nfs_bio.c revision 41791
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 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 37 * $Id: nfs_bio.c,v 1.64 1998/12/07 21:58:43 archie Exp $ 38 */ 39 40 41#include <sys/param.h> 42#include <sys/systm.h> 43#include <sys/resourcevar.h> 44#include <sys/signalvar.h> 45#include <sys/proc.h> 46#include <sys/buf.h> 47#include <sys/vnode.h> 48#include <sys/mount.h> 49#include <sys/kernel.h> 50 51#include <vm/vm.h> 52#include <vm/vm_extern.h> 53#include <vm/vm_prot.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 <nfs/rpcv2.h> 60#include <nfs/nfsproto.h> 61#include <nfs/nfs.h> 62#include <nfs/nfsmount.h> 63#include <nfs/nqnfs.h> 64#include <nfs/nfsnode.h> 65 66static struct buf *nfs_getcacheblk __P((struct vnode *vp, daddr_t bn, int size, 67 struct proc *p)); 68static void nfs_prot_buf __P((struct buf *bp, int off, int n)); 69 70extern int nfs_numasync; 71extern struct nfsstats nfsstats; 72 73/* 74 * Vnode op for VM getpages. 75 */ 76int 77nfs_getpages(ap) 78 struct vop_getpages_args /* { 79 struct vnode *a_vp; 80 vm_page_t *a_m; 81 int a_count; 82 int a_reqpage; 83 vm_ooffset_t a_offset; 84 } */ *ap; 85{ 86 int i, error, nextoff, size, toff, npages, count; 87 struct uio uio; 88 struct iovec iov; 89 vm_offset_t kva; 90 struct buf *bp; 91 struct vnode *vp; 92 struct proc *p; 93 struct ucred *cred; 94 struct nfsmount *nmp; 95 vm_page_t *pages; 96 97 vp = ap->a_vp; 98 p = curproc; /* XXX */ 99 cred = curproc->p_ucred; /* XXX */ 100 nmp = VFSTONFS(vp->v_mount); 101 pages = ap->a_m; 102 count = ap->a_count; 103 104 if (vp->v_object == NULL) { 105 printf("nfs_getpages: called with non-merged cache vnode??\n"); 106 return VM_PAGER_ERROR; 107 } 108 109 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 110 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 111 (void)nfs_fsinfo(nmp, vp, cred, p); 112 /* 113 * We use only the kva address for the buffer, but this is extremely 114 * convienient and fast. 115 */ 116 bp = getpbuf(); 117 118 npages = btoc(count); 119 kva = (vm_offset_t) bp->b_data; 120 pmap_qenter(kva, pages, npages); 121 122 iov.iov_base = (caddr_t) kva; 123 iov.iov_len = count; 124 uio.uio_iov = &iov; 125 uio.uio_iovcnt = 1; 126 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 127 uio.uio_resid = count; 128 uio.uio_segflg = UIO_SYSSPACE; 129 uio.uio_rw = UIO_READ; 130 uio.uio_procp = p; 131 132 error = nfs_readrpc(vp, &uio, cred); 133 pmap_qremove(kva, npages); 134 135 relpbuf(bp); 136 137 if (error && (uio.uio_resid == count)) 138 return VM_PAGER_ERROR; 139 140 size = count - uio.uio_resid; 141 142 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 143 vm_page_t m; 144 nextoff = toff + PAGE_SIZE; 145 m = pages[i]; 146 147 m->flags &= ~PG_ZERO; 148 149 if (nextoff <= size) { 150 m->valid = VM_PAGE_BITS_ALL; 151 m->dirty = 0; 152 } else { 153 int nvalid = ((size + DEV_BSIZE - 1) - toff) & ~(DEV_BSIZE - 1); 154 vm_page_set_validclean(m, 0, nvalid); 155 } 156 157 if (i != ap->a_reqpage) { 158 /* 159 * Whether or not to leave the page activated is up in 160 * the air, but we should put the page on a page queue 161 * somewhere (it already is in the object). Result: 162 * It appears that emperical results show that 163 * deactivating pages is best. 164 */ 165 166 /* 167 * Just in case someone was asking for this page we 168 * now tell them that it is ok to use. 169 */ 170 if (!error) { 171 if (m->flags & PG_WANTED) 172 vm_page_activate(m); 173 else 174 vm_page_deactivate(m); 175 vm_page_wakeup(m); 176 } else { 177 vnode_pager_freepage(m); 178 } 179 } 180 } 181 return 0; 182} 183 184/* 185 * Vnode op for VM putpages. 186 */ 187int 188nfs_putpages(ap) 189 struct vop_putpages_args /* { 190 struct vnode *a_vp; 191 vm_page_t *a_m; 192 int a_count; 193 int a_sync; 194 int *a_rtvals; 195 vm_ooffset_t a_offset; 196 } */ *ap; 197{ 198 struct uio uio; 199 struct iovec iov; 200 vm_offset_t kva; 201 struct buf *bp; 202 int iomode, must_commit, i, error, npages, count; 203 int *rtvals; 204 struct vnode *vp; 205 struct proc *p; 206 struct ucred *cred; 207 struct nfsmount *nmp; 208 vm_page_t *pages; 209 210 vp = ap->a_vp; 211 p = curproc; /* XXX */ 212 cred = curproc->p_ucred; /* XXX */ 213 nmp = VFSTONFS(vp->v_mount); 214 pages = ap->a_m; 215 count = ap->a_count; 216 rtvals = ap->a_rtvals; 217 npages = btoc(count); 218 219 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 220 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 221 (void)nfs_fsinfo(nmp, vp, cred, p); 222 223 for (i = 0; i < npages; i++) { 224 rtvals[i] = VM_PAGER_AGAIN; 225 } 226 227 /* 228 * We use only the kva address for the buffer, but this is extremely 229 * convienient and fast. 230 */ 231 bp = getpbuf(); 232 233 kva = (vm_offset_t) bp->b_data; 234 pmap_qenter(kva, pages, npages); 235 236 iov.iov_base = (caddr_t) kva; 237 iov.iov_len = count; 238 uio.uio_iov = &iov; 239 uio.uio_iovcnt = 1; 240 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 241 uio.uio_resid = count; 242 uio.uio_segflg = UIO_SYSSPACE; 243 uio.uio_rw = UIO_WRITE; 244 uio.uio_procp = p; 245 246 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 247 iomode = NFSV3WRITE_UNSTABLE; 248 else 249 iomode = NFSV3WRITE_FILESYNC; 250 251 error = nfs_writerpc(vp, &uio, cred, &iomode, &must_commit); 252 253 pmap_qremove(kva, npages); 254 relpbuf(bp); 255 256 if (!error) { 257 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 258 for (i = 0; i < nwritten; i++) { 259 rtvals[i] = VM_PAGER_OK; 260 pages[i]->dirty = 0; 261 } 262 if (must_commit) 263 nfs_clearcommit(vp->v_mount); 264 } 265 return rtvals[0]; 266} 267 268/* 269 * Vnode op for read using bio 270 */ 271int 272nfs_bioread(vp, uio, ioflag, cred, getpages) 273 register struct vnode *vp; 274 register struct uio *uio; 275 int ioflag; 276 struct ucred *cred; 277 int getpages; 278{ 279 register struct nfsnode *np = VTONFS(vp); 280 register int biosize, i; 281 off_t diff; 282 struct buf *bp = 0, *rabp; 283 struct vattr vattr; 284 struct proc *p; 285 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 286 daddr_t lbn, rabn; 287 int bufsize; 288 int nra, error = 0, n = 0, on = 0, not_readin; 289 290#ifdef DIAGNOSTIC 291 if (uio->uio_rw != UIO_READ) 292 panic("nfs_read mode"); 293#endif 294 if (uio->uio_resid == 0) 295 return (0); 296 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 297 return (EINVAL); 298 p = uio->uio_procp; 299 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 300 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 301 (void)nfs_fsinfo(nmp, vp, cred, p); 302 if (vp->v_type != VDIR && 303 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 304 return (EFBIG); 305 biosize = vp->v_mount->mnt_stat.f_iosize; 306 /* 307 * For nfs, cache consistency can only be maintained approximately. 308 * Although RFC1094 does not specify the criteria, the following is 309 * believed to be compatible with the reference port. 310 * For nqnfs, full cache consistency is maintained within the loop. 311 * For nfs: 312 * If the file's modify time on the server has changed since the 313 * last read rpc or you have written to the file, 314 * you may have lost data cache consistency with the 315 * server, so flush all of the file's data out of the cache. 316 * Then force a getattr rpc to ensure that you have up to date 317 * attributes. 318 * NB: This implies that cache data can be read when up to 319 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 320 * attributes this could be forced by setting n_attrstamp to 0 before 321 * the VOP_GETATTR() call. 322 */ 323 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) { 324 if (np->n_flag & NMODIFIED) { 325 if (vp->v_type != VREG) { 326 if (vp->v_type != VDIR) 327 panic("nfs: bioread, not dir"); 328 nfs_invaldir(vp); 329 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 330 if (error) 331 return (error); 332 } 333 np->n_attrstamp = 0; 334 error = VOP_GETATTR(vp, &vattr, cred, p); 335 if (error) 336 return (error); 337 np->n_mtime = vattr.va_mtime.tv_sec; 338 } else { 339 error = VOP_GETATTR(vp, &vattr, cred, p); 340 if (error) 341 return (error); 342 if (np->n_mtime != vattr.va_mtime.tv_sec) { 343 if (vp->v_type == VDIR) 344 nfs_invaldir(vp); 345 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 346 if (error) 347 return (error); 348 np->n_mtime = vattr.va_mtime.tv_sec; 349 } 350 } 351 } 352 do { 353 354 /* 355 * Get a valid lease. If cached data is stale, flush it. 356 */ 357 if (nmp->nm_flag & NFSMNT_NQNFS) { 358 if (NQNFS_CKINVALID(vp, np, ND_READ)) { 359 do { 360 error = nqnfs_getlease(vp, ND_READ, cred, p); 361 } while (error == NQNFS_EXPIRED); 362 if (error) 363 return (error); 364 if (np->n_lrev != np->n_brev || 365 (np->n_flag & NQNFSNONCACHE) || 366 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) { 367 if (vp->v_type == VDIR) 368 nfs_invaldir(vp); 369 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 370 if (error) 371 return (error); 372 np->n_brev = np->n_lrev; 373 } 374 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) { 375 nfs_invaldir(vp); 376 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 377 if (error) 378 return (error); 379 } 380 } 381 if (np->n_flag & NQNFSNONCACHE) { 382 switch (vp->v_type) { 383 case VREG: 384 return (nfs_readrpc(vp, uio, cred)); 385 case VLNK: 386 return (nfs_readlinkrpc(vp, uio, cred)); 387 case VDIR: 388 break; 389 default: 390 printf(" NQNFSNONCACHE: type %x unexpected\n", 391 vp->v_type); 392 }; 393 } 394 switch (vp->v_type) { 395 case VREG: 396 nfsstats.biocache_reads++; 397 lbn = uio->uio_offset / biosize; 398 on = uio->uio_offset & (biosize - 1); 399 not_readin = 1; 400 401 /* 402 * Start the read ahead(s), as required. 403 */ 404 if (nfs_numasync > 0 && nmp->nm_readahead > 0) { 405 for (nra = 0; nra < nmp->nm_readahead && 406 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 407 rabn = lbn + 1 + nra; 408 if (!incore(vp, rabn)) { 409 rabp = nfs_getcacheblk(vp, rabn, biosize, p); 410 if (!rabp) 411 return (EINTR); 412 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 413 rabp->b_flags |= (B_READ | B_ASYNC); 414 vfs_busy_pages(rabp, 0); 415 if (nfs_asyncio(rabp, cred)) { 416 rabp->b_flags |= B_INVAL|B_ERROR; 417 vfs_unbusy_pages(rabp); 418 brelse(rabp); 419 } 420 } else 421 brelse(rabp); 422 } 423 } 424 } 425 426 /* 427 * If the block is in the cache and has the required data 428 * in a valid region, just copy it out. 429 * Otherwise, get the block and write back/read in, 430 * as required. 431 */ 432again: 433 bufsize = biosize; 434 if ((off_t)(lbn + 1) * biosize > np->n_size && 435 (off_t)(lbn + 1) * biosize - np->n_size < biosize) { 436 bufsize = np->n_size - (off_t)lbn * biosize; 437 bufsize = (bufsize + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 438 } 439 bp = nfs_getcacheblk(vp, lbn, bufsize, p); 440 if (!bp) 441 return (EINTR); 442 /* 443 * If we are being called from nfs_getpages, we must 444 * make sure the buffer is a vmio buffer. The vp will 445 * already be setup for vmio but there may be some old 446 * non-vmio buffers attached to it. 447 */ 448 if (getpages && !(bp->b_flags & B_VMIO)) { 449#ifdef DIAGNOSTIC 450 printf("nfs_bioread: non vmio buf found, discarding\n"); 451#endif 452 bp->b_flags |= B_NOCACHE; 453 bp->b_flags |= B_INVAFTERWRITE; 454 if (bp->b_dirtyend > 0) { 455 if ((bp->b_flags & B_DELWRI) == 0) 456 panic("nfsbioread"); 457 if (VOP_BWRITE(bp) == EINTR) 458 return (EINTR); 459 } else 460 brelse(bp); 461 goto again; 462 } 463 if ((bp->b_flags & B_CACHE) == 0) { 464 bp->b_flags |= B_READ; 465 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 466 not_readin = 0; 467 vfs_busy_pages(bp, 0); 468 error = nfs_doio(bp, cred, p); 469 if (error) { 470 brelse(bp); 471 return (error); 472 } 473 } 474 if (bufsize > on) { 475 n = min((unsigned)(bufsize - on), uio->uio_resid); 476 } else { 477 n = 0; 478 } 479 diff = np->n_size - uio->uio_offset; 480 if (diff < n) 481 n = diff; 482 if (not_readin && n > 0) { 483 if (on < bp->b_validoff || (on + n) > bp->b_validend) { 484 bp->b_flags |= B_NOCACHE; 485 bp->b_flags |= B_INVAFTERWRITE; 486 if (bp->b_dirtyend > 0) { 487 if ((bp->b_flags & B_DELWRI) == 0) 488 panic("nfsbioread"); 489 if (VOP_BWRITE(bp) == EINTR) 490 return (EINTR); 491 } else 492 brelse(bp); 493 goto again; 494 } 495 } 496 vp->v_lastr = lbn; 497 diff = (on >= bp->b_validend) ? 0 : (bp->b_validend - on); 498 if (diff < n) 499 n = diff; 500 break; 501 case VLNK: 502 nfsstats.biocache_readlinks++; 503 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p); 504 if (!bp) 505 return (EINTR); 506 if ((bp->b_flags & B_CACHE) == 0) { 507 bp->b_flags |= B_READ; 508 vfs_busy_pages(bp, 0); 509 error = nfs_doio(bp, cred, p); 510 if (error) { 511 bp->b_flags |= B_ERROR; 512 brelse(bp); 513 return (error); 514 } 515 } 516 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 517 on = 0; 518 break; 519 case VDIR: 520 nfsstats.biocache_readdirs++; 521 if (np->n_direofoffset 522 && uio->uio_offset >= np->n_direofoffset) { 523 return (0); 524 } 525 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 526 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 527 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, p); 528 if (!bp) 529 return (EINTR); 530 if ((bp->b_flags & B_CACHE) == 0) { 531 bp->b_flags |= B_READ; 532 vfs_busy_pages(bp, 0); 533 error = nfs_doio(bp, cred, p); 534 if (error) { 535 brelse(bp); 536 } 537 while (error == NFSERR_BAD_COOKIE) { 538 nfs_invaldir(vp); 539 error = nfs_vinvalbuf(vp, 0, cred, p, 1); 540 /* 541 * Yuck! The directory has been modified on the 542 * server. The only way to get the block is by 543 * reading from the beginning to get all the 544 * offset cookies. 545 */ 546 for (i = 0; i <= lbn && !error; i++) { 547 if (np->n_direofoffset 548 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 549 return (0); 550 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, p); 551 if (!bp) 552 return (EINTR); 553 if ((bp->b_flags & B_DONE) == 0) { 554 bp->b_flags |= B_READ; 555 vfs_busy_pages(bp, 0); 556 error = nfs_doio(bp, cred, p); 557 if (error == 0 && (bp->b_flags & B_INVAL)) 558 break; 559 if (error) { 560 brelse(bp); 561 } else if (i < lbn) { 562 brelse(bp); 563 } 564 } 565 } 566 } 567 if (error) 568 return (error); 569 } 570 571 /* 572 * If not eof and read aheads are enabled, start one. 573 * (You need the current block first, so that you have the 574 * directory offset cookie of the next block.) 575 */ 576 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 577 (bp->b_flags & B_INVAL) == 0 && 578 (np->n_direofoffset == 0 || 579 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 580 !(np->n_flag & NQNFSNONCACHE) && 581 !incore(vp, lbn + 1)) { 582 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, p); 583 if (rabp) { 584 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 585 rabp->b_flags |= (B_READ | B_ASYNC); 586 vfs_busy_pages(rabp, 0); 587 if (nfs_asyncio(rabp, cred)) { 588 rabp->b_flags |= B_INVAL|B_ERROR; 589 vfs_unbusy_pages(rabp); 590 brelse(rabp); 591 } 592 } else { 593 brelse(rabp); 594 } 595 } 596 } 597 /* 598 * Make sure we use a signed variant of min() since 599 * the second term may be negative. 600 */ 601 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 602 break; 603 default: 604 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 605 break; 606 }; 607 608 if (n > 0) { 609 error = uiomove(bp->b_data + on, (int)n, uio); 610 } 611 switch (vp->v_type) { 612 case VREG: 613 break; 614 case VLNK: 615 n = 0; 616 break; 617 case VDIR: 618 if (np->n_flag & NQNFSNONCACHE) 619 bp->b_flags |= B_INVAL; 620 break; 621 default: 622 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 623 } 624 brelse(bp); 625 } while (error == 0 && uio->uio_resid > 0 && n > 0); 626 return (error); 627} 628 629static void 630nfs_prot_buf(bp, off, n) 631 struct buf *bp; 632 int off; 633 int n; 634{ 635 int pindex, boff, end; 636 637 if ((bp->b_flags & B_VMIO) == 0) 638 return; 639 640 end = round_page(off + n); 641 for (boff = trunc_page(off); boff < end; boff += PAGE_SIZE) { 642 pindex = boff >> PAGE_SHIFT; 643 vm_page_protect(bp->b_pages[pindex], VM_PROT_NONE); 644 } 645} 646 647/* 648 * Vnode op for write using bio 649 */ 650int 651nfs_write(ap) 652 struct vop_write_args /* { 653 struct vnode *a_vp; 654 struct uio *a_uio; 655 int a_ioflag; 656 struct ucred *a_cred; 657 } */ *ap; 658{ 659 register int biosize; 660 register struct uio *uio = ap->a_uio; 661 struct proc *p = uio->uio_procp; 662 register struct vnode *vp = ap->a_vp; 663 struct nfsnode *np = VTONFS(vp); 664 register struct ucred *cred = ap->a_cred; 665 int ioflag = ap->a_ioflag; 666 struct buf *bp; 667 struct vattr vattr; 668 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 669 daddr_t lbn; 670 int bufsize; 671 int n, on, error = 0, iomode, must_commit; 672 673#ifdef DIAGNOSTIC 674 if (uio->uio_rw != UIO_WRITE) 675 panic("nfs_write mode"); 676 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc) 677 panic("nfs_write proc"); 678#endif 679 if (vp->v_type != VREG) 680 return (EIO); 681 if (np->n_flag & NWRITEERR) { 682 np->n_flag &= ~NWRITEERR; 683 return (np->n_error); 684 } 685 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 686 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 687 (void)nfs_fsinfo(nmp, vp, cred, p); 688 if (ioflag & (IO_APPEND | IO_SYNC)) { 689 if (np->n_flag & NMODIFIED) { 690 np->n_attrstamp = 0; 691 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 692 if (error) 693 return (error); 694 } 695 if (ioflag & IO_APPEND) { 696 np->n_attrstamp = 0; 697 error = VOP_GETATTR(vp, &vattr, cred, p); 698 if (error) 699 return (error); 700 uio->uio_offset = np->n_size; 701 } 702 } 703 if (uio->uio_offset < 0) 704 return (EINVAL); 705 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 706 return (EFBIG); 707 if (uio->uio_resid == 0) 708 return (0); 709 /* 710 * Maybe this should be above the vnode op call, but so long as 711 * file servers have no limits, i don't think it matters 712 */ 713 if (p && uio->uio_offset + uio->uio_resid > 714 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 715 psignal(p, SIGXFSZ); 716 return (EFBIG); 717 } 718 /* 719 * I use nm_rsize, not nm_wsize so that all buffer cache blocks 720 * will be the same size within a filesystem. nfs_writerpc will 721 * still use nm_wsize when sizing the rpc's. 722 */ 723 biosize = vp->v_mount->mnt_stat.f_iosize; 724 do { 725 /* 726 * Check for a valid write lease. 727 */ 728 if ((nmp->nm_flag & NFSMNT_NQNFS) && 729 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 730 do { 731 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 732 } while (error == NQNFS_EXPIRED); 733 if (error) 734 return (error); 735 if (np->n_lrev != np->n_brev || 736 (np->n_flag & NQNFSNONCACHE)) { 737 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 738 if (error) 739 return (error); 740 np->n_brev = np->n_lrev; 741 } 742 } 743 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 744 iomode = NFSV3WRITE_FILESYNC; 745 error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit); 746 if (must_commit) 747 nfs_clearcommit(vp->v_mount); 748 return (error); 749 } 750 nfsstats.biocache_writes++; 751 lbn = uio->uio_offset / biosize; 752 on = uio->uio_offset & (biosize-1); 753 n = min((unsigned)(biosize - on), uio->uio_resid); 754again: 755 if (uio->uio_offset + n > np->n_size) { 756 np->n_size = uio->uio_offset + n; 757 np->n_flag |= NMODIFIED; 758 vnode_pager_setsize(vp, np->n_size); 759 } 760 bufsize = biosize; 761 if ((off_t)(lbn + 1) * biosize > np->n_size) { 762 bufsize = np->n_size - (off_t)lbn * biosize; 763 bufsize = (bufsize + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 764 } 765 bp = nfs_getcacheblk(vp, lbn, bufsize, p); 766 if (!bp) 767 return (EINTR); 768 if (bp->b_wcred == NOCRED) { 769 crhold(cred); 770 bp->b_wcred = cred; 771 } 772 np->n_flag |= NMODIFIED; 773 774 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 775 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 776 777 /* 778 * If the new write will leave a contiguous dirty 779 * area, just update the b_dirtyoff and b_dirtyend, 780 * otherwise force a write rpc of the old dirty area. 781 */ 782 if (bp->b_dirtyend > 0 && 783 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 784 bp->b_proc = p; 785 if (VOP_BWRITE(bp) == EINTR) 786 return (EINTR); 787 goto again; 788 } 789 790 /* 791 * Check for valid write lease and get one as required. 792 * In case getblk() and/or bwrite() delayed us. 793 */ 794 if ((nmp->nm_flag & NFSMNT_NQNFS) && 795 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 796 do { 797 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 798 } while (error == NQNFS_EXPIRED); 799 if (error) { 800 brelse(bp); 801 return (error); 802 } 803 if (np->n_lrev != np->n_brev || 804 (np->n_flag & NQNFSNONCACHE)) { 805 brelse(bp); 806 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 807 if (error) 808 return (error); 809 np->n_brev = np->n_lrev; 810 goto again; 811 } 812 } 813 814 error = uiomove((char *)bp->b_data + on, n, uio); 815 bp->b_flags &= ~B_NEEDCOMMIT; 816 if (error) { 817 bp->b_flags |= B_ERROR; 818 brelse(bp); 819 return (error); 820 } 821 822 /* 823 * This will keep the buffer and mmaped regions more coherent. 824 */ 825 nfs_prot_buf(bp, on, n); 826 827 if (bp->b_dirtyend > 0) { 828 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 829 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 830 } else { 831 bp->b_dirtyoff = on; 832 bp->b_dirtyend = on + n; 833 } 834 if (bp->b_validend == 0 || bp->b_validend < bp->b_dirtyoff || 835 bp->b_validoff > bp->b_dirtyend) { 836 bp->b_validoff = bp->b_dirtyoff; 837 bp->b_validend = bp->b_dirtyend; 838 } else { 839 bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff); 840 bp->b_validend = max(bp->b_validend, bp->b_dirtyend); 841 } 842 843 /* 844 * Since this block is being modified, it must be written 845 * again and not just committed. 846 */ 847 bp->b_flags &= ~B_NEEDCOMMIT; 848 849 /* 850 * If the lease is non-cachable or IO_SYNC do bwrite(). 851 */ 852 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 853 bp->b_proc = p; 854 if (ioflag & IO_INVAL) 855 bp->b_flags |= B_INVAL; 856 error = VOP_BWRITE(bp); 857 if (error) 858 return (error); 859 if (np->n_flag & NQNFSNONCACHE) { 860 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 861 if (error) 862 return (error); 863 } 864 } else if ((n + on) == biosize && 865 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 866 bp->b_proc = (struct proc *)0; 867 bp->b_flags |= B_ASYNC; 868 (void)nfs_writebp(bp, 0); 869 } else 870 bdwrite(bp); 871 } while (uio->uio_resid > 0 && n > 0); 872 return (0); 873} 874 875/* 876 * Get an nfs cache block. 877 * Allocate a new one if the block isn't currently in the cache 878 * and return the block marked busy. If the calling process is 879 * interrupted by a signal for an interruptible mount point, return 880 * NULL. 881 */ 882static struct buf * 883nfs_getcacheblk(vp, bn, size, p) 884 struct vnode *vp; 885 daddr_t bn; 886 int size; 887 struct proc *p; 888{ 889 register struct buf *bp; 890 struct mount *mp; 891 struct nfsmount *nmp; 892 893 mp = vp->v_mount; 894 nmp = VFSTONFS(mp); 895 896 if (nmp->nm_flag & NFSMNT_INT) { 897 bp = getblk(vp, bn, size, PCATCH, 0); 898 while (bp == (struct buf *)0) { 899 if (nfs_sigintr(nmp, (struct nfsreq *)0, p)) 900 return ((struct buf *)0); 901 bp = getblk(vp, bn, size, 0, 2 * hz); 902 } 903 } else 904 bp = getblk(vp, bn, size, 0, 0); 905 906 if (vp->v_type == VREG) { 907 int biosize; 908 biosize = mp->mnt_stat.f_iosize; 909 bp->b_blkno = bn * (biosize / DEV_BSIZE); 910 } 911 912 return (bp); 913} 914 915/* 916 * Flush and invalidate all dirty buffers. If another process is already 917 * doing the flush, just wait for completion. 918 */ 919int 920nfs_vinvalbuf(vp, flags, cred, p, intrflg) 921 struct vnode *vp; 922 int flags; 923 struct ucred *cred; 924 struct proc *p; 925 int intrflg; 926{ 927 register struct nfsnode *np = VTONFS(vp); 928 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 929 int error = 0, slpflag, slptimeo; 930 931 if (vp->v_flag & VXLOCK) { 932 return (0); 933 } 934 935 if ((nmp->nm_flag & NFSMNT_INT) == 0) 936 intrflg = 0; 937 if (intrflg) { 938 slpflag = PCATCH; 939 slptimeo = 2 * hz; 940 } else { 941 slpflag = 0; 942 slptimeo = 0; 943 } 944 /* 945 * First wait for any other process doing a flush to complete. 946 */ 947 while (np->n_flag & NFLUSHINPROG) { 948 np->n_flag |= NFLUSHWANT; 949 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval", 950 slptimeo); 951 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) 952 return (EINTR); 953 } 954 955 /* 956 * Now, flush as required. 957 */ 958 np->n_flag |= NFLUSHINPROG; 959 error = vinvalbuf(vp, flags, cred, p, slpflag, 0); 960 while (error) { 961 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) { 962 np->n_flag &= ~NFLUSHINPROG; 963 if (np->n_flag & NFLUSHWANT) { 964 np->n_flag &= ~NFLUSHWANT; 965 wakeup((caddr_t)&np->n_flag); 966 } 967 return (EINTR); 968 } 969 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo); 970 } 971 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 972 if (np->n_flag & NFLUSHWANT) { 973 np->n_flag &= ~NFLUSHWANT; 974 wakeup((caddr_t)&np->n_flag); 975 } 976 return (0); 977} 978 979/* 980 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 981 * This is mainly to avoid queueing async I/O requests when the nfsiods 982 * are all hung on a dead server. 983 */ 984int 985nfs_asyncio(bp, cred) 986 register struct buf *bp; 987 struct ucred *cred; 988{ 989 struct nfsmount *nmp; 990 int i; 991 int gotiod; 992 int slpflag = 0; 993 int slptimeo = 0; 994 int error; 995 996 if (nfs_numasync == 0) 997 return (EIO); 998 999 nmp = VFSTONFS(bp->b_vp->v_mount); 1000again: 1001 if (nmp->nm_flag & NFSMNT_INT) 1002 slpflag = PCATCH; 1003 gotiod = FALSE; 1004 1005 /* 1006 * Find a free iod to process this request. 1007 */ 1008 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 1009 if (nfs_iodwant[i]) { 1010 /* 1011 * Found one, so wake it up and tell it which 1012 * mount to process. 1013 */ 1014 NFS_DPF(ASYNCIO, 1015 ("nfs_asyncio: waking iod %d for mount %p\n", 1016 i, nmp)); 1017 nfs_iodwant[i] = (struct proc *)0; 1018 nfs_iodmount[i] = nmp; 1019 nmp->nm_bufqiods++; 1020 wakeup((caddr_t)&nfs_iodwant[i]); 1021 gotiod = TRUE; 1022 break; 1023 } 1024 1025 /* 1026 * If none are free, we may already have an iod working on this mount 1027 * point. If so, it will process our request. 1028 */ 1029 if (!gotiod) { 1030 if (nmp->nm_bufqiods > 0) { 1031 NFS_DPF(ASYNCIO, 1032 ("nfs_asyncio: %d iods are already processing mount %p\n", 1033 nmp->nm_bufqiods, nmp)); 1034 gotiod = TRUE; 1035 } 1036 } 1037 1038 /* 1039 * If we have an iod which can process the request, then queue 1040 * the buffer. 1041 */ 1042 if (gotiod) { 1043 /* 1044 * Ensure that the queue never grows too large. 1045 */ 1046 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1047 NFS_DPF(ASYNCIO, 1048 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1049 nmp->nm_bufqwant = TRUE; 1050 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO, 1051 "nfsaio", slptimeo); 1052 if (error) { 1053 if (nfs_sigintr(nmp, NULL, bp->b_proc)) 1054 return (EINTR); 1055 if (slpflag == PCATCH) { 1056 slpflag = 0; 1057 slptimeo = 2 * hz; 1058 } 1059 } 1060 /* 1061 * We might have lost our iod while sleeping, 1062 * so check and loop if nescessary. 1063 */ 1064 if (nmp->nm_bufqiods == 0) { 1065 NFS_DPF(ASYNCIO, 1066 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1067 goto again; 1068 } 1069 } 1070 1071 if (bp->b_flags & B_READ) { 1072 if (bp->b_rcred == NOCRED && cred != NOCRED) { 1073 crhold(cred); 1074 bp->b_rcred = cred; 1075 } 1076 } else { 1077 bp->b_flags |= B_WRITEINPROG; 1078 if (bp->b_wcred == NOCRED && cred != NOCRED) { 1079 crhold(cred); 1080 bp->b_wcred = cred; 1081 } 1082 } 1083 1084 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1085 nmp->nm_bufqlen++; 1086 return (0); 1087 } 1088 1089 /* 1090 * All the iods are busy on other mounts, so return EIO to 1091 * force the caller to process the i/o synchronously. 1092 */ 1093 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1094 return (EIO); 1095} 1096 1097/* 1098 * Do an I/O operation to/from a cache block. This may be called 1099 * synchronously or from an nfsiod. 1100 */ 1101int 1102nfs_doio(bp, cr, p) 1103 register struct buf *bp; 1104 struct ucred *cr; 1105 struct proc *p; 1106{ 1107 register struct uio *uiop; 1108 register struct vnode *vp; 1109 struct nfsnode *np; 1110 struct nfsmount *nmp; 1111 int error = 0, diff, len, iomode, must_commit = 0; 1112 struct uio uio; 1113 struct iovec io; 1114 1115 vp = bp->b_vp; 1116 np = VTONFS(vp); 1117 nmp = VFSTONFS(vp->v_mount); 1118 uiop = &uio; 1119 uiop->uio_iov = &io; 1120 uiop->uio_iovcnt = 1; 1121 uiop->uio_segflg = UIO_SYSSPACE; 1122 uiop->uio_procp = p; 1123 1124 /* 1125 * Historically, paging was done with physio, but no more. 1126 */ 1127 if (bp->b_flags & B_PHYS) { 1128 /* 1129 * ...though reading /dev/drum still gets us here. 1130 */ 1131 io.iov_len = uiop->uio_resid = bp->b_bcount; 1132 /* mapping was done by vmapbuf() */ 1133 io.iov_base = bp->b_data; 1134 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1135 if (bp->b_flags & B_READ) { 1136 uiop->uio_rw = UIO_READ; 1137 nfsstats.read_physios++; 1138 error = nfs_readrpc(vp, uiop, cr); 1139 } else { 1140 int com; 1141 1142 iomode = NFSV3WRITE_DATASYNC; 1143 uiop->uio_rw = UIO_WRITE; 1144 nfsstats.write_physios++; 1145 error = nfs_writerpc(vp, uiop, cr, &iomode, &com); 1146 } 1147 if (error) { 1148 bp->b_flags |= B_ERROR; 1149 bp->b_error = error; 1150 } 1151 } else if (bp->b_flags & B_READ) { 1152 io.iov_len = uiop->uio_resid = bp->b_bcount; 1153 io.iov_base = bp->b_data; 1154 uiop->uio_rw = UIO_READ; 1155 switch (vp->v_type) { 1156 case VREG: 1157 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1158 nfsstats.read_bios++; 1159 error = nfs_readrpc(vp, uiop, cr); 1160 if (!error) { 1161 bp->b_validoff = 0; 1162 if (uiop->uio_resid) { 1163 /* 1164 * If len > 0, there is a hole in the file and 1165 * no writes after the hole have been pushed to 1166 * the server yet. 1167 * Just zero fill the rest of the valid area. 1168 */ 1169 diff = bp->b_bcount - uiop->uio_resid; 1170 len = np->n_size - (((u_quad_t)bp->b_blkno) * DEV_BSIZE 1171 + diff); 1172 if (len > 0) { 1173 len = min(len, uiop->uio_resid); 1174 bzero((char *)bp->b_data + diff, len); 1175 bp->b_validend = diff + len; 1176 } else 1177 bp->b_validend = diff; 1178 } else 1179 bp->b_validend = bp->b_bcount; 1180 } 1181 if (p && (vp->v_flag & VTEXT) && 1182 (((nmp->nm_flag & NFSMNT_NQNFS) && 1183 NQNFS_CKINVALID(vp, np, ND_READ) && 1184 np->n_lrev != np->n_brev) || 1185 (!(nmp->nm_flag & NFSMNT_NQNFS) && 1186 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) { 1187 uprintf("Process killed due to text file modification\n"); 1188 psignal(p, SIGKILL); 1189 p->p_flag |= P_NOSWAP; 1190 } 1191 break; 1192 case VLNK: 1193 uiop->uio_offset = (off_t)0; 1194 nfsstats.readlink_bios++; 1195 error = nfs_readlinkrpc(vp, uiop, cr); 1196 break; 1197 case VDIR: 1198 nfsstats.readdir_bios++; 1199 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1200 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1201 error = nfs_readdirplusrpc(vp, uiop, cr); 1202 if (error == NFSERR_NOTSUPP) 1203 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1204 } 1205 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1206 error = nfs_readdirrpc(vp, uiop, cr); 1207 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1208 bp->b_flags |= B_INVAL; 1209 break; 1210 default: 1211 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1212 break; 1213 }; 1214 if (error) { 1215 bp->b_flags |= B_ERROR; 1216 bp->b_error = error; 1217 } 1218 } else { 1219 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1220 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1221 1222 if (bp->b_dirtyend > bp->b_dirtyoff) { 1223 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1224 - bp->b_dirtyoff; 1225 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1226 + bp->b_dirtyoff; 1227 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1228 uiop->uio_rw = UIO_WRITE; 1229 nfsstats.write_bios++; 1230 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1231 iomode = NFSV3WRITE_UNSTABLE; 1232 else 1233 iomode = NFSV3WRITE_FILESYNC; 1234 bp->b_flags |= B_WRITEINPROG; 1235 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit); 1236 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1237 bp->b_flags |= B_NEEDCOMMIT; 1238 if (bp->b_dirtyoff == 0 1239 && bp->b_dirtyend == bp->b_bufsize) 1240 bp->b_flags |= B_CLUSTEROK; 1241 } else 1242 bp->b_flags &= ~B_NEEDCOMMIT; 1243 bp->b_flags &= ~B_WRITEINPROG; 1244 1245 /* 1246 * For an interrupted write, the buffer is still valid 1247 * and the write hasn't been pushed to the server yet, 1248 * so we can't set B_ERROR and report the interruption 1249 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1250 * is not relevant, so the rpc attempt is essentially 1251 * a noop. For the case of a V3 write rpc not being 1252 * committed to stable storage, the block is still 1253 * dirty and requires either a commit rpc or another 1254 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1255 * the block is reused. This is indicated by setting 1256 * the B_DELWRI and B_NEEDCOMMIT flags. 1257 */ 1258 if (error == EINTR 1259 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1260 int s; 1261 1262 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1263 ++numdirtybuffers; 1264 bp->b_flags |= B_DELWRI; 1265 s = splbio(); 1266 reassignbuf(bp, vp); 1267 splx(s); 1268 if ((bp->b_flags & B_ASYNC) == 0) 1269 bp->b_flags |= B_EINTR; 1270 } else { 1271 if (error) { 1272 bp->b_flags |= B_ERROR; 1273 bp->b_error = np->n_error = error; 1274 np->n_flag |= NWRITEERR; 1275 } 1276 bp->b_dirtyoff = bp->b_dirtyend = 0; 1277 } 1278 } else { 1279 bp->b_resid = 0; 1280 biodone(bp); 1281 return (0); 1282 } 1283 } 1284 bp->b_resid = uiop->uio_resid; 1285 if (must_commit) 1286 nfs_clearcommit(vp->v_mount); 1287 biodone(bp); 1288 return (error); 1289} 1290