nfs_bio.c revision 1.151.2.1
1/* $NetBSD: nfs_bio.c,v 1.151.2.1 2007/03/13 17:51:12 ad Exp $ */ 2 3/* 4 * Copyright (c) 1989, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * Rick Macklem at The University of Guelph. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 35 */ 36 37#include <sys/cdefs.h> 38__KERNEL_RCSID(0, "$NetBSD: nfs_bio.c,v 1.151.2.1 2007/03/13 17:51:12 ad Exp $"); 39 40#include "opt_nfs.h" 41#include "opt_ddb.h" 42 43#include <sys/param.h> 44#include <sys/systm.h> 45#include <sys/resourcevar.h> 46#include <sys/signalvar.h> 47#include <sys/proc.h> 48#include <sys/buf.h> 49#include <sys/vnode.h> 50#include <sys/mount.h> 51#include <sys/kernel.h> 52#include <sys/namei.h> 53#include <sys/dirent.h> 54#include <sys/malloc.h> 55#include <sys/kauth.h> 56 57#include <uvm/uvm_extern.h> 58#include <uvm/uvm.h> 59 60#include <nfs/rpcv2.h> 61#include <nfs/nfsproto.h> 62#include <nfs/nfs.h> 63#include <nfs/nfsmount.h> 64#include <nfs/nfsnode.h> 65#include <nfs/nfs_var.h> 66 67extern int nfs_numasync; 68extern int nfs_commitsize; 69extern struct nfsstats nfsstats; 70 71static int nfs_doio_read __P((struct buf *, struct uio *)); 72static int nfs_doio_write __P((struct buf *, struct uio *)); 73static int nfs_doio_phys __P((struct buf *, struct uio *)); 74 75/* 76 * Vnode op for read using bio 77 * Any similarity to readip() is purely coincidental 78 */ 79int 80nfs_bioread(vp, uio, ioflag, cred, cflag) 81 struct vnode *vp; 82 struct uio *uio; 83 int ioflag, cflag; 84 kauth_cred_t cred; 85{ 86 struct nfsnode *np = VTONFS(vp); 87 struct buf *bp = NULL, *rabp; 88 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 89 struct nfsdircache *ndp = NULL, *nndp = NULL; 90 void *baddr; 91 int got_buf = 0, error = 0, n = 0, on = 0, en, enn; 92 int enough = 0; 93 struct dirent *dp, *pdp, *edp, *ep; 94 off_t curoff = 0; 95 int advice; 96 struct lwp *l = curlwp; 97 98#ifdef DIAGNOSTIC 99 if (uio->uio_rw != UIO_READ) 100 panic("nfs_read mode"); 101#endif 102 if (uio->uio_resid == 0) 103 return (0); 104 if (vp->v_type != VDIR && uio->uio_offset < 0) 105 return (EINVAL); 106#ifndef NFS_V2_ONLY 107 if ((nmp->nm_flag & NFSMNT_NFSV3) && 108 !(nmp->nm_iflag & NFSMNT_GOTFSINFO)) 109 (void)nfs_fsinfo(nmp, vp, cred, l); 110#endif 111 if (vp->v_type != VDIR && 112 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 113 return (EFBIG); 114 115 /* 116 * For nfs, cache consistency can only be maintained approximately. 117 * Although RFC1094 does not specify the criteria, the following is 118 * believed to be compatible with the reference port. 119 * 120 * If the file's modify time on the server has changed since the 121 * last read rpc or you have written to the file, 122 * you may have lost data cache consistency with the 123 * server, so flush all of the file's data out of the cache. 124 * Then force a getattr rpc to ensure that you have up to date 125 * attributes. 126 * NB: This implies that cache data can be read when up to 127 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 128 * attributes this could be forced by setting n_attrstamp to 0 before 129 * the VOP_GETATTR() call. 130 */ 131 132 if (vp->v_type != VLNK) { 133 error = nfs_flushstalebuf(vp, cred, l, 134 NFS_FLUSHSTALEBUF_MYWRITE); 135 if (error) 136 return error; 137 } 138 139 do { 140 /* 141 * Don't cache symlinks. 142 */ 143 if ((vp->v_flag & VROOT) && vp->v_type == VLNK) { 144 return (nfs_readlinkrpc(vp, uio, cred)); 145 } 146 baddr = (void *)0; 147 switch (vp->v_type) { 148 case VREG: 149 nfsstats.biocache_reads++; 150 151 advice = IO_ADV_DECODE(ioflag); 152 error = 0; 153 while (uio->uio_resid > 0) { 154 void *win; 155 int flags; 156 vsize_t bytelen; 157 158 nfs_delayedtruncate(vp); 159 if (np->n_size <= uio->uio_offset) { 160 break; 161 } 162 bytelen = 163 MIN(np->n_size - uio->uio_offset, uio->uio_resid); 164 win = ubc_alloc(&vp->v_uobj, uio->uio_offset, 165 &bytelen, advice, UBC_READ); 166 error = uiomove(win, bytelen, uio); 167 flags = UBC_WANT_UNMAP(vp) ? UBC_UNMAP : 0; 168 ubc_release(win, flags); 169 if (error) { 170 /* 171 * XXXkludge 172 * the file has been truncated on the server. 173 * there isn't much we can do. 174 */ 175 if (uio->uio_offset >= np->n_size) { 176 /* end of file */ 177 error = 0; 178 } else { 179 break; 180 } 181 } 182 } 183 break; 184 185 case VLNK: 186 nfsstats.biocache_readlinks++; 187 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, l); 188 if (!bp) 189 return (EINTR); 190 if ((bp->b_flags & B_DONE) == 0) { 191 bp->b_flags |= B_READ; 192 error = nfs_doio(bp); 193 if (error) { 194 brelse(bp); 195 return (error); 196 } 197 } 198 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 199 got_buf = 1; 200 on = 0; 201 break; 202 case VDIR: 203diragain: 204 nfsstats.biocache_readdirs++; 205 ndp = nfs_searchdircache(vp, uio->uio_offset, 206 (nmp->nm_flag & NFSMNT_XLATECOOKIE), 0); 207 if (!ndp) { 208 /* 209 * We've been handed a cookie that is not 210 * in the cache. If we're not translating 211 * 32 <-> 64, it may be a value that was 212 * flushed out of the cache because it grew 213 * too big. Let the server judge if it's 214 * valid or not. In the translation case, 215 * we have no way of validating this value, 216 * so punt. 217 */ 218 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) 219 return (EINVAL); 220 ndp = nfs_enterdircache(vp, uio->uio_offset, 221 uio->uio_offset, 0, 0); 222 } 223 224 if (NFS_EOFVALID(np) && 225 ndp->dc_cookie == np->n_direofoffset) { 226 nfs_putdircache(np, ndp); 227 nfsstats.direofcache_hits++; 228 return (0); 229 } 230 231 bp = nfs_getcacheblk(vp, NFSDC_BLKNO(ndp), NFS_DIRBLKSIZ, l); 232 if (!bp) 233 return (EINTR); 234 if ((bp->b_flags & B_DONE) == 0) { 235 bp->b_flags |= B_READ; 236 bp->b_dcookie = ndp->dc_blkcookie; 237 error = nfs_doio(bp); 238 if (error) { 239 /* 240 * Yuck! The directory has been modified on the 241 * server. Punt and let the userland code 242 * deal with it. 243 */ 244 nfs_putdircache(np, ndp); 245 brelse(bp); 246 /* 247 * nfs_request maps NFSERR_BAD_COOKIE to EINVAL. 248 */ 249 if (error == EINVAL) { /* NFSERR_BAD_COOKIE */ 250 nfs_invaldircache(vp, 0); 251 nfs_vinvalbuf(vp, 0, cred, l, 1); 252 } 253 return (error); 254 } 255 } 256 257 /* 258 * Just return if we hit EOF right away with this 259 * block. Always check here, because direofoffset 260 * may have been set by an nfsiod since the last 261 * check. 262 * 263 * also, empty block implies EOF. 264 */ 265 266 if (bp->b_bcount == bp->b_resid || 267 (NFS_EOFVALID(np) && 268 ndp->dc_blkcookie == np->n_direofoffset)) { 269 KASSERT(bp->b_bcount != bp->b_resid || 270 ndp->dc_blkcookie == bp->b_dcookie); 271 nfs_putdircache(np, ndp); 272 bp->b_flags |= B_NOCACHE; 273 brelse(bp); 274 return 0; 275 } 276 277 /* 278 * Find the entry we were looking for in the block. 279 */ 280 281 en = ndp->dc_entry; 282 283 pdp = dp = (struct dirent *)bp->b_data; 284 edp = (struct dirent *)(void *)((char *)bp->b_data + bp->b_bcount - 285 bp->b_resid); 286 enn = 0; 287 while (enn < en && dp < edp) { 288 pdp = dp; 289 dp = _DIRENT_NEXT(dp); 290 enn++; 291 } 292 293 /* 294 * If the entry number was bigger than the number of 295 * entries in the block, or the cookie of the previous 296 * entry doesn't match, the directory cache is 297 * stale. Flush it and try again (i.e. go to 298 * the server). 299 */ 300 if (dp >= edp || (struct dirent *)_DIRENT_NEXT(dp) > edp || 301 (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) { 302#ifdef DEBUG 303 printf("invalid cache: %p %p %p off %lx %lx\n", 304 pdp, dp, edp, 305 (unsigned long)uio->uio_offset, 306 (unsigned long)NFS_GETCOOKIE(pdp)); 307#endif 308 nfs_putdircache(np, ndp); 309 brelse(bp); 310 nfs_invaldircache(vp, 0); 311 nfs_vinvalbuf(vp, 0, cred, l, 0); 312 goto diragain; 313 } 314 315 on = (char *)dp - (char *)bp->b_data; 316 317 /* 318 * Cache all entries that may be exported to the 319 * user, as they may be thrown back at us. The 320 * NFSBIO_CACHECOOKIES flag indicates that all 321 * entries are being 'exported', so cache them all. 322 */ 323 324 if (en == 0 && pdp == dp) { 325 dp = _DIRENT_NEXT(dp); 326 enn++; 327 } 328 329 if (uio->uio_resid < (bp->b_bcount - bp->b_resid - on)) { 330 n = uio->uio_resid; 331 enough = 1; 332 } else 333 n = bp->b_bcount - bp->b_resid - on; 334 335 ep = (struct dirent *)(void *)((char *)bp->b_data + on + n); 336 337 /* 338 * Find last complete entry to copy, caching entries 339 * (if requested) as we go. 340 */ 341 342 while (dp < ep && (struct dirent *)_DIRENT_NEXT(dp) <= ep) { 343 if (cflag & NFSBIO_CACHECOOKIES) { 344 nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp), 345 ndp->dc_blkcookie, enn, bp->b_lblkno); 346 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) { 347 NFS_STASHCOOKIE32(pdp, 348 nndp->dc_cookie32); 349 } 350 nfs_putdircache(np, nndp); 351 } 352 pdp = dp; 353 dp = _DIRENT_NEXT(dp); 354 enn++; 355 } 356 nfs_putdircache(np, ndp); 357 358 /* 359 * If the last requested entry was not the last in the 360 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ), 361 * cache the cookie of the last requested one, and 362 * set of the offset to it. 363 */ 364 365 if ((on + n) < bp->b_bcount - bp->b_resid) { 366 curoff = NFS_GETCOOKIE(pdp); 367 nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie, 368 enn, bp->b_lblkno); 369 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) { 370 NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32); 371 curoff = nndp->dc_cookie32; 372 } 373 nfs_putdircache(np, nndp); 374 } else 375 curoff = bp->b_dcookie; 376 377 /* 378 * Always cache the entry for the next block, 379 * so that readaheads can use it. 380 */ 381 nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0); 382 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) { 383 if (curoff == bp->b_dcookie) { 384 NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32); 385 curoff = nndp->dc_cookie32; 386 } 387 } 388 389 n = (char *)_DIRENT_NEXT(pdp) - ((char *)bp->b_data + on); 390 391 /* 392 * If not eof and read aheads are enabled, start one. 393 * (You need the current block first, so that you have the 394 * directory offset cookie of the next block.) 395 */ 396 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 397 !NFS_EOFVALID(np)) { 398 rabp = nfs_getcacheblk(vp, NFSDC_BLKNO(nndp), 399 NFS_DIRBLKSIZ, l); 400 if (rabp) { 401 if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) { 402 rabp->b_dcookie = nndp->dc_cookie; 403 rabp->b_flags |= (B_READ | B_ASYNC); 404 if (nfs_asyncio(rabp)) { 405 rabp->b_flags |= B_INVAL; 406 brelse(rabp); 407 } 408 } else 409 brelse(rabp); 410 } 411 } 412 nfs_putdircache(np, nndp); 413 got_buf = 1; 414 break; 415 default: 416 printf(" nfsbioread: type %x unexpected\n",vp->v_type); 417 break; 418 } 419 420 if (n > 0) { 421 if (!baddr) 422 baddr = bp->b_data; 423 error = uiomove((char *)baddr + on, (int)n, uio); 424 } 425 switch (vp->v_type) { 426 case VREG: 427 break; 428 case VLNK: 429 n = 0; 430 break; 431 case VDIR: 432 uio->uio_offset = curoff; 433 if (enough) 434 n = 0; 435 break; 436 default: 437 printf(" nfsbioread: type %x unexpected\n",vp->v_type); 438 } 439 if (got_buf) 440 brelse(bp); 441 } while (error == 0 && uio->uio_resid > 0 && n > 0); 442 return (error); 443} 444 445/* 446 * Vnode op for write using bio 447 */ 448int 449nfs_write(v) 450 void *v; 451{ 452 struct vop_write_args /* { 453 struct vnode *a_vp; 454 struct uio *a_uio; 455 int a_ioflag; 456 kauth_cred_t a_cred; 457 } */ *ap = v; 458 struct uio *uio = ap->a_uio; 459 struct lwp *l = curlwp; 460 struct vnode *vp = ap->a_vp; 461 struct nfsnode *np = VTONFS(vp); 462 kauth_cred_t cred = ap->a_cred; 463 struct vattr vattr; 464 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 465 void *win; 466 voff_t oldoff, origoff; 467 vsize_t bytelen; 468 int flags, error = 0; 469 int ioflag = ap->a_ioflag; 470 int extended = 0, wrotedata = 0; 471 472#ifdef DIAGNOSTIC 473 if (uio->uio_rw != UIO_WRITE) 474 panic("nfs_write mode"); 475#endif 476 if (vp->v_type != VREG) 477 return (EIO); 478 if (np->n_flag & NWRITEERR) { 479 np->n_flag &= ~NWRITEERR; 480 return (np->n_error); 481 } 482#ifndef NFS_V2_ONLY 483 if ((nmp->nm_flag & NFSMNT_NFSV3) && 484 !(nmp->nm_iflag & NFSMNT_GOTFSINFO)) 485 (void)nfs_fsinfo(nmp, vp, cred, l); 486#endif 487 if (ioflag & (IO_APPEND | IO_SYNC)) { 488 if (np->n_flag & NMODIFIED) { 489 NFS_INVALIDATE_ATTRCACHE(np); 490 error = nfs_vinvalbuf(vp, V_SAVE, cred, l, 1); 491 if (error) 492 return (error); 493 } 494 if (ioflag & IO_APPEND) { 495 NFS_INVALIDATE_ATTRCACHE(np); 496 error = VOP_GETATTR(vp, &vattr, cred, l); 497 if (error) 498 return (error); 499 uio->uio_offset = np->n_size; 500 } 501 } 502 if (uio->uio_offset < 0) 503 return (EINVAL); 504 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 505 return (EFBIG); 506 if (uio->uio_resid == 0) 507 return (0); 508 /* 509 * Maybe this should be above the vnode op call, but so long as 510 * file servers have no limits, i don't think it matters 511 */ 512 if (l && l->l_proc && uio->uio_offset + uio->uio_resid > 513 l->l_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 514 psignal(l->l_proc, SIGXFSZ); 515 return (EFBIG); 516 } 517 518 origoff = uio->uio_offset; 519 do { 520 bool extending; /* if we are extending whole pages */ 521 u_quad_t oldsize; 522 oldoff = uio->uio_offset; 523 bytelen = uio->uio_resid; 524 525 nfsstats.biocache_writes++; 526 527 oldsize = np->n_size; 528 np->n_flag |= NMODIFIED; 529 if (np->n_size < uio->uio_offset + bytelen) { 530 np->n_size = uio->uio_offset + bytelen; 531 } 532 extending = ((uio->uio_offset & PAGE_MASK) == 0 && 533 (bytelen & PAGE_MASK) == 0 && 534 uio->uio_offset >= vp->v_size); 535 win = ubc_alloc(&vp->v_uobj, uio->uio_offset, &bytelen, 536 UVM_ADV_NORMAL, 537 UBC_WRITE | (extending ? UBC_FAULTBUSY : 0)); 538 error = uiomove(win, bytelen, uio); 539 flags = UBC_WANT_UNMAP(vp) ? UBC_UNMAP : 0; 540 ubc_release(win, flags); 541 if (error) { 542 if (extending) { 543 /* 544 * backout size and free pages past eof. 545 */ 546 np->n_size = oldsize; 547 mutex_enter(&vp->v_interlock); 548 (void)VOP_PUTPAGES(vp, round_page(vp->v_size), 549 0, PGO_SYNCIO | PGO_FREE); 550 } 551 break; 552 } 553 wrotedata = 1; 554 555 /* 556 * update UVM's notion of the size now that we've 557 * copied the data into the vnode's pages. 558 */ 559 560 if (vp->v_size < uio->uio_offset) { 561 uvm_vnp_setsize(vp, uio->uio_offset); 562 extended = 1; 563 } 564 565 if ((oldoff & ~(nmp->nm_wsize - 1)) != 566 (uio->uio_offset & ~(nmp->nm_wsize - 1))) { 567 mutex_enter(&vp->v_interlock); 568 error = VOP_PUTPAGES(vp, 569 trunc_page(oldoff & ~(nmp->nm_wsize - 1)), 570 round_page((uio->uio_offset + nmp->nm_wsize - 1) & 571 ~(nmp->nm_wsize - 1)), PGO_CLEANIT); 572 } 573 } while (uio->uio_resid > 0); 574 if (wrotedata) 575 VN_KNOTE(vp, NOTE_WRITE | (extended ? NOTE_EXTEND : 0)); 576 if (ioflag & IO_SYNC) { 577 mutex_enter(&vp->v_interlock); 578 error = VOP_PUTPAGES(vp, 579 trunc_page(origoff & ~(nmp->nm_wsize - 1)), 580 round_page((uio->uio_offset + nmp->nm_wsize - 1) & 581 ~(nmp->nm_wsize - 1)), 582 PGO_CLEANIT | PGO_SYNCIO); 583 } 584 return error; 585} 586 587/* 588 * Get an nfs cache block. 589 * Allocate a new one if the block isn't currently in the cache 590 * and return the block marked busy. If the calling process is 591 * interrupted by a signal for an interruptible mount point, return 592 * NULL. 593 */ 594struct buf * 595nfs_getcacheblk(vp, bn, size, l) 596 struct vnode *vp; 597 daddr_t bn; 598 int size; 599 struct lwp *l; 600{ 601 struct buf *bp; 602 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 603 604 if (nmp->nm_flag & NFSMNT_INT) { 605 bp = getblk(vp, bn, size, PCATCH, 0); 606 while (bp == NULL) { 607 if (nfs_sigintr(nmp, NULL, l)) 608 return (NULL); 609 bp = getblk(vp, bn, size, 0, 2 * hz); 610 } 611 } else 612 bp = getblk(vp, bn, size, 0, 0); 613 return (bp); 614} 615 616/* 617 * Flush and invalidate all dirty buffers. If another process is already 618 * doing the flush, just wait for completion. 619 */ 620int 621nfs_vinvalbuf(vp, flags, cred, l, intrflg) 622 struct vnode *vp; 623 int flags; 624 kauth_cred_t cred; 625 struct lwp *l; 626 int intrflg; 627{ 628 struct nfsnode *np = VTONFS(vp); 629 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 630 int error = 0, slpflag, slptimeo; 631 632 if ((nmp->nm_flag & NFSMNT_INT) == 0) 633 intrflg = 0; 634 if (intrflg) { 635 slpflag = PCATCH; 636 slptimeo = 2 * hz; 637 } else { 638 slpflag = 0; 639 slptimeo = 0; 640 } 641 /* 642 * First wait for any other process doing a flush to complete. 643 */ 644 mutex_enter(&vp->v_interlock); 645 while (np->n_flag & NFLUSHINPROG) { 646 np->n_flag |= NFLUSHWANT; 647 error = mtsleep(&np->n_flag, PRIBIO + 2, "nfsvinval", 648 slptimeo, &vp->v_interlock); 649 if (error && intrflg && nfs_sigintr(nmp, NULL, l)) { 650 mutex_exit(&vp->v_interlock); 651 return EINTR; 652 } 653 } 654 655 /* 656 * Now, flush as required. 657 */ 658 np->n_flag |= NFLUSHINPROG; 659 mutex_exit(&vp->v_interlock); 660 error = vinvalbuf(vp, flags, cred, l, slpflag, 0); 661 while (error) { 662 if (intrflg && nfs_sigintr(nmp, NULL, l)) { 663 error = EINTR; 664 break; 665 } 666 error = vinvalbuf(vp, flags, cred, l, 0, slptimeo); 667 } 668 mutex_enter(&vp->v_interlock); 669 if (error == 0) 670 np->n_flag &= ~NMODIFIED; 671 np->n_flag &= ~NFLUSHINPROG; 672 if (np->n_flag & NFLUSHWANT) { 673 np->n_flag &= ~NFLUSHWANT; 674 wakeup(&np->n_flag); 675 } 676 mutex_exit(&vp->v_interlock); 677 return error; 678} 679 680/* 681 * nfs_flushstalebuf: flush cache if it's stale. 682 * 683 * => caller shouldn't own any pages or buffers which belong to the vnode. 684 */ 685 686int 687nfs_flushstalebuf(struct vnode *vp, kauth_cred_t cred, struct lwp *l, 688 int flags) 689{ 690 struct nfsnode *np = VTONFS(vp); 691 struct vattr vattr; 692 int error; 693 694 if (np->n_flag & NMODIFIED) { 695 if ((flags & NFS_FLUSHSTALEBUF_MYWRITE) == 0 696 || vp->v_type != VREG) { 697 error = nfs_vinvalbuf(vp, V_SAVE, cred, l, 1); 698 if (error) 699 return error; 700 if (vp->v_type == VDIR) { 701 nfs_invaldircache(vp, 0); 702 } 703 } else { 704 /* 705 * XXX assuming writes are ours. 706 */ 707 } 708 NFS_INVALIDATE_ATTRCACHE(np); 709 error = VOP_GETATTR(vp, &vattr, cred, l); 710 if (error) 711 return error; 712 np->n_mtime = vattr.va_mtime; 713 } else { 714 error = VOP_GETATTR(vp, &vattr, cred, l); 715 if (error) 716 return error; 717 if (timespeccmp(&np->n_mtime, &vattr.va_mtime, !=)) { 718 if (vp->v_type == VDIR) { 719 nfs_invaldircache(vp, 0); 720 } 721 error = nfs_vinvalbuf(vp, V_SAVE, cred, l, 1); 722 if (error) 723 return error; 724 np->n_mtime = vattr.va_mtime; 725 } 726 } 727 728 return error; 729} 730 731/* 732 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 733 * This is mainly to avoid queueing async I/O requests when the nfsiods 734 * are all hung on a dead server. 735 */ 736 737int 738nfs_asyncio(bp) 739 struct buf *bp; 740{ 741 int i; 742 struct nfsmount *nmp; 743 int gotiod, slpflag = 0, slptimeo = 0, error; 744 745 if (nfs_numasync == 0) 746 return (EIO); 747 748 nmp = VFSTONFS(bp->b_vp->v_mount); 749again: 750 if (nmp->nm_flag & NFSMNT_INT) 751 slpflag = PCATCH; 752 gotiod = false; 753 754 /* 755 * Find a free iod to process this request. 756 */ 757 758 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) { 759 struct nfs_iod *iod = &nfs_asyncdaemon[i]; 760 761 mutex_enter(&iod->nid_lock); 762 if (iod->nid_want) { 763 /* 764 * Found one, so wake it up and tell it which 765 * mount to process. 766 */ 767 iod->nid_want = NULL; 768 iod->nid_mount = nmp; 769 wakeup(&iod->nid_want); 770 mutex_enter(&nmp->nm_lock); 771 mutex_exit(&iod->nid_lock); 772 nmp->nm_bufqiods++; 773 gotiod = true; 774 break; 775 } 776 mutex_exit(&iod->nid_lock); 777 } 778 779 /* 780 * If none are free, we may already have an iod working on this mount 781 * point. If so, it will process our request. 782 */ 783 784 if (!gotiod) { 785 mutex_enter(&nmp->nm_lock); 786 if (nmp->nm_bufqiods > 0) 787 gotiod = true; 788 } 789 790 KASSERT(mutex_owned(&nmp->nm_lock)); 791 792 /* 793 * If we have an iod which can process the request, then queue 794 * the buffer. However, even if we have an iod, do not initiate 795 * queue cleaning if curproc is the pageout daemon. if the NFS mount 796 * is via local loopback, we may put curproc (pagedaemon) to sleep 797 * waiting for the writes to complete. But the server (ourself) 798 * may block the write, waiting for its (ie., our) pagedaemon 799 * to produce clean pages to handle the write: deadlock. 800 * XXX: start non-loopback mounts straight away? If "lots free", 801 * let pagedaemon start loopback writes anyway? 802 */ 803 if (gotiod) { 804 805 /* 806 * Ensure that the queue never grows too large. 807 */ 808 if (curproc == uvm.pagedaemon_proc) { 809 /* Enque for later, to avoid free-page deadlock */ 810 (void) 0; 811 } else while (nmp->nm_bufqlen >= 2*nfs_numasync) { 812 nmp->nm_bufqwant = true; 813 error = mtsleep(&nmp->nm_bufq, 814 slpflag | PRIBIO | PNORELOCK, 815 "nfsaio", slptimeo, &nmp->nm_lock); 816 if (error) { 817 if (nfs_sigintr(nmp, NULL, curlwp)) 818 return (EINTR); 819 if (slpflag == PCATCH) { 820 slpflag = 0; 821 slptimeo = 2 * hz; 822 } 823 } 824 825 /* 826 * We might have lost our iod while sleeping, 827 * so check and loop if nescessary. 828 */ 829 830 if (nmp->nm_bufqiods == 0) 831 goto again; 832 833 mutex_enter(&nmp->nm_lock); 834 } 835 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 836 nmp->nm_bufqlen++; 837 mutex_exit(&nmp->nm_lock); 838 return (0); 839 } 840 mutex_exit(&nmp->nm_lock); 841 842 /* 843 * All the iods are busy on other mounts, so return EIO to 844 * force the caller to process the i/o synchronously. 845 */ 846 847 return (EIO); 848} 849 850/* 851 * nfs_doio for read. 852 */ 853static int 854nfs_doio_read(bp, uiop) 855 struct buf *bp; 856 struct uio *uiop; 857{ 858 struct vnode *vp = bp->b_vp; 859 struct nfsnode *np = VTONFS(vp); 860 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 861 int error = 0; 862 863 uiop->uio_rw = UIO_READ; 864 switch (vp->v_type) { 865 case VREG: 866 nfsstats.read_bios++; 867 error = nfs_readrpc(vp, uiop); 868 if (!error && uiop->uio_resid) { 869 int diff, len; 870 871 /* 872 * If uio_resid > 0, there is a hole in the file and 873 * no writes after the hole have been pushed to 874 * the server yet or the file has been truncated 875 * on the server. 876 * Just zero fill the rest of the valid area. 877 */ 878 879 KASSERT(vp->v_size >= 880 uiop->uio_offset + uiop->uio_resid); 881 diff = bp->b_bcount - uiop->uio_resid; 882 len = uiop->uio_resid; 883 memset((char *)bp->b_data + diff, 0, len); 884 uiop->uio_resid = 0; 885 } 886#if 0 887 if (uiop->uio_lwp && (vp->v_flag & VTEXT) && 888 timespeccmp(&np->n_mtime, &np->n_vattr->va_mtime, !=)) { 889 killproc(uiop->uio_lwp->l_proc, "process text file was modified"); 890#if 0 /* XXX NJWLWP */ 891 uiop->uio_lwp->l_proc->p_holdcnt++; 892#endif 893 } 894#endif 895 break; 896 case VLNK: 897 KASSERT(uiop->uio_offset == (off_t)0); 898 nfsstats.readlink_bios++; 899 error = nfs_readlinkrpc(vp, uiop, np->n_rcred); 900 break; 901 case VDIR: 902 nfsstats.readdir_bios++; 903 uiop->uio_offset = bp->b_dcookie; 904#ifndef NFS_V2_ONLY 905 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 906 error = nfs_readdirplusrpc(vp, uiop, 907 curlwp->l_cred); 908 /* 909 * nfs_request maps NFSERR_NOTSUPP to ENOTSUP. 910 */ 911 if (error == ENOTSUP) 912 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 913 } 914#else 915 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 916#endif 917 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 918 error = nfs_readdirrpc(vp, uiop, 919 curlwp->l_cred); 920 if (!error) { 921 bp->b_dcookie = uiop->uio_offset; 922 } 923 break; 924 default: 925 printf("nfs_doio: type %x unexpected\n", vp->v_type); 926 break; 927 } 928 if (error) { 929 bp->b_flags |= B_ERROR; 930 bp->b_error = error; 931 } 932 return error; 933} 934 935/* 936 * nfs_doio for write. 937 */ 938static int 939nfs_doio_write(bp, uiop) 940 struct buf *bp; 941 struct uio *uiop; 942{ 943 struct vnode *vp = bp->b_vp; 944 struct nfsnode *np = VTONFS(vp); 945 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 946 int iomode; 947 bool stalewriteverf = false; 948 int i, npages = (bp->b_bcount + PAGE_SIZE - 1) >> PAGE_SHIFT; 949 struct vm_page *pgs[npages]; 950#ifndef NFS_V2_ONLY 951 bool needcommit = true; /* need only COMMIT RPC */ 952#else 953 bool needcommit = false; /* need only COMMIT RPC */ 954#endif 955 bool pageprotected; 956 struct uvm_object *uobj = &vp->v_uobj; 957 int error; 958 off_t off, cnt; 959 960 if ((bp->b_flags & B_ASYNC) != 0 && NFS_ISV3(vp)) { 961 iomode = NFSV3WRITE_UNSTABLE; 962 } else { 963 iomode = NFSV3WRITE_FILESYNC; 964 } 965 966#ifndef NFS_V2_ONLY 967again: 968#endif 969 rw_enter(&nmp->nm_writeverflock, RW_READER); 970 971 for (i = 0; i < npages; i++) { 972 pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT)); 973 if (pgs[i]->uobject == uobj && 974 pgs[i]->offset == uiop->uio_offset + (i << PAGE_SHIFT)) { 975 KASSERT(pgs[i]->flags & PG_BUSY); 976 /* 977 * this page belongs to our object. 978 */ 979 mutex_enter(&uobj->vmobjlock); 980 /* 981 * write out the page stably if it's about to 982 * be released because we can't resend it 983 * on the server crash. 984 * 985 * XXX assuming PG_RELEASE|PG_PAGEOUT won't be 986 * changed until unbusy the page. 987 */ 988 if (pgs[i]->flags & (PG_RELEASED|PG_PAGEOUT)) 989 iomode = NFSV3WRITE_FILESYNC; 990 /* 991 * if we met a page which hasn't been sent yet, 992 * we need do WRITE RPC. 993 */ 994 if ((pgs[i]->flags & PG_NEEDCOMMIT) == 0) 995 needcommit = false; 996 mutex_exit(&uobj->vmobjlock); 997 } else { 998 iomode = NFSV3WRITE_FILESYNC; 999 needcommit = false; 1000 } 1001 } 1002 if (!needcommit && iomode == NFSV3WRITE_UNSTABLE) { 1003 mutex_enter(&uobj->vmobjlock); 1004 for (i = 0; i < npages; i++) { 1005 pgs[i]->flags |= PG_NEEDCOMMIT | PG_RDONLY; 1006 pmap_page_protect(pgs[i], VM_PROT_READ); 1007 } 1008 mutex_exit(&uobj->vmobjlock); 1009 pageprotected = true; /* pages can't be modified during i/o. */ 1010 } else 1011 pageprotected = false; 1012 1013 /* 1014 * Send the data to the server if necessary, 1015 * otherwise just send a commit rpc. 1016 */ 1017#ifndef NFS_V2_ONLY 1018 if (needcommit) { 1019 1020 /* 1021 * If the buffer is in the range that we already committed, 1022 * there's nothing to do. 1023 * 1024 * If it's in the range that we need to commit, push the 1025 * whole range at once, otherwise only push the buffer. 1026 * In both these cases, acquire the commit lock to avoid 1027 * other processes modifying the range. 1028 */ 1029 1030 off = uiop->uio_offset; 1031 cnt = bp->b_bcount; 1032 mutex_enter(&np->n_commitlock); 1033 if (!nfs_in_committed_range(vp, off, bp->b_bcount)) { 1034 bool pushedrange; 1035 if (nfs_in_tobecommitted_range(vp, off, bp->b_bcount)) { 1036 pushedrange = true; 1037 off = np->n_pushlo; 1038 cnt = np->n_pushhi - np->n_pushlo; 1039 } else { 1040 pushedrange = false; 1041 } 1042 error = nfs_commit(vp, off, cnt, curlwp); 1043 if (error == 0) { 1044 if (pushedrange) { 1045 nfs_merge_commit_ranges(vp); 1046 } else { 1047 nfs_add_committed_range(vp, off, cnt); 1048 } 1049 } 1050 } else { 1051 error = 0; 1052 } 1053 mutex_exit(&np->n_commitlock); 1054 rw_exit(&nmp->nm_writeverflock); 1055 if (!error) { 1056 /* 1057 * pages are now on stable storage. 1058 */ 1059 uiop->uio_resid = 0; 1060 mutex_enter(&uobj->vmobjlock); 1061 for (i = 0; i < npages; i++) { 1062 pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY); 1063 } 1064 mutex_exit(&uobj->vmobjlock); 1065 return 0; 1066 } else if (error == NFSERR_STALEWRITEVERF) { 1067 nfs_clearcommit(vp->v_mount); 1068 goto again; 1069 } 1070 if (error) { 1071 bp->b_flags |= B_ERROR; 1072 bp->b_error = np->n_error = error; 1073 np->n_flag |= NWRITEERR; 1074 } 1075 return error; 1076 } 1077#endif 1078 off = uiop->uio_offset; 1079 cnt = bp->b_bcount; 1080 uiop->uio_rw = UIO_WRITE; 1081 nfsstats.write_bios++; 1082 error = nfs_writerpc(vp, uiop, &iomode, pageprotected, &stalewriteverf); 1083#ifndef NFS_V2_ONLY 1084 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1085 /* 1086 * we need to commit pages later. 1087 */ 1088 mutex_enter(&np->n_commitlock); 1089 nfs_add_tobecommitted_range(vp, off, cnt); 1090 /* 1091 * if there can be too many uncommitted pages, commit them now. 1092 */ 1093 if (np->n_pushhi - np->n_pushlo > nfs_commitsize) { 1094 off = np->n_pushlo; 1095 cnt = nfs_commitsize >> 1; 1096 error = nfs_commit(vp, off, cnt, curlwp); 1097 if (!error) { 1098 nfs_add_committed_range(vp, off, cnt); 1099 nfs_del_tobecommitted_range(vp, off, cnt); 1100 } 1101 if (error == NFSERR_STALEWRITEVERF) { 1102 stalewriteverf = true; 1103 error = 0; /* it isn't a real error */ 1104 } 1105 } else { 1106 /* 1107 * re-dirty pages so that they will be passed 1108 * to us later again. 1109 */ 1110 mutex_enter(&uobj->vmobjlock); 1111 for (i = 0; i < npages; i++) { 1112 pgs[i]->flags &= ~PG_CLEAN; 1113 } 1114 mutex_exit(&uobj->vmobjlock); 1115 } 1116 mutex_exit(&np->n_commitlock); 1117 } else 1118#endif 1119 if (!error) { 1120 /* 1121 * pages are now on stable storage. 1122 */ 1123 mutex_enter(&np->n_commitlock); 1124 nfs_del_committed_range(vp, off, cnt); 1125 mutex_exit(&np->n_commitlock); 1126 mutex_enter(&uobj->vmobjlock); 1127 for (i = 0; i < npages; i++) { 1128 pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY); 1129 } 1130 mutex_exit(&uobj->vmobjlock); 1131 } else { 1132 /* 1133 * we got an error. 1134 */ 1135 bp->b_flags |= B_ERROR; 1136 bp->b_error = np->n_error = error; 1137 np->n_flag |= NWRITEERR; 1138 } 1139 1140 rw_exit(&nmp->nm_writeverflock); 1141 1142 if (stalewriteverf) { 1143 nfs_clearcommit(vp->v_mount); 1144 } 1145 return error; 1146} 1147 1148/* 1149 * nfs_doio for B_PHYS. 1150 */ 1151static int 1152nfs_doio_phys(bp, uiop) 1153 struct buf *bp; 1154 struct uio *uiop; 1155{ 1156 struct vnode *vp = bp->b_vp; 1157 int error; 1158 1159 uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT; 1160 if (bp->b_flags & B_READ) { 1161 uiop->uio_rw = UIO_READ; 1162 nfsstats.read_physios++; 1163 error = nfs_readrpc(vp, uiop); 1164 } else { 1165 int iomode = NFSV3WRITE_DATASYNC; 1166 bool stalewriteverf; 1167 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1168 1169 uiop->uio_rw = UIO_WRITE; 1170 nfsstats.write_physios++; 1171 rw_enter(&nmp->nm_writeverflock, RW_READER); 1172 error = nfs_writerpc(vp, uiop, &iomode, false, &stalewriteverf); 1173 rw_exit(&nmp->nm_writeverflock); 1174 if (stalewriteverf) { 1175 nfs_clearcommit(bp->b_vp->v_mount); 1176 } 1177 } 1178 if (error) { 1179 bp->b_flags |= B_ERROR; 1180 bp->b_error = error; 1181 } 1182 return error; 1183} 1184 1185/* 1186 * Do an I/O operation to/from a cache block. This may be called 1187 * synchronously or from an nfsiod. 1188 */ 1189int 1190nfs_doio(bp) 1191 struct buf *bp; 1192{ 1193 int error; 1194 struct uio uio; 1195 struct uio *uiop = &uio; 1196 struct iovec io; 1197 UVMHIST_FUNC("nfs_doio"); UVMHIST_CALLED(ubchist); 1198 1199 uiop->uio_iov = &io; 1200 uiop->uio_iovcnt = 1; 1201 uiop->uio_offset = (((off_t)bp->b_blkno) << DEV_BSHIFT); 1202 UIO_SETUP_SYSSPACE(uiop); 1203 io.iov_base = bp->b_data; 1204 io.iov_len = uiop->uio_resid = bp->b_bcount; 1205 1206 /* 1207 * Historically, paging was done with physio, but no more... 1208 */ 1209 if (bp->b_flags & B_PHYS) { 1210 /* 1211 * ...though reading /dev/drum still gets us here. 1212 */ 1213 error = nfs_doio_phys(bp, uiop); 1214 } else if (bp->b_flags & B_READ) { 1215 error = nfs_doio_read(bp, uiop); 1216 } else { 1217 error = nfs_doio_write(bp, uiop); 1218 } 1219 bp->b_resid = uiop->uio_resid; 1220 biodone(bp); 1221 return (error); 1222} 1223 1224/* 1225 * Vnode op for VM getpages. 1226 */ 1227 1228int 1229nfs_getpages(v) 1230 void *v; 1231{ 1232 struct vop_getpages_args /* { 1233 struct vnode *a_vp; 1234 voff_t a_offset; 1235 struct vm_page **a_m; 1236 int *a_count; 1237 int a_centeridx; 1238 vm_prot_t a_access_type; 1239 int a_advice; 1240 int a_flags; 1241 } */ *ap = v; 1242 1243 struct vnode *vp = ap->a_vp; 1244 struct uvm_object *uobj = &vp->v_uobj; 1245 struct nfsnode *np = VTONFS(vp); 1246 const int npages = *ap->a_count; 1247 struct vm_page *pg, **pgs, *opgs[npages]; 1248 off_t origoffset, len; 1249 int i, error; 1250 bool v3 = NFS_ISV3(vp); 1251 bool write = (ap->a_access_type & VM_PROT_WRITE) != 0; 1252 bool locked = (ap->a_flags & PGO_LOCKED) != 0; 1253 1254 /* 1255 * call the genfs code to get the pages. `pgs' may be NULL 1256 * when doing read-ahead. 1257 */ 1258 1259 pgs = ap->a_m; 1260 if (write && locked && v3) { 1261 KASSERT(pgs != NULL); 1262#ifdef DEBUG 1263 1264 /* 1265 * If PGO_LOCKED is set, real pages shouldn't exists 1266 * in the array. 1267 */ 1268 1269 for (i = 0; i < npages; i++) 1270 KDASSERT(pgs[i] == NULL || pgs[i] == PGO_DONTCARE); 1271#endif 1272 memcpy(opgs, pgs, npages * sizeof(struct vm_pages *)); 1273 } 1274 error = genfs_getpages(v); 1275 if (error) { 1276 return (error); 1277 } 1278 1279 /* 1280 * for read faults where the nfs node is not yet marked NMODIFIED, 1281 * set PG_RDONLY on the pages so that we come back here if someone 1282 * tries to modify later via the mapping that will be entered for 1283 * this fault. 1284 */ 1285 1286 if (!write && (np->n_flag & NMODIFIED) == 0 && pgs != NULL) { 1287 if (!locked) { 1288 mutex_enter(&uobj->vmobjlock); 1289 } 1290 for (i = 0; i < npages; i++) { 1291 pg = pgs[i]; 1292 if (pg == NULL || pg == PGO_DONTCARE) { 1293 continue; 1294 } 1295 pg->flags |= PG_RDONLY; 1296 } 1297 if (!locked) { 1298 mutex_exit(&uobj->vmobjlock); 1299 } 1300 } 1301 if (!write) { 1302 return (0); 1303 } 1304 1305 /* 1306 * this is a write fault, update the commit info. 1307 */ 1308 1309 origoffset = ap->a_offset; 1310 len = npages << PAGE_SHIFT; 1311 1312 if (v3) { 1313 if (!locked) { 1314 mutex_enter(&np->n_commitlock); 1315 } else { 1316 if (!mutex_tryenter(&np->n_commitlock)) { 1317 1318 /* 1319 * Since PGO_LOCKED is set, we need to unbusy 1320 * all pages fetched by genfs_getpages() above, 1321 * tell the caller that there are no pages 1322 * available and put back original pgs array. 1323 */ 1324 1325 mutex_enter(&uvm_pageqlock); 1326 uvm_page_unbusy(pgs, npages); 1327 mutex_exit(&uvm_pageqlock); 1328 *ap->a_count = 0; 1329 memcpy(pgs, opgs, 1330 npages * sizeof(struct vm_pages *)); 1331 return EBUSY; 1332 } 1333 } 1334 nfs_del_committed_range(vp, origoffset, len); 1335 nfs_del_tobecommitted_range(vp, origoffset, len); 1336 } 1337 np->n_flag |= NMODIFIED; 1338 if (!locked) { 1339 mutex_enter(&uobj->vmobjlock); 1340 } 1341 for (i = 0; i < npages; i++) { 1342 pg = pgs[i]; 1343 if (pg == NULL || pg == PGO_DONTCARE) { 1344 continue; 1345 } 1346 pg->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY); 1347 } 1348 if (!locked) { 1349 mutex_exit(&uobj->vmobjlock); 1350 } 1351 if (v3) { 1352 mutex_exit(&np->n_commitlock); 1353 } 1354 return (0); 1355} 1356