179 if (i != ap->a_reqpage) 180 vm_page_free(pages[i]); 181 } 182 vm_page_unlock_queues(); 183 VM_OBJECT_UNLOCK(object); 184 return VM_PAGER_ERROR; 185 } 186 187 /* 188 * Calculate the number of bytes read and validate only that number 189 * of bytes. Note that due to pending writes, size may be 0. This 190 * does not mean that the remaining data is invalid! 191 */ 192 193 size = count - uio.uio_resid; 194 VM_OBJECT_LOCK(object); 195 vm_page_lock_queues(); 196 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 197 vm_page_t m; 198 nextoff = toff + PAGE_SIZE; 199 m = pages[i]; 200 201 if (nextoff <= size) { 202 /* 203 * Read operation filled an entire page 204 */ 205 m->valid = VM_PAGE_BITS_ALL; 206 vm_page_undirty(m); 207 } else if (size > toff) { 208 /* 209 * Read operation filled a partial page. 210 */ 211 m->valid = 0; 212 vm_page_set_validclean(m, 0, size - toff); 213 /* handled by vm_fault now */ 214 /* vm_page_zero_invalid(m, TRUE); */ 215 } else { 216 /* 217 * Read operation was short. If no error occured 218 * we may have hit a zero-fill section. We simply 219 * leave valid set to 0. 220 */ 221 ; 222 } 223 if (i != ap->a_reqpage) { 224 /* 225 * Whether or not to leave the page activated is up in 226 * the air, but we should put the page on a page queue 227 * somewhere (it already is in the object). Result: 228 * It appears that emperical results show that 229 * deactivating pages is best. 230 */ 231 232 /* 233 * Just in case someone was asking for this page we 234 * now tell them that it is ok to use. 235 */ 236 if (!error) { 237 if (m->flags & PG_WANTED) 238 vm_page_activate(m); 239 else 240 vm_page_deactivate(m); 241 vm_page_wakeup(m); 242 } else { 243 vm_page_free(m); 244 } 245 } 246 } 247 vm_page_unlock_queues(); 248 VM_OBJECT_UNLOCK(object); 249 return 0; 250} 251 252/* 253 * Vnode op for VM putpages. 254 */ 255int 256nfs_putpages(struct vop_putpages_args *ap) 257{ 258 struct uio uio; 259 struct iovec iov; 260 vm_offset_t kva; 261 struct buf *bp; 262 int iomode, must_commit, i, error, npages, count; 263 off_t offset; 264 int *rtvals; 265 struct vnode *vp; 266 struct thread *td; 267 struct ucred *cred; 268 struct nfsmount *nmp; 269 struct nfsnode *np; 270 vm_page_t *pages; 271 272 GIANT_REQUIRED; 273 274 vp = ap->a_vp; 275 np = VTONFS(vp); 276 td = curthread; /* XXX */ 277 cred = curthread->td_ucred; /* XXX */ 278 nmp = VFSTONFS(vp->v_mount); 279 pages = ap->a_m; 280 count = ap->a_count; 281 rtvals = ap->a_rtvals; 282 npages = btoc(count); 283 offset = IDX_TO_OFF(pages[0]->pindex); 284 285 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 286 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 287 (void)nfs_fsinfo(nmp, vp, cred, td); 288 } 289 290 if (!nfs_directio_allow_mmap && (np->n_flag & NNONCACHE) && 291 (vp->v_type == VREG)) 292 printf("nfs_putpages: called on noncache-able vnode??\n"); 293 294 for (i = 0; i < npages; i++) 295 rtvals[i] = VM_PAGER_AGAIN; 296 297 /* 298 * When putting pages, do not extend file past EOF. 299 */ 300 301 if (offset + count > np->n_size) { 302 count = np->n_size - offset; 303 if (count < 0) 304 count = 0; 305 } 306 307 /* 308 * We use only the kva address for the buffer, but this is extremely 309 * convienient and fast. 310 */ 311 bp = getpbuf(&nfs_pbuf_freecnt); 312 313 kva = (vm_offset_t) bp->b_data; 314 pmap_qenter(kva, pages, npages); 315 cnt.v_vnodeout++; 316 cnt.v_vnodepgsout += count; 317 318 iov.iov_base = (caddr_t) kva; 319 iov.iov_len = count; 320 uio.uio_iov = &iov; 321 uio.uio_iovcnt = 1; 322 uio.uio_offset = offset; 323 uio.uio_resid = count; 324 uio.uio_segflg = UIO_SYSSPACE; 325 uio.uio_rw = UIO_WRITE; 326 uio.uio_td = td; 327 328 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 329 iomode = NFSV3WRITE_UNSTABLE; 330 else 331 iomode = NFSV3WRITE_FILESYNC; 332 333 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit); 334 335 pmap_qremove(kva, npages); 336 relpbuf(bp, &nfs_pbuf_freecnt); 337 338 if (!error) { 339 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 340 for (i = 0; i < nwritten; i++) { 341 rtvals[i] = VM_PAGER_OK; 342 vm_page_undirty(pages[i]); 343 } 344 if (must_commit) { 345 nfs_clearcommit(vp->v_mount); 346 } 347 } 348 return rtvals[0]; 349} 350 351/* 352 * Vnode op for read using bio 353 */ 354int 355nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 356{ 357 struct nfsnode *np = VTONFS(vp); 358 int biosize, i; 359 struct buf *bp, *rabp; 360 struct vattr vattr; 361 struct thread *td; 362 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 363 daddr_t lbn, rabn; 364 int bcount; 365 int seqcount; 366 int nra, error = 0, n = 0, on = 0; 367 368#ifdef DIAGNOSTIC 369 if (uio->uio_rw != UIO_READ) 370 panic("nfs_read mode"); 371#endif 372 if (uio->uio_resid == 0) 373 return (0); 374 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 375 return (EINVAL); 376 td = uio->uio_td; 377 378 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 379 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 380 (void)nfs_fsinfo(nmp, vp, cred, td); 381 if (vp->v_type != VDIR && 382 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 383 return (EFBIG); 384 385 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 386 /* No caching/ no readaheads. Just read data into the user buffer */ 387 return nfs_readrpc(vp, uio, cred); 388 389 biosize = vp->v_mount->mnt_stat.f_iosize; 390 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 391 /* 392 * For nfs, cache consistency can only be maintained approximately. 393 * Although RFC1094 does not specify the criteria, the following is 394 * believed to be compatible with the reference port. 395 * For nfs: 396 * If the file's modify time on the server has changed since the 397 * last read rpc or you have written to the file, 398 * you may have lost data cache consistency with the 399 * server, so flush all of the file's data out of the cache. 400 * Then force a getattr rpc to ensure that you have up to date 401 * attributes. 402 * NB: This implies that cache data can be read when up to 403 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 404 * attributes this could be forced by setting n_attrstamp to 0 before 405 * the VOP_GETATTR() call. 406 */ 407 if (np->n_flag & NMODIFIED) { 408 if (vp->v_type != VREG) { 409 if (vp->v_type != VDIR) 410 panic("nfs: bioread, not dir"); 411 (nmp->nm_rpcops->nr_invaldir)(vp); 412 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 413 if (error) 414 return (error); 415 } 416 np->n_attrstamp = 0; 417 error = VOP_GETATTR(vp, &vattr, cred, td); 418 if (error) 419 return (error); 420 np->n_mtime = vattr.va_mtime; 421 } else { 422 error = VOP_GETATTR(vp, &vattr, cred, td); 423 if (error) 424 return (error); 425 if ((np->n_flag & NSIZECHANGED) 426 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 427 if (vp->v_type == VDIR) 428 (nmp->nm_rpcops->nr_invaldir)(vp); 429 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 430 if (error) 431 return (error); 432 np->n_mtime = vattr.va_mtime; 433 np->n_flag &= ~NSIZECHANGED; 434 } 435 } 436 do { 437 switch (vp->v_type) { 438 case VREG: 439 nfsstats.biocache_reads++; 440 lbn = uio->uio_offset / biosize; 441 on = uio->uio_offset & (biosize - 1); 442 443 /* 444 * Start the read ahead(s), as required. 445 * The readahead is kicked off only if sequential access 446 * is detected, based on the readahead hint (ra_expect_lbn). 447 */ 448 if (nmp->nm_readahead > 0 && np->ra_expect_lbn == lbn) { 449 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 450 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 451 rabn = lbn + 1 + nra; 452 if (incore(&vp->v_bufobj, rabn) == NULL) { 453 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 454 if (!rabp) { 455 error = nfs_sigintr(nmp, NULL, td); 456 return (error ? error : EINTR); 457 } 458 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 459 rabp->b_flags |= B_ASYNC; 460 rabp->b_iocmd = BIO_READ; 461 vfs_busy_pages(rabp, 0); 462 if (nfs_asyncio(nmp, rabp, cred, td)) { 463 rabp->b_flags |= B_INVAL; 464 rabp->b_ioflags |= BIO_ERROR; 465 vfs_unbusy_pages(rabp); 466 brelse(rabp); 467 break; 468 } 469 } else { 470 brelse(rabp); 471 } 472 } 473 } 474 np->ra_expect_lbn = lbn + 1; 475 } 476 477 /* Note that bcount is *not* DEV_BSIZE aligned. */ 478 bcount = biosize; 479 if ((off_t)lbn * biosize >= np->n_size) { 480 bcount = 0; 481 } else if ((off_t)(lbn + 1) * biosize > np->n_size) { 482 bcount = np->n_size - (off_t)lbn * biosize; 483 } 484 bp = nfs_getcacheblk(vp, lbn, bcount, td); 485 486 if (!bp) { 487 error = nfs_sigintr(nmp, NULL, td); 488 return (error ? error : EINTR); 489 } 490 491 /* 492 * If B_CACHE is not set, we must issue the read. If this 493 * fails, we return an error. 494 */ 495 496 if ((bp->b_flags & B_CACHE) == 0) { 497 bp->b_iocmd = BIO_READ; 498 vfs_busy_pages(bp, 0); 499 error = nfs_doio(vp, bp, cred, td); 500 if (error) { 501 brelse(bp); 502 return (error); 503 } 504 } 505 506 /* 507 * on is the offset into the current bp. Figure out how many 508 * bytes we can copy out of the bp. Note that bcount is 509 * NOT DEV_BSIZE aligned. 510 * 511 * Then figure out how many bytes we can copy into the uio. 512 */ 513 514 n = 0; 515 if (on < bcount) 516 n = min((unsigned)(bcount - on), uio->uio_resid); 517 break; 518 case VLNK: 519 nfsstats.biocache_readlinks++; 520 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 521 if (!bp) { 522 error = nfs_sigintr(nmp, NULL, td); 523 return (error ? error : EINTR); 524 } 525 if ((bp->b_flags & B_CACHE) == 0) { 526 bp->b_iocmd = BIO_READ; 527 vfs_busy_pages(bp, 0); 528 error = nfs_doio(vp, bp, cred, td); 529 if (error) { 530 bp->b_ioflags |= BIO_ERROR; 531 brelse(bp); 532 return (error); 533 } 534 } 535 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 536 on = 0; 537 break; 538 case VDIR: 539 nfsstats.biocache_readdirs++; 540 if (np->n_direofoffset 541 && uio->uio_offset >= np->n_direofoffset) { 542 return (0); 543 } 544 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 545 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 546 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 547 if (!bp) { 548 error = nfs_sigintr(nmp, NULL, td); 549 return (error ? error : EINTR); 550 } 551 if ((bp->b_flags & B_CACHE) == 0) { 552 bp->b_iocmd = BIO_READ; 553 vfs_busy_pages(bp, 0); 554 error = nfs_doio(vp, bp, cred, td); 555 if (error) { 556 brelse(bp); 557 } 558 while (error == NFSERR_BAD_COOKIE) { 559 (nmp->nm_rpcops->nr_invaldir)(vp); 560 error = nfs_vinvalbuf(vp, 0, td, 1); 561 /* 562 * Yuck! The directory has been modified on the 563 * server. The only way to get the block is by 564 * reading from the beginning to get all the 565 * offset cookies. 566 * 567 * Leave the last bp intact unless there is an error. 568 * Loop back up to the while if the error is another 569 * NFSERR_BAD_COOKIE (double yuch!). 570 */ 571 for (i = 0; i <= lbn && !error; i++) { 572 if (np->n_direofoffset 573 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 574 return (0); 575 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 576 if (!bp) { 577 error = nfs_sigintr(nmp, NULL, td); 578 return (error ? error : EINTR); 579 } 580 if ((bp->b_flags & B_CACHE) == 0) { 581 bp->b_iocmd = BIO_READ; 582 vfs_busy_pages(bp, 0); 583 error = nfs_doio(vp, bp, cred, td); 584 /* 585 * no error + B_INVAL == directory EOF, 586 * use the block. 587 */ 588 if (error == 0 && (bp->b_flags & B_INVAL)) 589 break; 590 } 591 /* 592 * An error will throw away the block and the 593 * for loop will break out. If no error and this 594 * is not the block we want, we throw away the 595 * block and go for the next one via the for loop. 596 */ 597 if (error || i < lbn) 598 brelse(bp); 599 } 600 } 601 /* 602 * The above while is repeated if we hit another cookie 603 * error. If we hit an error and it wasn't a cookie error, 604 * we give up. 605 */ 606 if (error) 607 return (error); 608 } 609 610 /* 611 * If not eof and read aheads are enabled, start one. 612 * (You need the current block first, so that you have the 613 * directory offset cookie of the next block.) 614 */ 615 if (nmp->nm_readahead > 0 && 616 (bp->b_flags & B_INVAL) == 0 && 617 (np->n_direofoffset == 0 || 618 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 619 incore(&vp->v_bufobj, lbn + 1) == NULL) { 620 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 621 if (rabp) { 622 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 623 rabp->b_flags |= B_ASYNC; 624 rabp->b_iocmd = BIO_READ; 625 vfs_busy_pages(rabp, 0); 626 if (nfs_asyncio(nmp, rabp, cred, td)) { 627 rabp->b_flags |= B_INVAL; 628 rabp->b_ioflags |= BIO_ERROR; 629 vfs_unbusy_pages(rabp); 630 brelse(rabp); 631 } 632 } else { 633 brelse(rabp); 634 } 635 } 636 } 637 /* 638 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 639 * chopped for the EOF condition, we cannot tell how large 640 * NFS directories are going to be until we hit EOF. So 641 * an NFS directory buffer is *not* chopped to its EOF. Now, 642 * it just so happens that b_resid will effectively chop it 643 * to EOF. *BUT* this information is lost if the buffer goes 644 * away and is reconstituted into a B_CACHE state ( due to 645 * being VMIO ) later. So we keep track of the directory eof 646 * in np->n_direofoffset and chop it off as an extra step 647 * right here. 648 */ 649 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 650 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 651 n = np->n_direofoffset - uio->uio_offset; 652 break; 653 default: 654 printf(" nfs_bioread: type %x unexpected\n", vp->v_type); 655 bp = NULL; 656 break; 657 }; 658 659 if (n > 0) { 660 error = uiomove(bp->b_data + on, (int)n, uio); 661 } 662 if (vp->v_type == VLNK) 663 n = 0; 664 if (bp != NULL) 665 brelse(bp); 666 } while (error == 0 && uio->uio_resid > 0 && n > 0); 667 return (error); 668} 669 670/* 671 * The NFS write path cannot handle iovecs with len > 1. So we need to 672 * break up iovecs accordingly (restricting them to wsize). 673 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 674 * For the ASYNC case, 2 copies are needed. The first a copy from the 675 * user buffer to a staging buffer and then a second copy from the staging 676 * buffer to mbufs. This can be optimized by copying from the user buffer 677 * directly into mbufs and passing the chain down, but that requires a 678 * fair amount of re-working of the relevant codepaths (and can be done 679 * later). 680 */ 681static int 682nfs_directio_write(vp, uiop, cred, ioflag) 683 struct vnode *vp; 684 struct uio *uiop; 685 struct ucred *cred; 686 int ioflag; 687{ 688 int error; 689 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 690 struct thread *td = uiop->uio_td; 691 int size; 692 693 if (ioflag & IO_SYNC) { 694 int iomode, must_commit; 695 struct uio uio; 696 struct iovec iov; 697do_sync: 698 while (uiop->uio_resid > 0) { 699 size = min(uiop->uio_resid, nmp->nm_wsize); 700 size = min(uiop->uio_iov->iov_len, size); 701 iov.iov_base = uiop->uio_iov->iov_base; 702 iov.iov_len = size; 703 uio.uio_iov = &iov; 704 uio.uio_iovcnt = 1; 705 uio.uio_offset = uiop->uio_offset; 706 uio.uio_resid = size; 707 uio.uio_segflg = UIO_USERSPACE; 708 uio.uio_rw = UIO_WRITE; 709 uio.uio_td = td; 710 iomode = NFSV3WRITE_FILESYNC; 711 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, 712 &iomode, &must_commit); 713 KASSERT((must_commit == 0), 714 ("nfs_directio_write: Did not commit write")); 715 if (error) 716 return (error); 717 uiop->uio_offset += size; 718 uiop->uio_resid -= size; 719 if (uiop->uio_iov->iov_len <= size) { 720 uiop->uio_iovcnt--; 721 uiop->uio_iov++; 722 } else { 723 uiop->uio_iov->iov_base = 724 (char *)uiop->uio_iov->iov_base + size; 725 uiop->uio_iov->iov_len -= size; 726 } 727 } 728 } else { 729 struct uio *t_uio; 730 struct iovec *t_iov; 731 struct buf *bp; 732 733 /* 734 * Break up the write into blocksize chunks and hand these 735 * over to nfsiod's for write back. 736 * Unfortunately, this incurs a copy of the data. Since 737 * the user could modify the buffer before the write is 738 * initiated. 739 * 740 * The obvious optimization here is that one of the 2 copies 741 * in the async write path can be eliminated by copying the 742 * data here directly into mbufs and passing the mbuf chain 743 * down. But that will require a fair amount of re-working 744 * of the code and can be done if there's enough interest 745 * in NFS directio access. 746 */ 747 while (uiop->uio_resid > 0) { 748 size = min(uiop->uio_resid, nmp->nm_wsize); 749 size = min(uiop->uio_iov->iov_len, size); 750 bp = getpbuf(&nfs_pbuf_freecnt); 751 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 752 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 753 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 754 t_iov->iov_len = size; 755 t_uio->uio_iov = t_iov; 756 t_uio->uio_iovcnt = 1; 757 t_uio->uio_offset = uiop->uio_offset; 758 t_uio->uio_resid = size; 759 t_uio->uio_segflg = UIO_SYSSPACE; 760 t_uio->uio_rw = UIO_WRITE; 761 t_uio->uio_td = td; 762 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); 763 bp->b_flags |= B_DIRECT; 764 bp->b_iocmd = BIO_WRITE; 765 if (cred != NOCRED) { 766 crhold(cred); 767 bp->b_wcred = cred; 768 } else 769 bp->b_wcred = NOCRED; 770 bp->b_caller1 = (void *)t_uio; 771 bp->b_vp = vp; 772 vhold(vp); 773 error = nfs_asyncio(nmp, bp, NOCRED, td); 774 if (error) { 775 free(t_iov->iov_base, M_NFSDIRECTIO); 776 free(t_iov, M_NFSDIRECTIO); 777 free(t_uio, M_NFSDIRECTIO); 778 vdrop(bp->b_vp); 779 bp->b_vp = NULL; 780 relpbuf(bp, &nfs_pbuf_freecnt); 781 if (error == EINTR) 782 return (error); 783 goto do_sync; 784 } 785 uiop->uio_offset += size; 786 uiop->uio_resid -= size; 787 if (uiop->uio_iov->iov_len <= size) { 788 uiop->uio_iovcnt--; 789 uiop->uio_iov++; 790 } else { 791 uiop->uio_iov->iov_base = 792 (char *)uiop->uio_iov->iov_base + size; 793 uiop->uio_iov->iov_len -= size; 794 } 795 } 796 } 797 return (0); 798} 799 800/* 801 * Vnode op for write using bio 802 */ 803int 804nfs_write(struct vop_write_args *ap) 805{ 806 int biosize; 807 struct uio *uio = ap->a_uio; 808 struct thread *td = uio->uio_td; 809 struct vnode *vp = ap->a_vp; 810 struct nfsnode *np = VTONFS(vp); 811 struct ucred *cred = ap->a_cred; 812 int ioflag = ap->a_ioflag; 813 struct buf *bp; 814 struct vattr vattr; 815 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 816 daddr_t lbn; 817 int bcount; 818 int n, on, error = 0; 819 struct proc *p = td?td->td_proc:NULL; 820 821 GIANT_REQUIRED; 822 823#ifdef DIAGNOSTIC 824 if (uio->uio_rw != UIO_WRITE) 825 panic("nfs_write mode"); 826 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 827 panic("nfs_write proc"); 828#endif 829 if (vp->v_type != VREG) 830 return (EIO); 831 if (np->n_flag & NWRITEERR) { 832 np->n_flag &= ~NWRITEERR; 833 return (np->n_error); 834 } 835 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 836 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 837 (void)nfs_fsinfo(nmp, vp, cred, td); 838 839 /* 840 * Synchronously flush pending buffers if we are in synchronous 841 * mode or if we are appending. 842 */ 843 if (ioflag & (IO_APPEND | IO_SYNC)) { 844 if (np->n_flag & NMODIFIED) { 845#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 846 /* 847 * Require non-blocking, synchronous writes to 848 * dirty files to inform the program it needs 849 * to fsync(2) explicitly. 850 */ 851 if (ioflag & IO_NDELAY) 852 return (EAGAIN); 853#endif 854flush_and_restart: 855 np->n_attrstamp = 0; 856 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 857 if (error) 858 return (error); 859 } 860 } 861 862 /* 863 * If IO_APPEND then load uio_offset. We restart here if we cannot 864 * get the append lock. 865 */ 866 if (ioflag & IO_APPEND) { 867 np->n_attrstamp = 0; 868 error = VOP_GETATTR(vp, &vattr, cred, td); 869 if (error) 870 return (error); 871 uio->uio_offset = np->n_size; 872 } 873 874 if (uio->uio_offset < 0) 875 return (EINVAL); 876 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 877 return (EFBIG); 878 if (uio->uio_resid == 0) 879 return (0); 880 881 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 882 return nfs_directio_write(vp, uio, cred, ioflag); 883 884 /* 885 * Maybe this should be above the vnode op call, but so long as 886 * file servers have no limits, i don't think it matters 887 */ 888 if (p != NULL) { 889 PROC_LOCK(p); 890 if (uio->uio_offset + uio->uio_resid > 891 lim_cur(p, RLIMIT_FSIZE)) { 892 psignal(p, SIGXFSZ); 893 PROC_UNLOCK(p); 894 return (EFBIG); 895 } 896 PROC_UNLOCK(p); 897 } 898 899 biosize = vp->v_mount->mnt_stat.f_iosize; 900 /* 901 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 902 * would exceed the local maximum per-file write commit size when 903 * combined with those, we must decide whether to flush, 904 * go synchronous, or return error. We don't bother checking 905 * IO_UNIT -- we just make all writes atomic anyway, as there's 906 * no point optimizing for something that really won't ever happen. 907 */ 908 if (!(ioflag & IO_SYNC)) { 909 int needrestart = 0; 910 if (nmp->nm_wcommitsize < uio->uio_resid) { 911 /* 912 * If this request could not possibly be completed 913 * without exceeding the maximum outstanding write 914 * commit size, see if we can convert it into a 915 * synchronous write operation. 916 */ 917 if (ioflag & IO_NDELAY) 918 return (EAGAIN); 919 ioflag |= IO_SYNC; 920 if (np->n_flag & NMODIFIED) 921 needrestart = 1; 922 } else if (np->n_flag & NMODIFIED) { 923 int wouldcommit = 0; 924 BO_LOCK(&vp->v_bufobj); 925 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 926 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 927 b_bobufs) { 928 if (bp->b_flags & B_NEEDCOMMIT) 929 wouldcommit += bp->b_bcount; 930 } 931 } 932 BO_UNLOCK(&vp->v_bufobj); 933 /* 934 * Since we're not operating synchronously and 935 * bypassing the buffer cache, we are in a commit 936 * and holding all of these buffers whether 937 * transmitted or not. If not limited, this 938 * will lead to the buffer cache deadlocking, 939 * as no one else can flush our uncommitted buffers. 940 */ 941 wouldcommit += uio->uio_resid; 942 /* 943 * If we would initially exceed the maximum 944 * outstanding write commit size, flush and restart. 945 */ 946 if (wouldcommit > nmp->nm_wcommitsize) 947 needrestart = 1; 948 } 949 if (needrestart) 950 goto flush_and_restart; 951 } 952 953 do { 954 nfsstats.biocache_writes++; 955 lbn = uio->uio_offset / biosize; 956 on = uio->uio_offset & (biosize-1); 957 n = min((unsigned)(biosize - on), uio->uio_resid); 958again: 959 /* 960 * Handle direct append and file extension cases, calculate 961 * unaligned buffer size. 962 */ 963 964 if (uio->uio_offset == np->n_size && n) { 965 /* 966 * Get the buffer (in its pre-append state to maintain 967 * B_CACHE if it was previously set). Resize the 968 * nfsnode after we have locked the buffer to prevent 969 * readers from reading garbage. 970 */ 971 bcount = on; 972 bp = nfs_getcacheblk(vp, lbn, bcount, td); 973 974 if (bp != NULL) { 975 long save; 976 977 np->n_size = uio->uio_offset + n; 978 np->n_flag |= NMODIFIED; 979 vnode_pager_setsize(vp, np->n_size); 980 981 save = bp->b_flags & B_CACHE; 982 bcount += n; 983 allocbuf(bp, bcount); 984 bp->b_flags |= save; 985 } 986 } else { 987 /* 988 * Obtain the locked cache block first, and then 989 * adjust the file's size as appropriate. 990 */ 991 bcount = on + n; 992 if ((off_t)lbn * biosize + bcount < np->n_size) { 993 if ((off_t)(lbn + 1) * biosize < np->n_size) 994 bcount = biosize; 995 else 996 bcount = np->n_size - (off_t)lbn * biosize; 997 } 998 bp = nfs_getcacheblk(vp, lbn, bcount, td); 999 if (uio->uio_offset + n > np->n_size) { 1000 np->n_size = uio->uio_offset + n; 1001 np->n_flag |= NMODIFIED; 1002 vnode_pager_setsize(vp, np->n_size); 1003 } 1004 } 1005 1006 if (!bp) { 1007 error = nfs_sigintr(nmp, NULL, td); 1008 if (!error) 1009 error = EINTR; 1010 break; 1011 } 1012 1013 /* 1014 * Issue a READ if B_CACHE is not set. In special-append 1015 * mode, B_CACHE is based on the buffer prior to the write 1016 * op and is typically set, avoiding the read. If a read 1017 * is required in special append mode, the server will 1018 * probably send us a short-read since we extended the file 1019 * on our end, resulting in b_resid == 0 and, thusly, 1020 * B_CACHE getting set. 1021 * 1022 * We can also avoid issuing the read if the write covers 1023 * the entire buffer. We have to make sure the buffer state 1024 * is reasonable in this case since we will not be initiating 1025 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1026 * more information. 1027 * 1028 * B_CACHE may also be set due to the buffer being cached 1029 * normally. 1030 */ 1031 1032 if (on == 0 && n == bcount) { 1033 bp->b_flags |= B_CACHE; 1034 bp->b_flags &= ~B_INVAL; 1035 bp->b_ioflags &= ~BIO_ERROR; 1036 } 1037 1038 if ((bp->b_flags & B_CACHE) == 0) { 1039 bp->b_iocmd = BIO_READ; 1040 vfs_busy_pages(bp, 0); 1041 error = nfs_doio(vp, bp, cred, td); 1042 if (error) { 1043 brelse(bp); 1044 break; 1045 } 1046 } 1047 if (bp->b_wcred == NOCRED) 1048 bp->b_wcred = crhold(cred); 1049 np->n_flag |= NMODIFIED; 1050 1051 /* 1052 * If dirtyend exceeds file size, chop it down. This should 1053 * not normally occur but there is an append race where it 1054 * might occur XXX, so we log it. 1055 * 1056 * If the chopping creates a reverse-indexed or degenerate 1057 * situation with dirtyoff/end, we 0 both of them. 1058 */ 1059 1060 if (bp->b_dirtyend > bcount) { 1061 printf("NFS append race @%lx:%d\n", 1062 (long)bp->b_blkno * DEV_BSIZE, 1063 bp->b_dirtyend - bcount); 1064 bp->b_dirtyend = bcount; 1065 } 1066 1067 if (bp->b_dirtyoff >= bp->b_dirtyend) 1068 bp->b_dirtyoff = bp->b_dirtyend = 0; 1069 1070 /* 1071 * If the new write will leave a contiguous dirty 1072 * area, just update the b_dirtyoff and b_dirtyend, 1073 * otherwise force a write rpc of the old dirty area. 1074 * 1075 * While it is possible to merge discontiguous writes due to 1076 * our having a B_CACHE buffer ( and thus valid read data 1077 * for the hole), we don't because it could lead to 1078 * significant cache coherency problems with multiple clients, 1079 * especially if locking is implemented later on. 1080 * 1081 * as an optimization we could theoretically maintain 1082 * a linked list of discontinuous areas, but we would still 1083 * have to commit them separately so there isn't much 1084 * advantage to it except perhaps a bit of asynchronization. 1085 */ 1086 1087 if (bp->b_dirtyend > 0 && 1088 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1089 if (bwrite(bp) == EINTR) { 1090 error = EINTR; 1091 break; 1092 } 1093 goto again; 1094 } 1095 1096 error = uiomove((char *)bp->b_data + on, n, uio); 1097 1098 /* 1099 * Since this block is being modified, it must be written 1100 * again and not just committed. Since write clustering does 1101 * not work for the stage 1 data write, only the stage 2 1102 * commit rpc, we have to clear B_CLUSTEROK as well. 1103 */ 1104 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1105 1106 if (error) { 1107 bp->b_ioflags |= BIO_ERROR; 1108 brelse(bp); 1109 break; 1110 } 1111 1112 /* 1113 * Only update dirtyoff/dirtyend if not a degenerate 1114 * condition. 1115 */ 1116 if (n) { 1117 if (bp->b_dirtyend > 0) { 1118 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1119 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1120 } else { 1121 bp->b_dirtyoff = on; 1122 bp->b_dirtyend = on + n; 1123 } 1124 vfs_bio_set_validclean(bp, on, n); 1125 } 1126 1127 /* 1128 * If IO_SYNC do bwrite(). 1129 * 1130 * IO_INVAL appears to be unused. The idea appears to be 1131 * to turn off caching in this case. Very odd. XXX 1132 */ 1133 if ((ioflag & IO_SYNC)) { 1134 if (ioflag & IO_INVAL) 1135 bp->b_flags |= B_NOCACHE; 1136 error = bwrite(bp); 1137 if (error) 1138 break; 1139 } else if ((n + on) == biosize) { 1140 bp->b_flags |= B_ASYNC; 1141 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, 0); 1142 } else { 1143 bdwrite(bp); 1144 } 1145 } while (uio->uio_resid > 0 && n > 0); 1146 1147 return (error); 1148} 1149 1150/* 1151 * Get an nfs cache block. 1152 * 1153 * Allocate a new one if the block isn't currently in the cache 1154 * and return the block marked busy. If the calling process is 1155 * interrupted by a signal for an interruptible mount point, return 1156 * NULL. 1157 * 1158 * The caller must carefully deal with the possible B_INVAL state of 1159 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1160 * indirectly), so synchronous reads can be issued without worrying about 1161 * the B_INVAL state. We have to be a little more careful when dealing 1162 * with writes (see comments in nfs_write()) when extending a file past 1163 * its EOF. 1164 */ 1165static struct buf * 1166nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1167{ 1168 struct buf *bp; 1169 struct mount *mp; 1170 struct nfsmount *nmp; 1171 1172 mp = vp->v_mount; 1173 nmp = VFSTONFS(mp); 1174 1175 if (nmp->nm_flag & NFSMNT_INT) { 1176 sigset_t oldset; 1177 1178 nfs_set_sigmask(td, &oldset); 1179 bp = getblk(vp, bn, size, PCATCH, 0, 0); 1180 nfs_restore_sigmask(td, &oldset); 1181 while (bp == NULL) { 1182 if (nfs_sigintr(nmp, NULL, td)) 1183 return (NULL); 1184 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1185 } 1186 } else { 1187 bp = getblk(vp, bn, size, 0, 0, 0); 1188 } 1189 1190 if (vp->v_type == VREG) { 1191 int biosize; 1192 1193 biosize = mp->mnt_stat.f_iosize; 1194 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1195 } 1196 return (bp); 1197} 1198 1199/* 1200 * Flush and invalidate all dirty buffers. If another process is already 1201 * doing the flush, just wait for completion. 1202 */ 1203int 1204nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1205{ 1206 struct nfsnode *np = VTONFS(vp); 1207 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1208 int error = 0, slpflag, slptimeo; 1209 int old_lock = 0; 1210 1211 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf"); 1212 1213 /* 1214 * XXX This check stops us from needlessly doing a vinvalbuf when 1215 * being called through vclean(). It is not clear that this is 1216 * unsafe. 1217 */ 1218 if (vp->v_iflag & VI_DOOMED) 1219 return (0); 1220 1221 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1222 intrflg = 0; 1223 if (intrflg) { 1224 slpflag = PCATCH; 1225 slptimeo = 2 * hz; 1226 } else { 1227 slpflag = 0; 1228 slptimeo = 0; 1229 } 1230 1231 if ((old_lock = VOP_ISLOCKED(vp, td)) != LK_EXCLUSIVE) { 1232 if (old_lock == LK_SHARED) { 1233 /* Upgrade to exclusive lock, this might block */ 1234 vn_lock(vp, LK_UPGRADE | LK_RETRY, td); 1235 } else { 1236 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 1237 } 1238 } 1239 1240 /* 1241 * Now, flush as required. 1242 */ 1243 error = vinvalbuf(vp, flags, td, slpflag, 0); 1244 while (error) { 1245 if (intrflg && (error = nfs_sigintr(nmp, NULL, td))) 1246 goto out; 1247 error = vinvalbuf(vp, flags, td, 0, slptimeo); 1248 } 1249 np->n_flag &= ~NMODIFIED; 1250out: 1251 if (old_lock != LK_EXCLUSIVE) { 1252 if (old_lock == LK_SHARED) { 1253 /* Downgrade from exclusive lock, this might block */ 1254 vn_lock(vp, LK_DOWNGRADE, td); 1255 } else { 1256 VOP_UNLOCK(vp, 0, td); 1257 } 1258 } 1259 return error; 1260} 1261 1262/* 1263 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1264 * This is mainly to avoid queueing async I/O requests when the nfsiods 1265 * are all hung on a dead server. 1266 * 1267 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1268 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1269 */ 1270int 1271nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1272{ 1273 int iod; 1274 int gotiod; 1275 int slpflag = 0; 1276 int slptimeo = 0; 1277 int error, error2; 1278 1279 /* 1280 * Commits are usually short and sweet so lets save some cpu and 1281 * leave the async daemons for more important rpc's (such as reads 1282 * and writes). 1283 */ 1284 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1285 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1286 return(EIO); 1287 } 1288 1289again: 1290 if (nmp->nm_flag & NFSMNT_INT) 1291 slpflag = PCATCH; 1292 gotiod = FALSE; 1293 1294 /* 1295 * Find a free iod to process this request. 1296 */ 1297 for (iod = 0; iod < nfs_numasync; iod++) 1298 if (nfs_iodwant[iod]) { 1299 gotiod = TRUE; 1300 break; 1301 } 1302 1303 /* 1304 * Try to create one if none are free. 1305 */ 1306 if (!gotiod) { 1307 iod = nfs_nfsiodnew(); 1308 if (iod != -1) 1309 gotiod = TRUE; 1310 } 1311 1312 if (gotiod) { 1313 /* 1314 * Found one, so wake it up and tell it which 1315 * mount to process. 1316 */ 1317 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n", 1318 iod, nmp)); 1319 nfs_iodwant[iod] = NULL; 1320 nfs_iodmount[iod] = nmp; 1321 nmp->nm_bufqiods++; 1322 wakeup(&nfs_iodwant[iod]); 1323 } 1324 1325 /* 1326 * If none are free, we may already have an iod working on this mount 1327 * point. If so, it will process our request. 1328 */ 1329 if (!gotiod) { 1330 if (nmp->nm_bufqiods > 0) { 1331 NFS_DPF(ASYNCIO, 1332 ("nfs_asyncio: %d iods are already processing mount %p\n", 1333 nmp->nm_bufqiods, nmp)); 1334 gotiod = TRUE; 1335 } 1336 } 1337 1338 /* 1339 * If we have an iod which can process the request, then queue 1340 * the buffer. 1341 */ 1342 if (gotiod) { 1343 /* 1344 * Ensure that the queue never grows too large. We still want 1345 * to asynchronize so we block rather then return EIO. 1346 */ 1347 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1348 NFS_DPF(ASYNCIO, 1349 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1350 nmp->nm_bufqwant = TRUE; 1351 error = nfs_tsleep(td, &nmp->nm_bufq, slpflag | PRIBIO, 1352 "nfsaio", slptimeo); 1353 if (error) { 1354 error2 = nfs_sigintr(nmp, NULL, td); 1355 if (error2) 1356 return (error2); 1357 if (slpflag == PCATCH) { 1358 slpflag = 0; 1359 slptimeo = 2 * hz; 1360 } 1361 } 1362 /* 1363 * We might have lost our iod while sleeping, 1364 * so check and loop if nescessary. 1365 */ 1366 if (nmp->nm_bufqiods == 0) { 1367 NFS_DPF(ASYNCIO, 1368 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1369 goto again; 1370 } 1371 } 1372 1373 if (bp->b_iocmd == BIO_READ) { 1374 if (bp->b_rcred == NOCRED && cred != NOCRED) 1375 bp->b_rcred = crhold(cred); 1376 } else { 1377 if (bp->b_wcred == NOCRED && cred != NOCRED) 1378 bp->b_wcred = crhold(cred); 1379 } 1380 1381 if (bp->b_flags & B_REMFREE) 1382 bremfreef(bp); 1383 BUF_KERNPROC(bp); 1384 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1385 nmp->nm_bufqlen++; 1386 return (0); 1387 } 1388 1389 /* 1390 * All the iods are busy on other mounts, so return EIO to 1391 * force the caller to process the i/o synchronously. 1392 */ 1393 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1394 return (EIO); 1395} 1396 1397void 1398nfs_doio_directwrite(struct buf *bp) 1399{ 1400 int iomode, must_commit; 1401 struct uio *uiop = (struct uio *)bp->b_caller1; 1402 char *iov_base = uiop->uio_iov->iov_base; 1403 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount); 1404 1405 iomode = NFSV3WRITE_FILESYNC; 1406 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1407 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit); 1408 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write")); 1409 free(iov_base, M_NFSDIRECTIO); 1410 free(uiop->uio_iov, M_NFSDIRECTIO); 1411 free(uiop, M_NFSDIRECTIO); 1412 vdrop(bp->b_vp); 1413 bp->b_vp = NULL; 1414 relpbuf(bp, &nfs_pbuf_freecnt); 1415} 1416 1417/* 1418 * Do an I/O operation to/from a cache block. This may be called 1419 * synchronously or from an nfsiod. 1420 */ 1421int 1422nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) 1423{ 1424 struct uio *uiop; 1425 struct nfsnode *np; 1426 struct nfsmount *nmp; 1427 int error = 0, iomode, must_commit = 0; 1428 struct uio uio; 1429 struct iovec io; 1430 struct proc *p = td ? td->td_proc : NULL; 1431 1432 np = VTONFS(vp); 1433 nmp = VFSTONFS(vp->v_mount); 1434 uiop = &uio; 1435 uiop->uio_iov = &io; 1436 uiop->uio_iovcnt = 1; 1437 uiop->uio_segflg = UIO_SYSSPACE; 1438 uiop->uio_td = td; 1439 1440 /* 1441 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1442 * do this here so we do not have to do it in all the code that 1443 * calls us. 1444 */ 1445 bp->b_flags &= ~B_INVAL; 1446 bp->b_ioflags &= ~BIO_ERROR; 1447 1448 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1449 1450 if (bp->b_iocmd == BIO_READ) { 1451 io.iov_len = uiop->uio_resid = bp->b_bcount; 1452 io.iov_base = bp->b_data; 1453 uiop->uio_rw = UIO_READ; 1454 1455 switch (vp->v_type) { 1456 case VREG: 1457 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1458 nfsstats.read_bios++; 1459 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr); 1460 1461 if (!error) { 1462 if (uiop->uio_resid) { 1463 /* 1464 * If we had a short read with no error, we must have 1465 * hit a file hole. We should zero-fill the remainder. 1466 * This can also occur if the server hits the file EOF. 1467 * 1468 * Holes used to be able to occur due to pending 1469 * writes, but that is not possible any longer. 1470 */ 1471 int nread = bp->b_bcount - uiop->uio_resid; 1472 int left = uiop->uio_resid; 1473 1474 if (left > 0) 1475 bzero((char *)bp->b_data + nread, left); 1476 uiop->uio_resid = 0; 1477 } 1478 } 1479 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */ 1480 if (p && (vp->v_vflag & VV_TEXT) && 1481 (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime))) { 1482 PROC_LOCK(p); 1483 killproc(p, "text file modification"); 1484 PROC_UNLOCK(p); 1485 } 1486 break; 1487 case VLNK: 1488 uiop->uio_offset = (off_t)0; 1489 nfsstats.readlink_bios++; 1490 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr); 1491 break; 1492 case VDIR: 1493 nfsstats.readdir_bios++; 1494 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1495 if ((nmp->nm_flag & NFSMNT_NFSV4) != 0) 1496 error = nfs4_readdirrpc(vp, uiop, cr); 1497 else { 1498 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1499 error = nfs_readdirplusrpc(vp, uiop, cr); 1500 if (error == NFSERR_NOTSUPP) 1501 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1502 } 1503 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1504 error = nfs_readdirrpc(vp, uiop, cr); 1505 } 1506 /* 1507 * end-of-directory sets B_INVAL but does not generate an 1508 * error. 1509 */ 1510 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1511 bp->b_flags |= B_INVAL; 1512 break; 1513 default: 1514 printf("nfs_doio: type %x unexpected\n", vp->v_type); 1515 break; 1516 }; 1517 if (error) { 1518 bp->b_ioflags |= BIO_ERROR; 1519 bp->b_error = error; 1520 } 1521 } else { 1522 /* 1523 * If we only need to commit, try to commit 1524 */ 1525 if (bp->b_flags & B_NEEDCOMMIT) { 1526 int retv; 1527 off_t off; 1528 1529 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1530 retv = (nmp->nm_rpcops->nr_commit)( 1531 vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1532 bp->b_wcred, td); 1533 if (retv == 0) { 1534 bp->b_dirtyoff = bp->b_dirtyend = 0; 1535 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1536 bp->b_resid = 0; 1537 bufdone(bp); 1538 return (0); 1539 } 1540 if (retv == NFSERR_STALEWRITEVERF) { 1541 nfs_clearcommit(vp->v_mount); 1542 } 1543 } 1544 1545 /* 1546 * Setup for actual write 1547 */ 1548 1549 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1550 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1551 1552 if (bp->b_dirtyend > bp->b_dirtyoff) { 1553 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1554 - bp->b_dirtyoff; 1555 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1556 + bp->b_dirtyoff; 1557 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1558 uiop->uio_rw = UIO_WRITE; 1559 nfsstats.write_bios++; 1560 1561 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1562 iomode = NFSV3WRITE_UNSTABLE; 1563 else 1564 iomode = NFSV3WRITE_FILESYNC; 1565 1566 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit); 1567 1568 /* 1569 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1570 * to cluster the buffers needing commit. This will allow 1571 * the system to submit a single commit rpc for the whole 1572 * cluster. We can do this even if the buffer is not 100% 1573 * dirty (relative to the NFS blocksize), so we optimize the 1574 * append-to-file-case. 1575 * 1576 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1577 * cleared because write clustering only works for commit 1578 * rpc's, not for the data portion of the write). 1579 */ 1580 1581 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1582 bp->b_flags |= B_NEEDCOMMIT; 1583 if (bp->b_dirtyoff == 0 1584 && bp->b_dirtyend == bp->b_bcount) 1585 bp->b_flags |= B_CLUSTEROK; 1586 } else { 1587 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1588 } 1589 1590 /* 1591 * For an interrupted write, the buffer is still valid 1592 * and the write hasn't been pushed to the server yet, 1593 * so we can't set BIO_ERROR and report the interruption 1594 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1595 * is not relevant, so the rpc attempt is essentially 1596 * a noop. For the case of a V3 write rpc not being 1597 * committed to stable storage, the block is still 1598 * dirty and requires either a commit rpc or another 1599 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1600 * the block is reused. This is indicated by setting 1601 * the B_DELWRI and B_NEEDCOMMIT flags. 1602 * 1603 * If the buffer is marked B_PAGING, it does not reside on 1604 * the vp's paging queues so we cannot call bdirty(). The 1605 * bp in this case is not an NFS cache block so we should 1606 * be safe. XXX 1607 */ 1608 if (error == EINTR || error == EIO || error == ETIMEDOUT 1609 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1610 int s; 1611 1612 s = splbio(); 1613 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1614 if ((bp->b_flags & B_PAGING) == 0) { 1615 bdirty(bp); 1616 bp->b_flags &= ~B_DONE; 1617 } 1618 if (error && (bp->b_flags & B_ASYNC) == 0) 1619 bp->b_flags |= B_EINTR; 1620 splx(s); 1621 } else { 1622 if (error) { 1623 bp->b_ioflags |= BIO_ERROR; 1624 bp->b_error = np->n_error = error; 1625 np->n_flag |= NWRITEERR; 1626 } 1627 bp->b_dirtyoff = bp->b_dirtyend = 0; 1628 } 1629 } else { 1630 bp->b_resid = 0; 1631 bufdone(bp); 1632 return (0); 1633 } 1634 } 1635 bp->b_resid = uiop->uio_resid; 1636 if (must_commit) 1637 nfs_clearcommit(vp->v_mount); 1638 bufdone(bp); 1639 return (error); 1640} 1641 1642/* 1643 * Used to aid in handling ftruncate() operations on the NFS client side. 1644 * Truncation creates a number of special problems for NFS. We have to 1645 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1646 * we have to properly handle VM pages or (potentially dirty) buffers 1647 * that straddle the truncation point. 1648 */ 1649 1650int 1651nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1652{ 1653 struct nfsnode *np = VTONFS(vp); 1654 u_quad_t tsize = np->n_size; 1655 int biosize = vp->v_mount->mnt_stat.f_iosize; 1656 int error = 0; 1657 1658 np->n_size = nsize; 1659 1660 if (np->n_size < tsize) { 1661 struct buf *bp; 1662 daddr_t lbn; 1663 int bufsize; 1664 1665 /* 1666 * vtruncbuf() doesn't get the buffer overlapping the 1667 * truncation point. We may have a B_DELWRI and/or B_CACHE 1668 * buffer that now needs to be truncated. 1669 */ 1670 error = vtruncbuf(vp, cred, td, nsize, biosize); 1671 lbn = nsize / biosize; 1672 bufsize = nsize & (biosize - 1); 1673 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1674 if (!bp) 1675 return EINTR; 1676 if (bp->b_dirtyoff > bp->b_bcount) 1677 bp->b_dirtyoff = bp->b_bcount; 1678 if (bp->b_dirtyend > bp->b_bcount) 1679 bp->b_dirtyend = bp->b_bcount; 1680 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1681 brelse(bp); 1682 } else { 1683 vnode_pager_setsize(vp, nsize); 1684 } 1685 return(error); 1686} 1687 1688