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/proc.h> 45 #include <sys/resourcevar.h> 46 #include <sys/signalvar.h> 47 #include <sys/vmmeter.h> 48 #include <sys/vnode.h> 49 50 #include <vm/vm.h> 51 #include <vm/vm_extern.h> 52 #include <vm/vm_page.h> 53 #include <vm/vm_object.h> 54 #include <vm/vm_pager.h> 55 #include <vm/vnode_pager.h> 56 57 #include <fs/nfs/nfsport.h> 58 #include <fs/nfsclient/nfsmount.h> 59 #include <fs/nfsclient/nfs.h> 60 #include <fs/nfsclient/nfsnode.h> 61 62 extern int newnfs_directio_allow_mmap; 63 extern struct nfsstats newnfsstats; 64 extern struct mtx ncl_iod_mutex; 65 extern int ncl_numasync; 66 extern enum nfsiod_state ncl_iodwant[NFS_MAXRAHEAD]; 67 extern struct nfsmount *ncl_iodmount[NFS_MAXRAHEAD]; 68 extern int newnfs_directio_enable; 69 70 int ncl_pbuf_freecnt = -1; /* start out unlimited */ 71 72 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 73 struct thread *td); 74 static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 75 struct ucred *cred, int ioflag); 76 77 /* 78 * Vnode op for VM getpages. 79 */ 80 int 81 ncl_getpages(struct vop_getpages_args *ap) 82 { 83 int i, error, nextoff, size, toff, count, npages; 84 struct uio uio; 85 struct iovec iov; 86 vm_offset_t kva; 87 struct buf *bp; 88 struct vnode *vp; 89 struct thread *td; 90 struct ucred *cred; 91 struct nfsmount *nmp; 92 vm_object_t object; 93 vm_page_t *pages; 94 struct nfsnode *np; 95 96 vp = ap->a_vp; 97 np = VTONFS(vp); 98 td = curthread; /* XXX */ 99 cred = curthread->td_ucred; /* XXX */ 100 nmp = VFSTONFS(vp->v_mount); 101 pages = ap->a_m; 102 count = ap->a_count; 103 104 if ((object = vp->v_object) == NULL) { 105 ncl_printf("nfs_getpages: called with non-merged cache vnode??\n"); 106 return (VM_PAGER_ERROR); 107 } 108 109 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) { 110 mtx_lock(&np->n_mtx); 111 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 112 mtx_unlock(&np->n_mtx); 113 ncl_printf("nfs_getpages: called on non-cacheable vnode??\n"); 114 return (VM_PAGER_ERROR); 115 } else 116 mtx_unlock(&np->n_mtx); 117 } 118 119 mtx_lock(&nmp->nm_mtx); 120 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 121 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 122 mtx_unlock(&nmp->nm_mtx); 123 /* We'll never get here for v4, because we always have fsinfo */ 124 (void)ncl_fsinfo(nmp, vp, cred, td); 125 } else 126 mtx_unlock(&nmp->nm_mtx); 127 128 npages = btoc(count); 129 130 /* 131 * If the requested page is partially valid, just return it and 132 * allow the pager to zero-out the blanks. Partially valid pages 133 * can only occur at the file EOF. 134 */ 135 VM_OBJECT_LOCK(object); 136 if (pages[ap->a_reqpage]->valid != 0) { 137 vm_page_lock_queues(); 138 for (i = 0; i < npages; ++i) { 139 if (i != ap->a_reqpage) 140 vm_page_free(pages[i]); 141 } 142 vm_page_unlock_queues(); 143 VM_OBJECT_UNLOCK(object); 144 return (0); 145 } 146 VM_OBJECT_UNLOCK(object); 147 148 /* 149 * We use only the kva address for the buffer, but this is extremely 150 * convienient and fast. 151 */ 152 bp = getpbuf(&ncl_pbuf_freecnt); 153 154 kva = (vm_offset_t) bp->b_data; 155 pmap_qenter(kva, pages, npages); 156 PCPU_INC(cnt.v_vnodein); 157 PCPU_ADD(cnt.v_vnodepgsin, npages); 158 159 iov.iov_base = (caddr_t) kva; 160 iov.iov_len = count; 161 uio.uio_iov = &iov; 162 uio.uio_iovcnt = 1; 163 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 164 uio.uio_resid = count; 165 uio.uio_segflg = UIO_SYSSPACE; 166 uio.uio_rw = UIO_READ; 167 uio.uio_td = td; 168 169 error = ncl_readrpc(vp, &uio, cred); 170 pmap_qremove(kva, npages); 171 172 relpbuf(bp, &ncl_pbuf_freecnt); 173 174 if (error && (uio.uio_resid == count)) { 175 ncl_printf("nfs_getpages: error %d\n", error); 176 VM_OBJECT_LOCK(object); 177 vm_page_lock_queues(); 178 for (i = 0; i < npages; ++i) { 179 if (i != ap->a_reqpage) 180 vm_page_free(pages[i]); 181 } 182 vm_page_unlock_queues(); 183 VM_OBJECT_UNLOCK(object); 184 return (VM_PAGER_ERROR); 185 } 186 187 /* 188 * Calculate the number of bytes read and validate only that number 189 * of bytes. Note that due to pending writes, size may be 0. This 190 * does not mean that the remaining data is invalid! 191 */ 192 193 size = count - uio.uio_resid; 194 VM_OBJECT_LOCK(object); 195 vm_page_lock_queues(); 196 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 197 vm_page_t m; 198 nextoff = toff + PAGE_SIZE; 199 m = pages[i]; 200 201 if (nextoff <= size) { 202 /* 203 * Read operation filled an entire page 204 */ 205 m->valid = VM_PAGE_BITS_ALL; 206 KASSERT(m->dirty == 0, 207 ("nfs_getpages: page %p is dirty", m)); 208 } else if (size > toff) { 209 /* 210 * Read operation filled a partial page. 211 */ 212 m->valid = 0; 213 vm_page_set_valid(m, 0, size - toff); 214 KASSERT(m->dirty == 0, 215 ("nfs_getpages: page %p is dirty", m)); 216 } else { 217 /* 218 * Read operation was short. If no error occured 219 * we may have hit a zero-fill section. We simply 220 * leave valid set to 0. 221 */ 222 ; 223 } 224 if (i != ap->a_reqpage) { 225 /* 226 * Whether or not to leave the page activated is up in 227 * the air, but we should put the page on a page queue 228 * somewhere (it already is in the object). Result: 229 * It appears that emperical results show that 230 * deactivating pages is best. 231 */ 232 233 /* 234 * Just in case someone was asking for this page we 235 * now tell them that it is ok to use. 236 */ 237 if (!error) { 238 if (m->oflags & VPO_WANTED) 239 vm_page_activate(m); 240 else 241 vm_page_deactivate(m); 242 vm_page_wakeup(m); 243 } else { 244 vm_page_free(m); 245 } 246 } 247 } 248 vm_page_unlock_queues(); 249 VM_OBJECT_UNLOCK(object); 250 return (0); 251 } 252 253 /* 254 * Vnode op for VM putpages. 255 */ 256 int 257 ncl_putpages(struct vop_putpages_args *ap) 258 { 259 struct uio uio; 260 struct iovec iov; 261 vm_offset_t kva; 262 struct buf *bp; 263 int iomode, must_commit, i, error, npages, count; 264 off_t offset; 265 int *rtvals; 266 struct vnode *vp; 267 struct thread *td; 268 struct ucred *cred; 269 struct nfsmount *nmp; 270 struct nfsnode *np; 271 vm_page_t *pages; 272 273 vp = ap->a_vp; 274 np = VTONFS(vp); 275 td = curthread; /* XXX */ 276 cred = curthread->td_ucred; /* XXX */ 277 nmp = VFSTONFS(vp->v_mount); 278 pages = ap->a_m; 279 count = ap->a_count; 280 rtvals = ap->a_rtvals; 281 npages = btoc(count); 282 offset = IDX_TO_OFF(pages[0]->pindex); 283 284 mtx_lock(&nmp->nm_mtx); 285 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 286 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 287 mtx_unlock(&nmp->nm_mtx); 288 (void)ncl_fsinfo(nmp, vp, cred, td); 289 } else 290 mtx_unlock(&nmp->nm_mtx); 291 292 mtx_lock(&np->n_mtx); 293 if (newnfs_directio_enable && !newnfs_directio_allow_mmap && 294 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 295 mtx_unlock(&np->n_mtx); 296 ncl_printf("ncl_putpages: called on noncache-able vnode??\n"); 297 mtx_lock(&np->n_mtx); 298 } 299 300 for (i = 0; i < npages; i++) 301 rtvals[i] = VM_PAGER_AGAIN; 302 303 /* 304 * When putting pages, do not extend file past EOF. 305 */ 306 if (offset + count > np->n_size) { 307 count = np->n_size - offset; 308 if (count < 0) 309 count = 0; 310 } 311 mtx_unlock(&np->n_mtx); 312 313 /* 314 * We use only the kva address for the buffer, but this is extremely 315 * convienient and fast. 316 */ 317 bp = getpbuf(&ncl_pbuf_freecnt); 318 319 kva = (vm_offset_t) bp->b_data; 320 pmap_qenter(kva, pages, npages); 321 PCPU_INC(cnt.v_vnodeout); 322 PCPU_ADD(cnt.v_vnodepgsout, count); 323 324 iov.iov_base = (caddr_t) kva; 325 iov.iov_len = count; 326 uio.uio_iov = &iov; 327 uio.uio_iovcnt = 1; 328 uio.uio_offset = offset; 329 uio.uio_resid = count; 330 uio.uio_segflg = UIO_SYSSPACE; 331 uio.uio_rw = UIO_WRITE; 332 uio.uio_td = td; 333 334 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 335 iomode = NFSWRITE_UNSTABLE; 336 else 337 iomode = NFSWRITE_FILESYNC; 338 339 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit); 340 341 pmap_qremove(kva, npages); 342 relpbuf(bp, &ncl_pbuf_freecnt); 343 344 if (!error) { 345 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 346 for (i = 0; i < nwritten; i++) { 347 rtvals[i] = VM_PAGER_OK; 348 vm_page_undirty(pages[i]); 349 } 350 if (must_commit) { 351 ncl_clearcommit(vp->v_mount); 352 } 353 } 354 return rtvals[0]; 355 } 356 357 /* 358 * For nfs, cache consistency can only be maintained approximately. 359 * Although RFC1094 does not specify the criteria, the following is 360 * believed to be compatible with the reference port. 361 * For nfs: 362 * If the file's modify time on the server has changed since the 363 * last read rpc or you have written to the file, 364 * you may have lost data cache consistency with the 365 * server, so flush all of the file's data out of the cache. 366 * Then force a getattr rpc to ensure that you have up to date 367 * attributes. 368 * NB: This implies that cache data can be read when up to 369 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 370 * attributes this could be forced by setting n_attrstamp to 0 before 371 * the VOP_GETATTR() call. 372 */ 373 static inline int 374 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 375 { 376 int error = 0; 377 struct vattr vattr; 378 struct nfsnode *np = VTONFS(vp); 379 int old_lock; 380 381 /* 382 * Grab the exclusive lock before checking whether the cache is 383 * consistent. 384 * XXX - We can make this cheaper later (by acquiring cheaper locks). 385 * But for now, this suffices. 386 */ 387 old_lock = ncl_upgrade_vnlock(vp); 388 if (vp->v_iflag & VI_DOOMED) { 389 ncl_downgrade_vnlock(vp, old_lock); 390 return (EBADF); 391 } 392 393 mtx_lock(&np->n_mtx); 394 if (np->n_flag & NMODIFIED) { 395 mtx_unlock(&np->n_mtx); 396 if (vp->v_type != VREG) { 397 if (vp->v_type != VDIR) 398 panic("nfs: bioread, not dir"); 399 ncl_invaldir(vp); 400 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 401 if (error) 402 goto out; 403 } 404 np->n_attrstamp = 0; 405 error = VOP_GETATTR(vp, &vattr, cred); 406 if (error) 407 goto out; 408 mtx_lock(&np->n_mtx); 409 np->n_mtime = vattr.va_mtime; 410 mtx_unlock(&np->n_mtx); 411 } else { 412 mtx_unlock(&np->n_mtx); 413 error = VOP_GETATTR(vp, &vattr, cred); 414 if (error) 415 return (error); 416 mtx_lock(&np->n_mtx); 417 if ((np->n_flag & NSIZECHANGED) 418 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 419 mtx_unlock(&np->n_mtx); 420 if (vp->v_type == VDIR) 421 ncl_invaldir(vp); 422 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 423 if (error) 424 goto out; 425 mtx_lock(&np->n_mtx); 426 np->n_mtime = vattr.va_mtime; 427 np->n_flag &= ~NSIZECHANGED; 428 } 429 mtx_unlock(&np->n_mtx); 430 } 431 out: 432 ncl_downgrade_vnlock(vp, old_lock); 433 return error; 434 } 435 436 /* 437 * Vnode op for read using bio 438 */ 439 int 440 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 441 { 442 struct nfsnode *np = VTONFS(vp); 443 int biosize, i; 444 struct buf *bp, *rabp; 445 struct thread *td; 446 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 447 daddr_t lbn, rabn; 448 int bcount; 449 int seqcount; 450 int nra, error = 0, n = 0, on = 0; 451 452 #ifdef DIAGNOSTIC 453 if (uio->uio_rw != UIO_READ) 454 panic("ncl_read mode"); 455 #endif 456 if (uio->uio_resid == 0) 457 return (0); 458 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 459 return (EINVAL); 460 td = uio->uio_td; 461 462 mtx_lock(&nmp->nm_mtx); 463 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 464 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 465 mtx_unlock(&nmp->nm_mtx); 466 (void)ncl_fsinfo(nmp, vp, cred, td); 467 mtx_lock(&nmp->nm_mtx); 468 } 469 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0) 470 (void) newnfs_iosize(nmp); 471 mtx_unlock(&nmp->nm_mtx); 472 473 if (vp->v_type != VDIR && 474 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 475 return (EFBIG); 476 477 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 478 /* No caching/ no readaheads. Just read data into the user buffer */ 479 return ncl_readrpc(vp, uio, cred); 480 481 biosize = vp->v_mount->mnt_stat.f_iosize; 482 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 483 484 error = nfs_bioread_check_cons(vp, td, cred); 485 if (error) 486 return error; 487 488 do { 489 u_quad_t nsize; 490 491 mtx_lock(&np->n_mtx); 492 nsize = np->n_size; 493 mtx_unlock(&np->n_mtx); 494 495 switch (vp->v_type) { 496 case VREG: 497 NFSINCRGLOBAL(newnfsstats.biocache_reads); 498 lbn = uio->uio_offset / biosize; 499 on = uio->uio_offset & (biosize - 1); 500 501 /* 502 * Start the read ahead(s), as required. 503 */ 504 if (nmp->nm_readahead > 0) { 505 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 506 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 507 rabn = lbn + 1 + nra; 508 if (incore(&vp->v_bufobj, rabn) == NULL) { 509 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 510 if (!rabp) { 511 error = newnfs_sigintr(nmp, td); 512 if (error) 513 return (error); 514 else 515 break; 516 } 517 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 518 rabp->b_flags |= B_ASYNC; 519 rabp->b_iocmd = BIO_READ; 520 vfs_busy_pages(rabp, 0); 521 if (ncl_asyncio(nmp, rabp, cred, td)) { 522 rabp->b_flags |= B_INVAL; 523 rabp->b_ioflags |= BIO_ERROR; 524 vfs_unbusy_pages(rabp); 525 brelse(rabp); 526 break; 527 } 528 } else { 529 brelse(rabp); 530 } 531 } 532 } 533 } 534 535 /* Note that bcount is *not* DEV_BSIZE aligned. */ 536 bcount = biosize; 537 if ((off_t)lbn * biosize >= nsize) { 538 bcount = 0; 539 } else if ((off_t)(lbn + 1) * biosize > nsize) { 540 bcount = nsize - (off_t)lbn * biosize; 541 } 542 bp = nfs_getcacheblk(vp, lbn, bcount, td); 543 544 if (!bp) { 545 error = newnfs_sigintr(nmp, td); 546 return (error ? error : EINTR); 547 } 548 549 /* 550 * If B_CACHE is not set, we must issue the read. If this 551 * fails, we return an error. 552 */ 553 554 if ((bp->b_flags & B_CACHE) == 0) { 555 bp->b_iocmd = BIO_READ; 556 vfs_busy_pages(bp, 0); 557 error = ncl_doio(vp, bp, cred, td); 558 if (error) { 559 brelse(bp); 560 return (error); 561 } 562 } 563 564 /* 565 * on is the offset into the current bp. Figure out how many 566 * bytes we can copy out of the bp. Note that bcount is 567 * NOT DEV_BSIZE aligned. 568 * 569 * Then figure out how many bytes we can copy into the uio. 570 */ 571 572 n = 0; 573 if (on < bcount) 574 n = min((unsigned)(bcount - on), uio->uio_resid); 575 break; 576 case VLNK: 577 NFSINCRGLOBAL(newnfsstats.biocache_readlinks); 578 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 579 if (!bp) { 580 error = newnfs_sigintr(nmp, td); 581 return (error ? error : EINTR); 582 } 583 if ((bp->b_flags & B_CACHE) == 0) { 584 bp->b_iocmd = BIO_READ; 585 vfs_busy_pages(bp, 0); 586 error = ncl_doio(vp, bp, cred, td); 587 if (error) { 588 bp->b_ioflags |= BIO_ERROR; 589 brelse(bp); 590 return (error); 591 } 592 } 593 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 594 on = 0; 595 break; 596 case VDIR: 597 NFSINCRGLOBAL(newnfsstats.biocache_readdirs); 598 if (np->n_direofoffset 599 && uio->uio_offset >= np->n_direofoffset) { 600 return (0); 601 } 602 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 603 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 604 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 605 if (!bp) { 606 error = newnfs_sigintr(nmp, td); 607 return (error ? error : EINTR); 608 } 609 if ((bp->b_flags & B_CACHE) == 0) { 610 bp->b_iocmd = BIO_READ; 611 vfs_busy_pages(bp, 0); 612 error = ncl_doio(vp, bp, cred, td); 613 if (error) { 614 brelse(bp); 615 } 616 while (error == NFSERR_BAD_COOKIE) { 617 ncl_invaldir(vp); 618 error = ncl_vinvalbuf(vp, 0, td, 1); 619 /* 620 * Yuck! The directory has been modified on the 621 * server. The only way to get the block is by 622 * reading from the beginning to get all the 623 * offset cookies. 624 * 625 * Leave the last bp intact unless there is an error. 626 * Loop back up to the while if the error is another 627 * NFSERR_BAD_COOKIE (double yuch!). 628 */ 629 for (i = 0; i <= lbn && !error; i++) { 630 if (np->n_direofoffset 631 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 632 return (0); 633 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 634 if (!bp) { 635 error = newnfs_sigintr(nmp, td); 636 return (error ? error : EINTR); 637 } 638 if ((bp->b_flags & B_CACHE) == 0) { 639 bp->b_iocmd = BIO_READ; 640 vfs_busy_pages(bp, 0); 641 error = ncl_doio(vp, bp, cred, td); 642 /* 643 * no error + B_INVAL == directory EOF, 644 * use the block. 645 */ 646 if (error == 0 && (bp->b_flags & B_INVAL)) 647 break; 648 } 649 /* 650 * An error will throw away the block and the 651 * for loop will break out. If no error and this 652 * is not the block we want, we throw away the 653 * block and go for the next one via the for loop. 654 */ 655 if (error || i < lbn) 656 brelse(bp); 657 } 658 } 659 /* 660 * The above while is repeated if we hit another cookie 661 * error. If we hit an error and it wasn't a cookie error, 662 * we give up. 663 */ 664 if (error) 665 return (error); 666 } 667 668 /* 669 * If not eof and read aheads are enabled, start one. 670 * (You need the current block first, so that you have the 671 * directory offset cookie of the next block.) 672 */ 673 if (nmp->nm_readahead > 0 && 674 (bp->b_flags & B_INVAL) == 0 && 675 (np->n_direofoffset == 0 || 676 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 677 incore(&vp->v_bufobj, lbn + 1) == NULL) { 678 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 679 if (rabp) { 680 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 681 rabp->b_flags |= B_ASYNC; 682 rabp->b_iocmd = BIO_READ; 683 vfs_busy_pages(rabp, 0); 684 if (ncl_asyncio(nmp, rabp, cred, td)) { 685 rabp->b_flags |= B_INVAL; 686 rabp->b_ioflags |= BIO_ERROR; 687 vfs_unbusy_pages(rabp); 688 brelse(rabp); 689 } 690 } else { 691 brelse(rabp); 692 } 693 } 694 } 695 /* 696 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 697 * chopped for the EOF condition, we cannot tell how large 698 * NFS directories are going to be until we hit EOF. So 699 * an NFS directory buffer is *not* chopped to its EOF. Now, 700 * it just so happens that b_resid will effectively chop it 701 * to EOF. *BUT* this information is lost if the buffer goes 702 * away and is reconstituted into a B_CACHE state ( due to 703 * being VMIO ) later. So we keep track of the directory eof 704 * in np->n_direofoffset and chop it off as an extra step 705 * right here. 706 */ 707 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 708 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 709 n = np->n_direofoffset - uio->uio_offset; 710 break; 711 default: 712 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type); 713 bp = NULL; 714 break; 715 }; 716 717 if (n > 0) { 718 error = uiomove(bp->b_data + on, (int)n, uio); 719 } 720 if (vp->v_type == VLNK) 721 n = 0; 722 if (bp != NULL) 723 brelse(bp); 724 } while (error == 0 && uio->uio_resid > 0 && n > 0); 725 return (error); 726 } 727 728 /* 729 * The NFS write path cannot handle iovecs with len > 1. So we need to 730 * break up iovecs accordingly (restricting them to wsize). 731 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 732 * For the ASYNC case, 2 copies are needed. The first a copy from the 733 * user buffer to a staging buffer and then a second copy from the staging 734 * buffer to mbufs. This can be optimized by copying from the user buffer 735 * directly into mbufs and passing the chain down, but that requires a 736 * fair amount of re-working of the relevant codepaths (and can be done 737 * later). 738 */ 739 static int 740 nfs_directio_write(vp, uiop, cred, ioflag) 741 struct vnode *vp; 742 struct uio *uiop; 743 struct ucred *cred; 744 int ioflag; 745 { 746 int error; 747 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 748 struct thread *td = uiop->uio_td; 749 int size; 750 int wsize; 751 752 mtx_lock(&nmp->nm_mtx); 753 wsize = nmp->nm_wsize; 754 mtx_unlock(&nmp->nm_mtx); 755 if (ioflag & IO_SYNC) { 756 int iomode, must_commit; 757 struct uio uio; 758 struct iovec iov; 759 do_sync: 760 while (uiop->uio_resid > 0) { 761 size = min(uiop->uio_resid, wsize); 762 size = min(uiop->uio_iov->iov_len, size); 763 iov.iov_base = uiop->uio_iov->iov_base; 764 iov.iov_len = size; 765 uio.uio_iov = &iov; 766 uio.uio_iovcnt = 1; 767 uio.uio_offset = uiop->uio_offset; 768 uio.uio_resid = size; 769 uio.uio_segflg = UIO_USERSPACE; 770 uio.uio_rw = UIO_WRITE; 771 uio.uio_td = td; 772 iomode = NFSWRITE_FILESYNC; 773 error = ncl_writerpc(vp, &uio, cred, &iomode, 774 &must_commit); 775 KASSERT((must_commit == 0), 776 ("ncl_directio_write: Did not commit write")); 777 if (error) 778 return (error); 779 uiop->uio_offset += size; 780 uiop->uio_resid -= size; 781 if (uiop->uio_iov->iov_len <= size) { 782 uiop->uio_iovcnt--; 783 uiop->uio_iov++; 784 } else { 785 uiop->uio_iov->iov_base = 786 (char *)uiop->uio_iov->iov_base + size; 787 uiop->uio_iov->iov_len -= size; 788 } 789 } 790 } else { 791 struct uio *t_uio; 792 struct iovec *t_iov; 793 struct buf *bp; 794 795 /* 796 * Break up the write into blocksize chunks and hand these 797 * over to nfsiod's for write back. 798 * Unfortunately, this incurs a copy of the data. Since 799 * the user could modify the buffer before the write is 800 * initiated. 801 * 802 * The obvious optimization here is that one of the 2 copies 803 * in the async write path can be eliminated by copying the 804 * data here directly into mbufs and passing the mbuf chain 805 * down. But that will require a fair amount of re-working 806 * of the code and can be done if there's enough interest 807 * in NFS directio access. 808 */ 809 while (uiop->uio_resid > 0) { 810 size = min(uiop->uio_resid, wsize); 811 size = min(uiop->uio_iov->iov_len, size); 812 bp = getpbuf(&ncl_pbuf_freecnt); 813 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 814 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 815 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 816 t_iov->iov_len = size; 817 t_uio->uio_iov = t_iov; 818 t_uio->uio_iovcnt = 1; 819 t_uio->uio_offset = uiop->uio_offset; 820 t_uio->uio_resid = size; 821 t_uio->uio_segflg = UIO_SYSSPACE; 822 t_uio->uio_rw = UIO_WRITE; 823 t_uio->uio_td = td; 824 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); 825 bp->b_flags |= B_DIRECT; 826 bp->b_iocmd = BIO_WRITE; 827 if (cred != NOCRED) { 828 crhold(cred); 829 bp->b_wcred = cred; 830 } else 831 bp->b_wcred = NOCRED; 832 bp->b_caller1 = (void *)t_uio; 833 bp->b_vp = vp; 834 error = ncl_asyncio(nmp, bp, NOCRED, td); 835 if (error) { 836 free(t_iov->iov_base, M_NFSDIRECTIO); 837 free(t_iov, M_NFSDIRECTIO); 838 free(t_uio, M_NFSDIRECTIO); 839 bp->b_vp = NULL; 840 relpbuf(bp, &ncl_pbuf_freecnt); 841 if (error == EINTR) 842 return (error); 843 goto do_sync; 844 } 845 uiop->uio_offset += size; 846 uiop->uio_resid -= size; 847 if (uiop->uio_iov->iov_len <= size) { 848 uiop->uio_iovcnt--; 849 uiop->uio_iov++; 850 } else { 851 uiop->uio_iov->iov_base = 852 (char *)uiop->uio_iov->iov_base + size; 853 uiop->uio_iov->iov_len -= size; 854 } 855 } 856 } 857 return (0); 858 } 859 860 /* 861 * Vnode op for write using bio 862 */ 863 int 864 ncl_write(struct vop_write_args *ap) 865 { 866 int biosize; 867 struct uio *uio = ap->a_uio; 868 struct thread *td = uio->uio_td; 869 struct vnode *vp = ap->a_vp; 870 struct nfsnode *np = VTONFS(vp); 871 struct ucred *cred = ap->a_cred; 872 int ioflag = ap->a_ioflag; 873 struct buf *bp; 874 struct vattr vattr; 875 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 876 daddr_t lbn; 877 int bcount; 878 int n, on, error = 0; 879 struct proc *p = td?td->td_proc:NULL; 880 881 #ifdef DIAGNOSTIC 882 if (uio->uio_rw != UIO_WRITE) 883 panic("ncl_write mode"); 884 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 885 panic("ncl_write proc"); 886 #endif 887 if (vp->v_type != VREG) 888 return (EIO); 889 mtx_lock(&np->n_mtx); 890 if (np->n_flag & NWRITEERR) { 891 np->n_flag &= ~NWRITEERR; 892 mtx_unlock(&np->n_mtx); 893 return (np->n_error); 894 } else 895 mtx_unlock(&np->n_mtx); 896 mtx_lock(&nmp->nm_mtx); 897 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 898 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 899 mtx_unlock(&nmp->nm_mtx); 900 (void)ncl_fsinfo(nmp, vp, cred, td); 901 mtx_lock(&nmp->nm_mtx); 902 } 903 if (nmp->nm_wsize == 0) 904 (void) newnfs_iosize(nmp); 905 mtx_unlock(&nmp->nm_mtx); 906 907 /* 908 * Synchronously flush pending buffers if we are in synchronous 909 * mode or if we are appending. 910 */ 911 if (ioflag & (IO_APPEND | IO_SYNC)) { 912 mtx_lock(&np->n_mtx); 913 if (np->n_flag & NMODIFIED) { 914 mtx_unlock(&np->n_mtx); 915 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 916 /* 917 * Require non-blocking, synchronous writes to 918 * dirty files to inform the program it needs 919 * to fsync(2) explicitly. 920 */ 921 if (ioflag & IO_NDELAY) 922 return (EAGAIN); 923 #endif 924 flush_and_restart: 925 np->n_attrstamp = 0; 926 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 927 if (error) 928 return (error); 929 } else 930 mtx_unlock(&np->n_mtx); 931 } 932 933 /* 934 * If IO_APPEND then load uio_offset. We restart here if we cannot 935 * get the append lock. 936 */ 937 if (ioflag & IO_APPEND) { 938 np->n_attrstamp = 0; 939 error = VOP_GETATTR(vp, &vattr, cred); 940 if (error) 941 return (error); 942 mtx_lock(&np->n_mtx); 943 uio->uio_offset = np->n_size; 944 mtx_unlock(&np->n_mtx); 945 } 946 947 if (uio->uio_offset < 0) 948 return (EINVAL); 949 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 950 return (EFBIG); 951 if (uio->uio_resid == 0) 952 return (0); 953 954 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 955 return nfs_directio_write(vp, uio, cred, ioflag); 956 957 /* 958 * Maybe this should be above the vnode op call, but so long as 959 * file servers have no limits, i don't think it matters 960 */ 961 if (p != NULL) { 962 PROC_LOCK(p); 963 if (uio->uio_offset + uio->uio_resid > 964 lim_cur(p, RLIMIT_FSIZE)) { 965 psignal(p, SIGXFSZ); 966 PROC_UNLOCK(p); 967 return (EFBIG); 968 } 969 PROC_UNLOCK(p); 970 } 971 972 biosize = vp->v_mount->mnt_stat.f_iosize; 973 /* 974 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 975 * would exceed the local maximum per-file write commit size when 976 * combined with those, we must decide whether to flush, 977 * go synchronous, or return error. We don't bother checking 978 * IO_UNIT -- we just make all writes atomic anyway, as there's 979 * no point optimizing for something that really won't ever happen. 980 */ 981 if (!(ioflag & IO_SYNC)) { 982 int nflag; 983 984 mtx_lock(&np->n_mtx); 985 nflag = np->n_flag; 986 mtx_unlock(&np->n_mtx); 987 int needrestart = 0; 988 if (nmp->nm_wcommitsize < uio->uio_resid) { 989 /* 990 * If this request could not possibly be completed 991 * without exceeding the maximum outstanding write 992 * commit size, see if we can convert it into a 993 * synchronous write operation. 994 */ 995 if (ioflag & IO_NDELAY) 996 return (EAGAIN); 997 ioflag |= IO_SYNC; 998 if (nflag & NMODIFIED) 999 needrestart = 1; 1000 } else if (nflag & NMODIFIED) { 1001 int wouldcommit = 0; 1002 BO_LOCK(&vp->v_bufobj); 1003 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 1004 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 1005 b_bobufs) { 1006 if (bp->b_flags & B_NEEDCOMMIT) 1007 wouldcommit += bp->b_bcount; 1008 } 1009 } 1010 BO_UNLOCK(&vp->v_bufobj); 1011 /* 1012 * Since we're not operating synchronously and 1013 * bypassing the buffer cache, we are in a commit 1014 * and holding all of these buffers whether 1015 * transmitted or not. If not limited, this 1016 * will lead to the buffer cache deadlocking, 1017 * as no one else can flush our uncommitted buffers. 1018 */ 1019 wouldcommit += uio->uio_resid; 1020 /* 1021 * If we would initially exceed the maximum 1022 * outstanding write commit size, flush and restart. 1023 */ 1024 if (wouldcommit > nmp->nm_wcommitsize) 1025 needrestart = 1; 1026 } 1027 if (needrestart) 1028 goto flush_and_restart; 1029 } 1030 1031 do { 1032 NFSINCRGLOBAL(newnfsstats.biocache_writes); 1033 lbn = uio->uio_offset / biosize; 1034 on = uio->uio_offset & (biosize-1); 1035 n = min((unsigned)(biosize - on), uio->uio_resid); 1036 again: 1037 /* 1038 * Handle direct append and file extension cases, calculate 1039 * unaligned buffer size. 1040 */ 1041 mtx_lock(&np->n_mtx); 1042 if (uio->uio_offset == np->n_size && n) { 1043 mtx_unlock(&np->n_mtx); 1044 /* 1045 * Get the buffer (in its pre-append state to maintain 1046 * B_CACHE if it was previously set). Resize the 1047 * nfsnode after we have locked the buffer to prevent 1048 * readers from reading garbage. 1049 */ 1050 bcount = on; 1051 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1052 1053 if (bp != NULL) { 1054 long save; 1055 1056 mtx_lock(&np->n_mtx); 1057 np->n_size = uio->uio_offset + n; 1058 np->n_flag |= NMODIFIED; 1059 vnode_pager_setsize(vp, np->n_size); 1060 mtx_unlock(&np->n_mtx); 1061 1062 save = bp->b_flags & B_CACHE; 1063 bcount += n; 1064 allocbuf(bp, bcount); 1065 bp->b_flags |= save; 1066 } 1067 } else { 1068 /* 1069 * Obtain the locked cache block first, and then 1070 * adjust the file's size as appropriate. 1071 */ 1072 bcount = on + n; 1073 if ((off_t)lbn * biosize + bcount < np->n_size) { 1074 if ((off_t)(lbn + 1) * biosize < np->n_size) 1075 bcount = biosize; 1076 else 1077 bcount = np->n_size - (off_t)lbn * biosize; 1078 } 1079 mtx_unlock(&np->n_mtx); 1080 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1081 mtx_lock(&np->n_mtx); 1082 if (uio->uio_offset + n > np->n_size) { 1083 np->n_size = uio->uio_offset + n; 1084 np->n_flag |= NMODIFIED; 1085 vnode_pager_setsize(vp, np->n_size); 1086 } 1087 mtx_unlock(&np->n_mtx); 1088 } 1089 1090 if (!bp) { 1091 error = newnfs_sigintr(nmp, td); 1092 if (!error) 1093 error = EINTR; 1094 break; 1095 } 1096 1097 /* 1098 * Issue a READ if B_CACHE is not set. In special-append 1099 * mode, B_CACHE is based on the buffer prior to the write 1100 * op and is typically set, avoiding the read. If a read 1101 * is required in special append mode, the server will 1102 * probably send us a short-read since we extended the file 1103 * on our end, resulting in b_resid == 0 and, thusly, 1104 * B_CACHE getting set. 1105 * 1106 * We can also avoid issuing the read if the write covers 1107 * the entire buffer. We have to make sure the buffer state 1108 * is reasonable in this case since we will not be initiating 1109 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1110 * more information. 1111 * 1112 * B_CACHE may also be set due to the buffer being cached 1113 * normally. 1114 */ 1115 1116 if (on == 0 && n == bcount) { 1117 bp->b_flags |= B_CACHE; 1118 bp->b_flags &= ~B_INVAL; 1119 bp->b_ioflags &= ~BIO_ERROR; 1120 } 1121 1122 if ((bp->b_flags & B_CACHE) == 0) { 1123 bp->b_iocmd = BIO_READ; 1124 vfs_busy_pages(bp, 0); 1125 error = ncl_doio(vp, bp, cred, td); 1126 if (error) { 1127 brelse(bp); 1128 break; 1129 } 1130 } 1131 if (bp->b_wcred == NOCRED) 1132 bp->b_wcred = crhold(cred); 1133 mtx_lock(&np->n_mtx); 1134 np->n_flag |= NMODIFIED; 1135 mtx_unlock(&np->n_mtx); 1136 1137 /* 1138 * If dirtyend exceeds file size, chop it down. This should 1139 * not normally occur but there is an append race where it 1140 * might occur XXX, so we log it. 1141 * 1142 * If the chopping creates a reverse-indexed or degenerate 1143 * situation with dirtyoff/end, we 0 both of them. 1144 */ 1145 1146 if (bp->b_dirtyend > bcount) { 1147 ncl_printf("NFS append race @%lx:%d\n", 1148 (long)bp->b_blkno * DEV_BSIZE, 1149 bp->b_dirtyend - bcount); 1150 bp->b_dirtyend = bcount; 1151 } 1152 1153 if (bp->b_dirtyoff >= bp->b_dirtyend) 1154 bp->b_dirtyoff = bp->b_dirtyend = 0; 1155 1156 /* 1157 * If the new write will leave a contiguous dirty 1158 * area, just update the b_dirtyoff and b_dirtyend, 1159 * otherwise force a write rpc of the old dirty area. 1160 * 1161 * While it is possible to merge discontiguous writes due to 1162 * our having a B_CACHE buffer ( and thus valid read data 1163 * for the hole), we don't because it could lead to 1164 * significant cache coherency problems with multiple clients, 1165 * especially if locking is implemented later on. 1166 * 1167 * as an optimization we could theoretically maintain 1168 * a linked list of discontinuous areas, but we would still 1169 * have to commit them separately so there isn't much 1170 * advantage to it except perhaps a bit of asynchronization. 1171 */ 1172 1173 if (bp->b_dirtyend > 0 && 1174 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1175 if (bwrite(bp) == EINTR) { 1176 error = EINTR; 1177 break; 1178 } 1179 goto again; 1180 } 1181 1182 error = uiomove((char *)bp->b_data + on, n, uio); 1183 1184 /* 1185 * Since this block is being modified, it must be written 1186 * again and not just committed. Since write clustering does 1187 * not work for the stage 1 data write, only the stage 2 1188 * commit rpc, we have to clear B_CLUSTEROK as well. 1189 */ 1190 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1191 1192 if (error) { 1193 bp->b_ioflags |= BIO_ERROR; 1194 brelse(bp); 1195 break; 1196 } 1197 1198 /* 1199 * Only update dirtyoff/dirtyend if not a degenerate 1200 * condition. 1201 */ 1202 if (n) { 1203 if (bp->b_dirtyend > 0) { 1204 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1205 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1206 } else { 1207 bp->b_dirtyoff = on; 1208 bp->b_dirtyend = on + n; 1209 } 1210 vfs_bio_set_valid(bp, on, n); 1211 } 1212 1213 /* 1214 * If IO_SYNC do bwrite(). 1215 * 1216 * IO_INVAL appears to be unused. The idea appears to be 1217 * to turn off caching in this case. Very odd. XXX 1218 */ 1219 if ((ioflag & IO_SYNC)) { 1220 if (ioflag & IO_INVAL) 1221 bp->b_flags |= B_NOCACHE; 1222 error = bwrite(bp); 1223 if (error) 1224 break; 1225 } else if ((n + on) == biosize) { 1226 bp->b_flags |= B_ASYNC; 1227 (void) ncl_writebp(bp, 0, NULL); 1228 } else { 1229 bdwrite(bp); 1230 } 1231 } while (uio->uio_resid > 0 && n > 0); 1232 1233 return (error); 1234 } 1235 1236 /* 1237 * Get an nfs cache block. 1238 * 1239 * Allocate a new one if the block isn't currently in the cache 1240 * and return the block marked busy. If the calling process is 1241 * interrupted by a signal for an interruptible mount point, return 1242 * NULL. 1243 * 1244 * The caller must carefully deal with the possible B_INVAL state of 1245 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it 1246 * indirectly), so synchronous reads can be issued without worrying about 1247 * the B_INVAL state. We have to be a little more careful when dealing 1248 * with writes (see comments in nfs_write()) when extending a file past 1249 * its EOF. 1250 */ 1251 static struct buf * 1252 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1253 { 1254 struct buf *bp; 1255 struct mount *mp; 1256 struct nfsmount *nmp; 1257 1258 mp = vp->v_mount; 1259 nmp = VFSTONFS(mp); 1260 1261 if (nmp->nm_flag & NFSMNT_INT) { 1262 sigset_t oldset; 1263 1264 newnfs_set_sigmask(td, &oldset); 1265 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0); 1266 newnfs_restore_sigmask(td, &oldset); 1267 while (bp == NULL) { 1268 if (newnfs_sigintr(nmp, td)) 1269 return (NULL); 1270 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1271 } 1272 } else { 1273 bp = getblk(vp, bn, size, 0, 0, 0); 1274 } 1275 1276 if (vp->v_type == VREG) { 1277 int biosize; 1278 1279 biosize = mp->mnt_stat.f_iosize; 1280 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1281 } 1282 return (bp); 1283 } 1284 1285 /* 1286 * Flush and invalidate all dirty buffers. If another process is already 1287 * doing the flush, just wait for completion. 1288 */ 1289 int 1290 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1291 { 1292 struct nfsnode *np = VTONFS(vp); 1293 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1294 int error = 0, slpflag, slptimeo; 1295 int old_lock = 0; 1296 1297 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf"); 1298 1299 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1300 intrflg = 0; 1301 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)) 1302 intrflg = 1; 1303 if (intrflg) { 1304 slpflag = NFS_PCATCH; 1305 slptimeo = 2 * hz; 1306 } else { 1307 slpflag = 0; 1308 slptimeo = 0; 1309 } 1310 1311 old_lock = ncl_upgrade_vnlock(vp); 1312 if (vp->v_iflag & VI_DOOMED) { 1313 /* 1314 * Since vgonel() uses the generic vinvalbuf() to flush 1315 * dirty buffers and it does not call this function, it 1316 * is safe to just return OK when VI_DOOMED is set. 1317 */ 1318 ncl_downgrade_vnlock(vp, old_lock); 1319 return (0); 1320 } 1321 1322 /* 1323 * Now, flush as required. 1324 */ 1325 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1326 VM_OBJECT_LOCK(vp->v_bufobj.bo_object); 1327 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1328 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); 1329 /* 1330 * If the page clean was interrupted, fail the invalidation. 1331 * Not doing so, we run the risk of losing dirty pages in the 1332 * vinvalbuf() call below. 1333 */ 1334 if (intrflg && (error = newnfs_sigintr(nmp, td))) 1335 goto out; 1336 } 1337 1338 error = vinvalbuf(vp, flags, slpflag, 0); 1339 while (error) { 1340 if (intrflg && (error = newnfs_sigintr(nmp, td))) 1341 goto out; 1342 error = vinvalbuf(vp, flags, 0, slptimeo); 1343 } 1344 mtx_lock(&np->n_mtx); 1345 if (np->n_directio_asyncwr == 0) 1346 np->n_flag &= ~NMODIFIED; 1347 mtx_unlock(&np->n_mtx); 1348 out: 1349 ncl_downgrade_vnlock(vp, old_lock); 1350 return error; 1351 } 1352 1353 /* 1354 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1355 * This is mainly to avoid queueing async I/O requests when the nfsiods 1356 * are all hung on a dead server. 1357 * 1358 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1359 * is eventually dequeued by the async daemon, ncl_doio() *will*. 1360 */ 1361 int 1362 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1363 { 1364 int iod; 1365 int gotiod; 1366 int slpflag = 0; 1367 int slptimeo = 0; 1368 int error, error2; 1369 1370 /* 1371 * Unless iothreadcnt is set > 0, don't bother with async I/O 1372 * threads. For LAN environments, they don't buy any significant 1373 * performance improvement that you can't get with large block 1374 * sizes. 1375 */ 1376 if (nmp->nm_readahead == 0) 1377 return (EPERM); 1378 1379 /* 1380 * Commits are usually short and sweet so lets save some cpu and 1381 * leave the async daemons for more important rpc's (such as reads 1382 * and writes). 1383 */ 1384 mtx_lock(&ncl_iod_mutex); 1385 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1386 (nmp->nm_bufqiods > ncl_numasync / 2)) { 1387 mtx_unlock(&ncl_iod_mutex); 1388 return(EIO); 1389 } 1390 again: 1391 if (nmp->nm_flag & NFSMNT_INT) 1392 slpflag = NFS_PCATCH; 1393 gotiod = FALSE; 1394 1395 /* 1396 * Find a free iod to process this request. 1397 */ 1398 for (iod = 0; iod < ncl_numasync; iod++) 1399 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) { 1400 gotiod = TRUE; 1401 break; 1402 } 1403 1404 /* 1405 * Try to create one if none are free. 1406 */ 1407 if (!gotiod) { 1408 iod = ncl_nfsiodnew(1); 1409 if (iod != -1) 1410 gotiod = TRUE; 1411 } 1412 1413 if (gotiod) { 1414 /* 1415 * Found one, so wake it up and tell it which 1416 * mount to process. 1417 */ 1418 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n", 1419 iod, nmp)); 1420 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE; 1421 ncl_iodmount[iod] = nmp; 1422 nmp->nm_bufqiods++; 1423 wakeup(&ncl_iodwant[iod]); 1424 } 1425 1426 /* 1427 * If none are free, we may already have an iod working on this mount 1428 * point. If so, it will process our request. 1429 */ 1430 if (!gotiod) { 1431 if (nmp->nm_bufqiods > 0) { 1432 NFS_DPF(ASYNCIO, 1433 ("ncl_asyncio: %d iods are already processing mount %p\n", 1434 nmp->nm_bufqiods, nmp)); 1435 gotiod = TRUE; 1436 } 1437 } 1438 1439 /* 1440 * If we have an iod which can process the request, then queue 1441 * the buffer. 1442 */ 1443 if (gotiod) { 1444 /* 1445 * Ensure that the queue never grows too large. We still want 1446 * to asynchronize so we block rather then return EIO. 1447 */ 1448 while (nmp->nm_bufqlen >= 2*ncl_numasync) { 1449 NFS_DPF(ASYNCIO, 1450 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp)); 1451 nmp->nm_bufqwant = TRUE; 1452 error = newnfs_msleep(td, &nmp->nm_bufq, 1453 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio", 1454 slptimeo); 1455 if (error) { 1456 error2 = newnfs_sigintr(nmp, td); 1457 if (error2) { 1458 mtx_unlock(&ncl_iod_mutex); 1459 return (error2); 1460 } 1461 if (slpflag == NFS_PCATCH) { 1462 slpflag = 0; 1463 slptimeo = 2 * hz; 1464 } 1465 } 1466 /* 1467 * We might have lost our iod while sleeping, 1468 * so check and loop if nescessary. 1469 */ 1470 if (nmp->nm_bufqiods == 0) { 1471 NFS_DPF(ASYNCIO, 1472 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1473 goto again; 1474 } 1475 } 1476 1477 /* We might have lost our nfsiod */ 1478 if (nmp->nm_bufqiods == 0) { 1479 NFS_DPF(ASYNCIO, 1480 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1481 goto again; 1482 } 1483 1484 if (bp->b_iocmd == BIO_READ) { 1485 if (bp->b_rcred == NOCRED && cred != NOCRED) 1486 bp->b_rcred = crhold(cred); 1487 } else { 1488 if (bp->b_wcred == NOCRED && cred != NOCRED) 1489 bp->b_wcred = crhold(cred); 1490 } 1491 1492 if (bp->b_flags & B_REMFREE) 1493 bremfreef(bp); 1494 BUF_KERNPROC(bp); 1495 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1496 nmp->nm_bufqlen++; 1497 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1498 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1499 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1500 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1501 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1502 } 1503 mtx_unlock(&ncl_iod_mutex); 1504 return (0); 1505 } 1506 1507 mtx_unlock(&ncl_iod_mutex); 1508 1509 /* 1510 * All the iods are busy on other mounts, so return EIO to 1511 * force the caller to process the i/o synchronously. 1512 */ 1513 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n")); 1514 return (EIO); 1515 } 1516 1517 void 1518 ncl_doio_directwrite(struct buf *bp) 1519 { 1520 int iomode, must_commit; 1521 struct uio *uiop = (struct uio *)bp->b_caller1; 1522 char *iov_base = uiop->uio_iov->iov_base; 1523 1524 iomode = NFSWRITE_FILESYNC; 1525 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1526 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit); 1527 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write")); 1528 free(iov_base, M_NFSDIRECTIO); 1529 free(uiop->uio_iov, M_NFSDIRECTIO); 1530 free(uiop, M_NFSDIRECTIO); 1531 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1532 struct nfsnode *np = VTONFS(bp->b_vp); 1533 mtx_lock(&np->n_mtx); 1534 np->n_directio_asyncwr--; 1535 if (np->n_directio_asyncwr == 0) { 1536 np->n_flag &= ~NMODIFIED; 1537 if ((np->n_flag & NFSYNCWAIT)) { 1538 np->n_flag &= ~NFSYNCWAIT; 1539 wakeup((caddr_t)&np->n_directio_asyncwr); 1540 } 1541 } 1542 mtx_unlock(&np->n_mtx); 1543 } 1544 bp->b_vp = NULL; 1545 relpbuf(bp, &ncl_pbuf_freecnt); 1546 } 1547 1548 /* 1549 * Do an I/O operation to/from a cache block. This may be called 1550 * synchronously or from an nfsiod. 1551 */ 1552 int 1553 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) 1554 { 1555 struct uio *uiop; 1556 struct nfsnode *np; 1557 struct nfsmount *nmp; 1558 int error = 0, iomode, must_commit = 0; 1559 struct uio uio; 1560 struct iovec io; 1561 struct proc *p = td ? td->td_proc : NULL; 1562 uint8_t iocmd; 1563 1564 np = VTONFS(vp); 1565 nmp = VFSTONFS(vp->v_mount); 1566 uiop = &uio; 1567 uiop->uio_iov = &io; 1568 uiop->uio_iovcnt = 1; 1569 uiop->uio_segflg = UIO_SYSSPACE; 1570 uiop->uio_td = td; 1571 1572 /* 1573 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1574 * do this here so we do not have to do it in all the code that 1575 * calls us. 1576 */ 1577 bp->b_flags &= ~B_INVAL; 1578 bp->b_ioflags &= ~BIO_ERROR; 1579 1580 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp)); 1581 iocmd = bp->b_iocmd; 1582 if (iocmd == BIO_READ) { 1583 io.iov_len = uiop->uio_resid = bp->b_bcount; 1584 io.iov_base = bp->b_data; 1585 uiop->uio_rw = UIO_READ; 1586 1587 switch (vp->v_type) { 1588 case VREG: 1589 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1590 NFSINCRGLOBAL(newnfsstats.read_bios); 1591 error = ncl_readrpc(vp, uiop, cr); 1592 1593 if (!error) { 1594 if (uiop->uio_resid) { 1595 /* 1596 * If we had a short read with no error, we must have 1597 * hit a file hole. We should zero-fill the remainder. 1598 * This can also occur if the server hits the file EOF. 1599 * 1600 * Holes used to be able to occur due to pending 1601 * writes, but that is not possible any longer. 1602 */ 1603 int nread = bp->b_bcount - uiop->uio_resid; 1604 int left = uiop->uio_resid; 1605 1606 if (left > 0) 1607 bzero((char *)bp->b_data + nread, left); 1608 uiop->uio_resid = 0; 1609 } 1610 } 1611 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */ 1612 if (p && (vp->v_vflag & VV_TEXT)) { 1613 mtx_lock(&np->n_mtx); 1614 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) { 1615 mtx_unlock(&np->n_mtx); 1616 PROC_LOCK(p); 1617 killproc(p, "text file modification"); 1618 PROC_UNLOCK(p); 1619 } else 1620 mtx_unlock(&np->n_mtx); 1621 } 1622 break; 1623 case VLNK: 1624 uiop->uio_offset = (off_t)0; 1625 NFSINCRGLOBAL(newnfsstats.readlink_bios); 1626 error = ncl_readlinkrpc(vp, uiop, cr); 1627 break; 1628 case VDIR: 1629 NFSINCRGLOBAL(newnfsstats.readdir_bios); 1630 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1631 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1632 error = ncl_readdirplusrpc(vp, uiop, cr, td); 1633 if (error == NFSERR_NOTSUPP) 1634 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1635 } 1636 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1637 error = ncl_readdirrpc(vp, uiop, cr, td); 1638 /* 1639 * end-of-directory sets B_INVAL but does not generate an 1640 * error. 1641 */ 1642 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1643 bp->b_flags |= B_INVAL; 1644 break; 1645 default: 1646 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type); 1647 break; 1648 }; 1649 if (error) { 1650 bp->b_ioflags |= BIO_ERROR; 1651 bp->b_error = error; 1652 } 1653 } else { 1654 /* 1655 * If we only need to commit, try to commit 1656 */ 1657 if (bp->b_flags & B_NEEDCOMMIT) { 1658 int retv; 1659 off_t off; 1660 1661 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1662 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1663 bp->b_wcred, td); 1664 if (retv == 0) { 1665 bp->b_dirtyoff = bp->b_dirtyend = 0; 1666 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1667 bp->b_resid = 0; 1668 bufdone(bp); 1669 return (0); 1670 } 1671 if (retv == NFSERR_STALEWRITEVERF) { 1672 ncl_clearcommit(vp->v_mount); 1673 } 1674 } 1675 1676 /* 1677 * Setup for actual write 1678 */ 1679 mtx_lock(&np->n_mtx); 1680 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1681 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1682 mtx_unlock(&np->n_mtx); 1683 1684 if (bp->b_dirtyend > bp->b_dirtyoff) { 1685 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1686 - bp->b_dirtyoff; 1687 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1688 + bp->b_dirtyoff; 1689 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1690 uiop->uio_rw = UIO_WRITE; 1691 NFSINCRGLOBAL(newnfsstats.write_bios); 1692 1693 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1694 iomode = NFSWRITE_UNSTABLE; 1695 else 1696 iomode = NFSWRITE_FILESYNC; 1697 1698 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit); 1699 1700 /* 1701 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1702 * to cluster the buffers needing commit. This will allow 1703 * the system to submit a single commit rpc for the whole 1704 * cluster. We can do this even if the buffer is not 100% 1705 * dirty (relative to the NFS blocksize), so we optimize the 1706 * append-to-file-case. 1707 * 1708 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1709 * cleared because write clustering only works for commit 1710 * rpc's, not for the data portion of the write). 1711 */ 1712 1713 if (!error && iomode == NFSWRITE_UNSTABLE) { 1714 bp->b_flags |= B_NEEDCOMMIT; 1715 if (bp->b_dirtyoff == 0 1716 && bp->b_dirtyend == bp->b_bcount) 1717 bp->b_flags |= B_CLUSTEROK; 1718 } else { 1719 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1720 } 1721 1722 /* 1723 * For an interrupted write, the buffer is still valid 1724 * and the write hasn't been pushed to the server yet, 1725 * so we can't set BIO_ERROR and report the interruption 1726 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1727 * is not relevant, so the rpc attempt is essentially 1728 * a noop. For the case of a V3 write rpc not being 1729 * committed to stable storage, the block is still 1730 * dirty and requires either a commit rpc or another 1731 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1732 * the block is reused. This is indicated by setting 1733 * the B_DELWRI and B_NEEDCOMMIT flags. 1734 * 1735 * If the buffer is marked B_PAGING, it does not reside on 1736 * the vp's paging queues so we cannot call bdirty(). The 1737 * bp in this case is not an NFS cache block so we should 1738 * be safe. XXX 1739 * 1740 * The logic below breaks up errors into recoverable and 1741 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1742 * and keep the buffer around for potential write retries. 1743 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1744 * and save the error in the nfsnode. This is less than ideal 1745 * but necessary. Keeping such buffers around could potentially 1746 * cause buffer exhaustion eventually (they can never be written 1747 * out, so will get constantly be re-dirtied). It also causes 1748 * all sorts of vfs panics. For non-recoverable write errors, 1749 * also invalidate the attrcache, so we'll be forced to go over 1750 * the wire for this object, returning an error to user on next 1751 * call (most of the time). 1752 */ 1753 if (error == EINTR || error == EIO || error == ETIMEDOUT 1754 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1755 int s; 1756 1757 s = splbio(); 1758 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1759 if ((bp->b_flags & B_PAGING) == 0) { 1760 bdirty(bp); 1761 bp->b_flags &= ~B_DONE; 1762 } 1763 if (error && (bp->b_flags & B_ASYNC) == 0) 1764 bp->b_flags |= B_EINTR; 1765 splx(s); 1766 } else { 1767 if (error) { 1768 bp->b_ioflags |= BIO_ERROR; 1769 bp->b_flags |= B_INVAL; 1770 bp->b_error = np->n_error = error; 1771 mtx_lock(&np->n_mtx); 1772 np->n_flag |= NWRITEERR; 1773 np->n_attrstamp = 0; 1774 mtx_unlock(&np->n_mtx); 1775 } 1776 bp->b_dirtyoff = bp->b_dirtyend = 0; 1777 } 1778 } else { 1779 bp->b_resid = 0; 1780 bufdone(bp); 1781 return (0); 1782 } 1783 } 1784 bp->b_resid = uiop->uio_resid; 1785 if (must_commit) 1786 ncl_clearcommit(vp->v_mount); 1787 bufdone(bp); 1788 return (error); 1789 } 1790 1791 /* 1792 * Used to aid in handling ftruncate() operations on the NFS client side. 1793 * Truncation creates a number of special problems for NFS. We have to 1794 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1795 * we have to properly handle VM pages or (potentially dirty) buffers 1796 * that straddle the truncation point. 1797 */ 1798 1799 int 1800 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1801 { 1802 struct nfsnode *np = VTONFS(vp); 1803 u_quad_t tsize; 1804 int biosize = vp->v_mount->mnt_stat.f_iosize; 1805 int error = 0; 1806 1807 mtx_lock(&np->n_mtx); 1808 tsize = np->n_size; 1809 np->n_size = nsize; 1810 mtx_unlock(&np->n_mtx); 1811 1812 if (nsize < tsize) { 1813 struct buf *bp; 1814 daddr_t lbn; 1815 int bufsize; 1816 1817 /* 1818 * vtruncbuf() doesn't get the buffer overlapping the 1819 * truncation point. We may have a B_DELWRI and/or B_CACHE 1820 * buffer that now needs to be truncated. 1821 */ 1822 error = vtruncbuf(vp, cred, td, nsize, biosize); 1823 lbn = nsize / biosize; 1824 bufsize = nsize & (biosize - 1); 1825 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1826 if (!bp) 1827 return EINTR; 1828 if (bp->b_dirtyoff > bp->b_bcount) 1829 bp->b_dirtyoff = bp->b_bcount; 1830 if (bp->b_dirtyend > bp->b_bcount) 1831 bp->b_dirtyend = bp->b_bcount; 1832 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1833 brelse(bp); 1834 } else { 1835 vnode_pager_setsize(vp, nsize); 1836 } 1837 return(error); 1838 } 1839 1840