nfs_clbio.c revision 239554
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: stable/9/sys/fs/nfsclient/nfs_clbio.c 239554 2012-08-22 05:15:21Z kib $"); 37 38#include "opt_kdtrace.h" 39 40#include <sys/param.h> 41#include <sys/systm.h> 42#include <sys/bio.h> 43#include <sys/buf.h> 44#include <sys/kernel.h> 45#include <sys/mount.h> 46#include <sys/vmmeter.h> 47#include <sys/vnode.h> 48 49#include <vm/vm.h> 50#include <vm/vm_extern.h> 51#include <vm/vm_page.h> 52#include <vm/vm_object.h> 53#include <vm/vm_pager.h> 54#include <vm/vnode_pager.h> 55 56#include <fs/nfs/nfsport.h> 57#include <fs/nfsclient/nfsmount.h> 58#include <fs/nfsclient/nfs.h> 59#include <fs/nfsclient/nfsnode.h> 60#include <fs/nfsclient/nfs_kdtrace.h> 61 62extern int newnfs_directio_allow_mmap; 63extern struct nfsstats newnfsstats; 64extern struct mtx ncl_iod_mutex; 65extern int ncl_numasync; 66extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON]; 67extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON]; 68extern int newnfs_directio_enable; 69extern int nfs_keep_dirty_on_error; 70 71int ncl_pbuf_freecnt = -1; /* start out unlimited */ 72 73static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 74 struct thread *td); 75static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 76 struct ucred *cred, int ioflag); 77 78/* 79 * Vnode op for VM getpages. 80 */ 81int 82ncl_getpages(struct vop_getpages_args *ap) 83{ 84 int i, error, nextoff, size, toff, count, npages; 85 struct uio uio; 86 struct iovec iov; 87 vm_offset_t kva; 88 struct buf *bp; 89 struct vnode *vp; 90 struct thread *td; 91 struct ucred *cred; 92 struct nfsmount *nmp; 93 vm_object_t object; 94 vm_page_t *pages; 95 struct nfsnode *np; 96 97 vp = ap->a_vp; 98 np = VTONFS(vp); 99 td = curthread; /* XXX */ 100 cred = curthread->td_ucred; /* XXX */ 101 nmp = VFSTONFS(vp->v_mount); 102 pages = ap->a_m; 103 count = ap->a_count; 104 105 if ((object = vp->v_object) == NULL) { 106 ncl_printf("nfs_getpages: called with non-merged cache vnode??\n"); 107 return (VM_PAGER_ERROR); 108 } 109 110 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) { 111 mtx_lock(&np->n_mtx); 112 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 113 mtx_unlock(&np->n_mtx); 114 ncl_printf("nfs_getpages: called on non-cacheable vnode??\n"); 115 return (VM_PAGER_ERROR); 116 } else 117 mtx_unlock(&np->n_mtx); 118 } 119 120 mtx_lock(&nmp->nm_mtx); 121 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 122 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 123 mtx_unlock(&nmp->nm_mtx); 124 /* We'll never get here for v4, because we always have fsinfo */ 125 (void)ncl_fsinfo(nmp, vp, cred, td); 126 } else 127 mtx_unlock(&nmp->nm_mtx); 128 129 npages = btoc(count); 130 131 /* 132 * If the requested page is partially valid, just return it and 133 * allow the pager to zero-out the blanks. Partially valid pages 134 * can only occur at the file EOF. 135 */ 136 VM_OBJECT_LOCK(object); 137 if (pages[ap->a_reqpage]->valid != 0) { 138 for (i = 0; i < npages; ++i) { 139 if (i != ap->a_reqpage) { 140 vm_page_lock(pages[i]); 141 vm_page_free(pages[i]); 142 vm_page_unlock(pages[i]); 143 } 144 } 145 VM_OBJECT_UNLOCK(object); 146 return (0); 147 } 148 VM_OBJECT_UNLOCK(object); 149 150 /* 151 * We use only the kva address for the buffer, but this is extremely 152 * convienient and fast. 153 */ 154 bp = getpbuf(&ncl_pbuf_freecnt); 155 156 kva = (vm_offset_t) bp->b_data; 157 pmap_qenter(kva, pages, npages); 158 PCPU_INC(cnt.v_vnodein); 159 PCPU_ADD(cnt.v_vnodepgsin, npages); 160 161 iov.iov_base = (caddr_t) kva; 162 iov.iov_len = count; 163 uio.uio_iov = &iov; 164 uio.uio_iovcnt = 1; 165 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 166 uio.uio_resid = count; 167 uio.uio_segflg = UIO_SYSSPACE; 168 uio.uio_rw = UIO_READ; 169 uio.uio_td = td; 170 171 error = ncl_readrpc(vp, &uio, cred); 172 pmap_qremove(kva, npages); 173 174 relpbuf(bp, &ncl_pbuf_freecnt); 175 176 if (error && (uio.uio_resid == count)) { 177 ncl_printf("nfs_getpages: error %d\n", error); 178 VM_OBJECT_LOCK(object); 179 for (i = 0; i < npages; ++i) { 180 if (i != ap->a_reqpage) { 181 vm_page_lock(pages[i]); 182 vm_page_free(pages[i]); 183 vm_page_unlock(pages[i]); 184 } 185 } 186 VM_OBJECT_UNLOCK(object); 187 return (VM_PAGER_ERROR); 188 } 189 190 /* 191 * Calculate the number of bytes read and validate only that number 192 * of bytes. Note that due to pending writes, size may be 0. This 193 * does not mean that the remaining data is invalid! 194 */ 195 196 size = count - uio.uio_resid; 197 VM_OBJECT_LOCK(object); 198 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 199 vm_page_t m; 200 nextoff = toff + PAGE_SIZE; 201 m = pages[i]; 202 203 if (nextoff <= size) { 204 /* 205 * Read operation filled an entire page 206 */ 207 m->valid = VM_PAGE_BITS_ALL; 208 KASSERT(m->dirty == 0, 209 ("nfs_getpages: page %p is dirty", m)); 210 } else if (size > toff) { 211 /* 212 * Read operation filled a partial page. 213 */ 214 m->valid = 0; 215 vm_page_set_valid(m, 0, size - toff); 216 KASSERT(m->dirty == 0, 217 ("nfs_getpages: page %p is dirty", m)); 218 } else { 219 /* 220 * Read operation was short. If no error 221 * occured we may have hit a zero-fill 222 * section. We leave valid set to 0, and page 223 * is freed by vm_page_readahead_finish() if 224 * its index is not equal to requested, or 225 * page is zeroed and set valid by 226 * vm_pager_get_pages() for requested page. 227 */ 228 ; 229 } 230 if (i != ap->a_reqpage) 231 vm_page_readahead_finish(m); 232 } 233 VM_OBJECT_UNLOCK(object); 234 return (0); 235} 236 237/* 238 * Vnode op for VM putpages. 239 */ 240int 241ncl_putpages(struct vop_putpages_args *ap) 242{ 243 struct uio uio; 244 struct iovec iov; 245 vm_offset_t kva; 246 struct buf *bp; 247 int iomode, must_commit, i, error, npages, count; 248 off_t offset; 249 int *rtvals; 250 struct vnode *vp; 251 struct thread *td; 252 struct ucred *cred; 253 struct nfsmount *nmp; 254 struct nfsnode *np; 255 vm_page_t *pages; 256 257 vp = ap->a_vp; 258 np = VTONFS(vp); 259 td = curthread; /* XXX */ 260 /* Set the cred to n_writecred for the write rpcs. */ 261 if (np->n_writecred != NULL) 262 cred = crhold(np->n_writecred); 263 else 264 cred = crhold(curthread->td_ucred); /* XXX */ 265 nmp = VFSTONFS(vp->v_mount); 266 pages = ap->a_m; 267 count = ap->a_count; 268 rtvals = ap->a_rtvals; 269 npages = btoc(count); 270 offset = IDX_TO_OFF(pages[0]->pindex); 271 272 mtx_lock(&nmp->nm_mtx); 273 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 274 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 275 mtx_unlock(&nmp->nm_mtx); 276 (void)ncl_fsinfo(nmp, vp, cred, td); 277 } else 278 mtx_unlock(&nmp->nm_mtx); 279 280 mtx_lock(&np->n_mtx); 281 if (newnfs_directio_enable && !newnfs_directio_allow_mmap && 282 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 283 mtx_unlock(&np->n_mtx); 284 ncl_printf("ncl_putpages: called on noncache-able vnode??\n"); 285 mtx_lock(&np->n_mtx); 286 } 287 288 for (i = 0; i < npages; i++) 289 rtvals[i] = VM_PAGER_ERROR; 290 291 /* 292 * When putting pages, do not extend file past EOF. 293 */ 294 if (offset + count > np->n_size) { 295 count = np->n_size - offset; 296 if (count < 0) 297 count = 0; 298 } 299 mtx_unlock(&np->n_mtx); 300 301 /* 302 * We use only the kva address for the buffer, but this is extremely 303 * convienient and fast. 304 */ 305 bp = getpbuf(&ncl_pbuf_freecnt); 306 307 kva = (vm_offset_t) bp->b_data; 308 pmap_qenter(kva, pages, npages); 309 PCPU_INC(cnt.v_vnodeout); 310 PCPU_ADD(cnt.v_vnodepgsout, count); 311 312 iov.iov_base = (caddr_t) kva; 313 iov.iov_len = count; 314 uio.uio_iov = &iov; 315 uio.uio_iovcnt = 1; 316 uio.uio_offset = offset; 317 uio.uio_resid = count; 318 uio.uio_segflg = UIO_SYSSPACE; 319 uio.uio_rw = UIO_WRITE; 320 uio.uio_td = td; 321 322 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 323 iomode = NFSWRITE_UNSTABLE; 324 else 325 iomode = NFSWRITE_FILESYNC; 326 327 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0); 328 crfree(cred); 329 330 pmap_qremove(kva, npages); 331 relpbuf(bp, &ncl_pbuf_freecnt); 332 333 if (error == 0 || !nfs_keep_dirty_on_error) { 334 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid); 335 if (must_commit) 336 ncl_clearcommit(vp->v_mount); 337 } 338 return rtvals[0]; 339} 340 341/* 342 * For nfs, cache consistency can only be maintained approximately. 343 * Although RFC1094 does not specify the criteria, the following is 344 * believed to be compatible with the reference port. 345 * For nfs: 346 * If the file's modify time on the server has changed since the 347 * last read rpc or you have written to the file, 348 * you may have lost data cache consistency with the 349 * server, so flush all of the file's data out of the cache. 350 * Then force a getattr rpc to ensure that you have up to date 351 * attributes. 352 * NB: This implies that cache data can be read when up to 353 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 354 * attributes this could be forced by setting n_attrstamp to 0 before 355 * the VOP_GETATTR() call. 356 */ 357static inline int 358nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 359{ 360 int error = 0; 361 struct vattr vattr; 362 struct nfsnode *np = VTONFS(vp); 363 int old_lock; 364 365 /* 366 * Grab the exclusive lock before checking whether the cache is 367 * consistent. 368 * XXX - We can make this cheaper later (by acquiring cheaper locks). 369 * But for now, this suffices. 370 */ 371 old_lock = ncl_upgrade_vnlock(vp); 372 if (vp->v_iflag & VI_DOOMED) { 373 ncl_downgrade_vnlock(vp, old_lock); 374 return (EBADF); 375 } 376 377 mtx_lock(&np->n_mtx); 378 if (np->n_flag & NMODIFIED) { 379 mtx_unlock(&np->n_mtx); 380 if (vp->v_type != VREG) { 381 if (vp->v_type != VDIR) 382 panic("nfs: bioread, not dir"); 383 ncl_invaldir(vp); 384 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 385 if (error) 386 goto out; 387 } 388 np->n_attrstamp = 0; 389 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 390 error = VOP_GETATTR(vp, &vattr, cred); 391 if (error) 392 goto out; 393 mtx_lock(&np->n_mtx); 394 np->n_mtime = vattr.va_mtime; 395 mtx_unlock(&np->n_mtx); 396 } else { 397 mtx_unlock(&np->n_mtx); 398 error = VOP_GETATTR(vp, &vattr, cred); 399 if (error) 400 return (error); 401 mtx_lock(&np->n_mtx); 402 if ((np->n_flag & NSIZECHANGED) 403 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 404 mtx_unlock(&np->n_mtx); 405 if (vp->v_type == VDIR) 406 ncl_invaldir(vp); 407 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 408 if (error) 409 goto out; 410 mtx_lock(&np->n_mtx); 411 np->n_mtime = vattr.va_mtime; 412 np->n_flag &= ~NSIZECHANGED; 413 } 414 mtx_unlock(&np->n_mtx); 415 } 416out: 417 ncl_downgrade_vnlock(vp, old_lock); 418 return error; 419} 420 421/* 422 * Vnode op for read using bio 423 */ 424int 425ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 426{ 427 struct nfsnode *np = VTONFS(vp); 428 int biosize, i; 429 struct buf *bp, *rabp; 430 struct thread *td; 431 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 432 daddr_t lbn, rabn; 433 int bcount; 434 int seqcount; 435 int nra, error = 0, n = 0, on = 0; 436 off_t tmp_off; 437 438 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode")); 439 if (uio->uio_resid == 0) 440 return (0); 441 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 442 return (EINVAL); 443 td = uio->uio_td; 444 445 mtx_lock(&nmp->nm_mtx); 446 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 447 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 448 mtx_unlock(&nmp->nm_mtx); 449 (void)ncl_fsinfo(nmp, vp, cred, td); 450 mtx_lock(&nmp->nm_mtx); 451 } 452 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0) 453 (void) newnfs_iosize(nmp); 454 455 tmp_off = uio->uio_offset + uio->uio_resid; 456 if (vp->v_type != VDIR && 457 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) { 458 mtx_unlock(&nmp->nm_mtx); 459 return (EFBIG); 460 } 461 mtx_unlock(&nmp->nm_mtx); 462 463 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 464 /* No caching/ no readaheads. Just read data into the user buffer */ 465 return ncl_readrpc(vp, uio, cred); 466 467 biosize = vp->v_bufobj.bo_bsize; 468 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 469 470 error = nfs_bioread_check_cons(vp, td, cred); 471 if (error) 472 return error; 473 474 do { 475 u_quad_t nsize; 476 477 mtx_lock(&np->n_mtx); 478 nsize = np->n_size; 479 mtx_unlock(&np->n_mtx); 480 481 switch (vp->v_type) { 482 case VREG: 483 NFSINCRGLOBAL(newnfsstats.biocache_reads); 484 lbn = uio->uio_offset / biosize; 485 on = uio->uio_offset & (biosize - 1); 486 487 /* 488 * Start the read ahead(s), as required. 489 */ 490 if (nmp->nm_readahead > 0) { 491 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 492 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 493 rabn = lbn + 1 + nra; 494 if (incore(&vp->v_bufobj, rabn) == NULL) { 495 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 496 if (!rabp) { 497 error = newnfs_sigintr(nmp, td); 498 return (error ? error : EINTR); 499 } 500 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 501 rabp->b_flags |= B_ASYNC; 502 rabp->b_iocmd = BIO_READ; 503 vfs_busy_pages(rabp, 0); 504 if (ncl_asyncio(nmp, rabp, cred, td)) { 505 rabp->b_flags |= B_INVAL; 506 rabp->b_ioflags |= BIO_ERROR; 507 vfs_unbusy_pages(rabp); 508 brelse(rabp); 509 break; 510 } 511 } else { 512 brelse(rabp); 513 } 514 } 515 } 516 } 517 518 /* Note that bcount is *not* DEV_BSIZE aligned. */ 519 bcount = biosize; 520 if ((off_t)lbn * biosize >= nsize) { 521 bcount = 0; 522 } else if ((off_t)(lbn + 1) * biosize > nsize) { 523 bcount = nsize - (off_t)lbn * biosize; 524 } 525 bp = nfs_getcacheblk(vp, lbn, bcount, td); 526 527 if (!bp) { 528 error = newnfs_sigintr(nmp, td); 529 return (error ? error : EINTR); 530 } 531 532 /* 533 * If B_CACHE is not set, we must issue the read. If this 534 * fails, we return an error. 535 */ 536 537 if ((bp->b_flags & B_CACHE) == 0) { 538 bp->b_iocmd = BIO_READ; 539 vfs_busy_pages(bp, 0); 540 error = ncl_doio(vp, bp, cred, td, 0); 541 if (error) { 542 brelse(bp); 543 return (error); 544 } 545 } 546 547 /* 548 * on is the offset into the current bp. Figure out how many 549 * bytes we can copy out of the bp. Note that bcount is 550 * NOT DEV_BSIZE aligned. 551 * 552 * Then figure out how many bytes we can copy into the uio. 553 */ 554 555 n = 0; 556 if (on < bcount) 557 n = MIN((unsigned)(bcount - on), uio->uio_resid); 558 break; 559 case VLNK: 560 NFSINCRGLOBAL(newnfsstats.biocache_readlinks); 561 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 562 if (!bp) { 563 error = newnfs_sigintr(nmp, td); 564 return (error ? error : EINTR); 565 } 566 if ((bp->b_flags & B_CACHE) == 0) { 567 bp->b_iocmd = BIO_READ; 568 vfs_busy_pages(bp, 0); 569 error = ncl_doio(vp, bp, cred, td, 0); 570 if (error) { 571 bp->b_ioflags |= BIO_ERROR; 572 brelse(bp); 573 return (error); 574 } 575 } 576 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 577 on = 0; 578 break; 579 case VDIR: 580 NFSINCRGLOBAL(newnfsstats.biocache_readdirs); 581 if (np->n_direofoffset 582 && uio->uio_offset >= np->n_direofoffset) { 583 return (0); 584 } 585 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 586 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 587 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 588 if (!bp) { 589 error = newnfs_sigintr(nmp, td); 590 return (error ? error : EINTR); 591 } 592 if ((bp->b_flags & B_CACHE) == 0) { 593 bp->b_iocmd = BIO_READ; 594 vfs_busy_pages(bp, 0); 595 error = ncl_doio(vp, bp, cred, td, 0); 596 if (error) { 597 brelse(bp); 598 } 599 while (error == NFSERR_BAD_COOKIE) { 600 ncl_invaldir(vp); 601 error = ncl_vinvalbuf(vp, 0, td, 1); 602 /* 603 * Yuck! The directory has been modified on the 604 * server. The only way to get the block is by 605 * reading from the beginning to get all the 606 * offset cookies. 607 * 608 * Leave the last bp intact unless there is an error. 609 * Loop back up to the while if the error is another 610 * NFSERR_BAD_COOKIE (double yuch!). 611 */ 612 for (i = 0; i <= lbn && !error; i++) { 613 if (np->n_direofoffset 614 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 615 return (0); 616 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 617 if (!bp) { 618 error = newnfs_sigintr(nmp, td); 619 return (error ? error : EINTR); 620 } 621 if ((bp->b_flags & B_CACHE) == 0) { 622 bp->b_iocmd = BIO_READ; 623 vfs_busy_pages(bp, 0); 624 error = ncl_doio(vp, bp, cred, td, 0); 625 /* 626 * no error + B_INVAL == directory EOF, 627 * use the block. 628 */ 629 if (error == 0 && (bp->b_flags & B_INVAL)) 630 break; 631 } 632 /* 633 * An error will throw away the block and the 634 * for loop will break out. If no error and this 635 * is not the block we want, we throw away the 636 * block and go for the next one via the for loop. 637 */ 638 if (error || i < lbn) 639 brelse(bp); 640 } 641 } 642 /* 643 * The above while is repeated if we hit another cookie 644 * error. If we hit an error and it wasn't a cookie error, 645 * we give up. 646 */ 647 if (error) 648 return (error); 649 } 650 651 /* 652 * If not eof and read aheads are enabled, start one. 653 * (You need the current block first, so that you have the 654 * directory offset cookie of the next block.) 655 */ 656 if (nmp->nm_readahead > 0 && 657 (bp->b_flags & B_INVAL) == 0 && 658 (np->n_direofoffset == 0 || 659 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 660 incore(&vp->v_bufobj, lbn + 1) == NULL) { 661 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 662 if (rabp) { 663 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 664 rabp->b_flags |= B_ASYNC; 665 rabp->b_iocmd = BIO_READ; 666 vfs_busy_pages(rabp, 0); 667 if (ncl_asyncio(nmp, rabp, cred, td)) { 668 rabp->b_flags |= B_INVAL; 669 rabp->b_ioflags |= BIO_ERROR; 670 vfs_unbusy_pages(rabp); 671 brelse(rabp); 672 } 673 } else { 674 brelse(rabp); 675 } 676 } 677 } 678 /* 679 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 680 * chopped for the EOF condition, we cannot tell how large 681 * NFS directories are going to be until we hit EOF. So 682 * an NFS directory buffer is *not* chopped to its EOF. Now, 683 * it just so happens that b_resid will effectively chop it 684 * to EOF. *BUT* this information is lost if the buffer goes 685 * away and is reconstituted into a B_CACHE state ( due to 686 * being VMIO ) later. So we keep track of the directory eof 687 * in np->n_direofoffset and chop it off as an extra step 688 * right here. 689 */ 690 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 691 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 692 n = np->n_direofoffset - uio->uio_offset; 693 break; 694 default: 695 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type); 696 bp = NULL; 697 break; 698 }; 699 700 if (n > 0) { 701 error = uiomove(bp->b_data + on, (int)n, uio); 702 } 703 if (vp->v_type == VLNK) 704 n = 0; 705 if (bp != NULL) 706 brelse(bp); 707 } while (error == 0 && uio->uio_resid > 0 && n > 0); 708 return (error); 709} 710 711/* 712 * The NFS write path cannot handle iovecs with len > 1. So we need to 713 * break up iovecs accordingly (restricting them to wsize). 714 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 715 * For the ASYNC case, 2 copies are needed. The first a copy from the 716 * user buffer to a staging buffer and then a second copy from the staging 717 * buffer to mbufs. This can be optimized by copying from the user buffer 718 * directly into mbufs and passing the chain down, but that requires a 719 * fair amount of re-working of the relevant codepaths (and can be done 720 * later). 721 */ 722static int 723nfs_directio_write(vp, uiop, cred, ioflag) 724 struct vnode *vp; 725 struct uio *uiop; 726 struct ucred *cred; 727 int ioflag; 728{ 729 int error; 730 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 731 struct thread *td = uiop->uio_td; 732 int size; 733 int wsize; 734 735 mtx_lock(&nmp->nm_mtx); 736 wsize = nmp->nm_wsize; 737 mtx_unlock(&nmp->nm_mtx); 738 if (ioflag & IO_SYNC) { 739 int iomode, must_commit; 740 struct uio uio; 741 struct iovec iov; 742do_sync: 743 while (uiop->uio_resid > 0) { 744 size = MIN(uiop->uio_resid, wsize); 745 size = MIN(uiop->uio_iov->iov_len, size); 746 iov.iov_base = uiop->uio_iov->iov_base; 747 iov.iov_len = size; 748 uio.uio_iov = &iov; 749 uio.uio_iovcnt = 1; 750 uio.uio_offset = uiop->uio_offset; 751 uio.uio_resid = size; 752 uio.uio_segflg = UIO_USERSPACE; 753 uio.uio_rw = UIO_WRITE; 754 uio.uio_td = td; 755 iomode = NFSWRITE_FILESYNC; 756 error = ncl_writerpc(vp, &uio, cred, &iomode, 757 &must_commit, 0); 758 KASSERT((must_commit == 0), 759 ("ncl_directio_write: Did not commit write")); 760 if (error) 761 return (error); 762 uiop->uio_offset += size; 763 uiop->uio_resid -= size; 764 if (uiop->uio_iov->iov_len <= size) { 765 uiop->uio_iovcnt--; 766 uiop->uio_iov++; 767 } else { 768 uiop->uio_iov->iov_base = 769 (char *)uiop->uio_iov->iov_base + size; 770 uiop->uio_iov->iov_len -= size; 771 } 772 } 773 } else { 774 struct uio *t_uio; 775 struct iovec *t_iov; 776 struct buf *bp; 777 778 /* 779 * Break up the write into blocksize chunks and hand these 780 * over to nfsiod's for write back. 781 * Unfortunately, this incurs a copy of the data. Since 782 * the user could modify the buffer before the write is 783 * initiated. 784 * 785 * The obvious optimization here is that one of the 2 copies 786 * in the async write path can be eliminated by copying the 787 * data here directly into mbufs and passing the mbuf chain 788 * down. But that will require a fair amount of re-working 789 * of the code and can be done if there's enough interest 790 * in NFS directio access. 791 */ 792 while (uiop->uio_resid > 0) { 793 size = MIN(uiop->uio_resid, wsize); 794 size = MIN(uiop->uio_iov->iov_len, size); 795 bp = getpbuf(&ncl_pbuf_freecnt); 796 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 797 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 798 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 799 t_iov->iov_len = size; 800 t_uio->uio_iov = t_iov; 801 t_uio->uio_iovcnt = 1; 802 t_uio->uio_offset = uiop->uio_offset; 803 t_uio->uio_resid = size; 804 t_uio->uio_segflg = UIO_SYSSPACE; 805 t_uio->uio_rw = UIO_WRITE; 806 t_uio->uio_td = td; 807 KASSERT(uiop->uio_segflg == UIO_USERSPACE || 808 uiop->uio_segflg == UIO_SYSSPACE, 809 ("nfs_directio_write: Bad uio_segflg")); 810 if (uiop->uio_segflg == UIO_USERSPACE) { 811 error = copyin(uiop->uio_iov->iov_base, 812 t_iov->iov_base, size); 813 if (error != 0) 814 goto err_free; 815 } else 816 /* 817 * UIO_SYSSPACE may never happen, but handle 818 * it just in case it does. 819 */ 820 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, 821 size); 822 bp->b_flags |= B_DIRECT; 823 bp->b_iocmd = BIO_WRITE; 824 if (cred != NOCRED) { 825 crhold(cred); 826 bp->b_wcred = cred; 827 } else 828 bp->b_wcred = NOCRED; 829 bp->b_caller1 = (void *)t_uio; 830 bp->b_vp = vp; 831 error = ncl_asyncio(nmp, bp, NOCRED, td); 832err_free: 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 */ 861int 862ncl_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 off_t tmp_off; 878 879 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode")); 880 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, 881 ("ncl_write proc")); 882 if (vp->v_type != VREG) 883 return (EIO); 884 mtx_lock(&np->n_mtx); 885 if (np->n_flag & NWRITEERR) { 886 np->n_flag &= ~NWRITEERR; 887 mtx_unlock(&np->n_mtx); 888 return (np->n_error); 889 } else 890 mtx_unlock(&np->n_mtx); 891 mtx_lock(&nmp->nm_mtx); 892 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 893 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 894 mtx_unlock(&nmp->nm_mtx); 895 (void)ncl_fsinfo(nmp, vp, cred, td); 896 mtx_lock(&nmp->nm_mtx); 897 } 898 if (nmp->nm_wsize == 0) 899 (void) newnfs_iosize(nmp); 900 mtx_unlock(&nmp->nm_mtx); 901 902 /* 903 * Synchronously flush pending buffers if we are in synchronous 904 * mode or if we are appending. 905 */ 906 if (ioflag & (IO_APPEND | IO_SYNC)) { 907 mtx_lock(&np->n_mtx); 908 if (np->n_flag & NMODIFIED) { 909 mtx_unlock(&np->n_mtx); 910#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 911 /* 912 * Require non-blocking, synchronous writes to 913 * dirty files to inform the program it needs 914 * to fsync(2) explicitly. 915 */ 916 if (ioflag & IO_NDELAY) 917 return (EAGAIN); 918#endif 919flush_and_restart: 920 np->n_attrstamp = 0; 921 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 922 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 923 if (error) 924 return (error); 925 } else 926 mtx_unlock(&np->n_mtx); 927 } 928 929 /* 930 * If IO_APPEND then load uio_offset. We restart here if we cannot 931 * get the append lock. 932 */ 933 if (ioflag & IO_APPEND) { 934 np->n_attrstamp = 0; 935 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 936 error = VOP_GETATTR(vp, &vattr, cred); 937 if (error) 938 return (error); 939 mtx_lock(&np->n_mtx); 940 uio->uio_offset = np->n_size; 941 mtx_unlock(&np->n_mtx); 942 } 943 944 if (uio->uio_offset < 0) 945 return (EINVAL); 946 tmp_off = uio->uio_offset + uio->uio_resid; 947 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset) 948 return (EFBIG); 949 if (uio->uio_resid == 0) 950 return (0); 951 952 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 953 return nfs_directio_write(vp, uio, cred, ioflag); 954 955 /* 956 * Maybe this should be above the vnode op call, but so long as 957 * file servers have no limits, i don't think it matters 958 */ 959 if (vn_rlimit_fsize(vp, uio, td)) 960 return (EFBIG); 961 962 biosize = vp->v_bufobj.bo_bsize; 963 /* 964 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 965 * would exceed the local maximum per-file write commit size when 966 * combined with those, we must decide whether to flush, 967 * go synchronous, or return error. We don't bother checking 968 * IO_UNIT -- we just make all writes atomic anyway, as there's 969 * no point optimizing for something that really won't ever happen. 970 */ 971 if (!(ioflag & IO_SYNC)) { 972 int nflag; 973 974 mtx_lock(&np->n_mtx); 975 nflag = np->n_flag; 976 mtx_unlock(&np->n_mtx); 977 int needrestart = 0; 978 if (nmp->nm_wcommitsize < uio->uio_resid) { 979 /* 980 * If this request could not possibly be completed 981 * without exceeding the maximum outstanding write 982 * commit size, see if we can convert it into a 983 * synchronous write operation. 984 */ 985 if (ioflag & IO_NDELAY) 986 return (EAGAIN); 987 ioflag |= IO_SYNC; 988 if (nflag & NMODIFIED) 989 needrestart = 1; 990 } else if (nflag & NMODIFIED) { 991 int wouldcommit = 0; 992 BO_LOCK(&vp->v_bufobj); 993 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 994 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 995 b_bobufs) { 996 if (bp->b_flags & B_NEEDCOMMIT) 997 wouldcommit += bp->b_bcount; 998 } 999 } 1000 BO_UNLOCK(&vp->v_bufobj); 1001 /* 1002 * Since we're not operating synchronously and 1003 * bypassing the buffer cache, we are in a commit 1004 * and holding all of these buffers whether 1005 * transmitted or not. If not limited, this 1006 * will lead to the buffer cache deadlocking, 1007 * as no one else can flush our uncommitted buffers. 1008 */ 1009 wouldcommit += uio->uio_resid; 1010 /* 1011 * If we would initially exceed the maximum 1012 * outstanding write commit size, flush and restart. 1013 */ 1014 if (wouldcommit > nmp->nm_wcommitsize) 1015 needrestart = 1; 1016 } 1017 if (needrestart) 1018 goto flush_and_restart; 1019 } 1020 1021 do { 1022 NFSINCRGLOBAL(newnfsstats.biocache_writes); 1023 lbn = uio->uio_offset / biosize; 1024 on = uio->uio_offset & (biosize-1); 1025 n = MIN((unsigned)(biosize - on), uio->uio_resid); 1026again: 1027 /* 1028 * Handle direct append and file extension cases, calculate 1029 * unaligned buffer size. 1030 */ 1031 mtx_lock(&np->n_mtx); 1032 if (uio->uio_offset == np->n_size && n) { 1033 mtx_unlock(&np->n_mtx); 1034 /* 1035 * Get the buffer (in its pre-append state to maintain 1036 * B_CACHE if it was previously set). Resize the 1037 * nfsnode after we have locked the buffer to prevent 1038 * readers from reading garbage. 1039 */ 1040 bcount = on; 1041 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1042 1043 if (bp != NULL) { 1044 long save; 1045 1046 mtx_lock(&np->n_mtx); 1047 np->n_size = uio->uio_offset + n; 1048 np->n_flag |= NMODIFIED; 1049 vnode_pager_setsize(vp, np->n_size); 1050 mtx_unlock(&np->n_mtx); 1051 1052 save = bp->b_flags & B_CACHE; 1053 bcount += n; 1054 allocbuf(bp, bcount); 1055 bp->b_flags |= save; 1056 } 1057 } else { 1058 /* 1059 * Obtain the locked cache block first, and then 1060 * adjust the file's size as appropriate. 1061 */ 1062 bcount = on + n; 1063 if ((off_t)lbn * biosize + bcount < np->n_size) { 1064 if ((off_t)(lbn + 1) * biosize < np->n_size) 1065 bcount = biosize; 1066 else 1067 bcount = np->n_size - (off_t)lbn * biosize; 1068 } 1069 mtx_unlock(&np->n_mtx); 1070 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1071 mtx_lock(&np->n_mtx); 1072 if (uio->uio_offset + n > np->n_size) { 1073 np->n_size = uio->uio_offset + n; 1074 np->n_flag |= NMODIFIED; 1075 vnode_pager_setsize(vp, np->n_size); 1076 } 1077 mtx_unlock(&np->n_mtx); 1078 } 1079 1080 if (!bp) { 1081 error = newnfs_sigintr(nmp, td); 1082 if (!error) 1083 error = EINTR; 1084 break; 1085 } 1086 1087 /* 1088 * Issue a READ if B_CACHE is not set. In special-append 1089 * mode, B_CACHE is based on the buffer prior to the write 1090 * op and is typically set, avoiding the read. If a read 1091 * is required in special append mode, the server will 1092 * probably send us a short-read since we extended the file 1093 * on our end, resulting in b_resid == 0 and, thusly, 1094 * B_CACHE getting set. 1095 * 1096 * We can also avoid issuing the read if the write covers 1097 * the entire buffer. We have to make sure the buffer state 1098 * is reasonable in this case since we will not be initiating 1099 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1100 * more information. 1101 * 1102 * B_CACHE may also be set due to the buffer being cached 1103 * normally. 1104 */ 1105 1106 if (on == 0 && n == bcount) { 1107 bp->b_flags |= B_CACHE; 1108 bp->b_flags &= ~B_INVAL; 1109 bp->b_ioflags &= ~BIO_ERROR; 1110 } 1111 1112 if ((bp->b_flags & B_CACHE) == 0) { 1113 bp->b_iocmd = BIO_READ; 1114 vfs_busy_pages(bp, 0); 1115 error = ncl_doio(vp, bp, cred, td, 0); 1116 if (error) { 1117 brelse(bp); 1118 break; 1119 } 1120 } 1121 if (bp->b_wcred == NOCRED) 1122 bp->b_wcred = crhold(cred); 1123 mtx_lock(&np->n_mtx); 1124 np->n_flag |= NMODIFIED; 1125 mtx_unlock(&np->n_mtx); 1126 1127 /* 1128 * If dirtyend exceeds file size, chop it down. This should 1129 * not normally occur but there is an append race where it 1130 * might occur XXX, so we log it. 1131 * 1132 * If the chopping creates a reverse-indexed or degenerate 1133 * situation with dirtyoff/end, we 0 both of them. 1134 */ 1135 1136 if (bp->b_dirtyend > bcount) { 1137 ncl_printf("NFS append race @%lx:%d\n", 1138 (long)bp->b_blkno * DEV_BSIZE, 1139 bp->b_dirtyend - bcount); 1140 bp->b_dirtyend = bcount; 1141 } 1142 1143 if (bp->b_dirtyoff >= bp->b_dirtyend) 1144 bp->b_dirtyoff = bp->b_dirtyend = 0; 1145 1146 /* 1147 * If the new write will leave a contiguous dirty 1148 * area, just update the b_dirtyoff and b_dirtyend, 1149 * otherwise force a write rpc of the old dirty area. 1150 * 1151 * While it is possible to merge discontiguous writes due to 1152 * our having a B_CACHE buffer ( and thus valid read data 1153 * for the hole), we don't because it could lead to 1154 * significant cache coherency problems with multiple clients, 1155 * especially if locking is implemented later on. 1156 * 1157 * As an optimization we could theoretically maintain 1158 * a linked list of discontinuous areas, but we would still 1159 * have to commit them separately so there isn't much 1160 * advantage to it except perhaps a bit of asynchronization. 1161 */ 1162 1163 if (bp->b_dirtyend > 0 && 1164 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1165 if (bwrite(bp) == EINTR) { 1166 error = EINTR; 1167 break; 1168 } 1169 goto again; 1170 } 1171 1172 error = uiomove((char *)bp->b_data + on, n, uio); 1173 1174 /* 1175 * Since this block is being modified, it must be written 1176 * again and not just committed. Since write clustering does 1177 * not work for the stage 1 data write, only the stage 2 1178 * commit rpc, we have to clear B_CLUSTEROK as well. 1179 */ 1180 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1181 1182 if (error) { 1183 bp->b_ioflags |= BIO_ERROR; 1184 brelse(bp); 1185 break; 1186 } 1187 1188 /* 1189 * Only update dirtyoff/dirtyend if not a degenerate 1190 * condition. 1191 */ 1192 if (n) { 1193 if (bp->b_dirtyend > 0) { 1194 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1195 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1196 } else { 1197 bp->b_dirtyoff = on; 1198 bp->b_dirtyend = on + n; 1199 } 1200 vfs_bio_set_valid(bp, on, n); 1201 } 1202 1203 /* 1204 * If IO_SYNC do bwrite(). 1205 * 1206 * IO_INVAL appears to be unused. The idea appears to be 1207 * to turn off caching in this case. Very odd. XXX 1208 */ 1209 if ((ioflag & IO_SYNC)) { 1210 if (ioflag & IO_INVAL) 1211 bp->b_flags |= B_NOCACHE; 1212 error = bwrite(bp); 1213 if (error) 1214 break; 1215 } else if ((n + on) == biosize) { 1216 bp->b_flags |= B_ASYNC; 1217 (void) ncl_writebp(bp, 0, NULL); 1218 } else { 1219 bdwrite(bp); 1220 } 1221 } while (uio->uio_resid > 0 && n > 0); 1222 1223 return (error); 1224} 1225 1226/* 1227 * Get an nfs cache block. 1228 * 1229 * Allocate a new one if the block isn't currently in the cache 1230 * and return the block marked busy. If the calling process is 1231 * interrupted by a signal for an interruptible mount point, return 1232 * NULL. 1233 * 1234 * The caller must carefully deal with the possible B_INVAL state of 1235 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it 1236 * indirectly), so synchronous reads can be issued without worrying about 1237 * the B_INVAL state. We have to be a little more careful when dealing 1238 * with writes (see comments in nfs_write()) when extending a file past 1239 * its EOF. 1240 */ 1241static struct buf * 1242nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1243{ 1244 struct buf *bp; 1245 struct mount *mp; 1246 struct nfsmount *nmp; 1247 1248 mp = vp->v_mount; 1249 nmp = VFSTONFS(mp); 1250 1251 if (nmp->nm_flag & NFSMNT_INT) { 1252 sigset_t oldset; 1253 1254 newnfs_set_sigmask(td, &oldset); 1255 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0); 1256 newnfs_restore_sigmask(td, &oldset); 1257 while (bp == NULL) { 1258 if (newnfs_sigintr(nmp, td)) 1259 return (NULL); 1260 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1261 } 1262 } else { 1263 bp = getblk(vp, bn, size, 0, 0, 0); 1264 } 1265 1266 if (vp->v_type == VREG) 1267 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE); 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 */ 1275int 1276ncl_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); 1334out: 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 */ 1347int 1348ncl_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 * Commits are usually short and sweet so lets save some cpu and 1358 * leave the async daemons for more important rpc's (such as reads 1359 * and writes). 1360 */ 1361 mtx_lock(&ncl_iod_mutex); 1362 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1363 (nmp->nm_bufqiods > ncl_numasync / 2)) { 1364 mtx_unlock(&ncl_iod_mutex); 1365 return(EIO); 1366 } 1367again: 1368 if (nmp->nm_flag & NFSMNT_INT) 1369 slpflag = NFS_PCATCH; 1370 gotiod = FALSE; 1371 1372 /* 1373 * Find a free iod to process this request. 1374 */ 1375 for (iod = 0; iod < ncl_numasync; iod++) 1376 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) { 1377 gotiod = TRUE; 1378 break; 1379 } 1380 1381 /* 1382 * Try to create one if none are free. 1383 */ 1384 if (!gotiod) 1385 ncl_nfsiodnew(); 1386 else { 1387 /* 1388 * Found one, so wake it up and tell it which 1389 * mount to process. 1390 */ 1391 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n", 1392 iod, nmp)); 1393 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE; 1394 ncl_iodmount[iod] = nmp; 1395 nmp->nm_bufqiods++; 1396 wakeup(&ncl_iodwant[iod]); 1397 } 1398 1399 /* 1400 * If none are free, we may already have an iod working on this mount 1401 * point. If so, it will process our request. 1402 */ 1403 if (!gotiod) { 1404 if (nmp->nm_bufqiods > 0) { 1405 NFS_DPF(ASYNCIO, 1406 ("ncl_asyncio: %d iods are already processing mount %p\n", 1407 nmp->nm_bufqiods, nmp)); 1408 gotiod = TRUE; 1409 } 1410 } 1411 1412 /* 1413 * If we have an iod which can process the request, then queue 1414 * the buffer. 1415 */ 1416 if (gotiod) { 1417 /* 1418 * Ensure that the queue never grows too large. We still want 1419 * to asynchronize so we block rather then return EIO. 1420 */ 1421 while (nmp->nm_bufqlen >= 2*ncl_numasync) { 1422 NFS_DPF(ASYNCIO, 1423 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp)); 1424 nmp->nm_bufqwant = TRUE; 1425 error = newnfs_msleep(td, &nmp->nm_bufq, 1426 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio", 1427 slptimeo); 1428 if (error) { 1429 error2 = newnfs_sigintr(nmp, td); 1430 if (error2) { 1431 mtx_unlock(&ncl_iod_mutex); 1432 return (error2); 1433 } 1434 if (slpflag == NFS_PCATCH) { 1435 slpflag = 0; 1436 slptimeo = 2 * hz; 1437 } 1438 } 1439 /* 1440 * We might have lost our iod while sleeping, 1441 * so check and loop if nescessary. 1442 */ 1443 goto again; 1444 } 1445 1446 /* We might have lost our nfsiod */ 1447 if (nmp->nm_bufqiods == 0) { 1448 NFS_DPF(ASYNCIO, 1449 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1450 goto again; 1451 } 1452 1453 if (bp->b_iocmd == BIO_READ) { 1454 if (bp->b_rcred == NOCRED && cred != NOCRED) 1455 bp->b_rcred = crhold(cred); 1456 } else { 1457 if (bp->b_wcred == NOCRED && cred != NOCRED) 1458 bp->b_wcred = crhold(cred); 1459 } 1460 1461 if (bp->b_flags & B_REMFREE) 1462 bremfreef(bp); 1463 BUF_KERNPROC(bp); 1464 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1465 nmp->nm_bufqlen++; 1466 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1467 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1468 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1469 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1470 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1471 } 1472 mtx_unlock(&ncl_iod_mutex); 1473 return (0); 1474 } 1475 1476 mtx_unlock(&ncl_iod_mutex); 1477 1478 /* 1479 * All the iods are busy on other mounts, so return EIO to 1480 * force the caller to process the i/o synchronously. 1481 */ 1482 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n")); 1483 return (EIO); 1484} 1485 1486void 1487ncl_doio_directwrite(struct buf *bp) 1488{ 1489 int iomode, must_commit; 1490 struct uio *uiop = (struct uio *)bp->b_caller1; 1491 char *iov_base = uiop->uio_iov->iov_base; 1492 1493 iomode = NFSWRITE_FILESYNC; 1494 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1495 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0); 1496 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write")); 1497 free(iov_base, M_NFSDIRECTIO); 1498 free(uiop->uio_iov, M_NFSDIRECTIO); 1499 free(uiop, M_NFSDIRECTIO); 1500 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1501 struct nfsnode *np = VTONFS(bp->b_vp); 1502 mtx_lock(&np->n_mtx); 1503 np->n_directio_asyncwr--; 1504 if (np->n_directio_asyncwr == 0) { 1505 np->n_flag &= ~NMODIFIED; 1506 if ((np->n_flag & NFSYNCWAIT)) { 1507 np->n_flag &= ~NFSYNCWAIT; 1508 wakeup((caddr_t)&np->n_directio_asyncwr); 1509 } 1510 } 1511 mtx_unlock(&np->n_mtx); 1512 } 1513 bp->b_vp = NULL; 1514 relpbuf(bp, &ncl_pbuf_freecnt); 1515} 1516 1517/* 1518 * Do an I/O operation to/from a cache block. This may be called 1519 * synchronously or from an nfsiod. 1520 */ 1521int 1522ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td, 1523 int called_from_strategy) 1524{ 1525 struct uio *uiop; 1526 struct nfsnode *np; 1527 struct nfsmount *nmp; 1528 int error = 0, iomode, must_commit = 0; 1529 struct uio uio; 1530 struct iovec io; 1531 struct proc *p = td ? td->td_proc : NULL; 1532 uint8_t iocmd; 1533 1534 np = VTONFS(vp); 1535 nmp = VFSTONFS(vp->v_mount); 1536 uiop = &uio; 1537 uiop->uio_iov = &io; 1538 uiop->uio_iovcnt = 1; 1539 uiop->uio_segflg = UIO_SYSSPACE; 1540 uiop->uio_td = td; 1541 1542 /* 1543 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1544 * do this here so we do not have to do it in all the code that 1545 * calls us. 1546 */ 1547 bp->b_flags &= ~B_INVAL; 1548 bp->b_ioflags &= ~BIO_ERROR; 1549 1550 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp)); 1551 iocmd = bp->b_iocmd; 1552 if (iocmd == BIO_READ) { 1553 io.iov_len = uiop->uio_resid = bp->b_bcount; 1554 io.iov_base = bp->b_data; 1555 uiop->uio_rw = UIO_READ; 1556 1557 switch (vp->v_type) { 1558 case VREG: 1559 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1560 NFSINCRGLOBAL(newnfsstats.read_bios); 1561 error = ncl_readrpc(vp, uiop, cr); 1562 1563 if (!error) { 1564 if (uiop->uio_resid) { 1565 /* 1566 * If we had a short read with no error, we must have 1567 * hit a file hole. We should zero-fill the remainder. 1568 * This can also occur if the server hits the file EOF. 1569 * 1570 * Holes used to be able to occur due to pending 1571 * writes, but that is not possible any longer. 1572 */ 1573 int nread = bp->b_bcount - uiop->uio_resid; 1574 ssize_t left = uiop->uio_resid; 1575 1576 if (left > 0) 1577 bzero((char *)bp->b_data + nread, left); 1578 uiop->uio_resid = 0; 1579 } 1580 } 1581 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */ 1582 if (p && (vp->v_vflag & VV_TEXT)) { 1583 mtx_lock(&np->n_mtx); 1584 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) { 1585 mtx_unlock(&np->n_mtx); 1586 PROC_LOCK(p); 1587 killproc(p, "text file modification"); 1588 PROC_UNLOCK(p); 1589 } else 1590 mtx_unlock(&np->n_mtx); 1591 } 1592 break; 1593 case VLNK: 1594 uiop->uio_offset = (off_t)0; 1595 NFSINCRGLOBAL(newnfsstats.readlink_bios); 1596 error = ncl_readlinkrpc(vp, uiop, cr); 1597 break; 1598 case VDIR: 1599 NFSINCRGLOBAL(newnfsstats.readdir_bios); 1600 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1601 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1602 error = ncl_readdirplusrpc(vp, uiop, cr, td); 1603 if (error == NFSERR_NOTSUPP) 1604 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1605 } 1606 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1607 error = ncl_readdirrpc(vp, uiop, cr, td); 1608 /* 1609 * end-of-directory sets B_INVAL but does not generate an 1610 * error. 1611 */ 1612 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1613 bp->b_flags |= B_INVAL; 1614 break; 1615 default: 1616 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type); 1617 break; 1618 }; 1619 if (error) { 1620 bp->b_ioflags |= BIO_ERROR; 1621 bp->b_error = error; 1622 } 1623 } else { 1624 /* 1625 * If we only need to commit, try to commit 1626 */ 1627 if (bp->b_flags & B_NEEDCOMMIT) { 1628 int retv; 1629 off_t off; 1630 1631 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1632 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1633 bp->b_wcred, td); 1634 if (retv == 0) { 1635 bp->b_dirtyoff = bp->b_dirtyend = 0; 1636 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1637 bp->b_resid = 0; 1638 bufdone(bp); 1639 return (0); 1640 } 1641 if (retv == NFSERR_STALEWRITEVERF) { 1642 ncl_clearcommit(vp->v_mount); 1643 } 1644 } 1645 1646 /* 1647 * Setup for actual write 1648 */ 1649 mtx_lock(&np->n_mtx); 1650 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1651 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1652 mtx_unlock(&np->n_mtx); 1653 1654 if (bp->b_dirtyend > bp->b_dirtyoff) { 1655 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1656 - bp->b_dirtyoff; 1657 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1658 + bp->b_dirtyoff; 1659 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1660 uiop->uio_rw = UIO_WRITE; 1661 NFSINCRGLOBAL(newnfsstats.write_bios); 1662 1663 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1664 iomode = NFSWRITE_UNSTABLE; 1665 else 1666 iomode = NFSWRITE_FILESYNC; 1667 1668 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit, 1669 called_from_strategy); 1670 1671 /* 1672 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1673 * to cluster the buffers needing commit. This will allow 1674 * the system to submit a single commit rpc for the whole 1675 * cluster. We can do this even if the buffer is not 100% 1676 * dirty (relative to the NFS blocksize), so we optimize the 1677 * append-to-file-case. 1678 * 1679 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1680 * cleared because write clustering only works for commit 1681 * rpc's, not for the data portion of the write). 1682 */ 1683 1684 if (!error && iomode == NFSWRITE_UNSTABLE) { 1685 bp->b_flags |= B_NEEDCOMMIT; 1686 if (bp->b_dirtyoff == 0 1687 && bp->b_dirtyend == bp->b_bcount) 1688 bp->b_flags |= B_CLUSTEROK; 1689 } else { 1690 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1691 } 1692 1693 /* 1694 * For an interrupted write, the buffer is still valid 1695 * and the write hasn't been pushed to the server yet, 1696 * so we can't set BIO_ERROR and report the interruption 1697 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1698 * is not relevant, so the rpc attempt is essentially 1699 * a noop. For the case of a V3 write rpc not being 1700 * committed to stable storage, the block is still 1701 * dirty and requires either a commit rpc or another 1702 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1703 * the block is reused. This is indicated by setting 1704 * the B_DELWRI and B_NEEDCOMMIT flags. 1705 * 1706 * EIO is returned by ncl_writerpc() to indicate a recoverable 1707 * write error and is handled as above, except that 1708 * B_EINTR isn't set. One cause of this is a stale stateid 1709 * error for the RPC that indicates recovery is required, 1710 * when called with called_from_strategy != 0. 1711 * 1712 * If the buffer is marked B_PAGING, it does not reside on 1713 * the vp's paging queues so we cannot call bdirty(). The 1714 * bp in this case is not an NFS cache block so we should 1715 * be safe. XXX 1716 * 1717 * The logic below breaks up errors into recoverable and 1718 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1719 * and keep the buffer around for potential write retries. 1720 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1721 * and save the error in the nfsnode. This is less than ideal 1722 * but necessary. Keeping such buffers around could potentially 1723 * cause buffer exhaustion eventually (they can never be written 1724 * out, so will get constantly be re-dirtied). It also causes 1725 * all sorts of vfs panics. For non-recoverable write errors, 1726 * also invalidate the attrcache, so we'll be forced to go over 1727 * the wire for this object, returning an error to user on next 1728 * call (most of the time). 1729 */ 1730 if (error == EINTR || error == EIO || error == ETIMEDOUT 1731 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1732 int s; 1733 1734 s = splbio(); 1735 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1736 if ((bp->b_flags & B_PAGING) == 0) { 1737 bdirty(bp); 1738 bp->b_flags &= ~B_DONE; 1739 } 1740 if ((error == EINTR || error == ETIMEDOUT) && 1741 (bp->b_flags & B_ASYNC) == 0) 1742 bp->b_flags |= B_EINTR; 1743 splx(s); 1744 } else { 1745 if (error) { 1746 bp->b_ioflags |= BIO_ERROR; 1747 bp->b_flags |= B_INVAL; 1748 bp->b_error = np->n_error = error; 1749 mtx_lock(&np->n_mtx); 1750 np->n_flag |= NWRITEERR; 1751 np->n_attrstamp = 0; 1752 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1753 mtx_unlock(&np->n_mtx); 1754 } 1755 bp->b_dirtyoff = bp->b_dirtyend = 0; 1756 } 1757 } else { 1758 bp->b_resid = 0; 1759 bufdone(bp); 1760 return (0); 1761 } 1762 } 1763 bp->b_resid = uiop->uio_resid; 1764 if (must_commit) 1765 ncl_clearcommit(vp->v_mount); 1766 bufdone(bp); 1767 return (error); 1768} 1769 1770/* 1771 * Used to aid in handling ftruncate() operations on the NFS client side. 1772 * Truncation creates a number of special problems for NFS. We have to 1773 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1774 * we have to properly handle VM pages or (potentially dirty) buffers 1775 * that straddle the truncation point. 1776 */ 1777 1778int 1779ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1780{ 1781 struct nfsnode *np = VTONFS(vp); 1782 u_quad_t tsize; 1783 int biosize = vp->v_bufobj.bo_bsize; 1784 int error = 0; 1785 1786 mtx_lock(&np->n_mtx); 1787 tsize = np->n_size; 1788 np->n_size = nsize; 1789 mtx_unlock(&np->n_mtx); 1790 1791 if (nsize < tsize) { 1792 struct buf *bp; 1793 daddr_t lbn; 1794 int bufsize; 1795 1796 /* 1797 * vtruncbuf() doesn't get the buffer overlapping the 1798 * truncation point. We may have a B_DELWRI and/or B_CACHE 1799 * buffer that now needs to be truncated. 1800 */ 1801 error = vtruncbuf(vp, cred, td, nsize, biosize); 1802 lbn = nsize / biosize; 1803 bufsize = nsize & (biosize - 1); 1804 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1805 if (!bp) 1806 return EINTR; 1807 if (bp->b_dirtyoff > bp->b_bcount) 1808 bp->b_dirtyoff = bp->b_bcount; 1809 if (bp->b_dirtyend > bp->b_bcount) 1810 bp->b_dirtyend = bp->b_bcount; 1811 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1812 brelse(bp); 1813 } else { 1814 vnode_pager_setsize(vp, nsize); 1815 } 1816 return(error); 1817} 1818 1819