1/* $NetBSD$ */ 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, "$NetBSD$"); 39 40#ifdef _KERNEL_OPT 41#include "opt_nfs.h" 42#include "opt_ddb.h" 43#endif 44 45#include <sys/param.h> 46#include <sys/systm.h> 47#include <sys/resourcevar.h> 48#include <sys/signalvar.h> 49#include <sys/proc.h> 50#include <sys/buf.h> 51#include <sys/vnode.h> 52#include <sys/mount.h> 53#include <sys/kernel.h> 54#include <sys/namei.h> 55#include <sys/dirent.h> 56#include <sys/kauth.h> 57 58#include <uvm/uvm_extern.h> 59#include <uvm/uvm.h> 60 61#include <nfs/rpcv2.h> 62#include <nfs/nfsproto.h> 63#include <nfs/nfs.h> 64#include <nfs/nfsmount.h> 65#include <nfs/nfsnode.h> 66#include <nfs/nfs_var.h> 67 68extern int nfs_numasync; 69extern int nfs_commitsize; 70extern struct nfsstats nfsstats; 71 72static int nfs_doio_read(struct buf *, struct uio *); 73static int nfs_doio_write(struct buf *, struct uio *); 74static int nfs_doio_phys(struct buf *, struct uio *); 75 76/* 77 * Vnode op for read using bio 78 * Any similarity to readip() is purely coincidental 79 */ 80int 81nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, 82 kauth_cred_t cred, int cflag) 83{ 84 struct nfsnode *np = VTONFS(vp); 85 struct buf *bp = NULL, *rabp; 86 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 87 struct nfsdircache *ndp = NULL, *nndp = NULL; 88 void *baddr; 89 int got_buf = 0, error = 0, n = 0, on = 0, en, enn; 90 int enough = 0; 91 struct dirent *dp, *pdp, *edp, *ep; 92 off_t curoff = 0; 93 int advice; 94 struct lwp *l = curlwp; 95 96#ifdef DIAGNOSTIC 97 if (uio->uio_rw != UIO_READ) 98 panic("nfs_read mode"); 99#endif 100 if (uio->uio_resid == 0) 101 return (0); 102 if (vp->v_type != VDIR && uio->uio_offset < 0) 103 return (EINVAL); 104#ifndef NFS_V2_ONLY 105 if ((nmp->nm_flag & NFSMNT_NFSV3) && 106 !(nmp->nm_iflag & NFSMNT_GOTFSINFO)) 107 (void)nfs_fsinfo(nmp, vp, cred, l); 108#endif 109 if (vp->v_type != VDIR && 110 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 111 return (EFBIG); 112 113 /* 114 * For nfs, cache consistency can only be maintained approximately. 115 * Although RFC1094 does not specify the criteria, the following is 116 * believed to be compatible with the reference port. 117 * 118 * If the file's modify time on the server has changed since the 119 * last read rpc or you have written to the file, 120 * you may have lost data cache consistency with the 121 * server, so flush all of the file's data out of the cache. 122 * Then force a getattr rpc to ensure that you have up to date 123 * attributes. 124 * NB: This implies that cache data can be read when up to 125 * nfs_attrtimeo seconds out of date. If you find that you need current 126 * attributes this could be forced by setting n_attrstamp to 0 before 127 * the VOP_GETATTR() call. 128 */ 129 130 if (vp->v_type != VLNK) { 131 error = nfs_flushstalebuf(vp, cred, l, 132 NFS_FLUSHSTALEBUF_MYWRITE); 133 if (error) 134 return error; 135 } 136 137 do { 138 /* 139 * Don't cache symlinks. 140 */ 141 if ((vp->v_vflag & VV_ROOT) && vp->v_type == VLNK) { 142 return (nfs_readlinkrpc(vp, uio, cred)); 143 } 144 baddr = (void *)0; 145 switch (vp->v_type) { 146 case VREG: 147 nfsstats.biocache_reads++; 148 149 advice = IO_ADV_DECODE(ioflag); 150 error = 0; 151 while (uio->uio_resid > 0) { 152 vsize_t bytelen; 153 154 nfs_delayedtruncate(vp); 155 if (np->n_size <= uio->uio_offset) { 156 break; 157 } 158 bytelen = 159 MIN(np->n_size - uio->uio_offset, uio->uio_resid); 160 error = ubc_uiomove(&vp->v_uobj, uio, bytelen, advice, 161 UBC_READ | UBC_PARTIALOK | UBC_UNMAP_FLAG(vp)); 162 if (error) { 163 /* 164 * XXXkludge 165 * the file has been truncated on the server. 166 * there isn't much we can do. 167 */ 168 if (uio->uio_offset >= np->n_size) { 169 /* end of file */ 170 error = 0; 171 } else { 172 break; 173 } 174 } 175 } 176 break; 177 178 case VLNK: 179 nfsstats.biocache_readlinks++; 180 bp = nfs_getcacheblk(vp, (daddr_t)0, MAXPATHLEN, l); 181 if (!bp) 182 return (EINTR); 183 if ((bp->b_oflags & BO_DONE) == 0) { 184 bp->b_flags |= B_READ; 185 error = nfs_doio(bp); 186 if (error) { 187 brelse(bp, 0); 188 return (error); 189 } 190 } 191 n = MIN(uio->uio_resid, MAXPATHLEN - bp->b_resid); 192 got_buf = 1; 193 on = 0; 194 break; 195 case VDIR: 196diragain: 197 nfsstats.biocache_readdirs++; 198 ndp = nfs_searchdircache(vp, uio->uio_offset, 199 (nmp->nm_flag & NFSMNT_XLATECOOKIE), 0); 200 if (!ndp) { 201 /* 202 * We've been handed a cookie that is not 203 * in the cache. If we're not translating 204 * 32 <-> 64, it may be a value that was 205 * flushed out of the cache because it grew 206 * too big. Let the server judge if it's 207 * valid or not. In the translation case, 208 * we have no way of validating this value, 209 * so punt. 210 */ 211 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) 212 return (EINVAL); 213 ndp = nfs_enterdircache(vp, uio->uio_offset, 214 uio->uio_offset, 0, 0); 215 } 216 217 if (NFS_EOFVALID(np) && 218 ndp->dc_cookie == np->n_direofoffset) { 219 nfs_putdircache(np, ndp); 220 nfsstats.direofcache_hits++; 221 return (0); 222 } 223 224 bp = nfs_getcacheblk(vp, NFSDC_BLKNO(ndp), NFS_DIRBLKSIZ, l); 225 if (!bp) 226 return (EINTR); 227 if ((bp->b_oflags & BO_DONE) == 0) { 228 bp->b_flags |= B_READ; 229 bp->b_dcookie = ndp->dc_blkcookie; 230 error = nfs_doio(bp); 231 if (error) { 232 /* 233 * Yuck! The directory has been modified on the 234 * server. Punt and let the userland code 235 * deal with it. 236 */ 237 nfs_putdircache(np, ndp); 238 brelse(bp, 0); 239 /* 240 * nfs_request maps NFSERR_BAD_COOKIE to EINVAL. 241 */ 242 if (error == EINVAL) { /* NFSERR_BAD_COOKIE */ 243 nfs_invaldircache(vp, 0); 244 nfs_vinvalbuf(vp, 0, cred, l, 1); 245 } 246 return (error); 247 } 248 } 249 250 /* 251 * Just return if we hit EOF right away with this 252 * block. Always check here, because direofoffset 253 * may have been set by an nfsiod since the last 254 * check. 255 * 256 * also, empty block implies EOF. 257 */ 258 259 if (bp->b_bcount == bp->b_resid || 260 (NFS_EOFVALID(np) && 261 ndp->dc_blkcookie == np->n_direofoffset)) { 262 KASSERT(bp->b_bcount != bp->b_resid || 263 ndp->dc_blkcookie == bp->b_dcookie); 264 nfs_putdircache(np, ndp); 265 brelse(bp, BC_NOCACHE); 266 return 0; 267 } 268 269 /* 270 * Find the entry we were looking for in the block. 271 */ 272 273 en = ndp->dc_entry; 274 275 pdp = dp = (struct dirent *)bp->b_data; 276 edp = (struct dirent *)(void *)((char *)bp->b_data + bp->b_bcount - 277 bp->b_resid); 278 enn = 0; 279 while (enn < en && dp < edp) { 280 pdp = dp; 281 dp = _DIRENT_NEXT(dp); 282 enn++; 283 } 284 285 /* 286 * If the entry number was bigger than the number of 287 * entries in the block, or the cookie of the previous 288 * entry doesn't match, the directory cache is 289 * stale. Flush it and try again (i.e. go to 290 * the server). 291 */ 292 if (dp >= edp || (struct dirent *)_DIRENT_NEXT(dp) > edp || 293 (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) { 294#ifdef DEBUG 295 printf("invalid cache: %p %p %p off %jx %jx\n", 296 pdp, dp, edp, 297 (uintmax_t)uio->uio_offset, 298 (uintmax_t)NFS_GETCOOKIE(pdp)); 299#endif 300 nfs_putdircache(np, ndp); 301 brelse(bp, 0); 302 nfs_invaldircache(vp, 0); 303 nfs_vinvalbuf(vp, 0, cred, l, 0); 304 goto diragain; 305 } 306 307 on = (char *)dp - (char *)bp->b_data; 308 309 /* 310 * Cache all entries that may be exported to the 311 * user, as they may be thrown back at us. The 312 * NFSBIO_CACHECOOKIES flag indicates that all 313 * entries are being 'exported', so cache them all. 314 */ 315 316 if (en == 0 && pdp == dp) { 317 dp = _DIRENT_NEXT(dp); 318 enn++; 319 } 320 321 if (uio->uio_resid < (bp->b_bcount - bp->b_resid - on)) { 322 n = uio->uio_resid; 323 enough = 1; 324 } else 325 n = bp->b_bcount - bp->b_resid - on; 326 327 ep = (struct dirent *)(void *)((char *)bp->b_data + on + n); 328 329 /* 330 * Find last complete entry to copy, caching entries 331 * (if requested) as we go. 332 */ 333 334 while (dp < ep && (struct dirent *)_DIRENT_NEXT(dp) <= ep) { 335 if (cflag & NFSBIO_CACHECOOKIES) { 336 nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp), 337 ndp->dc_blkcookie, enn, bp->b_lblkno); 338 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) { 339 NFS_STASHCOOKIE32(pdp, 340 nndp->dc_cookie32); 341 } 342 nfs_putdircache(np, nndp); 343 } 344 pdp = dp; 345 dp = _DIRENT_NEXT(dp); 346 enn++; 347 } 348 nfs_putdircache(np, ndp); 349 350 /* 351 * If the last requested entry was not the last in the 352 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ), 353 * cache the cookie of the last requested one, and 354 * set of the offset to it. 355 */ 356 357 if ((on + n) < bp->b_bcount - bp->b_resid) { 358 curoff = NFS_GETCOOKIE(pdp); 359 nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie, 360 enn, bp->b_lblkno); 361 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) { 362 NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32); 363 curoff = nndp->dc_cookie32; 364 } 365 nfs_putdircache(np, nndp); 366 } else 367 curoff = bp->b_dcookie; 368 369 /* 370 * Always cache the entry for the next block, 371 * so that readaheads can use it. 372 */ 373 nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0); 374 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) { 375 if (curoff == bp->b_dcookie) { 376 NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32); 377 curoff = nndp->dc_cookie32; 378 } 379 } 380 381 n = (char *)_DIRENT_NEXT(pdp) - ((char *)bp->b_data + on); 382 383 /* 384 * If not eof and read aheads are enabled, start one. 385 * (You need the current block first, so that you have the 386 * directory offset cookie of the next block.) 387 */ 388 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 389 !NFS_EOFVALID(np)) { 390 rabp = nfs_getcacheblk(vp, NFSDC_BLKNO(nndp), 391 NFS_DIRBLKSIZ, l); 392 if (rabp) { 393 if ((rabp->b_oflags & (BO_DONE | BO_DELWRI)) == 0) { 394 rabp->b_dcookie = nndp->dc_cookie; 395 rabp->b_flags |= (B_READ | B_ASYNC); 396 if (nfs_asyncio(rabp)) { 397 brelse(rabp, BC_INVAL); 398 } 399 } else 400 brelse(rabp, 0); 401 } 402 } 403 nfs_putdircache(np, nndp); 404 got_buf = 1; 405 break; 406 default: 407 printf(" nfsbioread: type %x unexpected\n",vp->v_type); 408 break; 409 } 410 411 if (n > 0) { 412 if (!baddr) 413 baddr = bp->b_data; 414 error = uiomove((char *)baddr + on, (int)n, uio); 415 } 416 switch (vp->v_type) { 417 case VREG: 418 break; 419 case VLNK: 420 n = 0; 421 break; 422 case VDIR: 423 uio->uio_offset = curoff; 424 if (enough) 425 n = 0; 426 break; 427 default: 428 printf(" nfsbioread: type %x unexpected\n",vp->v_type); 429 } 430 if (got_buf) 431 brelse(bp, 0); 432 } while (error == 0 && uio->uio_resid > 0 && n > 0); 433 return (error); 434} 435 436/* 437 * Vnode op for write using bio 438 */ 439int 440nfs_write(void *v) 441{ 442 struct vop_write_args /* { 443 struct vnode *a_vp; 444 struct uio *a_uio; 445 int a_ioflag; 446 kauth_cred_t a_cred; 447 } */ *ap = v; 448 struct uio *uio = ap->a_uio; 449 struct lwp *l = curlwp; 450 struct vnode *vp = ap->a_vp; 451 struct nfsnode *np = VTONFS(vp); 452 kauth_cred_t cred = ap->a_cred; 453 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 454 voff_t oldoff, origoff; 455 vsize_t bytelen; 456 int error = 0; 457 int ioflag = ap->a_ioflag; 458 int extended = 0, wrotedata = 0; 459 460#ifdef DIAGNOSTIC 461 if (uio->uio_rw != UIO_WRITE) 462 panic("nfs_write mode"); 463#endif 464 if (vp->v_type != VREG) 465 return (EIO); 466 if (np->n_flag & NWRITEERR) { 467 np->n_flag &= ~NWRITEERR; 468 return (np->n_error); 469 } 470#ifndef NFS_V2_ONLY 471 if ((nmp->nm_flag & NFSMNT_NFSV3) && 472 !(nmp->nm_iflag & NFSMNT_GOTFSINFO)) 473 (void)nfs_fsinfo(nmp, vp, cred, l); 474#endif 475 if (ioflag & IO_APPEND) { 476 NFS_INVALIDATE_ATTRCACHE(np); 477 error = nfs_flushstalebuf(vp, cred, l, 478 NFS_FLUSHSTALEBUF_MYWRITE); 479 if (error) 480 return (error); 481 uio->uio_offset = np->n_size; 482 483 /* 484 * This is already checked above VOP_WRITE, but recheck 485 * the append case here to make sure our idea of the 486 * file size is as fresh as possible. 487 */ 488 if (uio->uio_offset + uio->uio_resid > 489 l->l_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 490 mutex_enter(proc_lock); 491 psignal(l->l_proc, SIGXFSZ); 492 mutex_exit(proc_lock); 493 return (EFBIG); 494 } 495 } 496 if (uio->uio_offset < 0) 497 return (EINVAL); 498 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 499 return (EFBIG); 500 if (uio->uio_resid == 0) 501 return (0); 502 503 origoff = uio->uio_offset; 504 do { 505 bool overwrite; /* if we are overwriting whole pages */ 506 u_quad_t oldsize; 507 oldoff = uio->uio_offset; 508 bytelen = uio->uio_resid; 509 510 nfsstats.biocache_writes++; 511 512 oldsize = np->n_size; 513 np->n_flag |= NMODIFIED; 514 if (np->n_size < uio->uio_offset + bytelen) { 515 np->n_size = uio->uio_offset + bytelen; 516 } 517 overwrite = false; 518 if ((uio->uio_offset & PAGE_MASK) == 0) { 519 if ((vp->v_vflag & VV_MAPPED) == 0 && 520 bytelen > PAGE_SIZE) { 521 bytelen = trunc_page(bytelen); 522 overwrite = true; 523 } else if ((bytelen & PAGE_MASK) == 0 && 524 uio->uio_offset >= vp->v_size) { 525 overwrite = true; 526 } 527 } 528 if (vp->v_size < uio->uio_offset + bytelen) { 529 uvm_vnp_setwritesize(vp, uio->uio_offset + bytelen); 530 } 531 error = ubc_uiomove(&vp->v_uobj, uio, bytelen, 532 UVM_ADV_RANDOM, UBC_WRITE | UBC_PARTIALOK | 533 (overwrite ? UBC_FAULTBUSY : 0) | 534 UBC_UNMAP_FLAG(vp)); 535 if (error) { 536 uvm_vnp_setwritesize(vp, vp->v_size); 537 if (overwrite && np->n_size != oldsize) { 538 /* 539 * backout size and free pages past eof. 540 */ 541 np->n_size = oldsize; 542 mutex_enter(vp->v_interlock); 543 (void)VOP_PUTPAGES(vp, round_page(vp->v_size), 544 0, PGO_SYNCIO | PGO_FREE); 545 } 546 break; 547 } 548 wrotedata = 1; 549 550 /* 551 * update UVM's notion of the size now that we've 552 * copied the data into the vnode's pages. 553 */ 554 555 if (vp->v_size < uio->uio_offset) { 556 uvm_vnp_setsize(vp, uio->uio_offset); 557 extended = 1; 558 } 559 560 if ((oldoff & ~(nmp->nm_wsize - 1)) != 561 (uio->uio_offset & ~(nmp->nm_wsize - 1))) { 562 mutex_enter(vp->v_interlock); 563 error = VOP_PUTPAGES(vp, 564 trunc_page(oldoff & ~(nmp->nm_wsize - 1)), 565 round_page((uio->uio_offset + nmp->nm_wsize - 1) & 566 ~(nmp->nm_wsize - 1)), PGO_CLEANIT); 567 } 568 } while (uio->uio_resid > 0); 569 if (wrotedata) 570 VN_KNOTE(vp, NOTE_WRITE | (extended ? NOTE_EXTEND : 0)); 571 if (error == 0 && (ioflag & IO_SYNC) != 0) { 572 mutex_enter(vp->v_interlock); 573 error = VOP_PUTPAGES(vp, 574 trunc_page(origoff & ~(nmp->nm_wsize - 1)), 575 round_page((uio->uio_offset + nmp->nm_wsize - 1) & 576 ~(nmp->nm_wsize - 1)), 577 PGO_CLEANIT | PGO_SYNCIO); 578 } 579 return error; 580} 581 582/* 583 * Get an nfs cache block. 584 * Allocate a new one if the block isn't currently in the cache 585 * and return the block marked busy. If the calling process is 586 * interrupted by a signal for an interruptible mount point, return 587 * NULL. 588 */ 589struct buf * 590nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct lwp *l) 591{ 592 struct buf *bp; 593 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 594 595 if (nmp->nm_flag & NFSMNT_INT) { 596 bp = getblk(vp, bn, size, PCATCH, 0); 597 while (bp == NULL) { 598 if (nfs_sigintr(nmp, NULL, l)) 599 return (NULL); 600 bp = getblk(vp, bn, size, 0, 2 * hz); 601 } 602 } else 603 bp = getblk(vp, bn, size, 0, 0); 604 return (bp); 605} 606 607/* 608 * Flush and invalidate all dirty buffers. If another process is already 609 * doing the flush, just wait for completion. 610 */ 611int 612nfs_vinvalbuf(struct vnode *vp, int flags, kauth_cred_t cred, 613 struct lwp *l, int intrflg) 614{ 615 struct nfsnode *np = VTONFS(vp); 616 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 617 int error = 0, allerror = 0, slptimeo; 618 bool catch; 619 620 if ((nmp->nm_flag & NFSMNT_INT) == 0) 621 intrflg = 0; 622 if (intrflg) { 623 catch = true; 624 slptimeo = 2 * hz; 625 } else { 626 catch = false; 627 slptimeo = 0; 628 } 629 /* 630 * First wait for any other process doing a flush to complete. 631 */ 632 mutex_enter(vp->v_interlock); 633 while (np->n_flag & NFLUSHINPROG) { 634 np->n_flag |= NFLUSHWANT; 635 error = mtsleep(&np->n_flag, PRIBIO + 2, "nfsvinval", 636 slptimeo, vp->v_interlock); 637 if (error && intrflg && nfs_sigintr(nmp, NULL, l)) { 638 mutex_exit(vp->v_interlock); 639 return EINTR; 640 } 641 } 642 643 /* 644 * Now, flush as required. 645 */ 646 np->n_flag |= NFLUSHINPROG; 647 mutex_exit(vp->v_interlock); 648 error = vinvalbuf(vp, flags, cred, l, catch, 0); 649 while (error) { 650 if (allerror == 0) 651 allerror = error; 652 if (intrflg && nfs_sigintr(nmp, NULL, l)) { 653 error = EINTR; 654 break; 655 } 656 error = vinvalbuf(vp, flags, cred, l, 0, slptimeo); 657 } 658 mutex_enter(vp->v_interlock); 659 if (allerror != 0) { 660 /* 661 * Keep error from vinvalbuf so fsync/close will know. 662 */ 663 np->n_error = allerror; 664 np->n_flag |= NWRITEERR; 665 } 666 if (error == 0) 667 np->n_flag &= ~NMODIFIED; 668 np->n_flag &= ~NFLUSHINPROG; 669 if (np->n_flag & NFLUSHWANT) { 670 np->n_flag &= ~NFLUSHWANT; 671 wakeup(&np->n_flag); 672 } 673 mutex_exit(vp->v_interlock); 674 return error; 675} 676 677/* 678 * nfs_flushstalebuf: flush cache if it's stale. 679 * 680 * => caller shouldn't own any pages or buffers which belong to the vnode. 681 */ 682 683int 684nfs_flushstalebuf(struct vnode *vp, kauth_cred_t cred, struct lwp *l, 685 int flags) 686{ 687 struct nfsnode *np = VTONFS(vp); 688 struct vattr vattr; 689 int error; 690 691 if (np->n_flag & NMODIFIED) { 692 if ((flags & NFS_FLUSHSTALEBUF_MYWRITE) == 0 693 || vp->v_type != VREG) { 694 error = nfs_vinvalbuf(vp, V_SAVE, cred, l, 1); 695 if (error) 696 return error; 697 if (vp->v_type == VDIR) { 698 nfs_invaldircache(vp, 0); 699 } 700 } else { 701 /* 702 * XXX assuming writes are ours. 703 */ 704 } 705 NFS_INVALIDATE_ATTRCACHE(np); 706 error = VOP_GETATTR(vp, &vattr, cred); 707 if (error) 708 return error; 709 np->n_mtime = vattr.va_mtime; 710 } else { 711 error = VOP_GETATTR(vp, &vattr, cred); 712 if (error) 713 return error; 714 if (timespeccmp(&np->n_mtime, &vattr.va_mtime, !=)) { 715 if (vp->v_type == VDIR) { 716 nfs_invaldircache(vp, 0); 717 } 718 error = nfs_vinvalbuf(vp, V_SAVE, cred, l, 1); 719 if (error) 720 return error; 721 np->n_mtime = vattr.va_mtime; 722 } 723 } 724 725 return error; 726} 727 728/* 729 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 730 * This is mainly to avoid queueing async I/O requests when the nfsiods 731 * are all hung on a dead server. 732 */ 733 734int 735nfs_asyncio(struct buf *bp) 736{ 737 struct nfs_iod *iod; 738 struct nfsmount *nmp; 739 int slptimeo = 0, error; 740 bool catch = false; 741 742 if (nfs_numasync == 0) 743 return (EIO); 744 745 nmp = VFSTONFS(bp->b_vp->v_mount); 746again: 747 if (nmp->nm_flag & NFSMNT_INT) 748 catch = true; 749 750 /* 751 * Find a free iod to process this request. 752 */ 753 754 mutex_enter(&nfs_iodlist_lock); 755 iod = LIST_FIRST(&nfs_iodlist_idle); 756 if (iod) { 757 /* 758 * Found one, so wake it up and tell it which 759 * mount to process. 760 */ 761 LIST_REMOVE(iod, nid_idle); 762 mutex_enter(&iod->nid_lock); 763 mutex_exit(&nfs_iodlist_lock); 764 KASSERT(iod->nid_mount == NULL); 765 iod->nid_mount = nmp; 766 cv_signal(&iod->nid_cv); 767 mutex_enter(&nmp->nm_lock); 768 mutex_exit(&iod->nid_lock); 769 nmp->nm_bufqiods++; 770 if (nmp->nm_bufqlen < 2 * nmp->nm_bufqiods) { 771 cv_broadcast(&nmp->nm_aiocv); 772 } 773 } else { 774 mutex_exit(&nfs_iodlist_lock); 775 mutex_enter(&nmp->nm_lock); 776 } 777 778 KASSERT(mutex_owned(&nmp->nm_lock)); 779 780 /* 781 * If we have an iod which can process the request, then queue 782 * the buffer. However, even if we have an iod, do not initiate 783 * queue cleaning if curproc is the pageout daemon. if the NFS mount 784 * is via local loopback, we may put curproc (pagedaemon) to sleep 785 * waiting for the writes to complete. But the server (ourself) 786 * may block the write, waiting for its (ie., our) pagedaemon 787 * to produce clean pages to handle the write: deadlock. 788 * XXX: start non-loopback mounts straight away? If "lots free", 789 * let pagedaemon start loopback writes anyway? 790 */ 791 if (nmp->nm_bufqiods > 0) { 792 793 /* 794 * Ensure that the queue never grows too large. 795 */ 796 if (curlwp == uvm.pagedaemon_lwp) { 797 /* Enque for later, to avoid free-page deadlock */ 798 } else while (nmp->nm_bufqlen >= 2 * nmp->nm_bufqiods) { 799 if (catch) { 800 error = cv_timedwait_sig(&nmp->nm_aiocv, 801 &nmp->nm_lock, slptimeo); 802 } else { 803 error = cv_timedwait(&nmp->nm_aiocv, 804 &nmp->nm_lock, slptimeo); 805 } 806 if (error) { 807 if (nfs_sigintr(nmp, NULL, curlwp)) { 808 mutex_exit(&nmp->nm_lock); 809 return (EINTR); 810 } 811 if (catch) { 812 catch = false; 813 slptimeo = 2 * hz; 814 } 815 } 816 817 /* 818 * We might have lost our iod while sleeping, 819 * so check and loop if necessary. 820 */ 821 822 if (nmp->nm_bufqiods == 0) { 823 mutex_exit(&nmp->nm_lock); 824 goto again; 825 } 826 } 827 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 828 nmp->nm_bufqlen++; 829 mutex_exit(&nmp->nm_lock); 830 return (0); 831 } 832 mutex_exit(&nmp->nm_lock); 833 834 /* 835 * All the iods are busy on other mounts, so return EIO to 836 * force the caller to process the i/o synchronously. 837 */ 838 839 return (EIO); 840} 841 842/* 843 * nfs_doio for read. 844 */ 845static int 846nfs_doio_read(struct buf *bp, struct uio *uiop) 847{ 848 struct vnode *vp = bp->b_vp; 849 struct nfsnode *np = VTONFS(vp); 850 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 851 int error = 0; 852 853 uiop->uio_rw = UIO_READ; 854 switch (vp->v_type) { 855 case VREG: 856 nfsstats.read_bios++; 857 error = nfs_readrpc(vp, uiop); 858 if (!error && uiop->uio_resid) { 859 int diff, len; 860 861 /* 862 * If uio_resid > 0, there is a hole in the file and 863 * no writes after the hole have been pushed to 864 * the server yet or the file has been truncated 865 * on the server. 866 * Just zero fill the rest of the valid area. 867 */ 868 869 KASSERT(vp->v_size >= 870 uiop->uio_offset + uiop->uio_resid); 871 diff = bp->b_bcount - uiop->uio_resid; 872 len = uiop->uio_resid; 873 memset((char *)bp->b_data + diff, 0, len); 874 uiop->uio_resid = 0; 875 } 876#if 0 877 if (uiop->uio_lwp && (vp->v_iflag & VI_TEXT) && 878 timespeccmp(&np->n_mtime, &np->n_vattr->va_mtime, !=)) { 879 mutex_enter(proc_lock); 880 killproc(uiop->uio_lwp->l_proc, "process text file was modified"); 881 mutex_exit(proc_lock); 882#if 0 /* XXX NJWLWP */ 883 uiop->uio_lwp->l_proc->p_holdcnt++; 884#endif 885 } 886#endif 887 break; 888 case VLNK: 889 KASSERT(uiop->uio_offset == (off_t)0); 890 nfsstats.readlink_bios++; 891 error = nfs_readlinkrpc(vp, uiop, np->n_rcred); 892 break; 893 case VDIR: 894 nfsstats.readdir_bios++; 895 uiop->uio_offset = bp->b_dcookie; 896#ifndef NFS_V2_ONLY 897 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 898 error = nfs_readdirplusrpc(vp, uiop, 899 curlwp->l_cred); 900 /* 901 * nfs_request maps NFSERR_NOTSUPP to ENOTSUP. 902 */ 903 if (error == ENOTSUP) 904 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 905 } 906#else 907 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 908#endif 909 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 910 error = nfs_readdirrpc(vp, uiop, 911 curlwp->l_cred); 912 if (!error) { 913 bp->b_dcookie = uiop->uio_offset; 914 } 915 break; 916 default: 917 printf("nfs_doio: type %x unexpected\n", vp->v_type); 918 break; 919 } 920 bp->b_error = error; 921 return error; 922} 923 924/* 925 * nfs_doio for write. 926 */ 927static int 928nfs_doio_write(struct buf *bp, struct uio *uiop) 929{ 930 struct vnode *vp = bp->b_vp; 931 struct nfsnode *np = VTONFS(vp); 932 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 933 int iomode; 934 bool stalewriteverf = false; 935 int i, npages = (bp->b_bcount + PAGE_SIZE - 1) >> PAGE_SHIFT; 936 struct vm_page **pgs, *spgs[UBC_MAX_PAGES]; 937#ifndef NFS_V2_ONLY 938 bool needcommit = true; /* need only COMMIT RPC */ 939#else 940 bool needcommit = false; /* need only COMMIT RPC */ 941#endif 942 bool pageprotected; 943 struct uvm_object *uobj = &vp->v_uobj; 944 int error; 945 off_t off, cnt; 946 947 if (npages < __arraycount(spgs)) 948 pgs = spgs; 949 else { 950 if ((pgs = kmem_alloc(sizeof(*pgs) * npages, KM_NOSLEEP)) == 951 NULL) 952 return ENOMEM; 953 } 954 955 if ((bp->b_flags & B_ASYNC) != 0 && NFS_ISV3(vp)) { 956 iomode = NFSV3WRITE_UNSTABLE; 957 } else { 958 iomode = NFSV3WRITE_FILESYNC; 959 } 960 961#ifndef NFS_V2_ONLY 962again: 963#endif 964 rw_enter(&nmp->nm_writeverflock, RW_READER); 965 966 for (i = 0; i < npages; i++) { 967 pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT)); 968 if (pgs[i]->uobject == uobj && 969 pgs[i]->offset == uiop->uio_offset + (i << PAGE_SHIFT)) { 970 KASSERT(pgs[i]->flags & PG_BUSY); 971 /* 972 * this page belongs to our object. 973 */ 974 mutex_enter(uobj->vmobjlock); 975 /* 976 * write out the page stably if it's about to 977 * be released because we can't resend it 978 * on the server crash. 979 * 980 * XXX assuming PG_RELEASE|PG_PAGEOUT won't be 981 * changed until unbusy the page. 982 */ 983 if (pgs[i]->flags & (PG_RELEASED|PG_PAGEOUT)) 984 iomode = NFSV3WRITE_FILESYNC; 985 /* 986 * if we met a page which hasn't been sent yet, 987 * we need do WRITE RPC. 988 */ 989 if ((pgs[i]->flags & PG_NEEDCOMMIT) == 0) 990 needcommit = false; 991 mutex_exit(uobj->vmobjlock); 992 } else { 993 iomode = NFSV3WRITE_FILESYNC; 994 needcommit = false; 995 } 996 } 997 if (!needcommit && iomode == NFSV3WRITE_UNSTABLE) { 998 mutex_enter(uobj->vmobjlock); 999 for (i = 0; i < npages; i++) { 1000 pgs[i]->flags |= PG_NEEDCOMMIT | PG_RDONLY; 1001 pmap_page_protect(pgs[i], VM_PROT_READ); 1002 } 1003 mutex_exit(uobj->vmobjlock); 1004 pageprotected = true; /* pages can't be modified during i/o. */ 1005 } else 1006 pageprotected = false; 1007 1008 /* 1009 * Send the data to the server if necessary, 1010 * otherwise just send a commit rpc. 1011 */ 1012#ifndef NFS_V2_ONLY 1013 if (needcommit) { 1014 1015 /* 1016 * If the buffer is in the range that we already committed, 1017 * there's nothing to do. 1018 * 1019 * If it's in the range that we need to commit, push the 1020 * whole range at once, otherwise only push the buffer. 1021 * In both these cases, acquire the commit lock to avoid 1022 * other processes modifying the range. 1023 */ 1024 1025 off = uiop->uio_offset; 1026 cnt = bp->b_bcount; 1027 mutex_enter(&np->n_commitlock); 1028 if (!nfs_in_committed_range(vp, off, bp->b_bcount)) { 1029 bool pushedrange; 1030 if (nfs_in_tobecommitted_range(vp, off, bp->b_bcount)) { 1031 pushedrange = true; 1032 off = np->n_pushlo; 1033 cnt = np->n_pushhi - np->n_pushlo; 1034 } else { 1035 pushedrange = false; 1036 } 1037 error = nfs_commit(vp, off, cnt, curlwp); 1038 if (error == 0) { 1039 if (pushedrange) { 1040 nfs_merge_commit_ranges(vp); 1041 } else { 1042 nfs_add_committed_range(vp, off, cnt); 1043 } 1044 } 1045 } else { 1046 error = 0; 1047 } 1048 mutex_exit(&np->n_commitlock); 1049 rw_exit(&nmp->nm_writeverflock); 1050 if (!error) { 1051 /* 1052 * pages are now on stable storage. 1053 */ 1054 uiop->uio_resid = 0; 1055 mutex_enter(uobj->vmobjlock); 1056 for (i = 0; i < npages; i++) { 1057 pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY); 1058 } 1059 mutex_exit(uobj->vmobjlock); 1060 goto out; 1061 } else if (error == NFSERR_STALEWRITEVERF) { 1062 nfs_clearcommit(vp->v_mount); 1063 goto again; 1064 } 1065 if (error) { 1066 bp->b_error = np->n_error = error; 1067 np->n_flag |= NWRITEERR; 1068 } 1069 goto out; 1070 } 1071#endif 1072 off = uiop->uio_offset; 1073 cnt = bp->b_bcount; 1074 uiop->uio_rw = UIO_WRITE; 1075 nfsstats.write_bios++; 1076 error = nfs_writerpc(vp, uiop, &iomode, pageprotected, &stalewriteverf); 1077#ifndef NFS_V2_ONLY 1078 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1079 /* 1080 * we need to commit pages later. 1081 */ 1082 mutex_enter(&np->n_commitlock); 1083 nfs_add_tobecommitted_range(vp, off, cnt); 1084 /* 1085 * if there can be too many uncommitted pages, commit them now. 1086 */ 1087 if (np->n_pushhi - np->n_pushlo > nfs_commitsize) { 1088 off = np->n_pushlo; 1089 cnt = nfs_commitsize >> 1; 1090 error = nfs_commit(vp, off, cnt, curlwp); 1091 if (!error) { 1092 nfs_add_committed_range(vp, off, cnt); 1093 nfs_del_tobecommitted_range(vp, off, cnt); 1094 } 1095 if (error == NFSERR_STALEWRITEVERF) { 1096 stalewriteverf = true; 1097 error = 0; /* it isn't a real error */ 1098 } 1099 } else { 1100 /* 1101 * re-dirty pages so that they will be passed 1102 * to us later again. 1103 */ 1104 mutex_enter(uobj->vmobjlock); 1105 for (i = 0; i < npages; i++) { 1106 pgs[i]->flags &= ~PG_CLEAN; 1107 } 1108 mutex_exit(uobj->vmobjlock); 1109 } 1110 mutex_exit(&np->n_commitlock); 1111 } else 1112#endif 1113 if (!error) { 1114 /* 1115 * pages are now on stable storage. 1116 */ 1117 mutex_enter(&np->n_commitlock); 1118 nfs_del_committed_range(vp, off, cnt); 1119 mutex_exit(&np->n_commitlock); 1120 mutex_enter(uobj->vmobjlock); 1121 for (i = 0; i < npages; i++) { 1122 pgs[i]->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY); 1123 } 1124 mutex_exit(uobj->vmobjlock); 1125 } else { 1126 /* 1127 * we got an error. 1128 */ 1129 bp->b_error = np->n_error = error; 1130 np->n_flag |= NWRITEERR; 1131 } 1132 1133 rw_exit(&nmp->nm_writeverflock); 1134 1135 1136 if (stalewriteverf) { 1137 nfs_clearcommit(vp->v_mount); 1138 } 1139#ifndef NFS_V2_ONLY 1140out: 1141#endif 1142 if (pgs != spgs) 1143 kmem_free(pgs, sizeof(*pgs) * npages); 1144 return error; 1145} 1146 1147/* 1148 * nfs_doio for B_PHYS. 1149 */ 1150static int 1151nfs_doio_phys(struct buf *bp, struct uio *uiop) 1152{ 1153 struct vnode *vp = bp->b_vp; 1154 int error; 1155 1156 uiop->uio_offset = ((off_t)bp->b_blkno) << DEV_BSHIFT; 1157 if (bp->b_flags & B_READ) { 1158 uiop->uio_rw = UIO_READ; 1159 nfsstats.read_physios++; 1160 error = nfs_readrpc(vp, uiop); 1161 } else { 1162 int iomode = NFSV3WRITE_DATASYNC; 1163 bool stalewriteverf; 1164 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1165 1166 uiop->uio_rw = UIO_WRITE; 1167 nfsstats.write_physios++; 1168 rw_enter(&nmp->nm_writeverflock, RW_READER); 1169 error = nfs_writerpc(vp, uiop, &iomode, false, &stalewriteverf); 1170 rw_exit(&nmp->nm_writeverflock); 1171 if (stalewriteverf) { 1172 nfs_clearcommit(bp->b_vp->v_mount); 1173 } 1174 } 1175 bp->b_error = error; 1176 return error; 1177} 1178 1179/* 1180 * Do an I/O operation to/from a cache block. This may be called 1181 * synchronously or from an nfsiod. 1182 */ 1183int 1184nfs_doio(struct buf *bp) 1185{ 1186 int error; 1187 struct uio uio; 1188 struct uio *uiop = &uio; 1189 struct iovec io; 1190 UVMHIST_FUNC("nfs_doio"); UVMHIST_CALLED(ubchist); 1191 1192 uiop->uio_iov = &io; 1193 uiop->uio_iovcnt = 1; 1194 uiop->uio_offset = (((off_t)bp->b_blkno) << DEV_BSHIFT); 1195 UIO_SETUP_SYSSPACE(uiop); 1196 io.iov_base = bp->b_data; 1197 io.iov_len = uiop->uio_resid = bp->b_bcount; 1198 1199 /* 1200 * Historically, paging was done with physio, but no more... 1201 */ 1202 if (bp->b_flags & B_PHYS) { 1203 /* 1204 * ...though reading /dev/drum still gets us here. 1205 */ 1206 error = nfs_doio_phys(bp, uiop); 1207 } else if (bp->b_flags & B_READ) { 1208 error = nfs_doio_read(bp, uiop); 1209 } else { 1210 error = nfs_doio_write(bp, uiop); 1211 } 1212 bp->b_resid = uiop->uio_resid; 1213 biodone(bp); 1214 return (error); 1215} 1216 1217/* 1218 * Vnode op for VM getpages. 1219 */ 1220 1221int 1222nfs_getpages(void *v) 1223{ 1224 struct vop_getpages_args /* { 1225 struct vnode *a_vp; 1226 voff_t a_offset; 1227 struct vm_page **a_m; 1228 int *a_count; 1229 int a_centeridx; 1230 vm_prot_t a_access_type; 1231 int a_advice; 1232 int a_flags; 1233 } */ *ap = v; 1234 1235 struct vnode *vp = ap->a_vp; 1236 struct uvm_object *uobj = &vp->v_uobj; 1237 struct nfsnode *np = VTONFS(vp); 1238 const int npages = *ap->a_count; 1239 struct vm_page *pg, **pgs, **opgs, *spgs[UBC_MAX_PAGES]; 1240 off_t origoffset, len; 1241 int i, error; 1242 bool v3 = NFS_ISV3(vp); 1243 bool write = (ap->a_access_type & VM_PROT_WRITE) != 0; 1244 bool locked = (ap->a_flags & PGO_LOCKED) != 0; 1245 1246 /* 1247 * If we are not locked we are not really using opgs, 1248 * so just initialize it 1249 */ 1250 if (!locked || npages < __arraycount(spgs)) 1251 opgs = spgs; 1252 else { 1253 if ((opgs = kmem_alloc(npages * sizeof(*opgs), KM_NOSLEEP)) == 1254 NULL) 1255 return ENOMEM; 1256 } 1257 1258 /* 1259 * call the genfs code to get the pages. `pgs' may be NULL 1260 * when doing read-ahead. 1261 */ 1262 pgs = ap->a_m; 1263 if (write && locked && v3) { 1264 KASSERT(pgs != NULL); 1265#ifdef DEBUG 1266 1267 /* 1268 * If PGO_LOCKED is set, real pages shouldn't exists 1269 * in the array. 1270 */ 1271 1272 for (i = 0; i < npages; i++) 1273 KDASSERT(pgs[i] == NULL || pgs[i] == PGO_DONTCARE); 1274#endif 1275 memcpy(opgs, pgs, npages * sizeof(struct vm_pages *)); 1276 } 1277 error = genfs_getpages(v); 1278 if (error) 1279 goto out; 1280 1281 /* 1282 * for read faults where the nfs node is not yet marked NMODIFIED, 1283 * set PG_RDONLY on the pages so that we come back here if someone 1284 * tries to modify later via the mapping that will be entered for 1285 * this fault. 1286 */ 1287 1288 if (!write && (np->n_flag & NMODIFIED) == 0 && pgs != NULL) { 1289 if (!locked) { 1290 mutex_enter(uobj->vmobjlock); 1291 } 1292 for (i = 0; i < npages; i++) { 1293 pg = pgs[i]; 1294 if (pg == NULL || pg == PGO_DONTCARE) { 1295 continue; 1296 } 1297 pg->flags |= PG_RDONLY; 1298 } 1299 if (!locked) { 1300 mutex_exit(uobj->vmobjlock); 1301 } 1302 } 1303 if (!write) 1304 goto out; 1305 1306 /* 1307 * this is a write fault, update the commit info. 1308 */ 1309 1310 origoffset = ap->a_offset; 1311 len = npages << PAGE_SHIFT; 1312 1313 if (v3) { 1314 if (!locked) { 1315 mutex_enter(&np->n_commitlock); 1316 } else { 1317 if (!mutex_tryenter(&np->n_commitlock)) { 1318 1319 /* 1320 * Since PGO_LOCKED is set, we need to unbusy 1321 * all pages fetched by genfs_getpages() above, 1322 * tell the caller that there are no pages 1323 * available and put back original pgs array. 1324 */ 1325 1326 mutex_enter(&uvm_pageqlock); 1327 uvm_page_unbusy(pgs, npages); 1328 mutex_exit(&uvm_pageqlock); 1329 *ap->a_count = 0; 1330 memcpy(pgs, opgs, 1331 npages * sizeof(struct vm_pages *)); 1332 error = EBUSY; 1333 goto out; 1334 } 1335 } 1336 nfs_del_committed_range(vp, origoffset, len); 1337 nfs_del_tobecommitted_range(vp, origoffset, len); 1338 } 1339 np->n_flag |= NMODIFIED; 1340 if (!locked) { 1341 mutex_enter(uobj->vmobjlock); 1342 } 1343 for (i = 0; i < npages; i++) { 1344 pg = pgs[i]; 1345 if (pg == NULL || pg == PGO_DONTCARE) { 1346 continue; 1347 } 1348 pg->flags &= ~(PG_NEEDCOMMIT | PG_RDONLY); 1349 } 1350 if (!locked) { 1351 mutex_exit(uobj->vmobjlock); 1352 } 1353 if (v3) { 1354 mutex_exit(&np->n_commitlock); 1355 } 1356out: 1357 if (opgs != spgs) 1358 kmem_free(opgs, sizeof(*opgs) * npages); 1359 return error; 1360} 1361