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 * 4. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 33 */ 34 35 #include <sys/cdefs.h> 36 __FBSDID("$FreeBSD$"); 37 38 #include <sys/param.h> 39 #include <sys/systm.h> 40 #include <sys/bio.h> 41 #include <sys/buf.h> 42 #include <sys/kernel.h> 43 #include <sys/mount.h> 44 #include <sys/vmmeter.h> 45 #include <sys/vnode.h> 46 47 #include <vm/vm.h> 48 #include <vm/vm_extern.h> 49 #include <vm/vm_page.h> 50 #include <vm/vm_object.h> 51 #include <vm/vm_pager.h> 52 #include <vm/vnode_pager.h> 53 54 #include <fs/nfs/nfsport.h> 55 #include <fs/nfsclient/nfsmount.h> 56 #include <fs/nfsclient/nfs.h> 57 #include <fs/nfsclient/nfsnode.h> 58 59 extern int newnfs_directio_allow_mmap; 60 extern struct nfsstats newnfsstats; 61 extern struct mtx ncl_iod_mutex; 62 extern int ncl_numasync; 63 extern enum nfsiod_state ncl_iodwant[NFS_MAXRAHEAD]; 64 extern struct nfsmount *ncl_iodmount[NFS_MAXRAHEAD]; 65 extern int newnfs_directio_enable; 66 67 int ncl_pbuf_freecnt = -1; /* start out unlimited */ 68 69 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 70 struct thread *td); 71 static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 72 struct ucred *cred, int ioflag); 73 74 /* 75 * Vnode op for VM getpages. 76 */ 77 int 78 ncl_getpages(struct vop_getpages_args *ap) 79 { 80 int i, error, nextoff, size, toff, count, npages; 81 struct uio uio; 82 struct iovec iov; 83 vm_offset_t kva; 84 struct buf *bp; 85 struct vnode *vp; 86 struct thread *td; 87 struct ucred *cred; 88 struct nfsmount *nmp; 89 vm_object_t object; 90 vm_page_t *pages; 91 struct nfsnode *np; 92 93 vp = ap->a_vp; 94 np = VTONFS(vp); 95 td = curthread; /* XXX */ 96 cred = curthread->td_ucred; /* XXX */ 97 nmp = VFSTONFS(vp->v_mount); 98 pages = ap->a_m; 99 count = ap->a_count; 100 101 if ((object = vp->v_object) == NULL) { 102 ncl_printf("nfs_getpages: called with non-merged cache vnode??\n"); 103 return (VM_PAGER_ERROR); 104 } 105 106 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) { 107 mtx_lock(&np->n_mtx); 108 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 109 mtx_unlock(&np->n_mtx); 110 ncl_printf("nfs_getpages: called on non-cacheable vnode??\n"); 111 return (VM_PAGER_ERROR); 112 } else 113 mtx_unlock(&np->n_mtx); 114 } 115 116 mtx_lock(&nmp->nm_mtx); 117 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 118 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 119 mtx_unlock(&nmp->nm_mtx); 120 /* We'll never get here for v4, because we always have fsinfo */ 121 (void)ncl_fsinfo(nmp, vp, cred, td); 122 } else 123 mtx_unlock(&nmp->nm_mtx); 124 125 npages = btoc(count); 126 127 /* 128 * If the requested page is partially valid, just return it and 129 * allow the pager to zero-out the blanks. Partially valid pages 130 * can only occur at the file EOF. 131 */ 132 VM_OBJECT_LOCK(object); 133 if (pages[ap->a_reqpage]->valid != 0) { 134 for (i = 0; i < npages; ++i) { 135 if (i != ap->a_reqpage) { 136 vm_page_lock(pages[i]); 137 vm_page_free(pages[i]); 138 vm_page_unlock(pages[i]); 139 } 140 } 141 VM_OBJECT_UNLOCK(object); 142 return (0); 143 } 144 VM_OBJECT_UNLOCK(object); 145 146 /* 147 * We use only the kva address for the buffer, but this is extremely 148 * convienient and fast. 149 */ 150 bp = getpbuf(&ncl_pbuf_freecnt); 151 152 kva = (vm_offset_t) bp->b_data; 153 pmap_qenter(kva, pages, npages); 154 PCPU_INC(cnt.v_vnodein); 155 PCPU_ADD(cnt.v_vnodepgsin, npages); 156 157 iov.iov_base = (caddr_t) 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 = ncl_readrpc(vp, &uio, cred); 168 pmap_qremove(kva, npages); 169 170 relpbuf(bp, &ncl_pbuf_freecnt); 171 172 if (error && (uio.uio_resid == count)) { 173 ncl_printf("nfs_getpages: error %d\n", error); 174 VM_OBJECT_LOCK(object); 175 for (i = 0; i < npages; ++i) { 176 if (i != ap->a_reqpage) { 177 vm_page_lock(pages[i]); 178 vm_page_free(pages[i]); 179 vm_page_unlock(pages[i]); 180 } 181 } 182 VM_OBJECT_UNLOCK(object); 183 return (VM_PAGER_ERROR); 184 } 185 186 /* 187 * Calculate the number of bytes read and validate only that number 188 * of bytes. Note that due to pending writes, size may be 0. This 189 * does not mean that the remaining data is invalid! 190 */ 191 192 size = count - uio.uio_resid; 193 VM_OBJECT_LOCK(object); 194 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 195 vm_page_t m; 196 nextoff = toff + PAGE_SIZE; 197 m = pages[i]; 198 199 if (nextoff <= size) { 200 /* 201 * Read operation filled an entire page 202 */ 203 m->valid = VM_PAGE_BITS_ALL; 204 KASSERT(m->dirty == 0, 205 ("nfs_getpages: page %p is dirty", m)); 206 } else if (size > toff) { 207 /* 208 * Read operation filled a partial page. 209 */ 210 m->valid = 0; 211 vm_page_set_valid(m, 0, size - toff); 212 KASSERT(m->dirty == 0, 213 ("nfs_getpages: page %p is dirty", m)); 214 } else { 215 /* 216 * Read operation was short. If no error occured 217 * we may have hit a zero-fill section. We simply 218 * leave valid set to 0. 219 */ 220 ; 221 } 222 if (i != ap->a_reqpage) { 223 /* 224 * Whether or not to leave the page activated is up in 225 * the air, but we should put the page on a page queue 226 * somewhere (it already is in the object). Result: 227 * It appears that emperical results show that 228 * deactivating pages is best. 229 */ 230 231 /* 232 * Just in case someone was asking for this page we 233 * now tell them that it is ok to use. 234 */ 235 if (!error) { 236 if (m->oflags & VPO_WANTED) { 237 vm_page_lock(m); 238 vm_page_activate(m); 239 vm_page_unlock(m); 240 } else { 241 vm_page_lock(m); 242 vm_page_deactivate(m); 243 vm_page_unlock(m); 244 } 245 vm_page_wakeup(m); 246 } else { 247 vm_page_lock(m); 248 vm_page_free(m); 249 vm_page_unlock(m); 250 } 251 } 252 } 253 VM_OBJECT_UNLOCK(object); 254 return (0); 255 } 256 257 /* 258 * Vnode op for VM putpages. 259 */ 260 int 261 ncl_putpages(struct vop_putpages_args *ap) 262 { 263 struct uio uio; 264 struct iovec iov; 265 vm_offset_t kva; 266 struct buf *bp; 267 int iomode, must_commit, i, error, npages, count; 268 off_t offset; 269 int *rtvals; 270 struct vnode *vp; 271 struct thread *td; 272 struct ucred *cred; 273 struct nfsmount *nmp; 274 struct nfsnode *np; 275 vm_page_t *pages; 276 277 vp = ap->a_vp; 278 np = VTONFS(vp); 279 td = curthread; /* XXX */ 280 cred = curthread->td_ucred; /* XXX */ 281 nmp = VFSTONFS(vp->v_mount); 282 pages = ap->a_m; 283 count = ap->a_count; 284 rtvals = ap->a_rtvals; 285 npages = btoc(count); 286 offset = IDX_TO_OFF(pages[0]->pindex); 287 288 mtx_lock(&nmp->nm_mtx); 289 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 290 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 291 mtx_unlock(&nmp->nm_mtx); 292 (void)ncl_fsinfo(nmp, vp, cred, td); 293 } else 294 mtx_unlock(&nmp->nm_mtx); 295 296 mtx_lock(&np->n_mtx); 297 if (newnfs_directio_enable && !newnfs_directio_allow_mmap && 298 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 299 mtx_unlock(&np->n_mtx); 300 ncl_printf("ncl_putpages: called on noncache-able vnode??\n"); 301 mtx_lock(&np->n_mtx); 302 } 303 304 for (i = 0; i < npages; i++) 305 rtvals[i] = VM_PAGER_AGAIN; 306 307 /* 308 * When putting pages, do not extend file past EOF. 309 */ 310 if (offset + count > np->n_size) { 311 count = np->n_size - offset; 312 if (count < 0) 313 count = 0; 314 } 315 mtx_unlock(&np->n_mtx); 316 317 /* 318 * We use only the kva address for the buffer, but this is extremely 319 * convienient and fast. 320 */ 321 bp = getpbuf(&ncl_pbuf_freecnt); 322 323 kva = (vm_offset_t) bp->b_data; 324 pmap_qenter(kva, pages, npages); 325 PCPU_INC(cnt.v_vnodeout); 326 PCPU_ADD(cnt.v_vnodepgsout, count); 327 328 iov.iov_base = (caddr_t) kva; 329 iov.iov_len = count; 330 uio.uio_iov = &iov; 331 uio.uio_iovcnt = 1; 332 uio.uio_offset = offset; 333 uio.uio_resid = count; 334 uio.uio_segflg = UIO_SYSSPACE; 335 uio.uio_rw = UIO_WRITE; 336 uio.uio_td = td; 337 338 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 339 iomode = NFSWRITE_UNSTABLE; 340 else 341 iomode = NFSWRITE_FILESYNC; 342 343 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0); 344 345 pmap_qremove(kva, npages); 346 relpbuf(bp, &ncl_pbuf_freecnt); 347 348 if (!error) { 349 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 350 for (i = 0; i < nwritten; i++) { 351 rtvals[i] = VM_PAGER_OK; 352 vm_page_undirty(pages[i]); 353 } 354 if (must_commit) { 355 ncl_clearcommit(vp->v_mount); 356 } 357 } 358 return rtvals[0]; 359 } 360 361 /* 362 * For nfs, cache consistency can only be maintained approximately. 363 * Although RFC1094 does not specify the criteria, the following is 364 * believed to be compatible with the reference port. 365 * For nfs: 366 * If the file's modify time on the server has changed since the 367 * last read rpc or you have written to the file, 368 * you may have lost data cache consistency with the 369 * server, so flush all of the file's data out of the cache. 370 * Then force a getattr rpc to ensure that you have up to date 371 * attributes. 372 * NB: This implies that cache data can be read when up to 373 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 374 * attributes this could be forced by setting n_attrstamp to 0 before 375 * the VOP_GETATTR() call. 376 */ 377 static inline int 378 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 379 { 380 int error = 0; 381 struct vattr vattr; 382 struct nfsnode *np = VTONFS(vp); 383 int old_lock; 384 385 /* 386 * Grab the exclusive lock before checking whether the cache is 387 * consistent. 388 * XXX - We can make this cheaper later (by acquiring cheaper locks). 389 * But for now, this suffices. 390 */ 391 old_lock = ncl_upgrade_vnlock(vp); 392 if (vp->v_iflag & VI_DOOMED) { 393 ncl_downgrade_vnlock(vp, old_lock); 394 return (EBADF); 395 } 396 397 mtx_lock(&np->n_mtx); 398 if (np->n_flag & NMODIFIED) { 399 mtx_unlock(&np->n_mtx); 400 if (vp->v_type != VREG) { 401 if (vp->v_type != VDIR) 402 panic("nfs: bioread, not dir"); 403 ncl_invaldir(vp); 404 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 405 if (error) 406 goto out; 407 } 408 np->n_attrstamp = 0; 409 error = VOP_GETATTR(vp, &vattr, cred); 410 if (error) 411 goto out; 412 mtx_lock(&np->n_mtx); 413 np->n_mtime = vattr.va_mtime; 414 mtx_unlock(&np->n_mtx); 415 } else { 416 mtx_unlock(&np->n_mtx); 417 error = VOP_GETATTR(vp, &vattr, cred); 418 if (error) 419 return (error); 420 mtx_lock(&np->n_mtx); 421 if ((np->n_flag & NSIZECHANGED) 422 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 423 mtx_unlock(&np->n_mtx); 424 if (vp->v_type == VDIR) 425 ncl_invaldir(vp); 426 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 427 if (error) 428 goto out; 429 mtx_lock(&np->n_mtx); 430 np->n_mtime = vattr.va_mtime; 431 np->n_flag &= ~NSIZECHANGED; 432 } 433 mtx_unlock(&np->n_mtx); 434 } 435 out: 436 ncl_downgrade_vnlock(vp, old_lock); 437 return error; 438 } 439 440 /* 441 * Vnode op for read using bio 442 */ 443 int 444 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 445 { 446 struct nfsnode *np = VTONFS(vp); 447 int biosize, i; 448 struct buf *bp, *rabp; 449 struct thread *td; 450 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 451 daddr_t lbn, rabn; 452 int bcount; 453 int seqcount; 454 int nra, error = 0, n = 0, on = 0; 455 456 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode")); 457 if (uio->uio_resid == 0) 458 return (0); 459 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 460 return (EINVAL); 461 td = uio->uio_td; 462 463 mtx_lock(&nmp->nm_mtx); 464 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 465 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 466 mtx_unlock(&nmp->nm_mtx); 467 (void)ncl_fsinfo(nmp, vp, cred, td); 468 mtx_lock(&nmp->nm_mtx); 469 } 470 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0) 471 (void) newnfs_iosize(nmp); 472 mtx_unlock(&nmp->nm_mtx); 473 474 if (vp->v_type != VDIR && 475 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 476 return (EFBIG); 477 478 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 479 /* No caching/ no readaheads. Just read data into the user buffer */ 480 return ncl_readrpc(vp, uio, cred); 481 482 biosize = vp->v_mount->mnt_stat.f_iosize; 483 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 484 485 error = nfs_bioread_check_cons(vp, td, cred); 486 if (error) 487 return error; 488 489 do { 490 u_quad_t nsize; 491 492 mtx_lock(&np->n_mtx); 493 nsize = np->n_size; 494 mtx_unlock(&np->n_mtx); 495 496 switch (vp->v_type) { 497 case VREG: 498 NFSINCRGLOBAL(newnfsstats.biocache_reads); 499 lbn = uio->uio_offset / biosize; 500 on = uio->uio_offset & (biosize - 1); 501 502 /* 503 * Start the read ahead(s), as required. 504 */ 505 if (nmp->nm_readahead > 0) { 506 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 507 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 508 rabn = lbn + 1 + nra; 509 if (incore(&vp->v_bufobj, rabn) == NULL) { 510 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 511 if (!rabp) { 512 error = newnfs_sigintr(nmp, td); 513 return (error ? error : EINTR); 514 } 515 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 516 rabp->b_flags |= B_ASYNC; 517 rabp->b_iocmd = BIO_READ; 518 vfs_busy_pages(rabp, 0); 519 if (ncl_asyncio(nmp, rabp, cred, td)) { 520 rabp->b_flags |= B_INVAL; 521 rabp->b_ioflags |= BIO_ERROR; 522 vfs_unbusy_pages(rabp); 523 brelse(rabp); 524 break; 525 } 526 } else { 527 brelse(rabp); 528 } 529 } 530 } 531 } 532 533 /* Note that bcount is *not* DEV_BSIZE aligned. */ 534 bcount = biosize; 535 if ((off_t)lbn * biosize >= nsize) { 536 bcount = 0; 537 } else if ((off_t)(lbn + 1) * biosize > nsize) { 538 bcount = nsize - (off_t)lbn * biosize; 539 } 540 bp = nfs_getcacheblk(vp, lbn, bcount, td); 541 542 if (!bp) { 543 error = newnfs_sigintr(nmp, td); 544 return (error ? error : EINTR); 545 } 546 547 /* 548 * If B_CACHE is not set, we must issue the read. If this 549 * fails, we return an error. 550 */ 551 552 if ((bp->b_flags & B_CACHE) == 0) { 553 bp->b_iocmd = BIO_READ; 554 vfs_busy_pages(bp, 0); 555 error = ncl_doio(vp, bp, cred, td, 0); 556 if (error) { 557 brelse(bp); 558 return (error); 559 } 560 } 561 562 /* 563 * on is the offset into the current bp. Figure out how many 564 * bytes we can copy out of the bp. Note that bcount is 565 * NOT DEV_BSIZE aligned. 566 * 567 * Then figure out how many bytes we can copy into the uio. 568 */ 569 570 n = 0; 571 if (on < bcount) 572 n = min((unsigned)(bcount - on), uio->uio_resid); 573 break; 574 case VLNK: 575 NFSINCRGLOBAL(newnfsstats.biocache_readlinks); 576 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 577 if (!bp) { 578 error = newnfs_sigintr(nmp, td); 579 return (error ? error : EINTR); 580 } 581 if ((bp->b_flags & B_CACHE) == 0) { 582 bp->b_iocmd = BIO_READ; 583 vfs_busy_pages(bp, 0); 584 error = ncl_doio(vp, bp, cred, td, 0); 585 if (error) { 586 bp->b_ioflags |= BIO_ERROR; 587 brelse(bp); 588 return (error); 589 } 590 } 591 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 592 on = 0; 593 break; 594 case VDIR: 595 NFSINCRGLOBAL(newnfsstats.biocache_readdirs); 596 if (np->n_direofoffset 597 && uio->uio_offset >= np->n_direofoffset) { 598 return (0); 599 } 600 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 601 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 602 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 603 if (!bp) { 604 error = newnfs_sigintr(nmp, td); 605 return (error ? error : EINTR); 606 } 607 if ((bp->b_flags & B_CACHE) == 0) { 608 bp->b_iocmd = BIO_READ; 609 vfs_busy_pages(bp, 0); 610 error = ncl_doio(vp, bp, cred, td, 0); 611 if (error) { 612 brelse(bp); 613 } 614 while (error == NFSERR_BAD_COOKIE) { 615 ncl_invaldir(vp); 616 error = ncl_vinvalbuf(vp, 0, td, 1); 617 /* 618 * Yuck! The directory has been modified on the 619 * server. The only way to get the block is by 620 * reading from the beginning to get all the 621 * offset cookies. 622 * 623 * Leave the last bp intact unless there is an error. 624 * Loop back up to the while if the error is another 625 * NFSERR_BAD_COOKIE (double yuch!). 626 */ 627 for (i = 0; i <= lbn && !error; i++) { 628 if (np->n_direofoffset 629 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 630 return (0); 631 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 632 if (!bp) { 633 error = newnfs_sigintr(nmp, td); 634 return (error ? error : EINTR); 635 } 636 if ((bp->b_flags & B_CACHE) == 0) { 637 bp->b_iocmd = BIO_READ; 638 vfs_busy_pages(bp, 0); 639 error = ncl_doio(vp, bp, cred, td, 0); 640 /* 641 * no error + B_INVAL == directory EOF, 642 * use the block. 643 */ 644 if (error == 0 && (bp->b_flags & B_INVAL)) 645 break; 646 } 647 /* 648 * An error will throw away the block and the 649 * for loop will break out. If no error and this 650 * is not the block we want, we throw away the 651 * block and go for the next one via the for loop. 652 */ 653 if (error || i < lbn) 654 brelse(bp); 655 } 656 } 657 /* 658 * The above while is repeated if we hit another cookie 659 * error. If we hit an error and it wasn't a cookie error, 660 * we give up. 661 */ 662 if (error) 663 return (error); 664 } 665 666 /* 667 * If not eof and read aheads are enabled, start one. 668 * (You need the current block first, so that you have the 669 * directory offset cookie of the next block.) 670 */ 671 if (nmp->nm_readahead > 0 && 672 (bp->b_flags & B_INVAL) == 0 && 673 (np->n_direofoffset == 0 || 674 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 675 incore(&vp->v_bufobj, lbn + 1) == NULL) { 676 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 677 if (rabp) { 678 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 679 rabp->b_flags |= B_ASYNC; 680 rabp->b_iocmd = BIO_READ; 681 vfs_busy_pages(rabp, 0); 682 if (ncl_asyncio(nmp, rabp, cred, td)) { 683 rabp->b_flags |= B_INVAL; 684 rabp->b_ioflags |= BIO_ERROR; 685 vfs_unbusy_pages(rabp); 686 brelse(rabp); 687 } 688 } else { 689 brelse(rabp); 690 } 691 } 692 } 693 /* 694 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 695 * chopped for the EOF condition, we cannot tell how large 696 * NFS directories are going to be until we hit EOF. So 697 * an NFS directory buffer is *not* chopped to its EOF. Now, 698 * it just so happens that b_resid will effectively chop it 699 * to EOF. *BUT* this information is lost if the buffer goes 700 * away and is reconstituted into a B_CACHE state ( due to 701 * being VMIO ) later. So we keep track of the directory eof 702 * in np->n_direofoffset and chop it off as an extra step 703 * right here. 704 */ 705 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 706 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 707 n = np->n_direofoffset - uio->uio_offset; 708 break; 709 default: 710 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type); 711 bp = NULL; 712 break; 713 }; 714 715 if (n > 0) { 716 error = uiomove(bp->b_data + on, (int)n, uio); 717 } 718 if (vp->v_type == VLNK) 719 n = 0; 720 if (bp != NULL) 721 brelse(bp); 722 } while (error == 0 && uio->uio_resid > 0 && n > 0); 723 return (error); 724 } 725 726 /* 727 * The NFS write path cannot handle iovecs with len > 1. So we need to 728 * break up iovecs accordingly (restricting them to wsize). 729 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 730 * For the ASYNC case, 2 copies are needed. The first a copy from the 731 * user buffer to a staging buffer and then a second copy from the staging 732 * buffer to mbufs. This can be optimized by copying from the user buffer 733 * directly into mbufs and passing the chain down, but that requires a 734 * fair amount of re-working of the relevant codepaths (and can be done 735 * later). 736 */ 737 static int 738 nfs_directio_write(vp, uiop, cred, ioflag) 739 struct vnode *vp; 740 struct uio *uiop; 741 struct ucred *cred; 742 int ioflag; 743 { 744 int error; 745 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 746 struct thread *td = uiop->uio_td; 747 int size; 748 int wsize; 749 750 mtx_lock(&nmp->nm_mtx); 751 wsize = nmp->nm_wsize; 752 mtx_unlock(&nmp->nm_mtx); 753 if (ioflag & IO_SYNC) { 754 int iomode, must_commit; 755 struct uio uio; 756 struct iovec iov; 757 do_sync: 758 while (uiop->uio_resid > 0) { 759 size = min(uiop->uio_resid, wsize); 760 size = min(uiop->uio_iov->iov_len, size); 761 iov.iov_base = uiop->uio_iov->iov_base; 762 iov.iov_len = size; 763 uio.uio_iov = &iov; 764 uio.uio_iovcnt = 1; 765 uio.uio_offset = uiop->uio_offset; 766 uio.uio_resid = size; 767 uio.uio_segflg = UIO_USERSPACE; 768 uio.uio_rw = UIO_WRITE; 769 uio.uio_td = td; 770 iomode = NFSWRITE_FILESYNC; 771 error = ncl_writerpc(vp, &uio, cred, &iomode, 772 &must_commit, 0); 773 KASSERT((must_commit == 0), 774 ("ncl_directio_write: Did not commit write")); 775 if (error) 776 return (error); 777 uiop->uio_offset += size; 778 uiop->uio_resid -= size; 779 if (uiop->uio_iov->iov_len <= size) { 780 uiop->uio_iovcnt--; 781 uiop->uio_iov++; 782 } else { 783 uiop->uio_iov->iov_base = 784 (char *)uiop->uio_iov->iov_base + size; 785 uiop->uio_iov->iov_len -= size; 786 } 787 } 788 } else { 789 struct uio *t_uio; 790 struct iovec *t_iov; 791 struct buf *bp; 792 793 /* 794 * Break up the write into blocksize chunks and hand these 795 * over to nfsiod's for write back. 796 * Unfortunately, this incurs a copy of the data. Since 797 * the user could modify the buffer before the write is 798 * initiated. 799 * 800 * The obvious optimization here is that one of the 2 copies 801 * in the async write path can be eliminated by copying the 802 * data here directly into mbufs and passing the mbuf chain 803 * down. But that will require a fair amount of re-working 804 * of the code and can be done if there's enough interest 805 * in NFS directio access. 806 */ 807 while (uiop->uio_resid > 0) { 808 size = min(uiop->uio_resid, wsize); 809 size = min(uiop->uio_iov->iov_len, size); 810 bp = getpbuf(&ncl_pbuf_freecnt); 811 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 812 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 813 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 814 t_iov->iov_len = size; 815 t_uio->uio_iov = t_iov; 816 t_uio->uio_iovcnt = 1; 817 t_uio->uio_offset = uiop->uio_offset; 818 t_uio->uio_resid = size; 819 t_uio->uio_segflg = UIO_SYSSPACE; 820 t_uio->uio_rw = UIO_WRITE; 821 t_uio->uio_td = td; 822 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); 823 bp->b_flags |= B_DIRECT; 824 bp->b_iocmd = BIO_WRITE; 825 if (cred != NOCRED) { 826 crhold(cred); 827 bp->b_wcred = cred; 828 } else 829 bp->b_wcred = NOCRED; 830 bp->b_caller1 = (void *)t_uio; 831 bp->b_vp = vp; 832 error = ncl_asyncio(nmp, bp, NOCRED, td); 833 if (error) { 834 free(t_iov->iov_base, M_NFSDIRECTIO); 835 free(t_iov, M_NFSDIRECTIO); 836 free(t_uio, M_NFSDIRECTIO); 837 bp->b_vp = NULL; 838 relpbuf(bp, &ncl_pbuf_freecnt); 839 if (error == EINTR) 840 return (error); 841 goto do_sync; 842 } 843 uiop->uio_offset += size; 844 uiop->uio_resid -= size; 845 if (uiop->uio_iov->iov_len <= size) { 846 uiop->uio_iovcnt--; 847 uiop->uio_iov++; 848 } else { 849 uiop->uio_iov->iov_base = 850 (char *)uiop->uio_iov->iov_base + size; 851 uiop->uio_iov->iov_len -= size; 852 } 853 } 854 } 855 return (0); 856 } 857 858 /* 859 * Vnode op for write using bio 860 */ 861 int 862 ncl_write(struct vop_write_args *ap) 863 { 864 int biosize; 865 struct uio *uio = ap->a_uio; 866 struct thread *td = uio->uio_td; 867 struct vnode *vp = ap->a_vp; 868 struct nfsnode *np = VTONFS(vp); 869 struct ucred *cred = ap->a_cred; 870 int ioflag = ap->a_ioflag; 871 struct buf *bp; 872 struct vattr vattr; 873 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 874 daddr_t lbn; 875 int bcount; 876 int n, on, error = 0; 877 878 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode")); 879 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, 880 ("ncl_write proc")); 881 if (vp->v_type != VREG) 882 return (EIO); 883 mtx_lock(&np->n_mtx); 884 if (np->n_flag & NWRITEERR) { 885 np->n_flag &= ~NWRITEERR; 886 mtx_unlock(&np->n_mtx); 887 return (np->n_error); 888 } else 889 mtx_unlock(&np->n_mtx); 890 mtx_lock(&nmp->nm_mtx); 891 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 892 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 893 mtx_unlock(&nmp->nm_mtx); 894 (void)ncl_fsinfo(nmp, vp, cred, td); 895 mtx_lock(&nmp->nm_mtx); 896 } 897 if (nmp->nm_wsize == 0) 898 (void) newnfs_iosize(nmp); 899 mtx_unlock(&nmp->nm_mtx); 900 901 /* 902 * Synchronously flush pending buffers if we are in synchronous 903 * mode or if we are appending. 904 */ 905 if (ioflag & (IO_APPEND | IO_SYNC)) { 906 mtx_lock(&np->n_mtx); 907 if (np->n_flag & NMODIFIED) { 908 mtx_unlock(&np->n_mtx); 909 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 910 /* 911 * Require non-blocking, synchronous writes to 912 * dirty files to inform the program it needs 913 * to fsync(2) explicitly. 914 */ 915 if (ioflag & IO_NDELAY) 916 return (EAGAIN); 917 #endif 918 flush_and_restart: 919 np->n_attrstamp = 0; 920 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 921 if (error) 922 return (error); 923 } else 924 mtx_unlock(&np->n_mtx); 925 } 926 927 /* 928 * If IO_APPEND then load uio_offset. We restart here if we cannot 929 * get the append lock. 930 */ 931 if (ioflag & IO_APPEND) { 932 np->n_attrstamp = 0; 933 error = VOP_GETATTR(vp, &vattr, cred); 934 if (error) 935 return (error); 936 mtx_lock(&np->n_mtx); 937 uio->uio_offset = np->n_size; 938 mtx_unlock(&np->n_mtx); 939 } 940 941 if (uio->uio_offset < 0) 942 return (EINVAL); 943 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 944 return (EFBIG); 945 if (uio->uio_resid == 0) 946 return (0); 947 948 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 949 return nfs_directio_write(vp, uio, cred, ioflag); 950 951 /* 952 * Maybe this should be above the vnode op call, but so long as 953 * file servers have no limits, i don't think it matters 954 */ 955 if (vn_rlimit_fsize(vp, uio, td)) 956 return (EFBIG); 957 958 biosize = vp->v_mount->mnt_stat.f_iosize; 959 /* 960 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 961 * would exceed the local maximum per-file write commit size when 962 * combined with those, we must decide whether to flush, 963 * go synchronous, or return error. We don't bother checking 964 * IO_UNIT -- we just make all writes atomic anyway, as there's 965 * no point optimizing for something that really won't ever happen. 966 */ 967 if (!(ioflag & IO_SYNC)) { 968 int nflag; 969 970 mtx_lock(&np->n_mtx); 971 nflag = np->n_flag; 972 mtx_unlock(&np->n_mtx); 973 int needrestart = 0; 974 if (nmp->nm_wcommitsize < uio->uio_resid) { 975 /* 976 * If this request could not possibly be completed 977 * without exceeding the maximum outstanding write 978 * commit size, see if we can convert it into a 979 * synchronous write operation. 980 */ 981 if (ioflag & IO_NDELAY) 982 return (EAGAIN); 983 ioflag |= IO_SYNC; 984 if (nflag & NMODIFIED) 985 needrestart = 1; 986 } else if (nflag & NMODIFIED) { 987 int wouldcommit = 0; 988 BO_LOCK(&vp->v_bufobj); 989 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 990 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 991 b_bobufs) { 992 if (bp->b_flags & B_NEEDCOMMIT) 993 wouldcommit += bp->b_bcount; 994 } 995 } 996 BO_UNLOCK(&vp->v_bufobj); 997 /* 998 * Since we're not operating synchronously and 999 * bypassing the buffer cache, we are in a commit 1000 * and holding all of these buffers whether 1001 * transmitted or not. If not limited, this 1002 * will lead to the buffer cache deadlocking, 1003 * as no one else can flush our uncommitted buffers. 1004 */ 1005 wouldcommit += uio->uio_resid; 1006 /* 1007 * If we would initially exceed the maximum 1008 * outstanding write commit size, flush and restart. 1009 */ 1010 if (wouldcommit > nmp->nm_wcommitsize) 1011 needrestart = 1; 1012 } 1013 if (needrestart) 1014 goto flush_and_restart; 1015 } 1016 1017 do { 1018 NFSINCRGLOBAL(newnfsstats.biocache_writes); 1019 lbn = uio->uio_offset / biosize; 1020 on = uio->uio_offset & (biosize-1); 1021 n = min((unsigned)(biosize - on), uio->uio_resid); 1022 again: 1023 /* 1024 * Handle direct append and file extension cases, calculate 1025 * unaligned buffer size. 1026 */ 1027 mtx_lock(&np->n_mtx); 1028 if (uio->uio_offset == np->n_size && n) { 1029 mtx_unlock(&np->n_mtx); 1030 /* 1031 * Get the buffer (in its pre-append state to maintain 1032 * B_CACHE if it was previously set). Resize the 1033 * nfsnode after we have locked the buffer to prevent 1034 * readers from reading garbage. 1035 */ 1036 bcount = on; 1037 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1038 1039 if (bp != NULL) { 1040 long save; 1041 1042 mtx_lock(&np->n_mtx); 1043 np->n_size = uio->uio_offset + n; 1044 np->n_flag |= NMODIFIED; 1045 vnode_pager_setsize(vp, np->n_size); 1046 mtx_unlock(&np->n_mtx); 1047 1048 save = bp->b_flags & B_CACHE; 1049 bcount += n; 1050 allocbuf(bp, bcount); 1051 bp->b_flags |= save; 1052 } 1053 } else { 1054 /* 1055 * Obtain the locked cache block first, and then 1056 * adjust the file's size as appropriate. 1057 */ 1058 bcount = on + n; 1059 if ((off_t)lbn * biosize + bcount < np->n_size) { 1060 if ((off_t)(lbn + 1) * biosize < np->n_size) 1061 bcount = biosize; 1062 else 1063 bcount = np->n_size - (off_t)lbn * biosize; 1064 } 1065 mtx_unlock(&np->n_mtx); 1066 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1067 mtx_lock(&np->n_mtx); 1068 if (uio->uio_offset + n > np->n_size) { 1069 np->n_size = uio->uio_offset + n; 1070 np->n_flag |= NMODIFIED; 1071 vnode_pager_setsize(vp, np->n_size); 1072 } 1073 mtx_unlock(&np->n_mtx); 1074 } 1075 1076 if (!bp) { 1077 error = newnfs_sigintr(nmp, td); 1078 if (!error) 1079 error = EINTR; 1080 break; 1081 } 1082 1083 /* 1084 * Issue a READ if B_CACHE is not set. In special-append 1085 * mode, B_CACHE is based on the buffer prior to the write 1086 * op and is typically set, avoiding the read. If a read 1087 * is required in special append mode, the server will 1088 * probably send us a short-read since we extended the file 1089 * on our end, resulting in b_resid == 0 and, thusly, 1090 * B_CACHE getting set. 1091 * 1092 * We can also avoid issuing the read if the write covers 1093 * the entire buffer. We have to make sure the buffer state 1094 * is reasonable in this case since we will not be initiating 1095 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1096 * more information. 1097 * 1098 * B_CACHE may also be set due to the buffer being cached 1099 * normally. 1100 */ 1101 1102 if (on == 0 && n == bcount) { 1103 bp->b_flags |= B_CACHE; 1104 bp->b_flags &= ~B_INVAL; 1105 bp->b_ioflags &= ~BIO_ERROR; 1106 } 1107 1108 if ((bp->b_flags & B_CACHE) == 0) { 1109 bp->b_iocmd = BIO_READ; 1110 vfs_busy_pages(bp, 0); 1111 error = ncl_doio(vp, bp, cred, td, 0); 1112 if (error) { 1113 brelse(bp); 1114 break; 1115 } 1116 } 1117 if (bp->b_wcred == NOCRED) 1118 bp->b_wcred = crhold(cred); 1119 mtx_lock(&np->n_mtx); 1120 np->n_flag |= NMODIFIED; 1121 mtx_unlock(&np->n_mtx); 1122 1123 /* 1124 * If dirtyend exceeds file size, chop it down. This should 1125 * not normally occur but there is an append race where it 1126 * might occur XXX, so we log it. 1127 * 1128 * If the chopping creates a reverse-indexed or degenerate 1129 * situation with dirtyoff/end, we 0 both of them. 1130 */ 1131 1132 if (bp->b_dirtyend > bcount) { 1133 ncl_printf("NFS append race @%lx:%d\n", 1134 (long)bp->b_blkno * DEV_BSIZE, 1135 bp->b_dirtyend - bcount); 1136 bp->b_dirtyend = bcount; 1137 } 1138 1139 if (bp->b_dirtyoff >= bp->b_dirtyend) 1140 bp->b_dirtyoff = bp->b_dirtyend = 0; 1141 1142 /* 1143 * If the new write will leave a contiguous dirty 1144 * area, just update the b_dirtyoff and b_dirtyend, 1145 * otherwise force a write rpc of the old dirty area. 1146 * 1147 * While it is possible to merge discontiguous writes due to 1148 * our having a B_CACHE buffer ( and thus valid read data 1149 * for the hole), we don't because it could lead to 1150 * significant cache coherency problems with multiple clients, 1151 * especially if locking is implemented later on. 1152 * 1153 * as an optimization we could theoretically maintain 1154 * a linked list of discontinuous areas, but we would still 1155 * have to commit them separately so there isn't much 1156 * advantage to it except perhaps a bit of asynchronization. 1157 */ 1158 1159 if (bp->b_dirtyend > 0 && 1160 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1161 if (bwrite(bp) == EINTR) { 1162 error = EINTR; 1163 break; 1164 } 1165 goto again; 1166 } 1167 1168 error = uiomove((char *)bp->b_data + on, n, uio); 1169 1170 /* 1171 * Since this block is being modified, it must be written 1172 * again and not just committed. Since write clustering does 1173 * not work for the stage 1 data write, only the stage 2 1174 * commit rpc, we have to clear B_CLUSTEROK as well. 1175 */ 1176 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1177 1178 if (error) { 1179 bp->b_ioflags |= BIO_ERROR; 1180 brelse(bp); 1181 break; 1182 } 1183 1184 /* 1185 * Only update dirtyoff/dirtyend if not a degenerate 1186 * condition. 1187 */ 1188 if (n) { 1189 if (bp->b_dirtyend > 0) { 1190 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1191 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1192 } else { 1193 bp->b_dirtyoff = on; 1194 bp->b_dirtyend = on + n; 1195 } 1196 vfs_bio_set_valid(bp, on, n); 1197 } 1198 1199 /* 1200 * If IO_SYNC do bwrite(). 1201 * 1202 * IO_INVAL appears to be unused. The idea appears to be 1203 * to turn off caching in this case. Very odd. XXX 1204 */ 1205 if ((ioflag & IO_SYNC)) { 1206 if (ioflag & IO_INVAL) 1207 bp->b_flags |= B_NOCACHE; 1208 error = bwrite(bp); 1209 if (error) 1210 break; 1211 } else if ((n + on) == biosize) { 1212 bp->b_flags |= B_ASYNC; 1213 (void) ncl_writebp(bp, 0, NULL); 1214 } else { 1215 bdwrite(bp); 1216 } 1217 } while (uio->uio_resid > 0 && n > 0); 1218 1219 return (error); 1220 } 1221 1222 /* 1223 * Get an nfs cache block. 1224 * 1225 * Allocate a new one if the block isn't currently in the cache 1226 * and return the block marked busy. If the calling process is 1227 * interrupted by a signal for an interruptible mount point, return 1228 * NULL. 1229 * 1230 * The caller must carefully deal with the possible B_INVAL state of 1231 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it 1232 * indirectly), so synchronous reads can be issued without worrying about 1233 * the B_INVAL state. We have to be a little more careful when dealing 1234 * with writes (see comments in nfs_write()) when extending a file past 1235 * its EOF. 1236 */ 1237 static struct buf * 1238 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1239 { 1240 struct buf *bp; 1241 struct mount *mp; 1242 struct nfsmount *nmp; 1243 1244 mp = vp->v_mount; 1245 nmp = VFSTONFS(mp); 1246 1247 if (nmp->nm_flag & NFSMNT_INT) { 1248 sigset_t oldset; 1249 1250 newnfs_set_sigmask(td, &oldset); 1251 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0); 1252 newnfs_restore_sigmask(td, &oldset); 1253 while (bp == NULL) { 1254 if (newnfs_sigintr(nmp, td)) 1255 return (NULL); 1256 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1257 } 1258 } else { 1259 bp = getblk(vp, bn, size, 0, 0, 0); 1260 } 1261 1262 if (vp->v_type == VREG) { 1263 int biosize; 1264 1265 biosize = mp->mnt_stat.f_iosize; 1266 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1267 } 1268 return (bp); 1269 } 1270 1271 /* 1272 * Flush and invalidate all dirty buffers. If another process is already 1273 * doing the flush, just wait for completion. 1274 */ 1275 int 1276 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1277 { 1278 struct nfsnode *np = VTONFS(vp); 1279 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1280 int error = 0, slpflag, slptimeo; 1281 int old_lock = 0; 1282 1283 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf"); 1284 1285 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1286 intrflg = 0; 1287 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)) 1288 intrflg = 1; 1289 if (intrflg) { 1290 slpflag = NFS_PCATCH; 1291 slptimeo = 2 * hz; 1292 } else { 1293 slpflag = 0; 1294 slptimeo = 0; 1295 } 1296 1297 old_lock = ncl_upgrade_vnlock(vp); 1298 if (vp->v_iflag & VI_DOOMED) { 1299 /* 1300 * Since vgonel() uses the generic vinvalbuf() to flush 1301 * dirty buffers and it does not call this function, it 1302 * is safe to just return OK when VI_DOOMED is set. 1303 */ 1304 ncl_downgrade_vnlock(vp, old_lock); 1305 return (0); 1306 } 1307 1308 /* 1309 * Now, flush as required. 1310 */ 1311 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1312 VM_OBJECT_LOCK(vp->v_bufobj.bo_object); 1313 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1314 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); 1315 /* 1316 * If the page clean was interrupted, fail the invalidation. 1317 * Not doing so, we run the risk of losing dirty pages in the 1318 * vinvalbuf() call below. 1319 */ 1320 if (intrflg && (error = newnfs_sigintr(nmp, td))) 1321 goto out; 1322 } 1323 1324 error = vinvalbuf(vp, flags, slpflag, 0); 1325 while (error) { 1326 if (intrflg && (error = newnfs_sigintr(nmp, td))) 1327 goto out; 1328 error = vinvalbuf(vp, flags, 0, slptimeo); 1329 } 1330 mtx_lock(&np->n_mtx); 1331 if (np->n_directio_asyncwr == 0) 1332 np->n_flag &= ~NMODIFIED; 1333 mtx_unlock(&np->n_mtx); 1334 out: 1335 ncl_downgrade_vnlock(vp, old_lock); 1336 return error; 1337 } 1338 1339 /* 1340 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1341 * This is mainly to avoid queueing async I/O requests when the nfsiods 1342 * are all hung on a dead server. 1343 * 1344 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1345 * is eventually dequeued by the async daemon, ncl_doio() *will*. 1346 */ 1347 int 1348 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1349 { 1350 int iod; 1351 int gotiod; 1352 int slpflag = 0; 1353 int slptimeo = 0; 1354 int error, error2; 1355 1356 /* 1357 * Unless iothreadcnt is set > 0, don't bother with async I/O 1358 * threads. For LAN environments, they don't buy any significant 1359 * performance improvement that you can't get with large block 1360 * sizes. 1361 */ 1362 if (nmp->nm_readahead == 0) 1363 return (EPERM); 1364 1365 /* 1366 * Commits are usually short and sweet so lets save some cpu and 1367 * leave the async daemons for more important rpc's (such as reads 1368 * and writes). 1369 */ 1370 mtx_lock(&ncl_iod_mutex); 1371 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1372 (nmp->nm_bufqiods > ncl_numasync / 2)) { 1373 mtx_unlock(&ncl_iod_mutex); 1374 return(EIO); 1375 } 1376 again: 1377 if (nmp->nm_flag & NFSMNT_INT) 1378 slpflag = NFS_PCATCH; 1379 gotiod = FALSE; 1380 1381 /* 1382 * Find a free iod to process this request. 1383 */ 1384 for (iod = 0; iod < ncl_numasync; iod++) 1385 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) { 1386 gotiod = TRUE; 1387 break; 1388 } 1389 1390 /* 1391 * Try to create one if none are free. 1392 */ 1393 if (!gotiod) { 1394 iod = ncl_nfsiodnew(1); 1395 if (iod != -1) 1396 gotiod = TRUE; 1397 } 1398 1399 if (gotiod) { 1400 /* 1401 * Found one, so wake it up and tell it which 1402 * mount to process. 1403 */ 1404 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n", 1405 iod, nmp)); 1406 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE; 1407 ncl_iodmount[iod] = nmp; 1408 nmp->nm_bufqiods++; 1409 wakeup(&ncl_iodwant[iod]); 1410 } 1411 1412 /* 1413 * If none are free, we may already have an iod working on this mount 1414 * point. If so, it will process our request. 1415 */ 1416 if (!gotiod) { 1417 if (nmp->nm_bufqiods > 0) { 1418 NFS_DPF(ASYNCIO, 1419 ("ncl_asyncio: %d iods are already processing mount %p\n", 1420 nmp->nm_bufqiods, nmp)); 1421 gotiod = TRUE; 1422 } 1423 } 1424 1425 /* 1426 * If we have an iod which can process the request, then queue 1427 * the buffer. 1428 */ 1429 if (gotiod) { 1430 /* 1431 * Ensure that the queue never grows too large. We still want 1432 * to asynchronize so we block rather then return EIO. 1433 */ 1434 while (nmp->nm_bufqlen >= 2*ncl_numasync) { 1435 NFS_DPF(ASYNCIO, 1436 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp)); 1437 nmp->nm_bufqwant = TRUE; 1438 error = newnfs_msleep(td, &nmp->nm_bufq, 1439 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio", 1440 slptimeo); 1441 if (error) { 1442 error2 = newnfs_sigintr(nmp, td); 1443 if (error2) { 1444 mtx_unlock(&ncl_iod_mutex); 1445 return (error2); 1446 } 1447 if (slpflag == NFS_PCATCH) { 1448 slpflag = 0; 1449 slptimeo = 2 * hz; 1450 } 1451 } 1452 /* 1453 * We might have lost our iod while sleeping, 1454 * so check and loop if nescessary. 1455 */ 1456 if (nmp->nm_bufqiods == 0) { 1457 NFS_DPF(ASYNCIO, 1458 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1459 goto again; 1460 } 1461 } 1462 1463 /* We might have lost our nfsiod */ 1464 if (nmp->nm_bufqiods == 0) { 1465 NFS_DPF(ASYNCIO, 1466 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1467 goto again; 1468 } 1469 1470 if (bp->b_iocmd == BIO_READ) { 1471 if (bp->b_rcred == NOCRED && cred != NOCRED) 1472 bp->b_rcred = crhold(cred); 1473 } else { 1474 if (bp->b_wcred == NOCRED && cred != NOCRED) 1475 bp->b_wcred = crhold(cred); 1476 } 1477 1478 if (bp->b_flags & B_REMFREE) 1479 bremfreef(bp); 1480 BUF_KERNPROC(bp); 1481 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1482 nmp->nm_bufqlen++; 1483 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1484 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1485 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1486 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1487 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1488 } 1489 mtx_unlock(&ncl_iod_mutex); 1490 return (0); 1491 } 1492 1493 mtx_unlock(&ncl_iod_mutex); 1494 1495 /* 1496 * All the iods are busy on other mounts, so return EIO to 1497 * force the caller to process the i/o synchronously. 1498 */ 1499 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n")); 1500 return (EIO); 1501 } 1502 1503 void 1504 ncl_doio_directwrite(struct buf *bp) 1505 { 1506 int iomode, must_commit; 1507 struct uio *uiop = (struct uio *)bp->b_caller1; 1508 char *iov_base = uiop->uio_iov->iov_base; 1509 1510 iomode = NFSWRITE_FILESYNC; 1511 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1512 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0); 1513 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write")); 1514 free(iov_base, M_NFSDIRECTIO); 1515 free(uiop->uio_iov, M_NFSDIRECTIO); 1516 free(uiop, M_NFSDIRECTIO); 1517 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1518 struct nfsnode *np = VTONFS(bp->b_vp); 1519 mtx_lock(&np->n_mtx); 1520 np->n_directio_asyncwr--; 1521 if (np->n_directio_asyncwr == 0) { 1522 np->n_flag &= ~NMODIFIED; 1523 if ((np->n_flag & NFSYNCWAIT)) { 1524 np->n_flag &= ~NFSYNCWAIT; 1525 wakeup((caddr_t)&np->n_directio_asyncwr); 1526 } 1527 } 1528 mtx_unlock(&np->n_mtx); 1529 } 1530 bp->b_vp = NULL; 1531 relpbuf(bp, &ncl_pbuf_freecnt); 1532 } 1533 1534 /* 1535 * Do an I/O operation to/from a cache block. This may be called 1536 * synchronously or from an nfsiod. 1537 */ 1538 int 1539 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td, 1540 int called_from_strategy) 1541 { 1542 struct uio *uiop; 1543 struct nfsnode *np; 1544 struct nfsmount *nmp; 1545 int error = 0, iomode, must_commit = 0; 1546 struct uio uio; 1547 struct iovec io; 1548 struct proc *p = td ? td->td_proc : NULL; 1549 uint8_t iocmd; 1550 1551 np = VTONFS(vp); 1552 nmp = VFSTONFS(vp->v_mount); 1553 uiop = &uio; 1554 uiop->uio_iov = &io; 1555 uiop->uio_iovcnt = 1; 1556 uiop->uio_segflg = UIO_SYSSPACE; 1557 uiop->uio_td = td; 1558 1559 /* 1560 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1561 * do this here so we do not have to do it in all the code that 1562 * calls us. 1563 */ 1564 bp->b_flags &= ~B_INVAL; 1565 bp->b_ioflags &= ~BIO_ERROR; 1566 1567 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp)); 1568 iocmd = bp->b_iocmd; 1569 if (iocmd == BIO_READ) { 1570 io.iov_len = uiop->uio_resid = bp->b_bcount; 1571 io.iov_base = bp->b_data; 1572 uiop->uio_rw = UIO_READ; 1573 1574 switch (vp->v_type) { 1575 case VREG: 1576 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1577 NFSINCRGLOBAL(newnfsstats.read_bios); 1578 error = ncl_readrpc(vp, uiop, cr); 1579 1580 if (!error) { 1581 if (uiop->uio_resid) { 1582 /* 1583 * If we had a short read with no error, we must have 1584 * hit a file hole. We should zero-fill the remainder. 1585 * This can also occur if the server hits the file EOF. 1586 * 1587 * Holes used to be able to occur due to pending 1588 * writes, but that is not possible any longer. 1589 */ 1590 int nread = bp->b_bcount - uiop->uio_resid; 1591 int left = uiop->uio_resid; 1592 1593 if (left > 0) 1594 bzero((char *)bp->b_data + nread, left); 1595 uiop->uio_resid = 0; 1596 } 1597 } 1598 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */ 1599 if (p && (vp->v_vflag & VV_TEXT)) { 1600 mtx_lock(&np->n_mtx); 1601 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) { 1602 mtx_unlock(&np->n_mtx); 1603 PROC_LOCK(p); 1604 killproc(p, "text file modification"); 1605 PROC_UNLOCK(p); 1606 } else 1607 mtx_unlock(&np->n_mtx); 1608 } 1609 break; 1610 case VLNK: 1611 uiop->uio_offset = (off_t)0; 1612 NFSINCRGLOBAL(newnfsstats.readlink_bios); 1613 error = ncl_readlinkrpc(vp, uiop, cr); 1614 break; 1615 case VDIR: 1616 NFSINCRGLOBAL(newnfsstats.readdir_bios); 1617 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1618 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1619 error = ncl_readdirplusrpc(vp, uiop, cr, td); 1620 if (error == NFSERR_NOTSUPP) 1621 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1622 } 1623 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1624 error = ncl_readdirrpc(vp, uiop, cr, td); 1625 /* 1626 * end-of-directory sets B_INVAL but does not generate an 1627 * error. 1628 */ 1629 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1630 bp->b_flags |= B_INVAL; 1631 break; 1632 default: 1633 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type); 1634 break; 1635 }; 1636 if (error) { 1637 bp->b_ioflags |= BIO_ERROR; 1638 bp->b_error = error; 1639 } 1640 } else { 1641 /* 1642 * If we only need to commit, try to commit 1643 */ 1644 if (bp->b_flags & B_NEEDCOMMIT) { 1645 int retv; 1646 off_t off; 1647 1648 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1649 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1650 bp->b_wcred, td); 1651 if (retv == 0) { 1652 bp->b_dirtyoff = bp->b_dirtyend = 0; 1653 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1654 bp->b_resid = 0; 1655 bufdone(bp); 1656 return (0); 1657 } 1658 if (retv == NFSERR_STALEWRITEVERF) { 1659 ncl_clearcommit(vp->v_mount); 1660 } 1661 } 1662 1663 /* 1664 * Setup for actual write 1665 */ 1666 mtx_lock(&np->n_mtx); 1667 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1668 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1669 mtx_unlock(&np->n_mtx); 1670 1671 if (bp->b_dirtyend > bp->b_dirtyoff) { 1672 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1673 - bp->b_dirtyoff; 1674 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1675 + bp->b_dirtyoff; 1676 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1677 uiop->uio_rw = UIO_WRITE; 1678 NFSINCRGLOBAL(newnfsstats.write_bios); 1679 1680 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1681 iomode = NFSWRITE_UNSTABLE; 1682 else 1683 iomode = NFSWRITE_FILESYNC; 1684 1685 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit, 1686 called_from_strategy); 1687 1688 /* 1689 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1690 * to cluster the buffers needing commit. This will allow 1691 * the system to submit a single commit rpc for the whole 1692 * cluster. We can do this even if the buffer is not 100% 1693 * dirty (relative to the NFS blocksize), so we optimize the 1694 * append-to-file-case. 1695 * 1696 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1697 * cleared because write clustering only works for commit 1698 * rpc's, not for the data portion of the write). 1699 */ 1700 1701 if (!error && iomode == NFSWRITE_UNSTABLE) { 1702 bp->b_flags |= B_NEEDCOMMIT; 1703 if (bp->b_dirtyoff == 0 1704 && bp->b_dirtyend == bp->b_bcount) 1705 bp->b_flags |= B_CLUSTEROK; 1706 } else { 1707 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1708 } 1709 1710 /* 1711 * For an interrupted write, the buffer is still valid 1712 * and the write hasn't been pushed to the server yet, 1713 * so we can't set BIO_ERROR and report the interruption 1714 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1715 * is not relevant, so the rpc attempt is essentially 1716 * a noop. For the case of a V3 write rpc not being 1717 * committed to stable storage, the block is still 1718 * dirty and requires either a commit rpc or another 1719 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1720 * the block is reused. This is indicated by setting 1721 * the B_DELWRI and B_NEEDCOMMIT flags. 1722 * 1723 * EIO is returned by ncl_writerpc() to indicate a recoverable 1724 * write error and is handled as above, except that 1725 * B_EINTR isn't set. One cause of this is a stale stateid 1726 * error for the RPC that indicates recovery is required, 1727 * when called with called_from_strategy != 0. 1728 * 1729 * If the buffer is marked B_PAGING, it does not reside on 1730 * the vp's paging queues so we cannot call bdirty(). The 1731 * bp in this case is not an NFS cache block so we should 1732 * be safe. XXX 1733 * 1734 * The logic below breaks up errors into recoverable and 1735 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1736 * and keep the buffer around for potential write retries. 1737 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1738 * and save the error in the nfsnode. This is less than ideal 1739 * but necessary. Keeping such buffers around could potentially 1740 * cause buffer exhaustion eventually (they can never be written 1741 * out, so will get constantly be re-dirtied). It also causes 1742 * all sorts of vfs panics. For non-recoverable write errors, 1743 * also invalidate the attrcache, so we'll be forced to go over 1744 * the wire for this object, returning an error to user on next 1745 * call (most of the time). 1746 */ 1747 if (error == EINTR || error == EIO || error == ETIMEDOUT 1748 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1749 int s; 1750 1751 s = splbio(); 1752 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1753 if ((bp->b_flags & B_PAGING) == 0) { 1754 bdirty(bp); 1755 bp->b_flags &= ~B_DONE; 1756 } 1757 if ((error == EINTR || error == ETIMEDOUT) && 1758 (bp->b_flags & B_ASYNC) == 0) 1759 bp->b_flags |= B_EINTR; 1760 splx(s); 1761 } else { 1762 if (error) { 1763 bp->b_ioflags |= BIO_ERROR; 1764 bp->b_flags |= B_INVAL; 1765 bp->b_error = np->n_error = error; 1766 mtx_lock(&np->n_mtx); 1767 np->n_flag |= NWRITEERR; 1768 np->n_attrstamp = 0; 1769 mtx_unlock(&np->n_mtx); 1770 } 1771 bp->b_dirtyoff = bp->b_dirtyend = 0; 1772 } 1773 } else { 1774 bp->b_resid = 0; 1775 bufdone(bp); 1776 return (0); 1777 } 1778 } 1779 bp->b_resid = uiop->uio_resid; 1780 if (must_commit) 1781 ncl_clearcommit(vp->v_mount); 1782 bufdone(bp); 1783 return (error); 1784 } 1785 1786 /* 1787 * Used to aid in handling ftruncate() operations on the NFS client side. 1788 * Truncation creates a number of special problems for NFS. We have to 1789 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1790 * we have to properly handle VM pages or (potentially dirty) buffers 1791 * that straddle the truncation point. 1792 */ 1793 1794 int 1795 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1796 { 1797 struct nfsnode *np = VTONFS(vp); 1798 u_quad_t tsize; 1799 int biosize = vp->v_mount->mnt_stat.f_iosize; 1800 int error = 0; 1801 1802 mtx_lock(&np->n_mtx); 1803 tsize = np->n_size; 1804 np->n_size = nsize; 1805 mtx_unlock(&np->n_mtx); 1806 1807 if (nsize < tsize) { 1808 struct buf *bp; 1809 daddr_t lbn; 1810 int bufsize; 1811 1812 /* 1813 * vtruncbuf() doesn't get the buffer overlapping the 1814 * truncation point. We may have a B_DELWRI and/or B_CACHE 1815 * buffer that now needs to be truncated. 1816 */ 1817 error = vtruncbuf(vp, cred, td, nsize, biosize); 1818 lbn = nsize / biosize; 1819 bufsize = nsize & (biosize - 1); 1820 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1821 if (!bp) 1822 return EINTR; 1823 if (bp->b_dirtyoff > bp->b_bcount) 1824 bp->b_dirtyoff = bp->b_bcount; 1825 if (bp->b_dirtyend > bp->b_bcount) 1826 bp->b_dirtyend = bp->b_bcount; 1827 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1828 brelse(bp); 1829 } else { 1830 vnode_pager_setsize(vp, nsize); 1831 } 1832 return(error); 1833 } 1834 1835