nfs_clbio.c revision 236446
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 236446 2012-06-02 11:44:50Z 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 occured 221 * we may have hit a zero-fill section. We simply 222 * leave valid set to 0. 223 */ 224 ; 225 } 226 if (i != ap->a_reqpage) { 227 /* 228 * Whether or not to leave the page activated is up in 229 * the air, but we should put the page on a page queue 230 * somewhere (it already is in the object). Result: 231 * It appears that emperical results show that 232 * deactivating pages is best. 233 */ 234 235 /* 236 * Just in case someone was asking for this page we 237 * now tell them that it is ok to use. 238 */ 239 if (!error) { 240 if (m->oflags & VPO_WANTED) { 241 vm_page_lock(m); 242 vm_page_activate(m); 243 vm_page_unlock(m); 244 } else { 245 vm_page_lock(m); 246 vm_page_deactivate(m); 247 vm_page_unlock(m); 248 } 249 vm_page_wakeup(m); 250 } else { 251 vm_page_lock(m); 252 vm_page_free(m); 253 vm_page_unlock(m); 254 } 255 } 256 } 257 VM_OBJECT_UNLOCK(object); 258 return (0); 259} 260 261/* 262 * Vnode op for VM putpages. 263 */ 264int 265ncl_putpages(struct vop_putpages_args *ap) 266{ 267 struct uio uio; 268 struct iovec iov; 269 vm_offset_t kva; 270 struct buf *bp; 271 int iomode, must_commit, i, error, npages, count; 272 off_t offset; 273 int *rtvals; 274 struct vnode *vp; 275 struct thread *td; 276 struct ucred *cred; 277 struct nfsmount *nmp; 278 struct nfsnode *np; 279 vm_page_t *pages; 280 281 vp = ap->a_vp; 282 np = VTONFS(vp); 283 td = curthread; /* XXX */ 284 /* Set the cred to n_writecred for the write rpcs. */ 285 if (np->n_writecred != NULL) 286 cred = crhold(np->n_writecred); 287 else 288 cred = crhold(curthread->td_ucred); /* XXX */ 289 nmp = VFSTONFS(vp->v_mount); 290 pages = ap->a_m; 291 count = ap->a_count; 292 rtvals = ap->a_rtvals; 293 npages = btoc(count); 294 offset = IDX_TO_OFF(pages[0]->pindex); 295 296 mtx_lock(&nmp->nm_mtx); 297 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 298 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 299 mtx_unlock(&nmp->nm_mtx); 300 (void)ncl_fsinfo(nmp, vp, cred, td); 301 } else 302 mtx_unlock(&nmp->nm_mtx); 303 304 mtx_lock(&np->n_mtx); 305 if (newnfs_directio_enable && !newnfs_directio_allow_mmap && 306 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 307 mtx_unlock(&np->n_mtx); 308 ncl_printf("ncl_putpages: called on noncache-able vnode??\n"); 309 mtx_lock(&np->n_mtx); 310 } 311 312 for (i = 0; i < npages; i++) 313 rtvals[i] = VM_PAGER_ERROR; 314 315 /* 316 * When putting pages, do not extend file past EOF. 317 */ 318 if (offset + count > np->n_size) { 319 count = np->n_size - offset; 320 if (count < 0) 321 count = 0; 322 } 323 mtx_unlock(&np->n_mtx); 324 325 /* 326 * We use only the kva address for the buffer, but this is extremely 327 * convienient and fast. 328 */ 329 bp = getpbuf(&ncl_pbuf_freecnt); 330 331 kva = (vm_offset_t) bp->b_data; 332 pmap_qenter(kva, pages, npages); 333 PCPU_INC(cnt.v_vnodeout); 334 PCPU_ADD(cnt.v_vnodepgsout, count); 335 336 iov.iov_base = (caddr_t) kva; 337 iov.iov_len = count; 338 uio.uio_iov = &iov; 339 uio.uio_iovcnt = 1; 340 uio.uio_offset = offset; 341 uio.uio_resid = count; 342 uio.uio_segflg = UIO_SYSSPACE; 343 uio.uio_rw = UIO_WRITE; 344 uio.uio_td = td; 345 346 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 347 iomode = NFSWRITE_UNSTABLE; 348 else 349 iomode = NFSWRITE_FILESYNC; 350 351 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0); 352 crfree(cred); 353 354 pmap_qremove(kva, npages); 355 relpbuf(bp, &ncl_pbuf_freecnt); 356 357 if (error == 0 || !nfs_keep_dirty_on_error) { 358 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid); 359 if (must_commit) 360 ncl_clearcommit(vp->v_mount); 361 } 362 return rtvals[0]; 363} 364 365/* 366 * For nfs, cache consistency can only be maintained approximately. 367 * Although RFC1094 does not specify the criteria, the following is 368 * believed to be compatible with the reference port. 369 * For nfs: 370 * If the file's modify time on the server has changed since the 371 * last read rpc or you have written to the file, 372 * you may have lost data cache consistency with the 373 * server, so flush all of the file's data out of the cache. 374 * Then force a getattr rpc to ensure that you have up to date 375 * attributes. 376 * NB: This implies that cache data can be read when up to 377 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 378 * attributes this could be forced by setting n_attrstamp to 0 before 379 * the VOP_GETATTR() call. 380 */ 381static inline int 382nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 383{ 384 int error = 0; 385 struct vattr vattr; 386 struct nfsnode *np = VTONFS(vp); 387 int old_lock; 388 389 /* 390 * Grab the exclusive lock before checking whether the cache is 391 * consistent. 392 * XXX - We can make this cheaper later (by acquiring cheaper locks). 393 * But for now, this suffices. 394 */ 395 old_lock = ncl_upgrade_vnlock(vp); 396 if (vp->v_iflag & VI_DOOMED) { 397 ncl_downgrade_vnlock(vp, old_lock); 398 return (EBADF); 399 } 400 401 mtx_lock(&np->n_mtx); 402 if (np->n_flag & NMODIFIED) { 403 mtx_unlock(&np->n_mtx); 404 if (vp->v_type != VREG) { 405 if (vp->v_type != VDIR) 406 panic("nfs: bioread, not dir"); 407 ncl_invaldir(vp); 408 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 409 if (error) 410 goto out; 411 } 412 np->n_attrstamp = 0; 413 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 414 error = VOP_GETATTR(vp, &vattr, cred); 415 if (error) 416 goto out; 417 mtx_lock(&np->n_mtx); 418 np->n_mtime = vattr.va_mtime; 419 mtx_unlock(&np->n_mtx); 420 } else { 421 mtx_unlock(&np->n_mtx); 422 error = VOP_GETATTR(vp, &vattr, cred); 423 if (error) 424 return (error); 425 mtx_lock(&np->n_mtx); 426 if ((np->n_flag & NSIZECHANGED) 427 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 428 mtx_unlock(&np->n_mtx); 429 if (vp->v_type == VDIR) 430 ncl_invaldir(vp); 431 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 432 if (error) 433 goto out; 434 mtx_lock(&np->n_mtx); 435 np->n_mtime = vattr.va_mtime; 436 np->n_flag &= ~NSIZECHANGED; 437 } 438 mtx_unlock(&np->n_mtx); 439 } 440out: 441 ncl_downgrade_vnlock(vp, old_lock); 442 return error; 443} 444 445/* 446 * Vnode op for read using bio 447 */ 448int 449ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 450{ 451 struct nfsnode *np = VTONFS(vp); 452 int biosize, i; 453 struct buf *bp, *rabp; 454 struct thread *td; 455 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 456 daddr_t lbn, rabn; 457 int bcount; 458 int seqcount; 459 int nra, error = 0, n = 0, on = 0; 460 off_t tmp_off; 461 462 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode")); 463 if (uio->uio_resid == 0) 464 return (0); 465 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 466 return (EINVAL); 467 td = uio->uio_td; 468 469 mtx_lock(&nmp->nm_mtx); 470 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 471 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 472 mtx_unlock(&nmp->nm_mtx); 473 (void)ncl_fsinfo(nmp, vp, cred, td); 474 mtx_lock(&nmp->nm_mtx); 475 } 476 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0) 477 (void) newnfs_iosize(nmp); 478 479 tmp_off = uio->uio_offset + uio->uio_resid; 480 if (vp->v_type != VDIR && 481 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) { 482 mtx_unlock(&nmp->nm_mtx); 483 return (EFBIG); 484 } 485 mtx_unlock(&nmp->nm_mtx); 486 487 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 488 /* No caching/ no readaheads. Just read data into the user buffer */ 489 return ncl_readrpc(vp, uio, cred); 490 491 biosize = vp->v_bufobj.bo_bsize; 492 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 493 494 error = nfs_bioread_check_cons(vp, td, cred); 495 if (error) 496 return error; 497 498 do { 499 u_quad_t nsize; 500 501 mtx_lock(&np->n_mtx); 502 nsize = np->n_size; 503 mtx_unlock(&np->n_mtx); 504 505 switch (vp->v_type) { 506 case VREG: 507 NFSINCRGLOBAL(newnfsstats.biocache_reads); 508 lbn = uio->uio_offset / biosize; 509 on = uio->uio_offset & (biosize - 1); 510 511 /* 512 * Start the read ahead(s), as required. 513 */ 514 if (nmp->nm_readahead > 0) { 515 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 516 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 517 rabn = lbn + 1 + nra; 518 if (incore(&vp->v_bufobj, rabn) == NULL) { 519 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 520 if (!rabp) { 521 error = newnfs_sigintr(nmp, td); 522 return (error ? error : EINTR); 523 } 524 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 525 rabp->b_flags |= B_ASYNC; 526 rabp->b_iocmd = BIO_READ; 527 vfs_busy_pages(rabp, 0); 528 if (ncl_asyncio(nmp, rabp, cred, td)) { 529 rabp->b_flags |= B_INVAL; 530 rabp->b_ioflags |= BIO_ERROR; 531 vfs_unbusy_pages(rabp); 532 brelse(rabp); 533 break; 534 } 535 } else { 536 brelse(rabp); 537 } 538 } 539 } 540 } 541 542 /* Note that bcount is *not* DEV_BSIZE aligned. */ 543 bcount = biosize; 544 if ((off_t)lbn * biosize >= nsize) { 545 bcount = 0; 546 } else if ((off_t)(lbn + 1) * biosize > nsize) { 547 bcount = nsize - (off_t)lbn * biosize; 548 } 549 bp = nfs_getcacheblk(vp, lbn, bcount, td); 550 551 if (!bp) { 552 error = newnfs_sigintr(nmp, td); 553 return (error ? error : EINTR); 554 } 555 556 /* 557 * If B_CACHE is not set, we must issue the read. If this 558 * fails, we return an error. 559 */ 560 561 if ((bp->b_flags & B_CACHE) == 0) { 562 bp->b_iocmd = BIO_READ; 563 vfs_busy_pages(bp, 0); 564 error = ncl_doio(vp, bp, cred, td, 0); 565 if (error) { 566 brelse(bp); 567 return (error); 568 } 569 } 570 571 /* 572 * on is the offset into the current bp. Figure out how many 573 * bytes we can copy out of the bp. Note that bcount is 574 * NOT DEV_BSIZE aligned. 575 * 576 * Then figure out how many bytes we can copy into the uio. 577 */ 578 579 n = 0; 580 if (on < bcount) 581 n = MIN((unsigned)(bcount - on), uio->uio_resid); 582 break; 583 case VLNK: 584 NFSINCRGLOBAL(newnfsstats.biocache_readlinks); 585 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 586 if (!bp) { 587 error = newnfs_sigintr(nmp, td); 588 return (error ? error : EINTR); 589 } 590 if ((bp->b_flags & B_CACHE) == 0) { 591 bp->b_iocmd = BIO_READ; 592 vfs_busy_pages(bp, 0); 593 error = ncl_doio(vp, bp, cred, td, 0); 594 if (error) { 595 bp->b_ioflags |= BIO_ERROR; 596 brelse(bp); 597 return (error); 598 } 599 } 600 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 601 on = 0; 602 break; 603 case VDIR: 604 NFSINCRGLOBAL(newnfsstats.biocache_readdirs); 605 if (np->n_direofoffset 606 && uio->uio_offset >= np->n_direofoffset) { 607 return (0); 608 } 609 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 610 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 611 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 612 if (!bp) { 613 error = newnfs_sigintr(nmp, td); 614 return (error ? error : EINTR); 615 } 616 if ((bp->b_flags & B_CACHE) == 0) { 617 bp->b_iocmd = BIO_READ; 618 vfs_busy_pages(bp, 0); 619 error = ncl_doio(vp, bp, cred, td, 0); 620 if (error) { 621 brelse(bp); 622 } 623 while (error == NFSERR_BAD_COOKIE) { 624 ncl_invaldir(vp); 625 error = ncl_vinvalbuf(vp, 0, td, 1); 626 /* 627 * Yuck! The directory has been modified on the 628 * server. The only way to get the block is by 629 * reading from the beginning to get all the 630 * offset cookies. 631 * 632 * Leave the last bp intact unless there is an error. 633 * Loop back up to the while if the error is another 634 * NFSERR_BAD_COOKIE (double yuch!). 635 */ 636 for (i = 0; i <= lbn && !error; i++) { 637 if (np->n_direofoffset 638 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 639 return (0); 640 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 641 if (!bp) { 642 error = newnfs_sigintr(nmp, td); 643 return (error ? error : EINTR); 644 } 645 if ((bp->b_flags & B_CACHE) == 0) { 646 bp->b_iocmd = BIO_READ; 647 vfs_busy_pages(bp, 0); 648 error = ncl_doio(vp, bp, cred, td, 0); 649 /* 650 * no error + B_INVAL == directory EOF, 651 * use the block. 652 */ 653 if (error == 0 && (bp->b_flags & B_INVAL)) 654 break; 655 } 656 /* 657 * An error will throw away the block and the 658 * for loop will break out. If no error and this 659 * is not the block we want, we throw away the 660 * block and go for the next one via the for loop. 661 */ 662 if (error || i < lbn) 663 brelse(bp); 664 } 665 } 666 /* 667 * The above while is repeated if we hit another cookie 668 * error. If we hit an error and it wasn't a cookie error, 669 * we give up. 670 */ 671 if (error) 672 return (error); 673 } 674 675 /* 676 * If not eof and read aheads are enabled, start one. 677 * (You need the current block first, so that you have the 678 * directory offset cookie of the next block.) 679 */ 680 if (nmp->nm_readahead > 0 && 681 (bp->b_flags & B_INVAL) == 0 && 682 (np->n_direofoffset == 0 || 683 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 684 incore(&vp->v_bufobj, lbn + 1) == NULL) { 685 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 686 if (rabp) { 687 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 688 rabp->b_flags |= B_ASYNC; 689 rabp->b_iocmd = BIO_READ; 690 vfs_busy_pages(rabp, 0); 691 if (ncl_asyncio(nmp, rabp, cred, td)) { 692 rabp->b_flags |= B_INVAL; 693 rabp->b_ioflags |= BIO_ERROR; 694 vfs_unbusy_pages(rabp); 695 brelse(rabp); 696 } 697 } else { 698 brelse(rabp); 699 } 700 } 701 } 702 /* 703 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 704 * chopped for the EOF condition, we cannot tell how large 705 * NFS directories are going to be until we hit EOF. So 706 * an NFS directory buffer is *not* chopped to its EOF. Now, 707 * it just so happens that b_resid will effectively chop it 708 * to EOF. *BUT* this information is lost if the buffer goes 709 * away and is reconstituted into a B_CACHE state ( due to 710 * being VMIO ) later. So we keep track of the directory eof 711 * in np->n_direofoffset and chop it off as an extra step 712 * right here. 713 */ 714 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 715 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 716 n = np->n_direofoffset - uio->uio_offset; 717 break; 718 default: 719 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type); 720 bp = NULL; 721 break; 722 }; 723 724 if (n > 0) { 725 error = uiomove(bp->b_data + on, (int)n, uio); 726 } 727 if (vp->v_type == VLNK) 728 n = 0; 729 if (bp != NULL) 730 brelse(bp); 731 } while (error == 0 && uio->uio_resid > 0 && n > 0); 732 return (error); 733} 734 735/* 736 * The NFS write path cannot handle iovecs with len > 1. So we need to 737 * break up iovecs accordingly (restricting them to wsize). 738 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 739 * For the ASYNC case, 2 copies are needed. The first a copy from the 740 * user buffer to a staging buffer and then a second copy from the staging 741 * buffer to mbufs. This can be optimized by copying from the user buffer 742 * directly into mbufs and passing the chain down, but that requires a 743 * fair amount of re-working of the relevant codepaths (and can be done 744 * later). 745 */ 746static int 747nfs_directio_write(vp, uiop, cred, ioflag) 748 struct vnode *vp; 749 struct uio *uiop; 750 struct ucred *cred; 751 int ioflag; 752{ 753 int error; 754 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 755 struct thread *td = uiop->uio_td; 756 int size; 757 int wsize; 758 759 mtx_lock(&nmp->nm_mtx); 760 wsize = nmp->nm_wsize; 761 mtx_unlock(&nmp->nm_mtx); 762 if (ioflag & IO_SYNC) { 763 int iomode, must_commit; 764 struct uio uio; 765 struct iovec iov; 766do_sync: 767 while (uiop->uio_resid > 0) { 768 size = MIN(uiop->uio_resid, wsize); 769 size = MIN(uiop->uio_iov->iov_len, size); 770 iov.iov_base = uiop->uio_iov->iov_base; 771 iov.iov_len = size; 772 uio.uio_iov = &iov; 773 uio.uio_iovcnt = 1; 774 uio.uio_offset = uiop->uio_offset; 775 uio.uio_resid = size; 776 uio.uio_segflg = UIO_USERSPACE; 777 uio.uio_rw = UIO_WRITE; 778 uio.uio_td = td; 779 iomode = NFSWRITE_FILESYNC; 780 error = ncl_writerpc(vp, &uio, cred, &iomode, 781 &must_commit, 0); 782 KASSERT((must_commit == 0), 783 ("ncl_directio_write: Did not commit write")); 784 if (error) 785 return (error); 786 uiop->uio_offset += size; 787 uiop->uio_resid -= size; 788 if (uiop->uio_iov->iov_len <= size) { 789 uiop->uio_iovcnt--; 790 uiop->uio_iov++; 791 } else { 792 uiop->uio_iov->iov_base = 793 (char *)uiop->uio_iov->iov_base + size; 794 uiop->uio_iov->iov_len -= size; 795 } 796 } 797 } else { 798 struct uio *t_uio; 799 struct iovec *t_iov; 800 struct buf *bp; 801 802 /* 803 * Break up the write into blocksize chunks and hand these 804 * over to nfsiod's for write back. 805 * Unfortunately, this incurs a copy of the data. Since 806 * the user could modify the buffer before the write is 807 * initiated. 808 * 809 * The obvious optimization here is that one of the 2 copies 810 * in the async write path can be eliminated by copying the 811 * data here directly into mbufs and passing the mbuf chain 812 * down. But that will require a fair amount of re-working 813 * of the code and can be done if there's enough interest 814 * in NFS directio access. 815 */ 816 while (uiop->uio_resid > 0) { 817 size = MIN(uiop->uio_resid, wsize); 818 size = MIN(uiop->uio_iov->iov_len, size); 819 bp = getpbuf(&ncl_pbuf_freecnt); 820 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 821 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 822 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 823 t_iov->iov_len = size; 824 t_uio->uio_iov = t_iov; 825 t_uio->uio_iovcnt = 1; 826 t_uio->uio_offset = uiop->uio_offset; 827 t_uio->uio_resid = size; 828 t_uio->uio_segflg = UIO_SYSSPACE; 829 t_uio->uio_rw = UIO_WRITE; 830 t_uio->uio_td = td; 831 KASSERT(uiop->uio_segflg == UIO_USERSPACE || 832 uiop->uio_segflg == UIO_SYSSPACE, 833 ("nfs_directio_write: Bad uio_segflg")); 834 if (uiop->uio_segflg == UIO_USERSPACE) { 835 error = copyin(uiop->uio_iov->iov_base, 836 t_iov->iov_base, size); 837 if (error != 0) 838 goto err_free; 839 } else 840 /* 841 * UIO_SYSSPACE may never happen, but handle 842 * it just in case it does. 843 */ 844 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, 845 size); 846 bp->b_flags |= B_DIRECT; 847 bp->b_iocmd = BIO_WRITE; 848 if (cred != NOCRED) { 849 crhold(cred); 850 bp->b_wcred = cred; 851 } else 852 bp->b_wcred = NOCRED; 853 bp->b_caller1 = (void *)t_uio; 854 bp->b_vp = vp; 855 error = ncl_asyncio(nmp, bp, NOCRED, td); 856err_free: 857 if (error) { 858 free(t_iov->iov_base, M_NFSDIRECTIO); 859 free(t_iov, M_NFSDIRECTIO); 860 free(t_uio, M_NFSDIRECTIO); 861 bp->b_vp = NULL; 862 relpbuf(bp, &ncl_pbuf_freecnt); 863 if (error == EINTR) 864 return (error); 865 goto do_sync; 866 } 867 uiop->uio_offset += size; 868 uiop->uio_resid -= size; 869 if (uiop->uio_iov->iov_len <= size) { 870 uiop->uio_iovcnt--; 871 uiop->uio_iov++; 872 } else { 873 uiop->uio_iov->iov_base = 874 (char *)uiop->uio_iov->iov_base + size; 875 uiop->uio_iov->iov_len -= size; 876 } 877 } 878 } 879 return (0); 880} 881 882/* 883 * Vnode op for write using bio 884 */ 885int 886ncl_write(struct vop_write_args *ap) 887{ 888 int biosize; 889 struct uio *uio = ap->a_uio; 890 struct thread *td = uio->uio_td; 891 struct vnode *vp = ap->a_vp; 892 struct nfsnode *np = VTONFS(vp); 893 struct ucred *cred = ap->a_cred; 894 int ioflag = ap->a_ioflag; 895 struct buf *bp; 896 struct vattr vattr; 897 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 898 daddr_t lbn; 899 int bcount; 900 int n, on, error = 0; 901 off_t tmp_off; 902 903 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode")); 904 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, 905 ("ncl_write proc")); 906 if (vp->v_type != VREG) 907 return (EIO); 908 mtx_lock(&np->n_mtx); 909 if (np->n_flag & NWRITEERR) { 910 np->n_flag &= ~NWRITEERR; 911 mtx_unlock(&np->n_mtx); 912 return (np->n_error); 913 } else 914 mtx_unlock(&np->n_mtx); 915 mtx_lock(&nmp->nm_mtx); 916 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 917 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 918 mtx_unlock(&nmp->nm_mtx); 919 (void)ncl_fsinfo(nmp, vp, cred, td); 920 mtx_lock(&nmp->nm_mtx); 921 } 922 if (nmp->nm_wsize == 0) 923 (void) newnfs_iosize(nmp); 924 mtx_unlock(&nmp->nm_mtx); 925 926 /* 927 * Synchronously flush pending buffers if we are in synchronous 928 * mode or if we are appending. 929 */ 930 if (ioflag & (IO_APPEND | IO_SYNC)) { 931 mtx_lock(&np->n_mtx); 932 if (np->n_flag & NMODIFIED) { 933 mtx_unlock(&np->n_mtx); 934#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 935 /* 936 * Require non-blocking, synchronous writes to 937 * dirty files to inform the program it needs 938 * to fsync(2) explicitly. 939 */ 940 if (ioflag & IO_NDELAY) 941 return (EAGAIN); 942#endif 943flush_and_restart: 944 np->n_attrstamp = 0; 945 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 946 error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 947 if (error) 948 return (error); 949 } else 950 mtx_unlock(&np->n_mtx); 951 } 952 953 /* 954 * If IO_APPEND then load uio_offset. We restart here if we cannot 955 * get the append lock. 956 */ 957 if (ioflag & IO_APPEND) { 958 np->n_attrstamp = 0; 959 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 960 error = VOP_GETATTR(vp, &vattr, cred); 961 if (error) 962 return (error); 963 mtx_lock(&np->n_mtx); 964 uio->uio_offset = np->n_size; 965 mtx_unlock(&np->n_mtx); 966 } 967 968 if (uio->uio_offset < 0) 969 return (EINVAL); 970 tmp_off = uio->uio_offset + uio->uio_resid; 971 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset) 972 return (EFBIG); 973 if (uio->uio_resid == 0) 974 return (0); 975 976 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 977 return nfs_directio_write(vp, uio, cred, ioflag); 978 979 /* 980 * Maybe this should be above the vnode op call, but so long as 981 * file servers have no limits, i don't think it matters 982 */ 983 if (vn_rlimit_fsize(vp, uio, td)) 984 return (EFBIG); 985 986 biosize = vp->v_bufobj.bo_bsize; 987 /* 988 * Find all of this file's B_NEEDCOMMIT buffers. If our writes 989 * would exceed the local maximum per-file write commit size when 990 * combined with those, we must decide whether to flush, 991 * go synchronous, or return error. We don't bother checking 992 * IO_UNIT -- we just make all writes atomic anyway, as there's 993 * no point optimizing for something that really won't ever happen. 994 */ 995 if (!(ioflag & IO_SYNC)) { 996 int nflag; 997 998 mtx_lock(&np->n_mtx); 999 nflag = np->n_flag; 1000 mtx_unlock(&np->n_mtx); 1001 int needrestart = 0; 1002 if (nmp->nm_wcommitsize < uio->uio_resid) { 1003 /* 1004 * If this request could not possibly be completed 1005 * without exceeding the maximum outstanding write 1006 * commit size, see if we can convert it into a 1007 * synchronous write operation. 1008 */ 1009 if (ioflag & IO_NDELAY) 1010 return (EAGAIN); 1011 ioflag |= IO_SYNC; 1012 if (nflag & NMODIFIED) 1013 needrestart = 1; 1014 } else if (nflag & NMODIFIED) { 1015 int wouldcommit = 0; 1016 BO_LOCK(&vp->v_bufobj); 1017 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 1018 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 1019 b_bobufs) { 1020 if (bp->b_flags & B_NEEDCOMMIT) 1021 wouldcommit += bp->b_bcount; 1022 } 1023 } 1024 BO_UNLOCK(&vp->v_bufobj); 1025 /* 1026 * Since we're not operating synchronously and 1027 * bypassing the buffer cache, we are in a commit 1028 * and holding all of these buffers whether 1029 * transmitted or not. If not limited, this 1030 * will lead to the buffer cache deadlocking, 1031 * as no one else can flush our uncommitted buffers. 1032 */ 1033 wouldcommit += uio->uio_resid; 1034 /* 1035 * If we would initially exceed the maximum 1036 * outstanding write commit size, flush and restart. 1037 */ 1038 if (wouldcommit > nmp->nm_wcommitsize) 1039 needrestart = 1; 1040 } 1041 if (needrestart) 1042 goto flush_and_restart; 1043 } 1044 1045 do { 1046 NFSINCRGLOBAL(newnfsstats.biocache_writes); 1047 lbn = uio->uio_offset / biosize; 1048 on = uio->uio_offset & (biosize-1); 1049 n = MIN((unsigned)(biosize - on), uio->uio_resid); 1050again: 1051 /* 1052 * Handle direct append and file extension cases, calculate 1053 * unaligned buffer size. 1054 */ 1055 mtx_lock(&np->n_mtx); 1056 if (uio->uio_offset == np->n_size && n) { 1057 mtx_unlock(&np->n_mtx); 1058 /* 1059 * Get the buffer (in its pre-append state to maintain 1060 * B_CACHE if it was previously set). Resize the 1061 * nfsnode after we have locked the buffer to prevent 1062 * readers from reading garbage. 1063 */ 1064 bcount = on; 1065 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1066 1067 if (bp != NULL) { 1068 long save; 1069 1070 mtx_lock(&np->n_mtx); 1071 np->n_size = uio->uio_offset + n; 1072 np->n_flag |= NMODIFIED; 1073 vnode_pager_setsize(vp, np->n_size); 1074 mtx_unlock(&np->n_mtx); 1075 1076 save = bp->b_flags & B_CACHE; 1077 bcount += n; 1078 allocbuf(bp, bcount); 1079 bp->b_flags |= save; 1080 } 1081 } else { 1082 /* 1083 * Obtain the locked cache block first, and then 1084 * adjust the file's size as appropriate. 1085 */ 1086 bcount = on + n; 1087 if ((off_t)lbn * biosize + bcount < np->n_size) { 1088 if ((off_t)(lbn + 1) * biosize < np->n_size) 1089 bcount = biosize; 1090 else 1091 bcount = np->n_size - (off_t)lbn * biosize; 1092 } 1093 mtx_unlock(&np->n_mtx); 1094 bp = nfs_getcacheblk(vp, lbn, bcount, td); 1095 mtx_lock(&np->n_mtx); 1096 if (uio->uio_offset + n > np->n_size) { 1097 np->n_size = uio->uio_offset + n; 1098 np->n_flag |= NMODIFIED; 1099 vnode_pager_setsize(vp, np->n_size); 1100 } 1101 mtx_unlock(&np->n_mtx); 1102 } 1103 1104 if (!bp) { 1105 error = newnfs_sigintr(nmp, td); 1106 if (!error) 1107 error = EINTR; 1108 break; 1109 } 1110 1111 /* 1112 * Issue a READ if B_CACHE is not set. In special-append 1113 * mode, B_CACHE is based on the buffer prior to the write 1114 * op and is typically set, avoiding the read. If a read 1115 * is required in special append mode, the server will 1116 * probably send us a short-read since we extended the file 1117 * on our end, resulting in b_resid == 0 and, thusly, 1118 * B_CACHE getting set. 1119 * 1120 * We can also avoid issuing the read if the write covers 1121 * the entire buffer. We have to make sure the buffer state 1122 * is reasonable in this case since we will not be initiating 1123 * I/O. See the comments in kern/vfs_bio.c's getblk() for 1124 * more information. 1125 * 1126 * B_CACHE may also be set due to the buffer being cached 1127 * normally. 1128 */ 1129 1130 if (on == 0 && n == bcount) { 1131 bp->b_flags |= B_CACHE; 1132 bp->b_flags &= ~B_INVAL; 1133 bp->b_ioflags &= ~BIO_ERROR; 1134 } 1135 1136 if ((bp->b_flags & B_CACHE) == 0) { 1137 bp->b_iocmd = BIO_READ; 1138 vfs_busy_pages(bp, 0); 1139 error = ncl_doio(vp, bp, cred, td, 0); 1140 if (error) { 1141 brelse(bp); 1142 break; 1143 } 1144 } 1145 if (bp->b_wcred == NOCRED) 1146 bp->b_wcred = crhold(cred); 1147 mtx_lock(&np->n_mtx); 1148 np->n_flag |= NMODIFIED; 1149 mtx_unlock(&np->n_mtx); 1150 1151 /* 1152 * If dirtyend exceeds file size, chop it down. This should 1153 * not normally occur but there is an append race where it 1154 * might occur XXX, so we log it. 1155 * 1156 * If the chopping creates a reverse-indexed or degenerate 1157 * situation with dirtyoff/end, we 0 both of them. 1158 */ 1159 1160 if (bp->b_dirtyend > bcount) { 1161 ncl_printf("NFS append race @%lx:%d\n", 1162 (long)bp->b_blkno * DEV_BSIZE, 1163 bp->b_dirtyend - bcount); 1164 bp->b_dirtyend = bcount; 1165 } 1166 1167 if (bp->b_dirtyoff >= bp->b_dirtyend) 1168 bp->b_dirtyoff = bp->b_dirtyend = 0; 1169 1170 /* 1171 * If the new write will leave a contiguous dirty 1172 * area, just update the b_dirtyoff and b_dirtyend, 1173 * otherwise force a write rpc of the old dirty area. 1174 * 1175 * While it is possible to merge discontiguous writes due to 1176 * our having a B_CACHE buffer ( and thus valid read data 1177 * for the hole), we don't because it could lead to 1178 * significant cache coherency problems with multiple clients, 1179 * especially if locking is implemented later on. 1180 * 1181 * As an optimization we could theoretically maintain 1182 * a linked list of discontinuous areas, but we would still 1183 * have to commit them separately so there isn't much 1184 * advantage to it except perhaps a bit of asynchronization. 1185 */ 1186 1187 if (bp->b_dirtyend > 0 && 1188 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1189 if (bwrite(bp) == EINTR) { 1190 error = EINTR; 1191 break; 1192 } 1193 goto again; 1194 } 1195 1196 error = uiomove((char *)bp->b_data + on, n, uio); 1197 1198 /* 1199 * Since this block is being modified, it must be written 1200 * again and not just committed. Since write clustering does 1201 * not work for the stage 1 data write, only the stage 2 1202 * commit rpc, we have to clear B_CLUSTEROK as well. 1203 */ 1204 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1205 1206 if (error) { 1207 bp->b_ioflags |= BIO_ERROR; 1208 brelse(bp); 1209 break; 1210 } 1211 1212 /* 1213 * Only update dirtyoff/dirtyend if not a degenerate 1214 * condition. 1215 */ 1216 if (n) { 1217 if (bp->b_dirtyend > 0) { 1218 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1219 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1220 } else { 1221 bp->b_dirtyoff = on; 1222 bp->b_dirtyend = on + n; 1223 } 1224 vfs_bio_set_valid(bp, on, n); 1225 } 1226 1227 /* 1228 * If IO_SYNC do bwrite(). 1229 * 1230 * IO_INVAL appears to be unused. The idea appears to be 1231 * to turn off caching in this case. Very odd. XXX 1232 */ 1233 if ((ioflag & IO_SYNC)) { 1234 if (ioflag & IO_INVAL) 1235 bp->b_flags |= B_NOCACHE; 1236 error = bwrite(bp); 1237 if (error) 1238 break; 1239 } else if ((n + on) == biosize) { 1240 bp->b_flags |= B_ASYNC; 1241 (void) ncl_writebp(bp, 0, NULL); 1242 } else { 1243 bdwrite(bp); 1244 } 1245 } while (uio->uio_resid > 0 && n > 0); 1246 1247 return (error); 1248} 1249 1250/* 1251 * Get an nfs cache block. 1252 * 1253 * Allocate a new one if the block isn't currently in the cache 1254 * and return the block marked busy. If the calling process is 1255 * interrupted by a signal for an interruptible mount point, return 1256 * NULL. 1257 * 1258 * The caller must carefully deal with the possible B_INVAL state of 1259 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it 1260 * indirectly), so synchronous reads can be issued without worrying about 1261 * the B_INVAL state. We have to be a little more careful when dealing 1262 * with writes (see comments in nfs_write()) when extending a file past 1263 * its EOF. 1264 */ 1265static struct buf * 1266nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1267{ 1268 struct buf *bp; 1269 struct mount *mp; 1270 struct nfsmount *nmp; 1271 1272 mp = vp->v_mount; 1273 nmp = VFSTONFS(mp); 1274 1275 if (nmp->nm_flag & NFSMNT_INT) { 1276 sigset_t oldset; 1277 1278 newnfs_set_sigmask(td, &oldset); 1279 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0); 1280 newnfs_restore_sigmask(td, &oldset); 1281 while (bp == NULL) { 1282 if (newnfs_sigintr(nmp, td)) 1283 return (NULL); 1284 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1285 } 1286 } else { 1287 bp = getblk(vp, bn, size, 0, 0, 0); 1288 } 1289 1290 if (vp->v_type == VREG) 1291 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE); 1292 return (bp); 1293} 1294 1295/* 1296 * Flush and invalidate all dirty buffers. If another process is already 1297 * doing the flush, just wait for completion. 1298 */ 1299int 1300ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1301{ 1302 struct nfsnode *np = VTONFS(vp); 1303 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1304 int error = 0, slpflag, slptimeo; 1305 int old_lock = 0; 1306 1307 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf"); 1308 1309 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1310 intrflg = 0; 1311 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)) 1312 intrflg = 1; 1313 if (intrflg) { 1314 slpflag = NFS_PCATCH; 1315 slptimeo = 2 * hz; 1316 } else { 1317 slpflag = 0; 1318 slptimeo = 0; 1319 } 1320 1321 old_lock = ncl_upgrade_vnlock(vp); 1322 if (vp->v_iflag & VI_DOOMED) { 1323 /* 1324 * Since vgonel() uses the generic vinvalbuf() to flush 1325 * dirty buffers and it does not call this function, it 1326 * is safe to just return OK when VI_DOOMED is set. 1327 */ 1328 ncl_downgrade_vnlock(vp, old_lock); 1329 return (0); 1330 } 1331 1332 /* 1333 * Now, flush as required. 1334 */ 1335 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1336 VM_OBJECT_LOCK(vp->v_bufobj.bo_object); 1337 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1338 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); 1339 /* 1340 * If the page clean was interrupted, fail the invalidation. 1341 * Not doing so, we run the risk of losing dirty pages in the 1342 * vinvalbuf() call below. 1343 */ 1344 if (intrflg && (error = newnfs_sigintr(nmp, td))) 1345 goto out; 1346 } 1347 1348 error = vinvalbuf(vp, flags, slpflag, 0); 1349 while (error) { 1350 if (intrflg && (error = newnfs_sigintr(nmp, td))) 1351 goto out; 1352 error = vinvalbuf(vp, flags, 0, slptimeo); 1353 } 1354 mtx_lock(&np->n_mtx); 1355 if (np->n_directio_asyncwr == 0) 1356 np->n_flag &= ~NMODIFIED; 1357 mtx_unlock(&np->n_mtx); 1358out: 1359 ncl_downgrade_vnlock(vp, old_lock); 1360 return error; 1361} 1362 1363/* 1364 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1365 * This is mainly to avoid queueing async I/O requests when the nfsiods 1366 * are all hung on a dead server. 1367 * 1368 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1369 * is eventually dequeued by the async daemon, ncl_doio() *will*. 1370 */ 1371int 1372ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1373{ 1374 int iod; 1375 int gotiod; 1376 int slpflag = 0; 1377 int slptimeo = 0; 1378 int error, error2; 1379 1380 /* 1381 * Commits are usually short and sweet so lets save some cpu and 1382 * leave the async daemons for more important rpc's (such as reads 1383 * and writes). 1384 */ 1385 mtx_lock(&ncl_iod_mutex); 1386 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1387 (nmp->nm_bufqiods > ncl_numasync / 2)) { 1388 mtx_unlock(&ncl_iod_mutex); 1389 return(EIO); 1390 } 1391again: 1392 if (nmp->nm_flag & NFSMNT_INT) 1393 slpflag = NFS_PCATCH; 1394 gotiod = FALSE; 1395 1396 /* 1397 * Find a free iod to process this request. 1398 */ 1399 for (iod = 0; iod < ncl_numasync; iod++) 1400 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) { 1401 gotiod = TRUE; 1402 break; 1403 } 1404 1405 /* 1406 * Try to create one if none are free. 1407 */ 1408 if (!gotiod) 1409 ncl_nfsiodnew(); 1410 else { 1411 /* 1412 * Found one, so wake it up and tell it which 1413 * mount to process. 1414 */ 1415 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n", 1416 iod, nmp)); 1417 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE; 1418 ncl_iodmount[iod] = nmp; 1419 nmp->nm_bufqiods++; 1420 wakeup(&ncl_iodwant[iod]); 1421 } 1422 1423 /* 1424 * If none are free, we may already have an iod working on this mount 1425 * point. If so, it will process our request. 1426 */ 1427 if (!gotiod) { 1428 if (nmp->nm_bufqiods > 0) { 1429 NFS_DPF(ASYNCIO, 1430 ("ncl_asyncio: %d iods are already processing mount %p\n", 1431 nmp->nm_bufqiods, nmp)); 1432 gotiod = TRUE; 1433 } 1434 } 1435 1436 /* 1437 * If we have an iod which can process the request, then queue 1438 * the buffer. 1439 */ 1440 if (gotiod) { 1441 /* 1442 * Ensure that the queue never grows too large. We still want 1443 * to asynchronize so we block rather then return EIO. 1444 */ 1445 while (nmp->nm_bufqlen >= 2*ncl_numasync) { 1446 NFS_DPF(ASYNCIO, 1447 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp)); 1448 nmp->nm_bufqwant = TRUE; 1449 error = newnfs_msleep(td, &nmp->nm_bufq, 1450 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio", 1451 slptimeo); 1452 if (error) { 1453 error2 = newnfs_sigintr(nmp, td); 1454 if (error2) { 1455 mtx_unlock(&ncl_iod_mutex); 1456 return (error2); 1457 } 1458 if (slpflag == NFS_PCATCH) { 1459 slpflag = 0; 1460 slptimeo = 2 * hz; 1461 } 1462 } 1463 /* 1464 * We might have lost our iod while sleeping, 1465 * so check and loop if nescessary. 1466 */ 1467 goto again; 1468 } 1469 1470 /* We might have lost our nfsiod */ 1471 if (nmp->nm_bufqiods == 0) { 1472 NFS_DPF(ASYNCIO, 1473 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1474 goto again; 1475 } 1476 1477 if (bp->b_iocmd == BIO_READ) { 1478 if (bp->b_rcred == NOCRED && cred != NOCRED) 1479 bp->b_rcred = crhold(cred); 1480 } else { 1481 if (bp->b_wcred == NOCRED && cred != NOCRED) 1482 bp->b_wcred = crhold(cred); 1483 } 1484 1485 if (bp->b_flags & B_REMFREE) 1486 bremfreef(bp); 1487 BUF_KERNPROC(bp); 1488 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1489 nmp->nm_bufqlen++; 1490 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1491 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1492 VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1493 VTONFS(bp->b_vp)->n_directio_asyncwr++; 1494 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1495 } 1496 mtx_unlock(&ncl_iod_mutex); 1497 return (0); 1498 } 1499 1500 mtx_unlock(&ncl_iod_mutex); 1501 1502 /* 1503 * All the iods are busy on other mounts, so return EIO to 1504 * force the caller to process the i/o synchronously. 1505 */ 1506 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n")); 1507 return (EIO); 1508} 1509 1510void 1511ncl_doio_directwrite(struct buf *bp) 1512{ 1513 int iomode, must_commit; 1514 struct uio *uiop = (struct uio *)bp->b_caller1; 1515 char *iov_base = uiop->uio_iov->iov_base; 1516 1517 iomode = NFSWRITE_FILESYNC; 1518 uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1519 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0); 1520 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write")); 1521 free(iov_base, M_NFSDIRECTIO); 1522 free(uiop->uio_iov, M_NFSDIRECTIO); 1523 free(uiop, M_NFSDIRECTIO); 1524 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1525 struct nfsnode *np = VTONFS(bp->b_vp); 1526 mtx_lock(&np->n_mtx); 1527 np->n_directio_asyncwr--; 1528 if (np->n_directio_asyncwr == 0) { 1529 np->n_flag &= ~NMODIFIED; 1530 if ((np->n_flag & NFSYNCWAIT)) { 1531 np->n_flag &= ~NFSYNCWAIT; 1532 wakeup((caddr_t)&np->n_directio_asyncwr); 1533 } 1534 } 1535 mtx_unlock(&np->n_mtx); 1536 } 1537 bp->b_vp = NULL; 1538 relpbuf(bp, &ncl_pbuf_freecnt); 1539} 1540 1541/* 1542 * Do an I/O operation to/from a cache block. This may be called 1543 * synchronously or from an nfsiod. 1544 */ 1545int 1546ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td, 1547 int called_from_strategy) 1548{ 1549 struct uio *uiop; 1550 struct nfsnode *np; 1551 struct nfsmount *nmp; 1552 int error = 0, iomode, must_commit = 0; 1553 struct uio uio; 1554 struct iovec io; 1555 struct proc *p = td ? td->td_proc : NULL; 1556 uint8_t iocmd; 1557 1558 np = VTONFS(vp); 1559 nmp = VFSTONFS(vp->v_mount); 1560 uiop = &uio; 1561 uiop->uio_iov = &io; 1562 uiop->uio_iovcnt = 1; 1563 uiop->uio_segflg = UIO_SYSSPACE; 1564 uiop->uio_td = td; 1565 1566 /* 1567 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1568 * do this here so we do not have to do it in all the code that 1569 * calls us. 1570 */ 1571 bp->b_flags &= ~B_INVAL; 1572 bp->b_ioflags &= ~BIO_ERROR; 1573 1574 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp)); 1575 iocmd = bp->b_iocmd; 1576 if (iocmd == BIO_READ) { 1577 io.iov_len = uiop->uio_resid = bp->b_bcount; 1578 io.iov_base = bp->b_data; 1579 uiop->uio_rw = UIO_READ; 1580 1581 switch (vp->v_type) { 1582 case VREG: 1583 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1584 NFSINCRGLOBAL(newnfsstats.read_bios); 1585 error = ncl_readrpc(vp, uiop, cr); 1586 1587 if (!error) { 1588 if (uiop->uio_resid) { 1589 /* 1590 * If we had a short read with no error, we must have 1591 * hit a file hole. We should zero-fill the remainder. 1592 * This can also occur if the server hits the file EOF. 1593 * 1594 * Holes used to be able to occur due to pending 1595 * writes, but that is not possible any longer. 1596 */ 1597 int nread = bp->b_bcount - uiop->uio_resid; 1598 ssize_t left = uiop->uio_resid; 1599 1600 if (left > 0) 1601 bzero((char *)bp->b_data + nread, left); 1602 uiop->uio_resid = 0; 1603 } 1604 } 1605 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */ 1606 if (p && (vp->v_vflag & VV_TEXT)) { 1607 mtx_lock(&np->n_mtx); 1608 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) { 1609 mtx_unlock(&np->n_mtx); 1610 PROC_LOCK(p); 1611 killproc(p, "text file modification"); 1612 PROC_UNLOCK(p); 1613 } else 1614 mtx_unlock(&np->n_mtx); 1615 } 1616 break; 1617 case VLNK: 1618 uiop->uio_offset = (off_t)0; 1619 NFSINCRGLOBAL(newnfsstats.readlink_bios); 1620 error = ncl_readlinkrpc(vp, uiop, cr); 1621 break; 1622 case VDIR: 1623 NFSINCRGLOBAL(newnfsstats.readdir_bios); 1624 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1625 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1626 error = ncl_readdirplusrpc(vp, uiop, cr, td); 1627 if (error == NFSERR_NOTSUPP) 1628 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1629 } 1630 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1631 error = ncl_readdirrpc(vp, uiop, cr, td); 1632 /* 1633 * end-of-directory sets B_INVAL but does not generate an 1634 * error. 1635 */ 1636 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1637 bp->b_flags |= B_INVAL; 1638 break; 1639 default: 1640 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type); 1641 break; 1642 }; 1643 if (error) { 1644 bp->b_ioflags |= BIO_ERROR; 1645 bp->b_error = error; 1646 } 1647 } else { 1648 /* 1649 * If we only need to commit, try to commit 1650 */ 1651 if (bp->b_flags & B_NEEDCOMMIT) { 1652 int retv; 1653 off_t off; 1654 1655 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1656 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1657 bp->b_wcred, td); 1658 if (retv == 0) { 1659 bp->b_dirtyoff = bp->b_dirtyend = 0; 1660 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1661 bp->b_resid = 0; 1662 bufdone(bp); 1663 return (0); 1664 } 1665 if (retv == NFSERR_STALEWRITEVERF) { 1666 ncl_clearcommit(vp->v_mount); 1667 } 1668 } 1669 1670 /* 1671 * Setup for actual write 1672 */ 1673 mtx_lock(&np->n_mtx); 1674 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1675 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1676 mtx_unlock(&np->n_mtx); 1677 1678 if (bp->b_dirtyend > bp->b_dirtyoff) { 1679 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1680 - bp->b_dirtyoff; 1681 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1682 + bp->b_dirtyoff; 1683 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1684 uiop->uio_rw = UIO_WRITE; 1685 NFSINCRGLOBAL(newnfsstats.write_bios); 1686 1687 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1688 iomode = NFSWRITE_UNSTABLE; 1689 else 1690 iomode = NFSWRITE_FILESYNC; 1691 1692 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit, 1693 called_from_strategy); 1694 1695 /* 1696 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1697 * to cluster the buffers needing commit. This will allow 1698 * the system to submit a single commit rpc for the whole 1699 * cluster. We can do this even if the buffer is not 100% 1700 * dirty (relative to the NFS blocksize), so we optimize the 1701 * append-to-file-case. 1702 * 1703 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1704 * cleared because write clustering only works for commit 1705 * rpc's, not for the data portion of the write). 1706 */ 1707 1708 if (!error && iomode == NFSWRITE_UNSTABLE) { 1709 bp->b_flags |= B_NEEDCOMMIT; 1710 if (bp->b_dirtyoff == 0 1711 && bp->b_dirtyend == bp->b_bcount) 1712 bp->b_flags |= B_CLUSTEROK; 1713 } else { 1714 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1715 } 1716 1717 /* 1718 * For an interrupted write, the buffer is still valid 1719 * and the write hasn't been pushed to the server yet, 1720 * so we can't set BIO_ERROR and report the interruption 1721 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1722 * is not relevant, so the rpc attempt is essentially 1723 * a noop. For the case of a V3 write rpc not being 1724 * committed to stable storage, the block is still 1725 * dirty and requires either a commit rpc or another 1726 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1727 * the block is reused. This is indicated by setting 1728 * the B_DELWRI and B_NEEDCOMMIT flags. 1729 * 1730 * EIO is returned by ncl_writerpc() to indicate a recoverable 1731 * write error and is handled as above, except that 1732 * B_EINTR isn't set. One cause of this is a stale stateid 1733 * error for the RPC that indicates recovery is required, 1734 * when called with called_from_strategy != 0. 1735 * 1736 * If the buffer is marked B_PAGING, it does not reside on 1737 * the vp's paging queues so we cannot call bdirty(). The 1738 * bp in this case is not an NFS cache block so we should 1739 * be safe. XXX 1740 * 1741 * The logic below breaks up errors into recoverable and 1742 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1743 * and keep the buffer around for potential write retries. 1744 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1745 * and save the error in the nfsnode. This is less than ideal 1746 * but necessary. Keeping such buffers around could potentially 1747 * cause buffer exhaustion eventually (they can never be written 1748 * out, so will get constantly be re-dirtied). It also causes 1749 * all sorts of vfs panics. For non-recoverable write errors, 1750 * also invalidate the attrcache, so we'll be forced to go over 1751 * the wire for this object, returning an error to user on next 1752 * call (most of the time). 1753 */ 1754 if (error == EINTR || error == EIO || error == ETIMEDOUT 1755 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1756 int s; 1757 1758 s = splbio(); 1759 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1760 if ((bp->b_flags & B_PAGING) == 0) { 1761 bdirty(bp); 1762 bp->b_flags &= ~B_DONE; 1763 } 1764 if ((error == EINTR || error == ETIMEDOUT) && 1765 (bp->b_flags & B_ASYNC) == 0) 1766 bp->b_flags |= B_EINTR; 1767 splx(s); 1768 } else { 1769 if (error) { 1770 bp->b_ioflags |= BIO_ERROR; 1771 bp->b_flags |= B_INVAL; 1772 bp->b_error = np->n_error = error; 1773 mtx_lock(&np->n_mtx); 1774 np->n_flag |= NWRITEERR; 1775 np->n_attrstamp = 0; 1776 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1777 mtx_unlock(&np->n_mtx); 1778 } 1779 bp->b_dirtyoff = bp->b_dirtyend = 0; 1780 } 1781 } else { 1782 bp->b_resid = 0; 1783 bufdone(bp); 1784 return (0); 1785 } 1786 } 1787 bp->b_resid = uiop->uio_resid; 1788 if (must_commit) 1789 ncl_clearcommit(vp->v_mount); 1790 bufdone(bp); 1791 return (error); 1792} 1793 1794/* 1795 * Used to aid in handling ftruncate() operations on the NFS client side. 1796 * Truncation creates a number of special problems for NFS. We have to 1797 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1798 * we have to properly handle VM pages or (potentially dirty) buffers 1799 * that straddle the truncation point. 1800 */ 1801 1802int 1803ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1804{ 1805 struct nfsnode *np = VTONFS(vp); 1806 u_quad_t tsize; 1807 int biosize = vp->v_bufobj.bo_bsize; 1808 int error = 0; 1809 1810 mtx_lock(&np->n_mtx); 1811 tsize = np->n_size; 1812 np->n_size = nsize; 1813 mtx_unlock(&np->n_mtx); 1814 1815 if (nsize < tsize) { 1816 struct buf *bp; 1817 daddr_t lbn; 1818 int bufsize; 1819 1820 /* 1821 * vtruncbuf() doesn't get the buffer overlapping the 1822 * truncation point. We may have a B_DELWRI and/or B_CACHE 1823 * buffer that now needs to be truncated. 1824 */ 1825 error = vtruncbuf(vp, cred, td, nsize, biosize); 1826 lbn = nsize / biosize; 1827 bufsize = nsize & (biosize - 1); 1828 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1829 if (!bp) 1830 return EINTR; 1831 if (bp->b_dirtyoff > bp->b_bcount) 1832 bp->b_dirtyoff = bp->b_bcount; 1833 if (bp->b_dirtyend > bp->b_bcount) 1834 bp->b_dirtyend = bp->b_bcount; 1835 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1836 brelse(bp); 1837 } else { 1838 vnode_pager_setsize(vp, nsize); 1839 } 1840 return(error); 1841} 1842 1843