| 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 | * $FreeBSD: /repoman/r/ncvs/src/sys/nfsclient/nfs_bio.c,v 1.130 2004/04/14 23:23:55 peadar Exp $ |
| 38 | * $DragonFly: src/sys/vfs/nfs/nfs_bio.c,v 1.42 2008/01/10 07:34:04 dillon Exp $ |
| 39 | */ |
| 40 | |
| 41 | |
| 42 | #include <sys/param.h> |
| 43 | #include <sys/systm.h> |
| 44 | #include <sys/resourcevar.h> |
| 45 | #include <sys/signalvar.h> |
| 46 | #include <sys/proc.h> |
| 47 | #include <sys/buf.h> |
| 48 | #include <sys/vnode.h> |
| 49 | #include <sys/mount.h> |
| 50 | #include <sys/kernel.h> |
| 51 | #include <sys/buf2.h> |
| 52 | #include <sys/msfbuf.h> |
| 53 | |
| 54 | #include <vm/vm.h> |
| 55 | #include <vm/vm_extern.h> |
| 56 | #include <vm/vm_page.h> |
| 57 | #include <vm/vm_object.h> |
| 58 | #include <vm/vm_pager.h> |
| 59 | #include <vm/vnode_pager.h> |
| 60 | |
| 61 | #include <sys/thread2.h> |
| 62 | |
| 63 | #include "rpcv2.h" |
| 64 | #include "nfsproto.h" |
| 65 | #include "nfs.h" |
| 66 | #include "nfsmount.h" |
| 67 | #include "nfsnode.h" |
| 68 | |
| 69 | static struct buf *nfs_getcacheblk(struct vnode *vp, off_t loffset, |
| 70 | int size, struct thread *td); |
| 71 | static int nfs_check_dirent(struct nfs_dirent *dp, int maxlen); |
| 72 | |
| 73 | extern int nfs_numasync; |
| 74 | extern int nfs_pbuf_freecnt; |
| 75 | extern struct nfsstats nfsstats; |
| 76 | |
| 77 | /* |
| 78 | * Vnode op for VM getpages. |
| 79 | * |
| 80 | * nfs_getpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, |
| 81 | * int a_reqpage, vm_ooffset_t a_offset) |
| 82 | */ |
| 83 | int |
| 84 | nfs_getpages(struct vop_getpages_args *ap) |
| 85 | { |
| 86 | struct thread *td = curthread; /* XXX */ |
| 87 | int i, error, nextoff, size, toff, count, npages; |
| 88 | struct uio uio; |
| 89 | struct iovec iov; |
| 90 | char *kva; |
| 91 | struct vnode *vp; |
| 92 | struct nfsmount *nmp; |
| 93 | vm_page_t *pages; |
| 94 | vm_page_t m; |
| 95 | struct msf_buf *msf; |
| 96 | |
| 97 | vp = ap->a_vp; |
| 98 | nmp = VFSTONFS(vp->v_mount); |
| 99 | pages = ap->a_m; |
| 100 | count = ap->a_count; |
| 101 | |
| 102 | if (vp->v_object == NULL) { |
| 103 | kprintf("nfs_getpages: called with non-merged cache vnode??\n"); |
| 104 | return VM_PAGER_ERROR; |
| 105 | } |
| 106 | |
| 107 | if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && |
| 108 | (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) |
| 109 | (void)nfs_fsinfo(nmp, vp, td); |
| 110 | |
| 111 | npages = btoc(count); |
| 112 | |
| 113 | /* |
| 114 | * NOTE that partially valid pages may occur in cases other |
| 115 | * then file EOF, such as when a file is partially written and |
| 116 | * ftruncate()-extended to a larger size. It is also possible |
| 117 | * for the valid bits to be set on garbage beyond the file EOF and |
| 118 | * clear in the area before EOF (e.g. m->valid == 0xfc), which can |
| 119 | * occur due to vtruncbuf() and the buffer cache's handling of |
| 120 | * pages which 'straddle' buffers or when b_bufsize is not a |
| 121 | * multiple of PAGE_SIZE.... the buffer cache cannot normally |
| 122 | * clear the extra bits. This kind of situation occurs when you |
| 123 | * make a small write() (m->valid == 0x03) and then mmap() and |
| 124 | * fault in the buffer(m->valid = 0xFF). When NFS flushes the |
| 125 | * buffer (vinvalbuf() m->valid = 0xFC) we are left with a mess. |
| 126 | * |
| 127 | * This is combined with the possibility that the pages are partially |
| 128 | * dirty or that there is a buffer backing the pages that is dirty |
| 129 | * (even if m->dirty is 0). |
| 130 | * |
| 131 | * To solve this problem several hacks have been made: (1) NFS |
| 132 | * guarentees that the IO block size is a multiple of PAGE_SIZE and |
| 133 | * (2) The buffer cache, when invalidating an NFS buffer, will |
| 134 | * disregard the buffer's fragmentory b_bufsize and invalidate |
| 135 | * the whole page rather then just the piece the buffer owns. |
| 136 | * |
| 137 | * This allows us to assume that a partially valid page found here |
| 138 | * is fully valid (vm_fault will zero'd out areas of the page not |
| 139 | * marked as valid). |
| 140 | */ |
| 141 | m = pages[ap->a_reqpage]; |
| 142 | if (m->valid != 0) { |
| 143 | for (i = 0; i < npages; ++i) { |
| 144 | if (i != ap->a_reqpage) |
| 145 | vnode_pager_freepage(pages[i]); |
| 146 | } |
| 147 | return(0); |
| 148 | } |
| 149 | |
| 150 | /* |
| 151 | * Use an MSF_BUF as a medium to retrieve data from the pages. |
| 152 | */ |
| 153 | msf_map_pagelist(&msf, pages, npages, 0); |
| 154 | KKASSERT(msf); |
| 155 | kva = msf_buf_kva(msf); |
| 156 | |
| 157 | iov.iov_base = kva; |
| 158 | iov.iov_len = count; |
| 159 | uio.uio_iov = &iov; |
| 160 | uio.uio_iovcnt = 1; |
| 161 | uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); |
| 162 | uio.uio_resid = count; |
| 163 | uio.uio_segflg = UIO_SYSSPACE; |
| 164 | uio.uio_rw = UIO_READ; |
| 165 | uio.uio_td = td; |
| 166 | |
| 167 | error = nfs_readrpc(vp, &uio); |
| 168 | msf_buf_free(msf); |
| 169 | |
| 170 | if (error && (uio.uio_resid == count)) { |
| 171 | kprintf("nfs_getpages: error %d\n", error); |
| 172 | for (i = 0; i < npages; ++i) { |
| 173 | if (i != ap->a_reqpage) |
| 174 | vnode_pager_freepage(pages[i]); |
| 175 | } |
| 176 | return VM_PAGER_ERROR; |
| 177 | } |
| 178 | |
| 179 | /* |
| 180 | * Calculate the number of bytes read and validate only that number |
| 181 | * of bytes. Note that due to pending writes, size may be 0. This |
| 182 | * does not mean that the remaining data is invalid! |
| 183 | */ |
| 184 | |
| 185 | size = count - uio.uio_resid; |
| 186 | |
| 187 | for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { |
| 188 | nextoff = toff + PAGE_SIZE; |
| 189 | m = pages[i]; |
| 190 | |
| 191 | m->flags &= ~PG_ZERO; |
| 192 | |
| 193 | if (nextoff <= size) { |
| 194 | /* |
| 195 | * Read operation filled an entire page |
| 196 | */ |
| 197 | m->valid = VM_PAGE_BITS_ALL; |
| 198 | vm_page_undirty(m); |
| 199 | } else if (size > toff) { |
| 200 | /* |
| 201 | * Read operation filled a partial page. |
| 202 | */ |
| 203 | m->valid = 0; |
| 204 | vm_page_set_validclean(m, 0, size - toff); |
| 205 | /* handled by vm_fault now */ |
| 206 | /* vm_page_zero_invalid(m, TRUE); */ |
| 207 | } else { |
| 208 | /* |
| 209 | * Read operation was short. If no error occured |
| 210 | * we may have hit a zero-fill section. We simply |
| 211 | * leave valid set to 0. |
| 212 | */ |
| 213 | ; |
| 214 | } |
| 215 | if (i != ap->a_reqpage) { |
| 216 | /* |
| 217 | * Whether or not to leave the page activated is up in |
| 218 | * the air, but we should put the page on a page queue |
| 219 | * somewhere (it already is in the object). Result: |
| 220 | * It appears that emperical results show that |
| 221 | * deactivating pages is best. |
| 222 | */ |
| 223 | |
| 224 | /* |
| 225 | * Just in case someone was asking for this page we |
| 226 | * now tell them that it is ok to use. |
| 227 | */ |
| 228 | if (!error) { |
| 229 | if (m->flags & PG_WANTED) |
| 230 | vm_page_activate(m); |
| 231 | else |
| 232 | vm_page_deactivate(m); |
| 233 | vm_page_wakeup(m); |
| 234 | } else { |
| 235 | vnode_pager_freepage(m); |
| 236 | } |
| 237 | } |
| 238 | } |
| 239 | return 0; |
| 240 | } |
| 241 | |
| 242 | /* |
| 243 | * Vnode op for VM putpages. |
| 244 | * |
| 245 | * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync, |
| 246 | * int *a_rtvals, vm_ooffset_t a_offset) |
| 247 | */ |
| 248 | int |
| 249 | nfs_putpages(struct vop_putpages_args *ap) |
| 250 | { |
| 251 | struct thread *td = curthread; |
| 252 | struct uio uio; |
| 253 | struct iovec iov; |
| 254 | char *kva; |
| 255 | int iomode, must_commit, i, error, npages, count; |
| 256 | off_t offset; |
| 257 | int *rtvals; |
| 258 | struct vnode *vp; |
| 259 | struct nfsmount *nmp; |
| 260 | struct nfsnode *np; |
| 261 | vm_page_t *pages; |
| 262 | struct msf_buf *msf; |
| 263 | |
| 264 | vp = ap->a_vp; |
| 265 | np = VTONFS(vp); |
| 266 | nmp = VFSTONFS(vp->v_mount); |
| 267 | pages = ap->a_m; |
| 268 | count = ap->a_count; |
| 269 | rtvals = ap->a_rtvals; |
| 270 | npages = btoc(count); |
| 271 | offset = IDX_TO_OFF(pages[0]->pindex); |
| 272 | |
| 273 | if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && |
| 274 | (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) |
| 275 | (void)nfs_fsinfo(nmp, vp, td); |
| 276 | |
| 277 | for (i = 0; i < npages; i++) { |
| 278 | rtvals[i] = VM_PAGER_AGAIN; |
| 279 | } |
| 280 | |
| 281 | /* |
| 282 | * When putting pages, do not extend file past EOF. |
| 283 | */ |
| 284 | |
| 285 | if (offset + count > np->n_size) { |
| 286 | count = np->n_size - offset; |
| 287 | if (count < 0) |
| 288 | count = 0; |
| 289 | } |
| 290 | |
| 291 | /* |
| 292 | * Use an MSF_BUF as a medium to retrieve data from the pages. |
| 293 | */ |
| 294 | msf_map_pagelist(&msf, pages, npages, 0); |
| 295 | KKASSERT(msf); |
| 296 | kva = msf_buf_kva(msf); |
| 297 | |
| 298 | iov.iov_base = kva; |
| 299 | iov.iov_len = count; |
| 300 | uio.uio_iov = &iov; |
| 301 | uio.uio_iovcnt = 1; |
| 302 | uio.uio_offset = offset; |
| 303 | uio.uio_resid = count; |
| 304 | uio.uio_segflg = UIO_SYSSPACE; |
| 305 | uio.uio_rw = UIO_WRITE; |
| 306 | uio.uio_td = td; |
| 307 | |
| 308 | if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) |
| 309 | iomode = NFSV3WRITE_UNSTABLE; |
| 310 | else |
| 311 | iomode = NFSV3WRITE_FILESYNC; |
| 312 | |
| 313 | error = nfs_writerpc(vp, &uio, &iomode, &must_commit); |
| 314 | |
| 315 | msf_buf_free(msf); |
| 316 | |
| 317 | if (!error) { |
| 318 | int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; |
| 319 | for (i = 0; i < nwritten; i++) { |
| 320 | rtvals[i] = VM_PAGER_OK; |
| 321 | vm_page_undirty(pages[i]); |
| 322 | } |
| 323 | if (must_commit) |
| 324 | nfs_clearcommit(vp->v_mount); |
| 325 | } |
| 326 | return rtvals[0]; |
| 327 | } |
| 328 | |
| 329 | /* |
| 330 | * Vnode op for read using bio |
| 331 | */ |
| 332 | int |
| 333 | nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag) |
| 334 | { |
| 335 | struct nfsnode *np = VTONFS(vp); |
| 336 | int biosize, i; |
| 337 | struct buf *bp = 0, *rabp; |
| 338 | struct vattr vattr; |
| 339 | struct thread *td; |
| 340 | struct nfsmount *nmp = VFSTONFS(vp->v_mount); |
| 341 | daddr_t lbn, rabn; |
| 342 | off_t raoffset; |
| 343 | off_t loffset; |
| 344 | int bcount; |
| 345 | int seqcount; |
| 346 | int nra, error = 0, n = 0, on = 0; |
| 347 | |
| 348 | #ifdef DIAGNOSTIC |
| 349 | if (uio->uio_rw != UIO_READ) |
| 350 | panic("nfs_read mode"); |
| 351 | #endif |
| 352 | if (uio->uio_resid == 0) |
| 353 | return (0); |
| 354 | if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ |
| 355 | return (EINVAL); |
| 356 | td = uio->uio_td; |
| 357 | |
| 358 | if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && |
| 359 | (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) |
| 360 | (void)nfs_fsinfo(nmp, vp, td); |
| 361 | if (vp->v_type != VDIR && |
| 362 | (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) |
| 363 | return (EFBIG); |
| 364 | biosize = vp->v_mount->mnt_stat.f_iosize; |
| 365 | seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); |
| 366 | |
| 367 | /* |
| 368 | * For nfs, cache consistency can only be maintained approximately. |
| 369 | * Although RFC1094 does not specify the criteria, the following is |
| 370 | * believed to be compatible with the reference port. |
| 371 | * |
| 372 | * NFS: If local changes have been made and this is a |
| 373 | * directory, the directory must be invalidated and |
| 374 | * the attribute cache must be cleared. |
| 375 | * |
| 376 | * GETATTR is called to synchronize the file size. |
| 377 | * |
| 378 | * If remote changes are detected local data is flushed |
| 379 | * and the cache is invalidated. |
| 380 | * |
| 381 | * NOTE: In the normal case the attribute cache is not |
| 382 | * cleared which means GETATTR may use cached data and |
| 383 | * not immediately detect changes made on the server. |
| 384 | */ |
| 385 | if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) { |
| 386 | nfs_invaldir(vp); |
| 387 | error = nfs_vinvalbuf(vp, V_SAVE, 1); |
| 388 | if (error) |
| 389 | return (error); |
| 390 | np->n_attrstamp = 0; |
| 391 | } |
| 392 | error = VOP_GETATTR(vp, &vattr); |
| 393 | if (error) |
| 394 | return (error); |
| 395 | if (np->n_flag & NRMODIFIED) { |
| 396 | if (vp->v_type == VDIR) |
| 397 | nfs_invaldir(vp); |
| 398 | error = nfs_vinvalbuf(vp, V_SAVE, 1); |
| 399 | if (error) |
| 400 | return (error); |
| 401 | np->n_flag &= ~NRMODIFIED; |
| 402 | } |
| 403 | do { |
| 404 | if (np->n_flag & NDONTCACHE) { |
| 405 | switch (vp->v_type) { |
| 406 | case VREG: |
| 407 | return (nfs_readrpc(vp, uio)); |
| 408 | case VLNK: |
| 409 | return (nfs_readlinkrpc(vp, uio)); |
| 410 | case VDIR: |
| 411 | break; |
| 412 | default: |
| 413 | kprintf(" NDONTCACHE: type %x unexpected\n", vp->v_type); |
| 414 | break; |
| 415 | }; |
| 416 | } |
| 417 | switch (vp->v_type) { |
| 418 | case VREG: |
| 419 | nfsstats.biocache_reads++; |
| 420 | lbn = uio->uio_offset / biosize; |
| 421 | on = uio->uio_offset & (biosize - 1); |
| 422 | loffset = (off_t)lbn * biosize; |
| 423 | |
| 424 | /* |
| 425 | * Start the read ahead(s), as required. |
| 426 | */ |
| 427 | if (nfs_numasync > 0 && nmp->nm_readahead > 0) { |
| 428 | for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && |
| 429 | (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { |
| 430 | rabn = lbn + 1 + nra; |
| 431 | raoffset = (off_t)rabn * biosize; |
| 432 | if (!findblk(vp, raoffset)) { |
| 433 | rabp = nfs_getcacheblk(vp, raoffset, biosize, td); |
| 434 | if (!rabp) |
| 435 | return (EINTR); |
| 436 | if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { |
| 437 | rabp->b_flags |= B_ASYNC; |
| 438 | rabp->b_cmd = BUF_CMD_READ; |
| 439 | vfs_busy_pages(vp, rabp); |
| 440 | if (nfs_asyncio(vp, &rabp->b_bio2, td)) { |
| 441 | rabp->b_flags |= B_INVAL|B_ERROR; |
| 442 | vfs_unbusy_pages(rabp); |
| 443 | brelse(rabp); |
| 444 | break; |
| 445 | } |
| 446 | } else { |
| 447 | brelse(rabp); |
| 448 | } |
| 449 | } |
| 450 | } |
| 451 | } |
| 452 | |
| 453 | /* |
| 454 | * Obtain the buffer cache block. Figure out the buffer size |
| 455 | * when we are at EOF. If we are modifying the size of the |
| 456 | * buffer based on an EOF condition we need to hold |
| 457 | * nfs_rslock() through obtaining the buffer to prevent |
| 458 | * a potential writer-appender from messing with n_size. |
| 459 | * Otherwise we may accidently truncate the buffer and |
| 460 | * lose dirty data. |
| 461 | * |
| 462 | * Note that bcount is *not* DEV_BSIZE aligned. |
| 463 | */ |
| 464 | |
| 465 | again: |
| 466 | bcount = biosize; |
| 467 | if (loffset >= np->n_size) { |
| 468 | bcount = 0; |
| 469 | } else if (loffset + biosize > np->n_size) { |
| 470 | bcount = np->n_size - loffset; |
| 471 | } |
| 472 | if (bcount != biosize) { |
| 473 | switch(nfs_rslock(np)) { |
| 474 | case ENOLCK: |
| 475 | goto again; |
| 476 | /* not reached */ |
| 477 | case EINTR: |
| 478 | case ERESTART: |
| 479 | return(EINTR); |
| 480 | /* not reached */ |
| 481 | default: |
| 482 | break; |
| 483 | } |
| 484 | } |
| 485 | |
| 486 | bp = nfs_getcacheblk(vp, loffset, bcount, td); |
| 487 | |
| 488 | if (bcount != biosize) |
| 489 | nfs_rsunlock(np); |
| 490 | if (!bp) |
| 491 | return (EINTR); |
| 492 | |
| 493 | /* |
| 494 | * If B_CACHE is not set, we must issue the read. If this |
| 495 | * fails, we return an error. |
| 496 | */ |
| 497 | |
| 498 | if ((bp->b_flags & B_CACHE) == 0) { |
| 499 | bp->b_cmd = BUF_CMD_READ; |
| 500 | vfs_busy_pages(vp, bp); |
| 501 | error = nfs_doio(vp, &bp->b_bio2, td); |
| 502 | if (error) { |
| 503 | brelse(bp); |
| 504 | return (error); |
| 505 | } |
| 506 | } |
| 507 | |
| 508 | /* |
| 509 | * on is the offset into the current bp. Figure out how many |
| 510 | * bytes we can copy out of the bp. Note that bcount is |
| 511 | * NOT DEV_BSIZE aligned. |
| 512 | * |
| 513 | * Then figure out how many bytes we can copy into the uio. |
| 514 | */ |
| 515 | |
| 516 | n = 0; |
| 517 | if (on < bcount) |
| 518 | n = min((unsigned)(bcount - on), uio->uio_resid); |
| 519 | break; |
| 520 | case VLNK: |
| 521 | biosize = min(NFS_MAXPATHLEN, np->n_size); |
| 522 | nfsstats.biocache_readlinks++; |
| 523 | bp = nfs_getcacheblk(vp, (off_t)0, biosize, td); |
| 524 | if (bp == NULL) |
| 525 | return (EINTR); |
| 526 | if ((bp->b_flags & B_CACHE) == 0) { |
| 527 | bp->b_cmd = BUF_CMD_READ; |
| 528 | vfs_busy_pages(vp, bp); |
| 529 | error = nfs_doio(vp, &bp->b_bio2, td); |
| 530 | if (error) { |
| 531 | bp->b_flags |= B_ERROR; |
| 532 | brelse(bp); |
| 533 | return (error); |
| 534 | } |
| 535 | } |
| 536 | n = min(uio->uio_resid, bp->b_bcount - bp->b_resid); |
| 537 | on = 0; |
| 538 | break; |
| 539 | case VDIR: |
| 540 | nfsstats.biocache_readdirs++; |
| 541 | if (np->n_direofoffset |
| 542 | && uio->uio_offset >= np->n_direofoffset) { |
| 543 | return (0); |
| 544 | } |
| 545 | lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; |
| 546 | on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); |
| 547 | loffset = uio->uio_offset - on; |
| 548 | bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td); |
| 549 | if (bp == NULL) |
| 550 | return (EINTR); |
| 551 | |
| 552 | if ((bp->b_flags & B_CACHE) == 0) { |
| 553 | bp->b_cmd = BUF_CMD_READ; |
| 554 | vfs_busy_pages(vp, bp); |
| 555 | error = nfs_doio(vp, &bp->b_bio2, td); |
| 556 | if (error) { |
| 557 | brelse(bp); |
| 558 | } |
| 559 | while (error == NFSERR_BAD_COOKIE) { |
| 560 | kprintf("got bad cookie vp %p bp %p\n", vp, bp); |
| 561 | nfs_invaldir(vp); |
| 562 | error = nfs_vinvalbuf(vp, 0, 1); |
| 563 | /* |
| 564 | * Yuck! The directory has been modified on the |
| 565 | * server. The only way to get the block is by |
| 566 | * reading from the beginning to get all the |
| 567 | * offset cookies. |
| 568 | * |
| 569 | * Leave the last bp intact unless there is an error. |
| 570 | * Loop back up to the while if the error is another |
| 571 | * NFSERR_BAD_COOKIE (double yuch!). |
| 572 | */ |
| 573 | for (i = 0; i <= lbn && !error; i++) { |
| 574 | if (np->n_direofoffset |
| 575 | && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) |
| 576 | return (0); |
| 577 | bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ, |
| 578 | NFS_DIRBLKSIZ, td); |
| 579 | if (!bp) |
| 580 | return (EINTR); |
| 581 | if ((bp->b_flags & B_CACHE) == 0) { |
| 582 | bp->b_cmd = BUF_CMD_READ; |
| 583 | vfs_busy_pages(vp, bp); |
| 584 | error = nfs_doio(vp, &bp->b_bio2, td); |
| 585 | /* |
| 586 | * no error + B_INVAL == directory EOF, |
| 587 | * use the block. |
| 588 | */ |
| 589 | if (error == 0 && (bp->b_flags & B_INVAL)) |
| 590 | break; |
| 591 | } |
| 592 | /* |
| 593 | * An error will throw away the block and the |
| 594 | * for loop will break out. If no error and this |
| 595 | * is not the block we want, we throw away the |
| 596 | * block and go for the next one via the for loop. |
| 597 | */ |
| 598 | if (error || i < lbn) |
| 599 | brelse(bp); |
| 600 | } |
| 601 | } |
| 602 | /* |
| 603 | * The above while is repeated if we hit another cookie |
| 604 | * error. If we hit an error and it wasn't a cookie error, |
| 605 | * we give up. |
| 606 | */ |
| 607 | if (error) |
| 608 | return (error); |
| 609 | } |
| 610 | |
| 611 | /* |
| 612 | * If not eof and read aheads are enabled, start one. |
| 613 | * (You need the current block first, so that you have the |
| 614 | * directory offset cookie of the next block.) |
| 615 | */ |
| 616 | if (nfs_numasync > 0 && nmp->nm_readahead > 0 && |
| 617 | (bp->b_flags & B_INVAL) == 0 && |
| 618 | (np->n_direofoffset == 0 || |
| 619 | loffset + NFS_DIRBLKSIZ < np->n_direofoffset) && |
| 620 | (np->n_flag & NDONTCACHE) == 0 && |
| 621 | !findblk(vp, loffset + NFS_DIRBLKSIZ)) { |
| 622 | rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ, |
| 623 | NFS_DIRBLKSIZ, td); |
| 624 | if (rabp) { |
| 625 | if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { |
| 626 | rabp->b_flags |= B_ASYNC; |
| 627 | rabp->b_cmd = BUF_CMD_READ; |
| 628 | vfs_busy_pages(vp, rabp); |
| 629 | if (nfs_asyncio(vp, &rabp->b_bio2, td)) { |
| 630 | rabp->b_flags |= B_INVAL|B_ERROR; |
| 631 | vfs_unbusy_pages(rabp); |
| 632 | brelse(rabp); |
| 633 | } |
| 634 | } else { |
| 635 | brelse(rabp); |
| 636 | } |
| 637 | } |
| 638 | } |
| 639 | /* |
| 640 | * Unlike VREG files, whos buffer size ( bp->b_bcount ) is |
| 641 | * chopped for the EOF condition, we cannot tell how large |
| 642 | * NFS directories are going to be until we hit EOF. So |
| 643 | * an NFS directory buffer is *not* chopped to its EOF. Now, |
| 644 | * it just so happens that b_resid will effectively chop it |
| 645 | * to EOF. *BUT* this information is lost if the buffer goes |
| 646 | * away and is reconstituted into a B_CACHE state ( due to |
| 647 | * being VMIO ) later. So we keep track of the directory eof |
| 648 | * in np->n_direofoffset and chop it off as an extra step |
| 649 | * right here. |
| 650 | */ |
| 651 | n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); |
| 652 | if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) |
| 653 | n = np->n_direofoffset - uio->uio_offset; |
| 654 | break; |
| 655 | default: |
| 656 | kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type); |
| 657 | break; |
| 658 | }; |
| 659 | |
| 660 | switch (vp->v_type) { |
| 661 | case VREG: |
| 662 | if (n > 0) |
| 663 | error = uiomove(bp->b_data + on, (int)n, uio); |
| 664 | break; |
| 665 | case VLNK: |
| 666 | if (n > 0) |
| 667 | error = uiomove(bp->b_data + on, (int)n, uio); |
| 668 | n = 0; |
| 669 | break; |
| 670 | case VDIR: |
| 671 | if (n > 0) { |
| 672 | off_t old_off = uio->uio_offset; |
| 673 | caddr_t cpos, epos; |
| 674 | struct nfs_dirent *dp; |
| 675 | |
| 676 | /* |
| 677 | * We are casting cpos to nfs_dirent, it must be |
| 678 | * int-aligned. |
| 679 | */ |
| 680 | if (on & 3) { |
| 681 | error = EINVAL; |
| 682 | break; |
| 683 | } |
| 684 | |
| 685 | cpos = bp->b_data + on; |
| 686 | epos = bp->b_data + on + n; |
| 687 | while (cpos < epos && error == 0 && uio->uio_resid > 0) { |
| 688 | dp = (struct nfs_dirent *)cpos; |
| 689 | error = nfs_check_dirent(dp, (int)(epos - cpos)); |
| 690 | if (error) |
| 691 | break; |
| 692 | if (vop_write_dirent(&error, uio, dp->nfs_ino, |
| 693 | dp->nfs_type, dp->nfs_namlen, dp->nfs_name)) { |
| 694 | break; |
| 695 | } |
| 696 | cpos += dp->nfs_reclen; |
| 697 | } |
| 698 | n = 0; |
| 699 | if (error == 0) |
| 700 | uio->uio_offset = old_off + cpos - bp->b_data - on; |
| 701 | } |
| 702 | /* |
| 703 | * Invalidate buffer if caching is disabled, forcing a |
| 704 | * re-read from the remote later. |
| 705 | */ |
| 706 | if (np->n_flag & NDONTCACHE) |
| 707 | bp->b_flags |= B_INVAL; |
| 708 | break; |
| 709 | default: |
| 710 | kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type); |
| 711 | } |
| 712 | brelse(bp); |
| 713 | } while (error == 0 && uio->uio_resid > 0 && n > 0); |
| 714 | return (error); |
| 715 | } |
| 716 | |
| 717 | /* |
| 718 | * Userland can supply any 'seek' offset when reading a NFS directory. |
| 719 | * Validate the structure so we don't panic the kernel. Note that |
| 720 | * the element name is nul terminated and the nul is not included |
| 721 | * in nfs_namlen. |
| 722 | */ |
| 723 | static |
| 724 | int |
| 725 | nfs_check_dirent(struct nfs_dirent *dp, int maxlen) |
| 726 | { |
| 727 | int nfs_name_off = offsetof(struct nfs_dirent, nfs_name[0]); |
| 728 | |
| 729 | if (nfs_name_off >= maxlen) |
| 730 | return (EINVAL); |
| 731 | if (dp->nfs_reclen < nfs_name_off || dp->nfs_reclen > maxlen) |
| 732 | return (EINVAL); |
| 733 | if (nfs_name_off + dp->nfs_namlen >= dp->nfs_reclen) |
| 734 | return (EINVAL); |
| 735 | if (dp->nfs_reclen & 3) |
| 736 | return (EINVAL); |
| 737 | return (0); |
| 738 | } |
| 739 | |
| 740 | /* |
| 741 | * Vnode op for write using bio |
| 742 | * |
| 743 | * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag, |
| 744 | * struct ucred *a_cred) |
| 745 | */ |
| 746 | int |
| 747 | nfs_write(struct vop_write_args *ap) |
| 748 | { |
| 749 | struct uio *uio = ap->a_uio; |
| 750 | struct thread *td = uio->uio_td; |
| 751 | struct vnode *vp = ap->a_vp; |
| 752 | struct nfsnode *np = VTONFS(vp); |
| 753 | int ioflag = ap->a_ioflag; |
| 754 | struct buf *bp; |
| 755 | struct vattr vattr; |
| 756 | struct nfsmount *nmp = VFSTONFS(vp->v_mount); |
| 757 | daddr_t lbn; |
| 758 | off_t loffset; |
| 759 | int n, on, error = 0, iomode, must_commit; |
| 760 | int haverslock = 0; |
| 761 | int bcount; |
| 762 | int biosize; |
| 763 | |
| 764 | #ifdef DIAGNOSTIC |
| 765 | if (uio->uio_rw != UIO_WRITE) |
| 766 | panic("nfs_write mode"); |
| 767 | if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) |
| 768 | panic("nfs_write proc"); |
| 769 | #endif |
| 770 | if (vp->v_type != VREG) |
| 771 | return (EIO); |
| 772 | if (np->n_flag & NWRITEERR) { |
| 773 | np->n_flag &= ~NWRITEERR; |
| 774 | return (np->n_error); |
| 775 | } |
| 776 | if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && |
| 777 | (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) |
| 778 | (void)nfs_fsinfo(nmp, vp, td); |
| 779 | |
| 780 | /* |
| 781 | * Synchronously flush pending buffers if we are in synchronous |
| 782 | * mode or if we are appending. |
| 783 | */ |
| 784 | if (ioflag & (IO_APPEND | IO_SYNC)) { |
| 785 | if (np->n_flag & NLMODIFIED) { |
| 786 | np->n_attrstamp = 0; |
| 787 | error = nfs_flush(vp, MNT_WAIT, td, 0); |
| 788 | /* error = nfs_vinvalbuf(vp, V_SAVE, 1); */ |
| 789 | if (error) |
| 790 | return (error); |
| 791 | } |
| 792 | } |
| 793 | |
| 794 | /* |
| 795 | * If IO_APPEND then load uio_offset. We restart here if we cannot |
| 796 | * get the append lock. |
| 797 | */ |
| 798 | restart: |
| 799 | if (ioflag & IO_APPEND) { |
| 800 | np->n_attrstamp = 0; |
| 801 | error = VOP_GETATTR(vp, &vattr); |
| 802 | if (error) |
| 803 | return (error); |
| 804 | uio->uio_offset = np->n_size; |
| 805 | } |
| 806 | |
| 807 | if (uio->uio_offset < 0) |
| 808 | return (EINVAL); |
| 809 | if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) |
| 810 | return (EFBIG); |
| 811 | if (uio->uio_resid == 0) |
| 812 | return (0); |
| 813 | |
| 814 | /* |
| 815 | * We need to obtain the rslock if we intend to modify np->n_size |
| 816 | * in order to guarentee the append point with multiple contending |
| 817 | * writers, to guarentee that no other appenders modify n_size |
| 818 | * while we are trying to obtain a truncated buffer (i.e. to avoid |
| 819 | * accidently truncating data written by another appender due to |
| 820 | * the race), and to ensure that the buffer is populated prior to |
| 821 | * our extending of the file. We hold rslock through the entire |
| 822 | * operation. |
| 823 | * |
| 824 | * Note that we do not synchronize the case where someone truncates |
| 825 | * the file while we are appending to it because attempting to lock |
| 826 | * this case may deadlock other parts of the system unexpectedly. |
| 827 | */ |
| 828 | if ((ioflag & IO_APPEND) || |
| 829 | uio->uio_offset + uio->uio_resid > np->n_size) { |
| 830 | switch(nfs_rslock(np)) { |
| 831 | case ENOLCK: |
| 832 | goto restart; |
| 833 | /* not reached */ |
| 834 | case EINTR: |
| 835 | case ERESTART: |
| 836 | return(EINTR); |
| 837 | /* not reached */ |
| 838 | default: |
| 839 | break; |
| 840 | } |
| 841 | haverslock = 1; |
| 842 | } |
| 843 | |
| 844 | /* |
| 845 | * Maybe this should be above the vnode op call, but so long as |
| 846 | * file servers have no limits, i don't think it matters |
| 847 | */ |
| 848 | if (td->td_proc && uio->uio_offset + uio->uio_resid > |
| 849 | td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) { |
| 850 | lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ); |
| 851 | if (haverslock) |
| 852 | nfs_rsunlock(np); |
| 853 | return (EFBIG); |
| 854 | } |
| 855 | |
| 856 | biosize = vp->v_mount->mnt_stat.f_iosize; |
| 857 | |
| 858 | do { |
| 859 | if ((np->n_flag & NDONTCACHE) && uio->uio_iovcnt == 1) { |
| 860 | iomode = NFSV3WRITE_FILESYNC; |
| 861 | error = nfs_writerpc(vp, uio, &iomode, &must_commit); |
| 862 | if (must_commit) |
| 863 | nfs_clearcommit(vp->v_mount); |
| 864 | break; |
| 865 | } |
| 866 | nfsstats.biocache_writes++; |
| 867 | lbn = uio->uio_offset / biosize; |
| 868 | on = uio->uio_offset & (biosize-1); |
| 869 | loffset = uio->uio_offset - on; |
| 870 | n = min((unsigned)(biosize - on), uio->uio_resid); |
| 871 | again: |
| 872 | /* |
| 873 | * Handle direct append and file extension cases, calculate |
| 874 | * unaligned buffer size. |
| 875 | */ |
| 876 | |
| 877 | if (uio->uio_offset == np->n_size && n) { |
| 878 | /* |
| 879 | * Get the buffer (in its pre-append state to maintain |
| 880 | * B_CACHE if it was previously set). Resize the |
| 881 | * nfsnode after we have locked the buffer to prevent |
| 882 | * readers from reading garbage. |
| 883 | */ |
| 884 | bcount = on; |
| 885 | bp = nfs_getcacheblk(vp, loffset, bcount, td); |
| 886 | |
| 887 | if (bp != NULL) { |
| 888 | long save; |
| 889 | |
| 890 | np->n_size = uio->uio_offset + n; |
| 891 | np->n_flag |= NLMODIFIED; |
| 892 | vnode_pager_setsize(vp, np->n_size); |
| 893 | |
| 894 | save = bp->b_flags & B_CACHE; |
| 895 | bcount += n; |
| 896 | allocbuf(bp, bcount); |
| 897 | bp->b_flags |= save; |
| 898 | } |
| 899 | } else { |
| 900 | /* |
| 901 | * Obtain the locked cache block first, and then |
| 902 | * adjust the file's size as appropriate. |
| 903 | */ |
| 904 | bcount = on + n; |
| 905 | if (loffset + bcount < np->n_size) { |
| 906 | if (loffset + biosize < np->n_size) |
| 907 | bcount = biosize; |
| 908 | else |
| 909 | bcount = np->n_size - loffset; |
| 910 | } |
| 911 | bp = nfs_getcacheblk(vp, loffset, bcount, td); |
| 912 | if (uio->uio_offset + n > np->n_size) { |
| 913 | np->n_size = uio->uio_offset + n; |
| 914 | np->n_flag |= NLMODIFIED; |
| 915 | vnode_pager_setsize(vp, np->n_size); |
| 916 | } |
| 917 | } |
| 918 | |
| 919 | if (bp == NULL) { |
| 920 | error = EINTR; |
| 921 | break; |
| 922 | } |
| 923 | |
| 924 | /* |
| 925 | * Issue a READ if B_CACHE is not set. In special-append |
| 926 | * mode, B_CACHE is based on the buffer prior to the write |
| 927 | * op and is typically set, avoiding the read. If a read |
| 928 | * is required in special append mode, the server will |
| 929 | * probably send us a short-read since we extended the file |
| 930 | * on our end, resulting in b_resid == 0 and, thusly, |
| 931 | * B_CACHE getting set. |
| 932 | * |
| 933 | * We can also avoid issuing the read if the write covers |
| 934 | * the entire buffer. We have to make sure the buffer state |
| 935 | * is reasonable in this case since we will not be initiating |
| 936 | * I/O. See the comments in kern/vfs_bio.c's getblk() for |
| 937 | * more information. |
| 938 | * |
| 939 | * B_CACHE may also be set due to the buffer being cached |
| 940 | * normally. |
| 941 | * |
| 942 | * When doing a UIO_NOCOPY write the buffer is not |
| 943 | * overwritten and we cannot just set B_CACHE unconditionally |
| 944 | * for full-block writes. |
| 945 | */ |
| 946 | |
| 947 | if (on == 0 && n == bcount && uio->uio_segflg != UIO_NOCOPY) { |
| 948 | bp->b_flags |= B_CACHE; |
| 949 | bp->b_flags &= ~(B_ERROR | B_INVAL); |
| 950 | } |
| 951 | |
| 952 | if ((bp->b_flags & B_CACHE) == 0) { |
| 953 | bp->b_cmd = BUF_CMD_READ; |
| 954 | vfs_busy_pages(vp, bp); |
| 955 | error = nfs_doio(vp, &bp->b_bio2, td); |
| 956 | if (error) { |
| 957 | brelse(bp); |
| 958 | break; |
| 959 | } |
| 960 | } |
| 961 | if (!bp) { |
| 962 | error = EINTR; |
| 963 | break; |
| 964 | } |
| 965 | np->n_flag |= NLMODIFIED; |
| 966 | |
| 967 | /* |
| 968 | * If dirtyend exceeds file size, chop it down. This should |
| 969 | * not normally occur but there is an append race where it |
| 970 | * might occur XXX, so we log it. |
| 971 | * |
| 972 | * If the chopping creates a reverse-indexed or degenerate |
| 973 | * situation with dirtyoff/end, we 0 both of them. |
| 974 | */ |
| 975 | |
| 976 | if (bp->b_dirtyend > bcount) { |
| 977 | kprintf("NFS append race @%08llx:%d\n", |
| 978 | bp->b_bio2.bio_offset, |
| 979 | bp->b_dirtyend - bcount); |
| 980 | bp->b_dirtyend = bcount; |
| 981 | } |
| 982 | |
| 983 | if (bp->b_dirtyoff >= bp->b_dirtyend) |
| 984 | bp->b_dirtyoff = bp->b_dirtyend = 0; |
| 985 | |
| 986 | /* |
| 987 | * If the new write will leave a contiguous dirty |
| 988 | * area, just update the b_dirtyoff and b_dirtyend, |
| 989 | * otherwise force a write rpc of the old dirty area. |
| 990 | * |
| 991 | * While it is possible to merge discontiguous writes due to |
| 992 | * our having a B_CACHE buffer ( and thus valid read data |
| 993 | * for the hole), we don't because it could lead to |
| 994 | * significant cache coherency problems with multiple clients, |
| 995 | * especially if locking is implemented later on. |
| 996 | * |
| 997 | * as an optimization we could theoretically maintain |
| 998 | * a linked list of discontinuous areas, but we would still |
| 999 | * have to commit them separately so there isn't much |
| 1000 | * advantage to it except perhaps a bit of asynchronization. |
| 1001 | */ |
| 1002 | |
| 1003 | if (bp->b_dirtyend > 0 && |
| 1004 | (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { |
| 1005 | if (bwrite(bp) == EINTR) { |
| 1006 | error = EINTR; |
| 1007 | break; |
| 1008 | } |
| 1009 | goto again; |
| 1010 | } |
| 1011 | |
| 1012 | error = uiomove((char *)bp->b_data + on, n, uio); |
| 1013 | |
| 1014 | /* |
| 1015 | * Since this block is being modified, it must be written |
| 1016 | * again and not just committed. Since write clustering does |
| 1017 | * not work for the stage 1 data write, only the stage 2 |
| 1018 | * commit rpc, we have to clear B_CLUSTEROK as well. |
| 1019 | */ |
| 1020 | bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); |
| 1021 | |
| 1022 | if (error) { |
| 1023 | bp->b_flags |= B_ERROR; |
| 1024 | brelse(bp); |
| 1025 | break; |
| 1026 | } |
| 1027 | |
| 1028 | /* |
| 1029 | * Only update dirtyoff/dirtyend if not a degenerate |
| 1030 | * condition. |
| 1031 | */ |
| 1032 | if (n) { |
| 1033 | if (bp->b_dirtyend > 0) { |
| 1034 | bp->b_dirtyoff = min(on, bp->b_dirtyoff); |
| 1035 | bp->b_dirtyend = max((on + n), bp->b_dirtyend); |
| 1036 | } else { |
| 1037 | bp->b_dirtyoff = on; |
| 1038 | bp->b_dirtyend = on + n; |
| 1039 | } |
| 1040 | vfs_bio_set_validclean(bp, on, n); |
| 1041 | } |
| 1042 | /* |
| 1043 | * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN |
| 1044 | * filesystem). XXX also use for loopback NFS mounts. |
| 1045 | */ |
| 1046 | if (ioflag & IO_NOWDRAIN) |
| 1047 | bp->b_flags |= B_NOWDRAIN; |
| 1048 | |
| 1049 | /* |
| 1050 | * If the lease is non-cachable or IO_SYNC do bwrite(). |
| 1051 | * |
| 1052 | * IO_INVAL appears to be unused. The idea appears to be |
| 1053 | * to turn off caching in this case. Very odd. XXX |
| 1054 | */ |
| 1055 | if ((np->n_flag & NDONTCACHE) || (ioflag & IO_SYNC)) { |
| 1056 | if (ioflag & IO_INVAL) |
| 1057 | bp->b_flags |= B_NOCACHE; |
| 1058 | error = bwrite(bp); |
| 1059 | if (error) |
| 1060 | break; |
| 1061 | if (np->n_flag & NDONTCACHE) { |
| 1062 | error = nfs_vinvalbuf(vp, V_SAVE, 1); |
| 1063 | if (error) |
| 1064 | break; |
| 1065 | } |
| 1066 | } else if ((n + on) == biosize) { |
| 1067 | bp->b_flags |= B_ASYNC; |
| 1068 | bwrite(bp); |
| 1069 | } else { |
| 1070 | bdwrite(bp); |
| 1071 | } |
| 1072 | } while (uio->uio_resid > 0 && n > 0); |
| 1073 | |
| 1074 | if (haverslock) |
| 1075 | nfs_rsunlock(np); |
| 1076 | |
| 1077 | return (error); |
| 1078 | } |
| 1079 | |
| 1080 | /* |
| 1081 | * Get an nfs cache block. |
| 1082 | * |
| 1083 | * Allocate a new one if the block isn't currently in the cache |
| 1084 | * and return the block marked busy. If the calling process is |
| 1085 | * interrupted by a signal for an interruptible mount point, return |
| 1086 | * NULL. |
| 1087 | * |
| 1088 | * The caller must carefully deal with the possible B_INVAL state of |
| 1089 | * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it |
| 1090 | * indirectly), so synchronous reads can be issued without worrying about |
| 1091 | * the B_INVAL state. We have to be a little more careful when dealing |
| 1092 | * with writes (see comments in nfs_write()) when extending a file past |
| 1093 | * its EOF. |
| 1094 | */ |
| 1095 | static struct buf * |
| 1096 | nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td) |
| 1097 | { |
| 1098 | struct buf *bp; |
| 1099 | struct mount *mp; |
| 1100 | struct nfsmount *nmp; |
| 1101 | |
| 1102 | mp = vp->v_mount; |
| 1103 | nmp = VFSTONFS(mp); |
| 1104 | |
| 1105 | if (nmp->nm_flag & NFSMNT_INT) { |
| 1106 | bp = getblk(vp, loffset, size, GETBLK_PCATCH, 0); |
| 1107 | while (bp == NULL) { |
| 1108 | if (nfs_sigintr(nmp, (struct nfsreq *)0, td)) |
| 1109 | return (NULL); |
| 1110 | bp = getblk(vp, loffset, size, 0, 2 * hz); |
| 1111 | } |
| 1112 | } else { |
| 1113 | bp = getblk(vp, loffset, size, 0, 0); |
| 1114 | } |
| 1115 | |
| 1116 | /* |
| 1117 | * bio2, the 'device' layer. Since BIOs use 64 bit byte offsets |
| 1118 | * now, no translation is necessary. |
| 1119 | */ |
| 1120 | bp->b_bio2.bio_offset = loffset; |
| 1121 | return (bp); |
| 1122 | } |
| 1123 | |
| 1124 | /* |
| 1125 | * Flush and invalidate all dirty buffers. If another process is already |
| 1126 | * doing the flush, just wait for completion. |
| 1127 | */ |
| 1128 | int |
| 1129 | nfs_vinvalbuf(struct vnode *vp, int flags, int intrflg) |
| 1130 | { |
| 1131 | struct nfsnode *np = VTONFS(vp); |
| 1132 | struct nfsmount *nmp = VFSTONFS(vp->v_mount); |
| 1133 | int error = 0, slpflag, slptimeo; |
| 1134 | thread_t td = curthread; |
| 1135 | |
| 1136 | if (vp->v_flag & VRECLAIMED) |
| 1137 | return (0); |
| 1138 | |
| 1139 | if ((nmp->nm_flag & NFSMNT_INT) == 0) |
| 1140 | intrflg = 0; |
| 1141 | if (intrflg) { |
| 1142 | slpflag = PCATCH; |
| 1143 | slptimeo = 2 * hz; |
| 1144 | } else { |
| 1145 | slpflag = 0; |
| 1146 | slptimeo = 0; |
| 1147 | } |
| 1148 | /* |
| 1149 | * First wait for any other process doing a flush to complete. |
| 1150 | */ |
| 1151 | while (np->n_flag & NFLUSHINPROG) { |
| 1152 | np->n_flag |= NFLUSHWANT; |
| 1153 | error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo); |
| 1154 | if (error && intrflg && nfs_sigintr(nmp, NULL, td)) |
| 1155 | return (EINTR); |
| 1156 | } |
| 1157 | |
| 1158 | /* |
| 1159 | * Now, flush as required. |
| 1160 | */ |
| 1161 | np->n_flag |= NFLUSHINPROG; |
| 1162 | error = vinvalbuf(vp, flags, slpflag, 0); |
| 1163 | while (error) { |
| 1164 | if (intrflg && nfs_sigintr(nmp, NULL, td)) { |
| 1165 | np->n_flag &= ~NFLUSHINPROG; |
| 1166 | if (np->n_flag & NFLUSHWANT) { |
| 1167 | np->n_flag &= ~NFLUSHWANT; |
| 1168 | wakeup((caddr_t)&np->n_flag); |
| 1169 | } |
| 1170 | return (EINTR); |
| 1171 | } |
| 1172 | error = vinvalbuf(vp, flags, 0, slptimeo); |
| 1173 | } |
| 1174 | np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG); |
| 1175 | if (np->n_flag & NFLUSHWANT) { |
| 1176 | np->n_flag &= ~NFLUSHWANT; |
| 1177 | wakeup((caddr_t)&np->n_flag); |
| 1178 | } |
| 1179 | return (0); |
| 1180 | } |
| 1181 | |
| 1182 | /* |
| 1183 | * Initiate asynchronous I/O. Return an error if no nfsiods are available. |
| 1184 | * This is mainly to avoid queueing async I/O requests when the nfsiods |
| 1185 | * are all hung on a dead server. |
| 1186 | * |
| 1187 | * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp |
| 1188 | * is eventually dequeued by the async daemon, nfs_doio() *will*. |
| 1189 | */ |
| 1190 | int |
| 1191 | nfs_asyncio(struct vnode *vp, struct bio *bio, struct thread *td) |
| 1192 | { |
| 1193 | struct buf *bp = bio->bio_buf; |
| 1194 | struct nfsmount *nmp; |
| 1195 | int i; |
| 1196 | int gotiod; |
| 1197 | int slpflag = 0; |
| 1198 | int slptimeo = 0; |
| 1199 | int error; |
| 1200 | |
| 1201 | /* |
| 1202 | * If no async daemons then return EIO to force caller to run the rpc |
| 1203 | * synchronously. |
| 1204 | */ |
| 1205 | if (nfs_numasync == 0) |
| 1206 | return (EIO); |
| 1207 | |
| 1208 | KKASSERT(vp->v_tag == VT_NFS); |
| 1209 | nmp = VFSTONFS(vp->v_mount); |
| 1210 | |
| 1211 | /* |
| 1212 | * Commits are usually short and sweet so lets save some cpu and |
| 1213 | * leave the async daemons for more important rpc's (such as reads |
| 1214 | * and writes). |
| 1215 | */ |
| 1216 | if (bp->b_cmd == BUF_CMD_WRITE && (bp->b_flags & B_NEEDCOMMIT) && |
| 1217 | (nmp->nm_bioqiods > nfs_numasync / 2)) { |
| 1218 | return(EIO); |
| 1219 | } |
| 1220 | |
| 1221 | again: |
| 1222 | if (nmp->nm_flag & NFSMNT_INT) |
| 1223 | slpflag = PCATCH; |
| 1224 | gotiod = FALSE; |
| 1225 | |
| 1226 | /* |
| 1227 | * Find a free iod to process this request. |
| 1228 | */ |
| 1229 | for (i = 0; i < NFS_MAXASYNCDAEMON; i++) |
| 1230 | if (nfs_iodwant[i]) { |
| 1231 | /* |
| 1232 | * Found one, so wake it up and tell it which |
| 1233 | * mount to process. |
| 1234 | */ |
| 1235 | NFS_DPF(ASYNCIO, |
| 1236 | ("nfs_asyncio: waking iod %d for mount %p\n", |
| 1237 | i, nmp)); |
| 1238 | nfs_iodwant[i] = NULL; |
| 1239 | nfs_iodmount[i] = nmp; |
| 1240 | nmp->nm_bioqiods++; |
| 1241 | wakeup((caddr_t)&nfs_iodwant[i]); |
| 1242 | gotiod = TRUE; |
| 1243 | break; |
| 1244 | } |
| 1245 | |
| 1246 | /* |
| 1247 | * If none are free, we may already have an iod working on this mount |
| 1248 | * point. If so, it will process our request. |
| 1249 | */ |
| 1250 | if (!gotiod) { |
| 1251 | if (nmp->nm_bioqiods > 0) { |
| 1252 | NFS_DPF(ASYNCIO, |
| 1253 | ("nfs_asyncio: %d iods are already processing mount %p\n", |
| 1254 | nmp->nm_bioqiods, nmp)); |
| 1255 | gotiod = TRUE; |
| 1256 | } |
| 1257 | } |
| 1258 | |
| 1259 | /* |
| 1260 | * If we have an iod which can process the request, then queue |
| 1261 | * the buffer. |
| 1262 | */ |
| 1263 | if (gotiod) { |
| 1264 | /* |
| 1265 | * Ensure that the queue never grows too large. We still want |
| 1266 | * to asynchronize so we block rather then return EIO. |
| 1267 | */ |
| 1268 | while (nmp->nm_bioqlen >= 2*nfs_numasync) { |
| 1269 | NFS_DPF(ASYNCIO, |
| 1270 | ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); |
| 1271 | nmp->nm_bioqwant = TRUE; |
| 1272 | error = tsleep(&nmp->nm_bioq, slpflag, |
| 1273 | "nfsaio", slptimeo); |
| 1274 | if (error) { |
| 1275 | if (nfs_sigintr(nmp, NULL, td)) |
| 1276 | return (EINTR); |
| 1277 | if (slpflag == PCATCH) { |
| 1278 | slpflag = 0; |
| 1279 | slptimeo = 2 * hz; |
| 1280 | } |
| 1281 | } |
| 1282 | /* |
| 1283 | * We might have lost our iod while sleeping, |
| 1284 | * so check and loop if nescessary. |
| 1285 | */ |
| 1286 | if (nmp->nm_bioqiods == 0) { |
| 1287 | NFS_DPF(ASYNCIO, |
| 1288 | ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); |
| 1289 | goto again; |
| 1290 | } |
| 1291 | } |
| 1292 | BUF_KERNPROC(bp); |
| 1293 | |
| 1294 | /* |
| 1295 | * The passed bio's buffer is not necessary associated with |
| 1296 | * the NFS vnode it is being written to. Store the NFS vnode |
| 1297 | * in the BIO driver info. |
| 1298 | */ |
| 1299 | bio->bio_driver_info = vp; |
| 1300 | TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act); |
| 1301 | nmp->nm_bioqlen++; |
| 1302 | return (0); |
| 1303 | } |
| 1304 | |
| 1305 | /* |
| 1306 | * All the iods are busy on other mounts, so return EIO to |
| 1307 | * force the caller to process the i/o synchronously. |
| 1308 | */ |
| 1309 | NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); |
| 1310 | return (EIO); |
| 1311 | } |
| 1312 | |
| 1313 | /* |
| 1314 | * Do an I/O operation to/from a cache block. This may be called |
| 1315 | * synchronously or from an nfsiod. The BIO is normalized for DEV_BSIZE. |
| 1316 | * |
| 1317 | * NOTE! TD MIGHT BE NULL |
| 1318 | */ |
| 1319 | int |
| 1320 | nfs_doio(struct vnode *vp, struct bio *bio, struct thread *td) |
| 1321 | { |
| 1322 | struct buf *bp = bio->bio_buf; |
| 1323 | struct uio *uiop; |
| 1324 | struct nfsnode *np; |
| 1325 | struct nfsmount *nmp; |
| 1326 | int error = 0, iomode, must_commit = 0; |
| 1327 | struct uio uio; |
| 1328 | struct iovec io; |
| 1329 | |
| 1330 | KKASSERT(vp->v_tag == VT_NFS); |
| 1331 | np = VTONFS(vp); |
| 1332 | nmp = VFSTONFS(vp->v_mount); |
| 1333 | uiop = &uio; |
| 1334 | uiop->uio_iov = &io; |
| 1335 | uiop->uio_iovcnt = 1; |
| 1336 | uiop->uio_segflg = UIO_SYSSPACE; |
| 1337 | uiop->uio_td = td; |
| 1338 | |
| 1339 | /* |
| 1340 | * clear B_ERROR and B_INVAL state prior to initiating the I/O. We |
| 1341 | * do this here so we do not have to do it in all the code that |
| 1342 | * calls us. |
| 1343 | */ |
| 1344 | bp->b_flags &= ~(B_ERROR | B_INVAL); |
| 1345 | |
| 1346 | |
| 1347 | KASSERT(bp->b_cmd != BUF_CMD_DONE, |
| 1348 | ("nfs_doio: bp %p already marked done!", bp)); |
| 1349 | |
| 1350 | if (bp->b_cmd == BUF_CMD_READ) { |
| 1351 | io.iov_len = uiop->uio_resid = bp->b_bcount; |
| 1352 | io.iov_base = bp->b_data; |
| 1353 | uiop->uio_rw = UIO_READ; |
| 1354 | |
| 1355 | switch (vp->v_type) { |
| 1356 | case VREG: |
| 1357 | uiop->uio_offset = bio->bio_offset; |
| 1358 | nfsstats.read_bios++; |
| 1359 | error = nfs_readrpc(vp, uiop); |
| 1360 | |
| 1361 | if (!error) { |
| 1362 | if (uiop->uio_resid) { |
| 1363 | /* |
| 1364 | * If we had a short read with no error, we must have |
| 1365 | * hit a file hole. We should zero-fill the remainder. |
| 1366 | * This can also occur if the server hits the file EOF. |
| 1367 | * |
| 1368 | * Holes used to be able to occur due to pending |
| 1369 | * writes, but that is not possible any longer. |
| 1370 | */ |
| 1371 | int nread = bp->b_bcount - uiop->uio_resid; |
| 1372 | int left = uiop->uio_resid; |
| 1373 | |
| 1374 | if (left > 0) |
| 1375 | bzero((char *)bp->b_data + nread, left); |
| 1376 | uiop->uio_resid = 0; |
| 1377 | } |
| 1378 | } |
| 1379 | if (td && td->td_proc && (vp->v_flag & VTEXT) && |
| 1380 | np->n_mtime != np->n_vattr.va_mtime.tv_sec) { |
| 1381 | uprintf("Process killed due to text file modification\n"); |
| 1382 | ksignal(td->td_proc, SIGKILL); |
| 1383 | } |
| 1384 | break; |
| 1385 | case VLNK: |
| 1386 | uiop->uio_offset = 0; |
| 1387 | nfsstats.readlink_bios++; |
| 1388 | error = nfs_readlinkrpc(vp, uiop); |
| 1389 | break; |
| 1390 | case VDIR: |
| 1391 | nfsstats.readdir_bios++; |
| 1392 | uiop->uio_offset = bio->bio_offset; |
| 1393 | if (nmp->nm_flag & NFSMNT_RDIRPLUS) { |
| 1394 | error = nfs_readdirplusrpc(vp, uiop); |
| 1395 | if (error == NFSERR_NOTSUPP) |
| 1396 | nmp->nm_flag &= ~NFSMNT_RDIRPLUS; |
| 1397 | } |
| 1398 | if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) |
| 1399 | error = nfs_readdirrpc(vp, uiop); |
| 1400 | /* |
| 1401 | * end-of-directory sets B_INVAL but does not generate an |
| 1402 | * error. |
| 1403 | */ |
| 1404 | if (error == 0 && uiop->uio_resid == bp->b_bcount) |
| 1405 | bp->b_flags |= B_INVAL; |
| 1406 | break; |
| 1407 | default: |
| 1408 | kprintf("nfs_doio: type %x unexpected\n",vp->v_type); |
| 1409 | break; |
| 1410 | }; |
| 1411 | if (error) { |
| 1412 | bp->b_flags |= B_ERROR; |
| 1413 | bp->b_error = error; |
| 1414 | } |
| 1415 | } else { |
| 1416 | /* |
| 1417 | * If we only need to commit, try to commit |
| 1418 | */ |
| 1419 | KKASSERT(bp->b_cmd == BUF_CMD_WRITE); |
| 1420 | if (bp->b_flags & B_NEEDCOMMIT) { |
| 1421 | int retv; |
| 1422 | off_t off; |
| 1423 | |
| 1424 | off = bio->bio_offset + bp->b_dirtyoff; |
| 1425 | retv = nfs_commit(vp, off, |
| 1426 | bp->b_dirtyend - bp->b_dirtyoff, td); |
| 1427 | if (retv == 0) { |
| 1428 | bp->b_dirtyoff = bp->b_dirtyend = 0; |
| 1429 | bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); |
| 1430 | bp->b_resid = 0; |
| 1431 | biodone(bio); |
| 1432 | return (0); |
| 1433 | } |
| 1434 | if (retv == NFSERR_STALEWRITEVERF) { |
| 1435 | nfs_clearcommit(vp->v_mount); |
| 1436 | } |
| 1437 | } |
| 1438 | |
| 1439 | /* |
| 1440 | * Setup for actual write |
| 1441 | */ |
| 1442 | |
| 1443 | if (bio->bio_offset + bp->b_dirtyend > np->n_size) |
| 1444 | bp->b_dirtyend = np->n_size - bio->bio_offset; |
| 1445 | |
| 1446 | if (bp->b_dirtyend > bp->b_dirtyoff) { |
| 1447 | io.iov_len = uiop->uio_resid = bp->b_dirtyend |
| 1448 | - bp->b_dirtyoff; |
| 1449 | uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff; |
| 1450 | io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; |
| 1451 | uiop->uio_rw = UIO_WRITE; |
| 1452 | nfsstats.write_bios++; |
| 1453 | |
| 1454 | if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) |
| 1455 | iomode = NFSV3WRITE_UNSTABLE; |
| 1456 | else |
| 1457 | iomode = NFSV3WRITE_FILESYNC; |
| 1458 | |
| 1459 | error = nfs_writerpc(vp, uiop, &iomode, &must_commit); |
| 1460 | |
| 1461 | /* |
| 1462 | * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try |
| 1463 | * to cluster the buffers needing commit. This will allow |
| 1464 | * the system to submit a single commit rpc for the whole |
| 1465 | * cluster. We can do this even if the buffer is not 100% |
| 1466 | * dirty (relative to the NFS blocksize), so we optimize the |
| 1467 | * append-to-file-case. |
| 1468 | * |
| 1469 | * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be |
| 1470 | * cleared because write clustering only works for commit |
| 1471 | * rpc's, not for the data portion of the write). |
| 1472 | */ |
| 1473 | |
| 1474 | if (!error && iomode == NFSV3WRITE_UNSTABLE) { |
| 1475 | bp->b_flags |= B_NEEDCOMMIT; |
| 1476 | if (bp->b_dirtyoff == 0 |
| 1477 | && bp->b_dirtyend == bp->b_bcount) |
| 1478 | bp->b_flags |= B_CLUSTEROK; |
| 1479 | } else { |
| 1480 | bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); |
| 1481 | } |
| 1482 | |
| 1483 | /* |
| 1484 | * For an interrupted write, the buffer is still valid |
| 1485 | * and the write hasn't been pushed to the server yet, |
| 1486 | * so we can't set B_ERROR and report the interruption |
| 1487 | * by setting B_EINTR. For the B_ASYNC case, B_EINTR |
| 1488 | * is not relevant, so the rpc attempt is essentially |
| 1489 | * a noop. For the case of a V3 write rpc not being |
| 1490 | * committed to stable storage, the block is still |
| 1491 | * dirty and requires either a commit rpc or another |
| 1492 | * write rpc with iomode == NFSV3WRITE_FILESYNC before |
| 1493 | * the block is reused. This is indicated by setting |
| 1494 | * the B_DELWRI and B_NEEDCOMMIT flags. |
| 1495 | * |
| 1496 | * If the buffer is marked B_PAGING, it does not reside on |
| 1497 | * the vp's paging queues so we cannot call bdirty(). The |
| 1498 | * bp in this case is not an NFS cache block so we should |
| 1499 | * be safe. XXX |
| 1500 | */ |
| 1501 | if (error == EINTR |
| 1502 | || (!error && (bp->b_flags & B_NEEDCOMMIT))) { |
| 1503 | crit_enter(); |
| 1504 | bp->b_flags &= ~(B_INVAL|B_NOCACHE); |
| 1505 | if ((bp->b_flags & B_PAGING) == 0) |
| 1506 | bdirty(bp); |
| 1507 | if (error && (bp->b_flags & B_ASYNC) == 0) |
| 1508 | bp->b_flags |= B_EINTR; |
| 1509 | crit_exit(); |
| 1510 | } else { |
| 1511 | if (error) { |
| 1512 | bp->b_flags |= B_ERROR; |
| 1513 | bp->b_error = np->n_error = error; |
| 1514 | np->n_flag |= NWRITEERR; |
| 1515 | } |
| 1516 | bp->b_dirtyoff = bp->b_dirtyend = 0; |
| 1517 | } |
| 1518 | } else { |
| 1519 | bp->b_resid = 0; |
| 1520 | biodone(bio); |
| 1521 | return (0); |
| 1522 | } |
| 1523 | } |
| 1524 | bp->b_resid = uiop->uio_resid; |
| 1525 | if (must_commit) |
| 1526 | nfs_clearcommit(vp->v_mount); |
| 1527 | biodone(bio); |
| 1528 | return (error); |
| 1529 | } |
| 1530 | |
| 1531 | /* |
| 1532 | * Used to aid in handling ftruncate() operations on the NFS client side. |
| 1533 | * Truncation creates a number of special problems for NFS. We have to |
| 1534 | * throw away VM pages and buffer cache buffers that are beyond EOF, and |
| 1535 | * we have to properly handle VM pages or (potentially dirty) buffers |
| 1536 | * that straddle the truncation point. |
| 1537 | */ |
| 1538 | |
| 1539 | int |
| 1540 | nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize) |
| 1541 | { |
| 1542 | struct nfsnode *np = VTONFS(vp); |
| 1543 | u_quad_t tsize = np->n_size; |
| 1544 | int biosize = vp->v_mount->mnt_stat.f_iosize; |
| 1545 | int error = 0; |
| 1546 | |
| 1547 | np->n_size = nsize; |
| 1548 | |
| 1549 | if (np->n_size < tsize) { |
| 1550 | struct buf *bp; |
| 1551 | daddr_t lbn; |
| 1552 | off_t loffset; |
| 1553 | int bufsize; |
| 1554 | |
| 1555 | /* |
| 1556 | * vtruncbuf() doesn't get the buffer overlapping the |
| 1557 | * truncation point. We may have a B_DELWRI and/or B_CACHE |
| 1558 | * buffer that now needs to be truncated. |
| 1559 | */ |
| 1560 | error = vtruncbuf(vp, nsize, biosize); |
| 1561 | lbn = nsize / biosize; |
| 1562 | bufsize = nsize & (biosize - 1); |
| 1563 | loffset = nsize - bufsize; |
| 1564 | bp = nfs_getcacheblk(vp, loffset, bufsize, td); |
| 1565 | if (bp->b_dirtyoff > bp->b_bcount) |
| 1566 | bp->b_dirtyoff = bp->b_bcount; |
| 1567 | if (bp->b_dirtyend > bp->b_bcount) |
| 1568 | bp->b_dirtyend = bp->b_bcount; |
| 1569 | bp->b_flags |= B_RELBUF; /* don't leave garbage around */ |
| 1570 | brelse(bp); |
| 1571 | } else { |
| 1572 | vnode_pager_setsize(vp, nsize); |
| 1573 | } |
| 1574 | return(error); |
| 1575 | } |
| 1576 | |