nfs_bio.c revision 33134
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 1.49 1998/02/04 22:33:13 eivind 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)); 68 69extern int nfs_numasync; 70extern struct nfsstats nfsstats; 71 72/* 73 * Vnode op for VM getpages. 74 */ 75int 76nfs_getpages(ap) 77 struct vop_getpages_args *ap; 78{ 79 int i, pcount, error; 80 struct uio uio; 81 struct iovec iov; 82 vm_page_t m; 83 vm_offset_t kva; 84 85 if ((ap->a_vp->v_object) == NULL) { 86 printf("nfs_getpages: called with non-merged cache vnode??\n"); 87 return EOPNOTSUPP; 88 } 89 90 m = ap->a_m[ap->a_reqpage]; 91 kva = vm_pager_map_page(m); 92 93 iov.iov_base = (caddr_t) kva; 94 iov.iov_len = PAGE_SIZE; 95 uio.uio_iov = &iov; 96 uio.uio_iovcnt = 1; 97 uio.uio_offset = IDX_TO_OFF(m->pindex); 98 uio.uio_resid = PAGE_SIZE; 99 uio.uio_segflg = UIO_SYSSPACE; 100 uio.uio_rw = UIO_READ; 101 uio.uio_procp = curproc; 102 103 error = nfs_readrpc(ap->a_vp, &uio, curproc->p_ucred); 104 vm_pager_unmap_page(kva); 105 106 pcount = round_page(ap->a_count) / PAGE_SIZE; 107 for (i = 0; i < pcount; i++) { 108 if (i != ap->a_reqpage) { 109 vnode_pager_freepage(ap->a_m[i]); 110 } 111 } 112 113 if (error && (uio.uio_resid == PAGE_SIZE)) 114 return VM_PAGER_ERROR; 115 return 0; 116} 117 118/* 119 * Vnode op for read using bio 120 * Any similarity to readip() is purely coincidental 121 */ 122int 123nfs_bioread(vp, uio, ioflag, cred, getpages) 124 register struct vnode *vp; 125 register struct uio *uio; 126 int ioflag; 127 struct ucred *cred; 128 int getpages; 129{ 130 register struct nfsnode *np = VTONFS(vp); 131 register int biosize, diff, i; 132 struct buf *bp = 0, *rabp; 133 struct vattr vattr; 134 struct proc *p; 135 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 136 daddr_t lbn, rabn; 137 int bufsize; 138 int nra, error = 0, n = 0, on = 0, not_readin; 139 140#ifdef DIAGNOSTIC 141 if (uio->uio_rw != UIO_READ) 142 panic("nfs_read mode"); 143#endif 144 if (uio->uio_resid == 0) 145 return (0); 146 if (uio->uio_offset < 0) 147 return (EINVAL); 148 p = uio->uio_procp; 149 if ((nmp->nm_flag & (NFSMNT_NFSV3 | NFSMNT_GOTFSINFO)) == NFSMNT_NFSV3) 150 (void)nfs_fsinfo(nmp, vp, cred, p); 151 biosize = vp->v_mount->mnt_stat.f_iosize; 152 /* 153 * For nfs, cache consistency can only be maintained approximately. 154 * Although RFC1094 does not specify the criteria, the following is 155 * believed to be compatible with the reference port. 156 * For nqnfs, full cache consistency is maintained within the loop. 157 * For nfs: 158 * If the file's modify time on the server has changed since the 159 * last read rpc or you have written to the file, 160 * you may have lost data cache consistency with the 161 * server, so flush all of the file's data out of the cache. 162 * Then force a getattr rpc to ensure that you have up to date 163 * attributes. 164 * NB: This implies that cache data can be read when up to 165 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 166 * attributes this could be forced by setting n_attrstamp to 0 before 167 * the VOP_GETATTR() call. 168 */ 169 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) { 170 if (np->n_flag & NMODIFIED) { 171 if (vp->v_type != VREG) { 172 if (vp->v_type != VDIR) 173 panic("nfs: bioread, not dir"); 174 nfs_invaldir(vp); 175 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 176 if (error) 177 return (error); 178 } 179 np->n_attrstamp = 0; 180 error = VOP_GETATTR(vp, &vattr, cred, p); 181 if (error) 182 return (error); 183 np->n_mtime = vattr.va_mtime.tv_sec; 184 } else { 185 error = VOP_GETATTR(vp, &vattr, cred, p); 186 if (error) 187 return (error); 188 if (np->n_mtime != vattr.va_mtime.tv_sec) { 189 if (vp->v_type == VDIR) 190 nfs_invaldir(vp); 191 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 192 if (error) 193 return (error); 194 np->n_mtime = vattr.va_mtime.tv_sec; 195 } 196 } 197 } 198 do { 199 200 /* 201 * Get a valid lease. If cached data is stale, flush it. 202 */ 203 if (nmp->nm_flag & NFSMNT_NQNFS) { 204 if (NQNFS_CKINVALID(vp, np, ND_READ)) { 205 do { 206 error = nqnfs_getlease(vp, ND_READ, cred, p); 207 } while (error == NQNFS_EXPIRED); 208 if (error) 209 return (error); 210 if (np->n_lrev != np->n_brev || 211 (np->n_flag & NQNFSNONCACHE) || 212 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) { 213 if (vp->v_type == VDIR) 214 nfs_invaldir(vp); 215 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 216 if (error) 217 return (error); 218 np->n_brev = np->n_lrev; 219 } 220 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) { 221 nfs_invaldir(vp); 222 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 223 if (error) 224 return (error); 225 } 226 } 227 if (np->n_flag & NQNFSNONCACHE) { 228 switch (vp->v_type) { 229 case VREG: 230 return (nfs_readrpc(vp, uio, cred)); 231 case VLNK: 232 return (nfs_readlinkrpc(vp, uio, cred)); 233 case VDIR: 234 break; 235 default: 236 printf(" NQNFSNONCACHE: type %x unexpected\n", 237 vp->v_type); 238 }; 239 } 240 switch (vp->v_type) { 241 case VREG: 242 nfsstats.biocache_reads++; 243 lbn = uio->uio_offset / biosize; 244 on = uio->uio_offset & (biosize - 1); 245 not_readin = 1; 246 247 /* 248 * Start the read ahead(s), as required. 249 */ 250 if (nfs_numasync > 0 && nmp->nm_readahead > 0) { 251 for (nra = 0; nra < nmp->nm_readahead && 252 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 253 rabn = lbn + 1 + nra; 254 if (!incore(vp, rabn)) { 255 rabp = nfs_getcacheblk(vp, rabn, biosize, p); 256 if (!rabp) 257 return (EINTR); 258 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 259 rabp->b_flags |= (B_READ | B_ASYNC); 260 vfs_busy_pages(rabp, 0); 261 if (nfs_asyncio(rabp, cred)) { 262 rabp->b_flags |= B_INVAL|B_ERROR; 263 vfs_unbusy_pages(rabp); 264 brelse(rabp); 265 } 266 } else 267 brelse(rabp); 268 } 269 } 270 } 271 272 /* 273 * If the block is in the cache and has the required data 274 * in a valid region, just copy it out. 275 * Otherwise, get the block and write back/read in, 276 * as required. 277 */ 278again: 279 bufsize = biosize; 280 if ((off_t)(lbn + 1) * biosize > np->n_size && 281 (off_t)(lbn + 1) * biosize - np->n_size < biosize) { 282 bufsize = np->n_size - lbn * biosize; 283 bufsize = (bufsize + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 284 } 285 bp = nfs_getcacheblk(vp, lbn, bufsize, p); 286 if (!bp) 287 return (EINTR); 288 /* 289 * If we are being called from nfs_getpages, we must 290 * make sure the buffer is a vmio buffer. The vp will 291 * already be setup for vmio but there may be some old 292 * non-vmio buffers attached to it. 293 */ 294 if (getpages && !(bp->b_flags & B_VMIO)) { 295#ifdef DIAGNOSTIC 296 printf("nfs_bioread: non vmio buf found, discarding\n"); 297#endif 298 bp->b_flags |= B_NOCACHE; 299 bp->b_flags |= B_INVAFTERWRITE; 300 if (bp->b_dirtyend > 0) { 301 if ((bp->b_flags & B_DELWRI) == 0) 302 panic("nfsbioread"); 303 if (VOP_BWRITE(bp) == EINTR) 304 return (EINTR); 305 } else 306 brelse(bp); 307 goto again; 308 } 309 if ((bp->b_flags & B_CACHE) == 0) { 310 bp->b_flags |= B_READ; 311 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 312 not_readin = 0; 313 vfs_busy_pages(bp, 0); 314 error = nfs_doio(bp, cred, p); 315 if (error) { 316 brelse(bp); 317 return (error); 318 } 319 } 320 if (bufsize > on) { 321 n = min((unsigned)(bufsize - on), uio->uio_resid); 322 } else { 323 n = 0; 324 } 325 diff = np->n_size - uio->uio_offset; 326 if (diff < n) 327 n = diff; 328 if (not_readin && n > 0) { 329 if (on < bp->b_validoff || (on + n) > bp->b_validend) { 330 bp->b_flags |= B_NOCACHE; 331 bp->b_flags |= B_INVAFTERWRITE; 332 if (bp->b_dirtyend > 0) { 333 if ((bp->b_flags & B_DELWRI) == 0) 334 panic("nfsbioread"); 335 if (VOP_BWRITE(bp) == EINTR) 336 return (EINTR); 337 } else 338 brelse(bp); 339 goto again; 340 } 341 } 342 vp->v_lastr = lbn; 343 diff = (on >= bp->b_validend) ? 0 : (bp->b_validend - on); 344 if (diff < n) 345 n = diff; 346 break; 347 case VLNK: 348 nfsstats.biocache_readlinks++; 349 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p); 350 if (!bp) 351 return (EINTR); 352 if ((bp->b_flags & B_CACHE) == 0) { 353 bp->b_flags |= B_READ; 354 vfs_busy_pages(bp, 0); 355 error = nfs_doio(bp, cred, p); 356 if (error) { 357 bp->b_flags |= B_ERROR; 358 brelse(bp); 359 return (error); 360 } 361 } 362 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 363 on = 0; 364 break; 365 case VDIR: 366 nfsstats.biocache_readdirs++; 367 if (np->n_direofoffset 368 && uio->uio_offset >= np->n_direofoffset) { 369 return (0); 370 } 371 lbn = uio->uio_offset / NFS_DIRBLKSIZ; 372 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 373 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, p); 374 if (!bp) 375 return (EINTR); 376 if ((bp->b_flags & B_CACHE) == 0) { 377 bp->b_flags |= B_READ; 378 vfs_busy_pages(bp, 0); 379 error = nfs_doio(bp, cred, p); 380 if (error) { 381 brelse(bp); 382 } 383 while (error == NFSERR_BAD_COOKIE) { 384 nfs_invaldir(vp); 385 error = nfs_vinvalbuf(vp, 0, cred, p, 1); 386 /* 387 * Yuck! The directory has been modified on the 388 * server. The only way to get the block is by 389 * reading from the beginning to get all the 390 * offset cookies. 391 */ 392 for (i = 0; i <= lbn && !error; i++) { 393 if (np->n_direofoffset 394 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 395 return (0); 396 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, p); 397 if (!bp) 398 return (EINTR); 399 if ((bp->b_flags & B_DONE) == 0) { 400 bp->b_flags |= B_READ; 401 vfs_busy_pages(bp, 0); 402 error = nfs_doio(bp, cred, p); 403 if (error) { 404 brelse(bp); 405 } else if (i < lbn) { 406 brelse(bp); 407 } 408 } 409 } 410 } 411 if (error) 412 return (error); 413 } 414 415 /* 416 * If not eof and read aheads are enabled, start one. 417 * (You need the current block first, so that you have the 418 * directory offset cookie of the next block.) 419 */ 420 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 421 (np->n_direofoffset == 0 || 422 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 423 !(np->n_flag & NQNFSNONCACHE) && 424 !incore(vp, lbn + 1)) { 425 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, p); 426 if (rabp) { 427 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 428 rabp->b_flags |= (B_READ | B_ASYNC); 429 vfs_busy_pages(rabp, 0); 430 if (nfs_asyncio(rabp, cred)) { 431 rabp->b_flags |= B_INVAL|B_ERROR; 432 vfs_unbusy_pages(rabp); 433 brelse(rabp); 434 } 435 } else { 436 brelse(rabp); 437 } 438 } 439 } 440 /* 441 * Make sure we use a signed variant of min() since 442 * the second term may be negative. 443 */ 444 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 445 break; 446 default: 447 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 448 break; 449 }; 450 451 if (n > 0) { 452 error = uiomove(bp->b_data + on, (int)n, uio); 453 } 454 switch (vp->v_type) { 455 case VREG: 456 break; 457 case VLNK: 458 n = 0; 459 break; 460 case VDIR: 461 if (np->n_flag & NQNFSNONCACHE) 462 bp->b_flags |= B_INVAL; 463 break; 464 default: 465 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 466 } 467 brelse(bp); 468 } while (error == 0 && uio->uio_resid > 0 && n > 0); 469 return (error); 470} 471 472/* 473 * Vnode op for write using bio 474 */ 475int 476nfs_write(ap) 477 struct vop_write_args /* { 478 struct vnode *a_vp; 479 struct uio *a_uio; 480 int a_ioflag; 481 struct ucred *a_cred; 482 } */ *ap; 483{ 484 register int biosize; 485 register struct uio *uio = ap->a_uio; 486 struct proc *p = uio->uio_procp; 487 register struct vnode *vp = ap->a_vp; 488 struct nfsnode *np = VTONFS(vp); 489 register struct ucred *cred = ap->a_cred; 490 int ioflag = ap->a_ioflag; 491 struct buf *bp; 492 struct vattr vattr; 493 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 494 daddr_t lbn; 495 int bufsize; 496 int n, on, error = 0, iomode, must_commit; 497 498#ifdef DIAGNOSTIC 499 if (uio->uio_rw != UIO_WRITE) 500 panic("nfs_write mode"); 501 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc) 502 panic("nfs_write proc"); 503#endif 504 if (vp->v_type != VREG) 505 return (EIO); 506 if (np->n_flag & NWRITEERR) { 507 np->n_flag &= ~NWRITEERR; 508 return (np->n_error); 509 } 510 if ((nmp->nm_flag & (NFSMNT_NFSV3 | NFSMNT_GOTFSINFO)) == NFSMNT_NFSV3) 511 (void)nfs_fsinfo(nmp, vp, cred, p); 512 if (ioflag & (IO_APPEND | IO_SYNC)) { 513 if (np->n_flag & NMODIFIED) { 514 np->n_attrstamp = 0; 515 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 516 if (error) 517 return (error); 518 } 519 if (ioflag & IO_APPEND) { 520 np->n_attrstamp = 0; 521 error = VOP_GETATTR(vp, &vattr, cred, p); 522 if (error) 523 return (error); 524 uio->uio_offset = np->n_size; 525 } 526 } 527 if (uio->uio_offset < 0) 528 return (EINVAL); 529 if (uio->uio_resid == 0) 530 return (0); 531 /* 532 * Maybe this should be above the vnode op call, but so long as 533 * file servers have no limits, i don't think it matters 534 */ 535 if (p && uio->uio_offset + uio->uio_resid > 536 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 537 psignal(p, SIGXFSZ); 538 return (EFBIG); 539 } 540 /* 541 * I use nm_rsize, not nm_wsize so that all buffer cache blocks 542 * will be the same size within a filesystem. nfs_writerpc will 543 * still use nm_wsize when sizing the rpc's. 544 */ 545 biosize = vp->v_mount->mnt_stat.f_iosize; 546 do { 547 /* 548 * Check for a valid write lease. 549 */ 550 if ((nmp->nm_flag & NFSMNT_NQNFS) && 551 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 552 do { 553 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 554 } while (error == NQNFS_EXPIRED); 555 if (error) 556 return (error); 557 if (np->n_lrev != np->n_brev || 558 (np->n_flag & NQNFSNONCACHE)) { 559 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 560 if (error) 561 return (error); 562 np->n_brev = np->n_lrev; 563 } 564 } 565 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 566 iomode = NFSV3WRITE_FILESYNC; 567 error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit); 568 if (must_commit) 569 nfs_clearcommit(vp->v_mount); 570 return (error); 571 } 572 nfsstats.biocache_writes++; 573 lbn = uio->uio_offset / biosize; 574 on = uio->uio_offset & (biosize-1); 575 n = min((unsigned)(biosize - on), uio->uio_resid); 576again: 577 if (uio->uio_offset + n > np->n_size) { 578 np->n_size = uio->uio_offset + n; 579 np->n_flag |= NMODIFIED; 580 vnode_pager_setsize(vp, (u_long)np->n_size); 581 } 582 bufsize = biosize; 583 if ((lbn + 1) * biosize > np->n_size) { 584 bufsize = np->n_size - lbn * biosize; 585 bufsize = (bufsize + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 586 } 587 bp = nfs_getcacheblk(vp, lbn, bufsize, p); 588 if (!bp) 589 return (EINTR); 590 if (bp->b_wcred == NOCRED) { 591 crhold(cred); 592 bp->b_wcred = cred; 593 } 594 np->n_flag |= NMODIFIED; 595 596 if ((bp->b_blkno * DEV_BSIZE) + bp->b_dirtyend > np->n_size) { 597 bp->b_dirtyend = np->n_size - (bp->b_blkno * DEV_BSIZE); 598 } 599 600 /* 601 * If the new write will leave a contiguous dirty 602 * area, just update the b_dirtyoff and b_dirtyend, 603 * otherwise force a write rpc of the old dirty area. 604 */ 605 if (bp->b_dirtyend > 0 && 606 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 607 bp->b_proc = p; 608 if (VOP_BWRITE(bp) == EINTR) 609 return (EINTR); 610 goto again; 611 } 612 613 /* 614 * Check for valid write lease and get one as required. 615 * In case getblk() and/or bwrite() delayed us. 616 */ 617 if ((nmp->nm_flag & NFSMNT_NQNFS) && 618 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 619 do { 620 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 621 } while (error == NQNFS_EXPIRED); 622 if (error) { 623 brelse(bp); 624 return (error); 625 } 626 if (np->n_lrev != np->n_brev || 627 (np->n_flag & NQNFSNONCACHE)) { 628 brelse(bp); 629 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 630 if (error) 631 return (error); 632 np->n_brev = np->n_lrev; 633 goto again; 634 } 635 } 636 error = uiomove((char *)bp->b_data + on, n, uio); 637 if (error) { 638 bp->b_flags |= B_ERROR; 639 brelse(bp); 640 return (error); 641 } 642 if (bp->b_dirtyend > 0) { 643 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 644 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 645 } else { 646 bp->b_dirtyoff = on; 647 bp->b_dirtyend = on + n; 648 } 649 if (bp->b_validend == 0 || bp->b_validend < bp->b_dirtyoff || 650 bp->b_validoff > bp->b_dirtyend) { 651 bp->b_validoff = bp->b_dirtyoff; 652 bp->b_validend = bp->b_dirtyend; 653 } else { 654 bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff); 655 bp->b_validend = max(bp->b_validend, bp->b_dirtyend); 656 } 657 658 /* 659 * Since this block is being modified, it must be written 660 * again and not just committed. 661 */ 662 bp->b_flags &= ~B_NEEDCOMMIT; 663 664 /* 665 * If the lease is non-cachable or IO_SYNC do bwrite(). 666 */ 667 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 668 bp->b_proc = p; 669 error = VOP_BWRITE(bp); 670 if (error) 671 return (error); 672 if (np->n_flag & NQNFSNONCACHE) { 673 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 674 if (error) 675 return (error); 676 } 677 } else if ((n + on) == biosize && 678 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 679 bp->b_proc = (struct proc *)0; 680 bp->b_flags |= B_ASYNC; 681 (void)nfs_writebp(bp, 0); 682 } else 683 bdwrite(bp); 684 } while (uio->uio_resid > 0 && n > 0); 685 return (0); 686} 687 688/* 689 * Get an nfs cache block. 690 * Allocate a new one if the block isn't currently in the cache 691 * and return the block marked busy. If the calling process is 692 * interrupted by a signal for an interruptible mount point, return 693 * NULL. 694 */ 695static struct buf * 696nfs_getcacheblk(vp, bn, size, p) 697 struct vnode *vp; 698 daddr_t bn; 699 int size; 700 struct proc *p; 701{ 702 register struct buf *bp; 703 struct mount *mp; 704 struct nfsmount *nmp; 705 706 mp = vp->v_mount; 707 nmp = VFSTONFS(mp); 708 709 if (nmp->nm_flag & NFSMNT_INT) { 710 bp = getblk(vp, bn, size, PCATCH, 0); 711 while (bp == (struct buf *)0) { 712 if (nfs_sigintr(nmp, (struct nfsreq *)0, p)) 713 return ((struct buf *)0); 714 bp = getblk(vp, bn, size, 0, 2 * hz); 715 } 716 } else 717 bp = getblk(vp, bn, size, 0, 0); 718 719 if( vp->v_type == VREG) { 720 int biosize; 721 biosize = mp->mnt_stat.f_iosize; 722 bp->b_blkno = (bn * biosize) / DEV_BSIZE; 723 } 724 725 return (bp); 726} 727 728/* 729 * Flush and invalidate all dirty buffers. If another process is already 730 * doing the flush, just wait for completion. 731 */ 732int 733nfs_vinvalbuf(vp, flags, cred, p, intrflg) 734 struct vnode *vp; 735 int flags; 736 struct ucred *cred; 737 struct proc *p; 738 int intrflg; 739{ 740 register struct nfsnode *np = VTONFS(vp); 741 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 742 int error = 0, slpflag, slptimeo; 743 744 if (vp->v_flag & VXLOCK) { 745 return (0); 746 } 747 748 if ((nmp->nm_flag & NFSMNT_INT) == 0) 749 intrflg = 0; 750 if (intrflg) { 751 slpflag = PCATCH; 752 slptimeo = 2 * hz; 753 } else { 754 slpflag = 0; 755 slptimeo = 0; 756 } 757 /* 758 * First wait for any other process doing a flush to complete. 759 */ 760 while (np->n_flag & NFLUSHINPROG) { 761 np->n_flag |= NFLUSHWANT; 762 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval", 763 slptimeo); 764 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) 765 return (EINTR); 766 } 767 768 /* 769 * Now, flush as required. 770 */ 771 np->n_flag |= NFLUSHINPROG; 772 error = vinvalbuf(vp, flags, cred, p, slpflag, 0); 773 while (error) { 774 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) { 775 np->n_flag &= ~NFLUSHINPROG; 776 if (np->n_flag & NFLUSHWANT) { 777 np->n_flag &= ~NFLUSHWANT; 778 wakeup((caddr_t)&np->n_flag); 779 } 780 return (EINTR); 781 } 782 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo); 783 } 784 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 785 if (np->n_flag & NFLUSHWANT) { 786 np->n_flag &= ~NFLUSHWANT; 787 wakeup((caddr_t)&np->n_flag); 788 } 789 return (0); 790} 791 792/* 793 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 794 * This is mainly to avoid queueing async I/O requests when the nfsiods 795 * are all hung on a dead server. 796 */ 797int 798nfs_asyncio(bp, cred) 799 register struct buf *bp; 800 struct ucred *cred; 801{ 802 struct nfsmount *nmp; 803 int i; 804 int gotiod; 805 int slpflag = 0; 806 int slptimeo = 0; 807 int error; 808 809 if (nfs_numasync == 0) 810 return (EIO); 811 812 nmp = VFSTONFS(bp->b_vp->v_mount); 813again: 814 if (nmp->nm_flag & NFSMNT_INT) 815 slpflag = PCATCH; 816 gotiod = FALSE; 817 818 /* 819 * Find a free iod to process this request. 820 */ 821 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 822 if (nfs_iodwant[i]) { 823 /* 824 * Found one, so wake it up and tell it which 825 * mount to process. 826 */ 827 NFS_DPF(ASYNCIO, 828 ("nfs_asyncio: waking iod %d for mount %p\n", 829 i, nmp)); 830 nfs_iodwant[i] = (struct proc *)0; 831 nfs_iodmount[i] = nmp; 832 nmp->nm_bufqiods++; 833 wakeup((caddr_t)&nfs_iodwant[i]); 834 gotiod = TRUE; 835 break; 836 } 837 838 /* 839 * If none are free, we may already have an iod working on this mount 840 * point. If so, it will process our request. 841 */ 842 if (!gotiod) { 843 if (nmp->nm_bufqiods > 0) { 844 NFS_DPF(ASYNCIO, 845 ("nfs_asyncio: %d iods are already processing mount %p\n", 846 nmp->nm_bufqiods, nmp)); 847 gotiod = TRUE; 848 } 849 } 850 851 /* 852 * If we have an iod which can process the request, then queue 853 * the buffer. 854 */ 855 if (gotiod) { 856 /* 857 * Ensure that the queue never grows too large. 858 */ 859 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 860 NFS_DPF(ASYNCIO, 861 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 862 nmp->nm_bufqwant = TRUE; 863 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO, 864 "nfsaio", slptimeo); 865 if (error) { 866 if (nfs_sigintr(nmp, NULL, bp->b_proc)) 867 return (EINTR); 868 if (slpflag == PCATCH) { 869 slpflag = 0; 870 slptimeo = 2 * hz; 871 } 872 } 873 /* 874 * We might have lost our iod while sleeping, 875 * so check and loop if nescessary. 876 */ 877 if (nmp->nm_bufqiods == 0) { 878 NFS_DPF(ASYNCIO, 879 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 880 goto again; 881 } 882 } 883 884 if (bp->b_flags & B_READ) { 885 if (bp->b_rcred == NOCRED && cred != NOCRED) { 886 crhold(cred); 887 bp->b_rcred = cred; 888 } 889 } else { 890 bp->b_flags |= B_WRITEINPROG; 891 if (bp->b_wcred == NOCRED && cred != NOCRED) { 892 crhold(cred); 893 bp->b_wcred = cred; 894 } 895 } 896 897 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 898 nmp->nm_bufqlen++; 899 return (0); 900 } 901 902 /* 903 * All the iods are busy on other mounts, so return EIO to 904 * force the caller to process the i/o synchronously. 905 */ 906 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 907 return (EIO); 908} 909 910/* 911 * Do an I/O operation to/from a cache block. This may be called 912 * synchronously or from an nfsiod. 913 */ 914int 915nfs_doio(bp, cr, p) 916 register struct buf *bp; 917 struct ucred *cr; 918 struct proc *p; 919{ 920 register struct uio *uiop; 921 register struct vnode *vp; 922 struct nfsnode *np; 923 struct nfsmount *nmp; 924 int error = 0, diff, len, iomode, must_commit = 0; 925 struct uio uio; 926 struct iovec io; 927 928 vp = bp->b_vp; 929 np = VTONFS(vp); 930 nmp = VFSTONFS(vp->v_mount); 931 uiop = &uio; 932 uiop->uio_iov = &io; 933 uiop->uio_iovcnt = 1; 934 uiop->uio_segflg = UIO_SYSSPACE; 935 uiop->uio_procp = p; 936 937 /* 938 * Historically, paging was done with physio, but no more. 939 */ 940 if (bp->b_flags & B_PHYS) { 941 /* 942 * ...though reading /dev/drum still gets us here. 943 */ 944 io.iov_len = uiop->uio_resid = bp->b_bcount; 945 /* mapping was done by vmapbuf() */ 946 io.iov_base = bp->b_data; 947 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 948 if (bp->b_flags & B_READ) { 949 uiop->uio_rw = UIO_READ; 950 nfsstats.read_physios++; 951 error = nfs_readrpc(vp, uiop, cr); 952 } else { 953 int com; 954 955 iomode = NFSV3WRITE_DATASYNC; 956 uiop->uio_rw = UIO_WRITE; 957 nfsstats.write_physios++; 958 error = nfs_writerpc(vp, uiop, cr, &iomode, &com); 959 } 960 if (error) { 961 bp->b_flags |= B_ERROR; 962 bp->b_error = error; 963 } 964 } else if (bp->b_flags & B_READ) { 965 io.iov_len = uiop->uio_resid = bp->b_bcount; 966 io.iov_base = bp->b_data; 967 uiop->uio_rw = UIO_READ; 968 switch (vp->v_type) { 969 case VREG: 970 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 971 nfsstats.read_bios++; 972 error = nfs_readrpc(vp, uiop, cr); 973 if (!error) { 974 bp->b_validoff = 0; 975 if (uiop->uio_resid) { 976 /* 977 * If len > 0, there is a hole in the file and 978 * no writes after the hole have been pushed to 979 * the server yet. 980 * Just zero fill the rest of the valid area. 981 */ 982 diff = bp->b_bcount - uiop->uio_resid; 983 len = np->n_size - (((u_quad_t)bp->b_blkno) * DEV_BSIZE 984 + diff); 985 if (len > 0) { 986 len = min(len, uiop->uio_resid); 987 bzero((char *)bp->b_data + diff, len); 988 bp->b_validend = diff + len; 989 } else 990 bp->b_validend = diff; 991 } else 992 bp->b_validend = bp->b_bcount; 993 } 994 if (p && (vp->v_flag & VTEXT) && 995 (((nmp->nm_flag & NFSMNT_NQNFS) && 996 NQNFS_CKINVALID(vp, np, ND_READ) && 997 np->n_lrev != np->n_brev) || 998 (!(nmp->nm_flag & NFSMNT_NQNFS) && 999 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) { 1000 uprintf("Process killed due to text file modification\n"); 1001 psignal(p, SIGKILL); 1002 p->p_flag |= P_NOSWAP; 1003 } 1004 break; 1005 case VLNK: 1006 uiop->uio_offset = (off_t)0; 1007 nfsstats.readlink_bios++; 1008 error = nfs_readlinkrpc(vp, uiop, cr); 1009 break; 1010 case VDIR: 1011 nfsstats.readdir_bios++; 1012 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1013 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1014 error = nfs_readdirplusrpc(vp, uiop, cr); 1015 if (error == NFSERR_NOTSUPP) 1016 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1017 } 1018 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1019 error = nfs_readdirrpc(vp, uiop, cr); 1020 break; 1021 default: 1022 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1023 break; 1024 }; 1025 if (error) { 1026 bp->b_flags |= B_ERROR; 1027 bp->b_error = error; 1028 } 1029 } else { 1030 if (((bp->b_blkno * DEV_BSIZE) + bp->b_dirtyend) > np->n_size) 1031 bp->b_dirtyend = np->n_size - (bp->b_blkno * DEV_BSIZE); 1032 1033 if (bp->b_dirtyend > bp->b_dirtyoff) { 1034 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1035 - bp->b_dirtyoff; 1036 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE 1037 + bp->b_dirtyoff; 1038 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1039 uiop->uio_rw = UIO_WRITE; 1040 nfsstats.write_bios++; 1041 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1042 iomode = NFSV3WRITE_UNSTABLE; 1043 else 1044 iomode = NFSV3WRITE_FILESYNC; 1045 bp->b_flags |= B_WRITEINPROG; 1046 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit); 1047 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1048 bp->b_flags |= B_NEEDCOMMIT; 1049 if (bp->b_dirtyoff == 0 1050 && bp->b_dirtyend == bp->b_bufsize) 1051 bp->b_flags |= B_CLUSTEROK; 1052 } else 1053 bp->b_flags &= ~B_NEEDCOMMIT; 1054 bp->b_flags &= ~B_WRITEINPROG; 1055 1056 /* 1057 * For an interrupted write, the buffer is still valid 1058 * and the write hasn't been pushed to the server yet, 1059 * so we can't set B_ERROR and report the interruption 1060 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1061 * is not relevant, so the rpc attempt is essentially 1062 * a noop. For the case of a V3 write rpc not being 1063 * committed to stable storage, the block is still 1064 * dirty and requires either a commit rpc or another 1065 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1066 * the block is reused. This is indicated by setting 1067 * the B_DELWRI and B_NEEDCOMMIT flags. 1068 */ 1069 if (error == EINTR 1070 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1071 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1072 ++numdirtybuffers; 1073 bp->b_flags |= B_DELWRI; 1074 reassignbuf(bp, vp); 1075 if ((bp->b_flags & B_ASYNC) == 0) 1076 bp->b_flags |= B_EINTR; 1077 } else { 1078 if (error) { 1079 bp->b_flags |= B_ERROR; 1080 bp->b_error = np->n_error = error; 1081 np->n_flag |= NWRITEERR; 1082 } 1083 bp->b_dirtyoff = bp->b_dirtyend = 0; 1084 } 1085 } else { 1086 bp->b_resid = 0; 1087 biodone(bp); 1088 return (0); 1089 } 1090 } 1091 bp->b_resid = uiop->uio_resid; 1092 if (must_commit) 1093 nfs_clearcommit(vp->v_mount); 1094 biodone(bp); 1095 return (error); 1096} 1097