1 /* 2 * Copyright (c) 1989, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 35 * $FreeBSD: src/sys/kern/vfs_subr.c,v 1.249.2.30 2003/04/04 20:35:57 tegge Exp $ 36 */ 37 38 /* 39 * External virtual filesystem routines 40 */ 41 #include "opt_ddb.h" 42 #include "opt_inet.h" 43 #include "opt_inet6.h" 44 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/uio.h> 48 #include <sys/buf.h> 49 #include <sys/conf.h> 50 #include <sys/dirent.h> 51 #include <sys/endian.h> 52 #include <sys/eventhandler.h> 53 #include <sys/fcntl.h> 54 #include <sys/file.h> 55 #include <sys/kernel.h> 56 #include <sys/kthread.h> 57 #include <sys/malloc.h> 58 #include <sys/mbuf.h> 59 #include <sys/mount.h> 60 #include <sys/priv.h> 61 #include <sys/proc.h> 62 #include <sys/reboot.h> 63 #include <sys/socket.h> 64 #include <sys/stat.h> 65 #include <sys/sysctl.h> 66 #include <sys/syslog.h> 67 #include <sys/unistd.h> 68 #include <sys/vmmeter.h> 69 #include <sys/vnode.h> 70 71 #include <machine/limits.h> 72 73 #include <vm/vm.h> 74 #include <vm/vm_object.h> 75 #include <vm/vm_extern.h> 76 #include <vm/vm_kern.h> 77 #include <vm/pmap.h> 78 #include <vm/vm_map.h> 79 #include <vm/vm_page.h> 80 #include <vm/vm_pager.h> 81 #include <vm/vnode_pager.h> 82 #include <vm/vm_zone.h> 83 84 #include <sys/buf2.h> 85 #include <vm/vm_page2.h> 86 87 #include <netinet/in.h> 88 89 static MALLOC_DEFINE(M_NETCRED, "Export Host", "Export host address structure"); 90 91 __read_mostly int numvnodes; 92 SYSCTL_INT(_debug, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, 93 "Number of vnodes allocated"); 94 __read_mostly int verbose_reclaims; 95 SYSCTL_INT(_debug, OID_AUTO, verbose_reclaims, CTLFLAG_RD, &verbose_reclaims, 0, 96 "Output filename of reclaimed vnode(s)"); 97 98 __read_mostly enum vtype iftovt_tab[16] = { 99 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 100 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 101 }; 102 __read_mostly int vttoif_tab[9] = { 103 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 104 S_IFSOCK, S_IFIFO, S_IFMT, 105 }; 106 107 static int reassignbufcalls; 108 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 109 0, "Number of times buffers have been reassigned to the proper list"); 110 111 __read_mostly static int check_buf_overlap = 2; /* invasive check */ 112 SYSCTL_INT(_vfs, OID_AUTO, check_buf_overlap, CTLFLAG_RW, &check_buf_overlap, 113 0, "Enable overlapping buffer checks"); 114 115 int nfs_mount_type = -1; 116 static struct lwkt_token spechash_token; 117 struct nfs_public nfs_pub; /* publicly exported FS */ 118 119 __read_mostly int maxvnodes; 120 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 121 &maxvnodes, 0, "Maximum number of vnodes"); 122 123 static struct radix_node_head *vfs_create_addrlist_af(int af, 124 struct netexport *nep); 125 static void vfs_free_addrlist (struct netexport *nep); 126 static int vfs_free_netcred (struct radix_node *rn, void *w); 127 static void vfs_free_addrlist_af (struct radix_node_head **prnh); 128 static int vfs_hang_addrlist (struct mount *mp, struct netexport *nep, 129 const struct export_args *argp); 130 static void vclean_vxlocked(struct vnode *vp, int flags); 131 132 __read_mostly int prtactive = 0; /* 1 => print out reclaim of active vnodes */ 133 134 /* 135 * Red black tree functions 136 */ 137 static int rb_buf_compare(struct buf *b1, struct buf *b2); 138 RB_GENERATE2(buf_rb_tree, buf, b_rbnode, rb_buf_compare, off_t, b_loffset); 139 RB_GENERATE2(buf_rb_hash, buf, b_rbhash, rb_buf_compare, off_t, b_loffset); 140 141 static int 142 rb_buf_compare(struct buf *b1, struct buf *b2) 143 { 144 if (b1->b_loffset < b2->b_loffset) 145 return(-1); 146 if (b1->b_loffset > b2->b_loffset) 147 return(1); 148 return(0); 149 } 150 151 /* 152 * Initialize the vnode management data structures. 153 * 154 * Called from vfsinit() 155 */ 156 #define MAXVNBREAKMEM (1L * 1024 * 1024 * 1024) 157 #define MINVNODES 2000 158 #define MAXVNODES 4000000 159 160 void 161 vfs_subr_init(void) 162 { 163 int factor1; /* Limit based on ram (x 2 above 1GB) */ 164 int factor2; /* Limit based on available KVM */ 165 size_t freemem; 166 167 /* 168 * Size maxvnodes to available memory. Size significantly 169 * smaller on low-memory systems (calculations for the first 170 * 1GB of ram), and pump it up a bit when free memory is 171 * above 1GB. 172 * 173 * The general minimum is maxproc * 8 (we want someone pushing 174 * up maxproc a lot to also get more vnodes). Usually maxproc 175 * does not affect this calculation. 176 * 177 * There isn't much of a point allowing maxvnodes to exceed a 178 * few million as our modern filesystems cache pages in the 179 * underlying block device and not so much hanging off of VM 180 * objects. 181 */ 182 factor1 = 50 * (sizeof(struct vm_object) + sizeof(struct vnode)); 183 factor2 = 30 * (sizeof(struct vm_object) + sizeof(struct vnode)); 184 185 freemem = (int64_t)vmstats.v_page_count * PAGE_SIZE; 186 187 maxvnodes = freemem / factor1; 188 if (freemem > MAXVNBREAKMEM) 189 maxvnodes += (freemem - MAXVNBREAKMEM) / factor1; 190 maxvnodes = imax(maxvnodes, maxproc * 8); 191 maxvnodes = imin(maxvnodes, KvaSize / factor2); 192 maxvnodes = imin(maxvnodes, MAXVNODES); 193 maxvnodes = imax(maxvnodes, MINVNODES); 194 195 lwkt_token_init(&spechash_token, "spechash"); 196 } 197 198 /* 199 * Knob to control the precision of file timestamps: 200 * 201 * 0 = seconds only; nanoseconds zeroed. 202 * 1 = seconds and nanoseconds, accurate within 1/HZ. 203 * 2 = seconds and nanoseconds, truncated to microseconds. 204 * >=3 = seconds and nanoseconds, maximum precision. 205 * 206 * Note that utimes() precision is microseconds because it takes a timeval 207 * structure, so its probably best to default to USEC and not NSEC. 208 */ 209 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC, 210 TSP_USEC_PRECISE, TSP_NSEC_PRECISE }; 211 212 __read_mostly static int timestamp_precision = TSP_USEC; 213 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 214 ×tamp_precision, 0, "Precision of file timestamps"); 215 216 /* 217 * Get a current timestamp. 218 * 219 * MPSAFE 220 */ 221 void 222 vfs_timestamp(struct timespec *tsp) 223 { 224 switch (timestamp_precision) { 225 case TSP_SEC: /* seconds precision */ 226 getnanotime(tsp); 227 tsp->tv_nsec = 0; 228 break; 229 default: 230 case TSP_HZ: /* ticks precision (limit to microseconds) */ 231 getnanotime(tsp); 232 tsp->tv_nsec -= tsp->tv_nsec % 1000; 233 break; 234 case TSP_USEC: /* microseconds (ticks precision) */ 235 getnanotime(tsp); 236 tsp->tv_nsec -= tsp->tv_nsec % 1000; 237 break; 238 case TSP_NSEC: /* nanoseconds (ticks precision) */ 239 getnanotime(tsp); 240 break; 241 case TSP_USEC_PRECISE: /* microseconds (high preceision) */ 242 nanotime(tsp); 243 tsp->tv_nsec -= tsp->tv_nsec % 1000; 244 break; 245 case TSP_NSEC_PRECISE: /* nanoseconds (high precision) */ 246 nanotime(tsp); 247 break; 248 } 249 } 250 251 /* 252 * Set vnode attributes to VNOVAL 253 */ 254 void 255 vattr_null(struct vattr *vap) 256 { 257 vap->va_type = VNON; 258 vap->va_size = VNOVAL; 259 vap->va_bytes = VNOVAL; 260 vap->va_mode = VNOVAL; 261 vap->va_nlink = VNOVAL; 262 vap->va_uid = VNOVAL; 263 vap->va_gid = VNOVAL; 264 vap->va_fsid = VNOVAL; 265 vap->va_fileid = VNOVAL; 266 vap->va_blocksize = VNOVAL; 267 vap->va_rmajor = VNOVAL; 268 vap->va_rminor = VNOVAL; 269 vap->va_atime.tv_sec = VNOVAL; 270 vap->va_atime.tv_nsec = VNOVAL; 271 vap->va_mtime.tv_sec = VNOVAL; 272 vap->va_mtime.tv_nsec = VNOVAL; 273 vap->va_ctime.tv_sec = VNOVAL; 274 vap->va_ctime.tv_nsec = VNOVAL; 275 vap->va_flags = VNOVAL; 276 vap->va_gen = VNOVAL; 277 vap->va_vaflags = 0; 278 /* va_*_uuid fields are only valid if related flags are set */ 279 } 280 281 /* 282 * Flush out and invalidate all buffers associated with a vnode. 283 * 284 * vp must be locked. 285 */ 286 static int vinvalbuf_bp(struct buf *bp, void *data); 287 288 struct vinvalbuf_bp_info { 289 struct vnode *vp; 290 int slptimeo; 291 int lkflags; 292 int flags; 293 int clean; 294 }; 295 296 int 297 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 298 { 299 struct vinvalbuf_bp_info info; 300 vm_object_t object; 301 int error; 302 303 lwkt_gettoken(&vp->v_token); 304 305 /* 306 * If we are being asked to save, call fsync to ensure that the inode 307 * is updated. 308 */ 309 if (flags & V_SAVE) { 310 error = bio_track_wait(&vp->v_track_write, slpflag, slptimeo); 311 if (error) 312 goto done; 313 if (!RB_EMPTY(&vp->v_rbdirty_tree)) { 314 if ((error = VOP_FSYNC(vp, MNT_WAIT, 0)) != 0) 315 goto done; 316 #if 0 317 /* 318 * Dirty bufs may be left or generated via races 319 * in circumstances where vinvalbuf() is called on 320 * a vnode not undergoing reclamation. Only 321 * panic if we are trying to reclaim the vnode. 322 */ 323 if ((vp->v_flag & VRECLAIMED) && 324 (bio_track_active(&vp->v_track_write) || 325 !RB_EMPTY(&vp->v_rbdirty_tree))) { 326 panic("vinvalbuf: dirty bufs"); 327 } 328 #endif 329 } 330 } 331 info.slptimeo = slptimeo; 332 info.lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 333 if (slpflag & PCATCH) 334 info.lkflags |= LK_PCATCH; 335 info.flags = flags; 336 info.vp = vp; 337 338 /* 339 * Flush the buffer cache until nothing is left, wait for all I/O 340 * to complete. At least one pass is required. We might block 341 * in the pip code so we have to re-check. Order is important. 342 */ 343 do { 344 /* 345 * Flush buffer cache 346 */ 347 if (!RB_EMPTY(&vp->v_rbclean_tree)) { 348 info.clean = 1; 349 error = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 350 NULL, vinvalbuf_bp, &info); 351 } 352 if (!RB_EMPTY(&vp->v_rbdirty_tree)) { 353 info.clean = 0; 354 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 355 NULL, vinvalbuf_bp, &info); 356 } 357 358 /* 359 * Wait for I/O completion. 360 */ 361 bio_track_wait(&vp->v_track_write, 0, 0); 362 if ((object = vp->v_object) != NULL) 363 refcount_wait(&object->paging_in_progress, "vnvlbx"); 364 } while (bio_track_active(&vp->v_track_write) || 365 !RB_EMPTY(&vp->v_rbclean_tree) || 366 !RB_EMPTY(&vp->v_rbdirty_tree)); 367 368 /* 369 * Destroy the copy in the VM cache, too. 370 */ 371 if ((object = vp->v_object) != NULL) { 372 vm_object_page_remove(object, 0, 0, 373 (flags & V_SAVE) ? TRUE : FALSE); 374 } 375 376 if (!RB_EMPTY(&vp->v_rbdirty_tree) || !RB_EMPTY(&vp->v_rbclean_tree)) 377 panic("vinvalbuf: flush failed"); 378 if (!RB_EMPTY(&vp->v_rbhash_tree)) 379 panic("vinvalbuf: flush failed, buffers still present"); 380 error = 0; 381 done: 382 lwkt_reltoken(&vp->v_token); 383 return (error); 384 } 385 386 static int 387 vinvalbuf_bp(struct buf *bp, void *data) 388 { 389 struct vinvalbuf_bp_info *info = data; 390 int error; 391 392 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 393 atomic_add_int(&bp->b_refs, 1); 394 error = BUF_TIMELOCK(bp, info->lkflags, 395 "vinvalbuf", info->slptimeo); 396 atomic_subtract_int(&bp->b_refs, 1); 397 if (error == 0) { 398 BUF_UNLOCK(bp); 399 error = ENOLCK; 400 } 401 if (error == ENOLCK) 402 return(0); 403 return (-error); 404 } 405 KKASSERT(bp->b_vp == info->vp); 406 407 /* 408 * Must check clean/dirty status after successfully locking as 409 * it may race. 410 */ 411 if ((info->clean && (bp->b_flags & B_DELWRI)) || 412 (info->clean == 0 && (bp->b_flags & B_DELWRI) == 0)) { 413 BUF_UNLOCK(bp); 414 return(0); 415 } 416 417 /* 418 * NOTE: NO B_LOCKED CHECK. Also no buf_checkwrite() 419 * check. This code will write out the buffer, period. 420 */ 421 bremfree(bp); 422 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 423 (info->flags & V_SAVE)) { 424 cluster_awrite(bp); 425 } else if (info->flags & V_SAVE) { 426 /* 427 * Cannot set B_NOCACHE on a clean buffer as this will 428 * destroy the VM backing store which might actually 429 * be dirty (and unsynchronized). 430 */ 431 bp->b_flags |= (B_INVAL | B_RELBUF); 432 brelse(bp); 433 } else { 434 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 435 brelse(bp); 436 } 437 return(0); 438 } 439 440 /* 441 * Truncate a file's buffer and pages to a specified length. This 442 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 443 * sync activity. 444 * 445 * The vnode must be locked. 446 */ 447 static int vtruncbuf_bp_trunc_cmp(struct buf *bp, void *data); 448 static int vtruncbuf_bp_trunc(struct buf *bp, void *data); 449 static int vtruncbuf_bp_metasync_cmp(struct buf *bp, void *data); 450 static int vtruncbuf_bp_metasync(struct buf *bp, void *data); 451 452 struct vtruncbuf_info { 453 struct vnode *vp; 454 off_t truncloffset; 455 int clean; 456 }; 457 458 int 459 vtruncbuf(struct vnode *vp, off_t length, int blksize) 460 { 461 struct vtruncbuf_info info; 462 const char *filename; 463 int count; 464 465 /* 466 * Round up to the *next* block, then destroy the buffers in question. 467 * Since we are only removing some of the buffers we must rely on the 468 * scan count to determine whether a loop is necessary. 469 */ 470 if ((count = (int)(length % blksize)) != 0) 471 info.truncloffset = length + (blksize - count); 472 else 473 info.truncloffset = length; 474 info.vp = vp; 475 476 lwkt_gettoken(&vp->v_token); 477 do { 478 info.clean = 1; 479 count = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 480 vtruncbuf_bp_trunc_cmp, 481 vtruncbuf_bp_trunc, &info); 482 info.clean = 0; 483 count += RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 484 vtruncbuf_bp_trunc_cmp, 485 vtruncbuf_bp_trunc, &info); 486 } while(count); 487 488 /* 489 * For safety, fsync any remaining metadata if the file is not being 490 * truncated to 0. Since the metadata does not represent the entire 491 * dirty list we have to rely on the hit count to ensure that we get 492 * all of it. 493 */ 494 if (length > 0) { 495 do { 496 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 497 vtruncbuf_bp_metasync_cmp, 498 vtruncbuf_bp_metasync, &info); 499 } while (count); 500 } 501 502 /* 503 * Clean out any left over VM backing store. 504 * 505 * It is possible to have in-progress I/O from buffers that were 506 * not part of the truncation. This should not happen if we 507 * are truncating to 0-length. 508 */ 509 vnode_pager_setsize(vp, length); 510 bio_track_wait(&vp->v_track_write, 0, 0); 511 512 /* 513 * Debugging only 514 */ 515 spin_lock(&vp->v_spin); 516 filename = TAILQ_FIRST(&vp->v_namecache) ? 517 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"; 518 spin_unlock(&vp->v_spin); 519 520 /* 521 * Make sure no buffers were instantiated while we were trying 522 * to clean out the remaining VM pages. This could occur due 523 * to busy dirty VM pages being flushed out to disk. 524 */ 525 do { 526 info.clean = 1; 527 count = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 528 vtruncbuf_bp_trunc_cmp, 529 vtruncbuf_bp_trunc, &info); 530 info.clean = 0; 531 count += RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 532 vtruncbuf_bp_trunc_cmp, 533 vtruncbuf_bp_trunc, &info); 534 if (count) { 535 kprintf("Warning: vtruncbuf(): Had to re-clean %d " 536 "left over buffers in %s\n", count, filename); 537 } 538 } while(count); 539 540 lwkt_reltoken(&vp->v_token); 541 542 return (0); 543 } 544 545 /* 546 * The callback buffer is beyond the new file EOF and must be destroyed. 547 * Note that the compare function must conform to the RB_SCAN's requirements. 548 */ 549 static 550 int 551 vtruncbuf_bp_trunc_cmp(struct buf *bp, void *data) 552 { 553 struct vtruncbuf_info *info = data; 554 555 if (bp->b_loffset >= info->truncloffset) 556 return(0); 557 return(-1); 558 } 559 560 static 561 int 562 vtruncbuf_bp_trunc(struct buf *bp, void *data) 563 { 564 struct vtruncbuf_info *info = data; 565 566 /* 567 * Do not try to use a buffer we cannot immediately lock, but sleep 568 * anyway to prevent a livelock. The code will loop until all buffers 569 * can be acted upon. 570 * 571 * We must always revalidate the buffer after locking it to deal 572 * with MP races. 573 */ 574 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 575 atomic_add_int(&bp->b_refs, 1); 576 if (BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL) == 0) 577 BUF_UNLOCK(bp); 578 atomic_subtract_int(&bp->b_refs, 1); 579 } else if ((info->clean && (bp->b_flags & B_DELWRI)) || 580 (info->clean == 0 && (bp->b_flags & B_DELWRI) == 0) || 581 bp->b_vp != info->vp || 582 vtruncbuf_bp_trunc_cmp(bp, data)) { 583 BUF_UNLOCK(bp); 584 } else { 585 bremfree(bp); 586 bp->b_flags |= (B_INVAL | B_RELBUF | B_NOCACHE); 587 brelse(bp); 588 } 589 return(1); 590 } 591 592 /* 593 * Fsync all meta-data after truncating a file to be non-zero. Only metadata 594 * blocks (with a negative loffset) are scanned. 595 * Note that the compare function must conform to the RB_SCAN's requirements. 596 */ 597 static int 598 vtruncbuf_bp_metasync_cmp(struct buf *bp, void *data __unused) 599 { 600 if (bp->b_loffset < 0) 601 return(0); 602 return(1); 603 } 604 605 static int 606 vtruncbuf_bp_metasync(struct buf *bp, void *data) 607 { 608 struct vtruncbuf_info *info = data; 609 610 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 611 atomic_add_int(&bp->b_refs, 1); 612 if (BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL) == 0) 613 BUF_UNLOCK(bp); 614 atomic_subtract_int(&bp->b_refs, 1); 615 } else if ((bp->b_flags & B_DELWRI) == 0 || 616 bp->b_vp != info->vp || 617 vtruncbuf_bp_metasync_cmp(bp, data)) { 618 BUF_UNLOCK(bp); 619 } else { 620 bremfree(bp); 621 if (bp->b_vp == info->vp) 622 bawrite(bp); 623 else 624 bwrite(bp); 625 } 626 return(1); 627 } 628 629 /* 630 * vfsync - implements a multipass fsync on a file which understands 631 * dependancies and meta-data. The passed vnode must be locked. The 632 * waitfor argument may be MNT_WAIT or MNT_NOWAIT, or MNT_LAZY. 633 * 634 * When fsyncing data asynchronously just do one consolidated pass starting 635 * with the most negative block number. This may not get all the data due 636 * to dependancies. 637 * 638 * When fsyncing data synchronously do a data pass, then a metadata pass, 639 * then do additional data+metadata passes to try to get all the data out. 640 * 641 * Caller must ref the vnode but does not have to lock it. 642 */ 643 static int vfsync_wait_output(struct vnode *vp, 644 int (*waitoutput)(struct vnode *, struct thread *)); 645 static int vfsync_dummy_cmp(struct buf *bp __unused, void *data __unused); 646 static int vfsync_data_only_cmp(struct buf *bp, void *data); 647 static int vfsync_meta_only_cmp(struct buf *bp, void *data); 648 static int vfsync_lazy_range_cmp(struct buf *bp, void *data); 649 static int vfsync_bp(struct buf *bp, void *data); 650 651 struct vfsync_info { 652 struct vnode *vp; 653 int fastpass; 654 int synchronous; 655 int syncdeps; 656 int lazycount; 657 int lazylimit; 658 int skippedbufs; 659 int (*checkdef)(struct buf *); 660 int (*cmpfunc)(struct buf *, void *); 661 }; 662 663 int 664 vfsync(struct vnode *vp, int waitfor, int passes, 665 int (*checkdef)(struct buf *), 666 int (*waitoutput)(struct vnode *, struct thread *)) 667 { 668 struct vfsync_info info; 669 int error; 670 671 bzero(&info, sizeof(info)); 672 info.vp = vp; 673 if ((info.checkdef = checkdef) == NULL) 674 info.syncdeps = 1; 675 676 lwkt_gettoken(&vp->v_token); 677 678 switch(waitfor) { 679 case MNT_LAZY | MNT_NOWAIT: 680 case MNT_LAZY: 681 /* 682 * Lazy (filesystem syncer typ) Asynchronous plus limit the 683 * number of data (not meta) pages we try to flush to 1MB. 684 * A non-zero return means that lazy limit was reached. 685 */ 686 info.lazylimit = 1024 * 1024; 687 info.syncdeps = 1; 688 info.cmpfunc = vfsync_lazy_range_cmp; 689 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 690 vfsync_lazy_range_cmp, vfsync_bp, &info); 691 info.cmpfunc = vfsync_meta_only_cmp; 692 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 693 vfsync_meta_only_cmp, vfsync_bp, &info); 694 if (error == 0) 695 vp->v_lazyw = 0; 696 else if (!RB_EMPTY(&vp->v_rbdirty_tree)) 697 vn_syncer_add(vp, 1); 698 error = 0; 699 break; 700 case MNT_NOWAIT: 701 /* 702 * Asynchronous. Do a data-only pass and a meta-only pass. 703 */ 704 info.syncdeps = 1; 705 info.cmpfunc = vfsync_data_only_cmp; 706 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, vfsync_data_only_cmp, 707 vfsync_bp, &info); 708 info.cmpfunc = vfsync_meta_only_cmp; 709 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, vfsync_meta_only_cmp, 710 vfsync_bp, &info); 711 error = 0; 712 break; 713 default: 714 /* 715 * Synchronous. Do a data-only pass, then a meta-data+data 716 * pass, then additional integrated passes to try to get 717 * all the dependancies flushed. 718 */ 719 info.cmpfunc = vfsync_data_only_cmp; 720 info.fastpass = 1; 721 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, vfsync_data_only_cmp, 722 vfsync_bp, &info); 723 info.fastpass = 0; 724 error = vfsync_wait_output(vp, waitoutput); 725 if (error == 0) { 726 info.skippedbufs = 0; 727 info.cmpfunc = vfsync_dummy_cmp; 728 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 729 vfsync_bp, &info); 730 error = vfsync_wait_output(vp, waitoutput); 731 if (info.skippedbufs) { 732 kprintf("Warning: vfsync skipped %d dirty " 733 "buf%s in pass2!\n", 734 info.skippedbufs, 735 ((info.skippedbufs > 1) ? "s" : "")); 736 } 737 } 738 while (error == 0 && passes > 0 && 739 !RB_EMPTY(&vp->v_rbdirty_tree) 740 ) { 741 info.skippedbufs = 0; 742 if (--passes == 0) { 743 info.synchronous = 1; 744 info.syncdeps = 1; 745 } 746 info.cmpfunc = vfsync_dummy_cmp; 747 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 748 vfsync_bp, &info); 749 if (error < 0) 750 error = -error; 751 info.syncdeps = 1; 752 if (error == 0) 753 error = vfsync_wait_output(vp, waitoutput); 754 if (info.skippedbufs && passes == 0) { 755 kprintf("Warning: vfsync skipped %d dirty " 756 "buf%s in final pass!\n", 757 info.skippedbufs, 758 ((info.skippedbufs > 1) ? "s" : "")); 759 } 760 } 761 #if 0 762 /* 763 * This case can occur normally because vnode lock might 764 * not be held. 765 */ 766 if (!RB_EMPTY(&vp->v_rbdirty_tree)) 767 kprintf("dirty bufs left after final pass\n"); 768 #endif 769 break; 770 } 771 lwkt_reltoken(&vp->v_token); 772 773 return(error); 774 } 775 776 static int 777 vfsync_wait_output(struct vnode *vp, 778 int (*waitoutput)(struct vnode *, struct thread *)) 779 { 780 int error; 781 782 error = bio_track_wait(&vp->v_track_write, 0, 0); 783 if (waitoutput) 784 error = waitoutput(vp, curthread); 785 return(error); 786 } 787 788 static int 789 vfsync_dummy_cmp(struct buf *bp __unused, void *data __unused) 790 { 791 return(0); 792 } 793 794 static int 795 vfsync_data_only_cmp(struct buf *bp, void *data) 796 { 797 if (bp->b_loffset < 0) 798 return(-1); 799 return(0); 800 } 801 802 static int 803 vfsync_meta_only_cmp(struct buf *bp, void *data) 804 { 805 if (bp->b_loffset < 0) 806 return(0); 807 return(1); 808 } 809 810 static int 811 vfsync_lazy_range_cmp(struct buf *bp, void *data) 812 { 813 struct vfsync_info *info = data; 814 815 if (bp->b_loffset < info->vp->v_lazyw) 816 return(-1); 817 return(0); 818 } 819 820 static int 821 vfsync_bp(struct buf *bp, void *data) 822 { 823 struct vfsync_info *info = data; 824 struct vnode *vp = info->vp; 825 int error; 826 827 if (info->fastpass) { 828 /* 829 * Ignore buffers that we cannot immediately lock. 830 */ 831 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 832 /* 833 * Removed BUF_TIMELOCK(..., 1), even a 1-tick 834 * delay can mess up performance 835 * 836 * Another reason is that during a dirty-buffer 837 * scan a clustered write can start I/O on buffers 838 * ahead of the scan, causing the scan to not 839 * get a lock here. Usually this means the write 840 * is already in progress so, in fact, we *want* 841 * to skip the buffer. 842 */ 843 ++info->skippedbufs; 844 return(0); 845 } 846 } else if (info->synchronous == 0) { 847 /* 848 * Normal pass, give the buffer a little time to become 849 * available to us. 850 */ 851 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE, "bflst2", hz / 10)) { 852 ++info->skippedbufs; 853 return(0); 854 } 855 } else { 856 /* 857 * Synchronous pass, give the buffer a lot of time before 858 * giving up. 859 */ 860 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE, "bflst3", hz * 10)) { 861 ++info->skippedbufs; 862 return(0); 863 } 864 } 865 866 /* 867 * We must revalidate the buffer after locking. 868 */ 869 if ((bp->b_flags & B_DELWRI) == 0 || 870 bp->b_vp != info->vp || 871 info->cmpfunc(bp, data)) { 872 BUF_UNLOCK(bp); 873 return(0); 874 } 875 876 /* 877 * If syncdeps is not set we do not try to write buffers which have 878 * dependancies. 879 */ 880 if (!info->synchronous && info->syncdeps == 0 && info->checkdef(bp)) { 881 BUF_UNLOCK(bp); 882 return(0); 883 } 884 885 /* 886 * B_NEEDCOMMIT (primarily used by NFS) is a state where the buffer 887 * has been written but an additional handshake with the device 888 * is required before we can dispose of the buffer. We have no idea 889 * how to do this so we have to skip these buffers. 890 */ 891 if (bp->b_flags & B_NEEDCOMMIT) { 892 BUF_UNLOCK(bp); 893 return(0); 894 } 895 896 /* 897 * Ask bioops if it is ok to sync. If not the VFS may have 898 * set B_LOCKED so we have to cycle the buffer. 899 */ 900 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 901 bremfree(bp); 902 brelse(bp); 903 return(0); 904 } 905 906 if (info->synchronous) { 907 /* 908 * Synchronous flush. An error may be returned and will 909 * stop the scan. 910 */ 911 bremfree(bp); 912 error = bwrite(bp); 913 } else { 914 /* 915 * Asynchronous flush. We use the error return to support 916 * MNT_LAZY flushes. 917 * 918 * In low-memory situations we revert to synchronous 919 * operation. This should theoretically prevent the I/O 920 * path from exhausting memory in a non-recoverable way. 921 */ 922 vp->v_lazyw = bp->b_loffset; 923 bremfree(bp); 924 if (vm_page_count_min(0)) { 925 /* low memory */ 926 info->lazycount += bp->b_bufsize; 927 bwrite(bp); 928 } else { 929 /* normal */ 930 info->lazycount += cluster_awrite(bp); 931 waitrunningbufspace(); 932 /*vm_wait_nominal();*/ 933 } 934 if (info->lazylimit && info->lazycount >= info->lazylimit) 935 error = 1; 936 else 937 error = 0; 938 } 939 return(-error); 940 } 941 942 /* 943 * Associate a buffer with a vnode. 944 * 945 * MPSAFE 946 */ 947 int 948 bgetvp(struct vnode *vp, struct buf *bp, int testsize) 949 { 950 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 951 KKASSERT((bp->b_flags & (B_HASHED|B_DELWRI|B_VNCLEAN|B_VNDIRTY)) == 0); 952 953 /* 954 * Insert onto list for new vnode. 955 */ 956 lwkt_gettoken(&vp->v_token); 957 958 if (buf_rb_hash_RB_INSERT(&vp->v_rbhash_tree, bp)) { 959 lwkt_reltoken(&vp->v_token); 960 return (EEXIST); 961 } 962 963 /* 964 * Diagnostics (mainly for HAMMER debugging). Check for 965 * overlapping buffers. 966 */ 967 if (check_buf_overlap) { 968 struct buf *bx; 969 bx = buf_rb_hash_RB_PREV(bp); 970 if (bx) { 971 if (bx->b_loffset + bx->b_bufsize > bp->b_loffset) { 972 kprintf("bgetvp: overlapl %016jx/%d %016jx " 973 "bx %p bp %p\n", 974 (intmax_t)bx->b_loffset, 975 bx->b_bufsize, 976 (intmax_t)bp->b_loffset, 977 bx, bp); 978 if (check_buf_overlap > 1) 979 panic("bgetvp - overlapping buffer"); 980 } 981 } 982 bx = buf_rb_hash_RB_NEXT(bp); 983 if (bx) { 984 if (bp->b_loffset + testsize > bx->b_loffset) { 985 kprintf("bgetvp: overlapr %016jx/%d %016jx " 986 "bp %p bx %p\n", 987 (intmax_t)bp->b_loffset, 988 testsize, 989 (intmax_t)bx->b_loffset, 990 bp, bx); 991 if (check_buf_overlap > 1) 992 panic("bgetvp - overlapping buffer"); 993 } 994 } 995 } 996 bp->b_vp = vp; 997 bp->b_flags |= B_HASHED; 998 bp->b_flags |= B_VNCLEAN; 999 if (buf_rb_tree_RB_INSERT(&vp->v_rbclean_tree, bp)) 1000 panic("reassignbuf: dup lblk/clean vp %p bp %p", vp, bp); 1001 /*vhold(vp);*/ 1002 lwkt_reltoken(&vp->v_token); 1003 return(0); 1004 } 1005 1006 /* 1007 * Disassociate a buffer from a vnode. 1008 * 1009 * MPSAFE 1010 */ 1011 void 1012 brelvp(struct buf *bp) 1013 { 1014 struct vnode *vp; 1015 1016 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1017 1018 /* 1019 * Delete from old vnode list, if on one. 1020 */ 1021 vp = bp->b_vp; 1022 lwkt_gettoken(&vp->v_token); 1023 if (bp->b_flags & (B_VNDIRTY | B_VNCLEAN)) { 1024 if (bp->b_flags & B_VNDIRTY) 1025 buf_rb_tree_RB_REMOVE(&vp->v_rbdirty_tree, bp); 1026 else 1027 buf_rb_tree_RB_REMOVE(&vp->v_rbclean_tree, bp); 1028 bp->b_flags &= ~(B_VNDIRTY | B_VNCLEAN); 1029 } 1030 if (bp->b_flags & B_HASHED) { 1031 buf_rb_hash_RB_REMOVE(&vp->v_rbhash_tree, bp); 1032 bp->b_flags &= ~B_HASHED; 1033 } 1034 1035 /* 1036 * Only remove from synclist when no dirty buffers are left AND 1037 * the VFS has not flagged the vnode's inode as being dirty. 1038 */ 1039 if ((vp->v_flag & (VONWORKLST | VISDIRTY | VOBJDIRTY)) == VONWORKLST && 1040 RB_EMPTY(&vp->v_rbdirty_tree)) { 1041 vn_syncer_remove(vp, 0); 1042 } 1043 bp->b_vp = NULL; 1044 1045 lwkt_reltoken(&vp->v_token); 1046 1047 /*vdrop(vp);*/ 1048 } 1049 1050 /* 1051 * Reassign the buffer to the proper clean/dirty list based on B_DELWRI. 1052 * This routine is called when the state of the B_DELWRI bit is changed. 1053 * 1054 * Must be called with vp->v_token held. 1055 * MPSAFE 1056 */ 1057 void 1058 reassignbuf(struct buf *bp) 1059 { 1060 struct vnode *vp = bp->b_vp; 1061 int delay; 1062 1063 ASSERT_LWKT_TOKEN_HELD(&vp->v_token); 1064 ++reassignbufcalls; 1065 1066 /* 1067 * B_PAGING flagged buffers cannot be reassigned because their vp 1068 * is not fully linked in. 1069 */ 1070 if (bp->b_flags & B_PAGING) 1071 panic("cannot reassign paging buffer"); 1072 1073 if (bp->b_flags & B_DELWRI) { 1074 /* 1075 * Move to the dirty list, add the vnode to the worklist 1076 */ 1077 if (bp->b_flags & B_VNCLEAN) { 1078 buf_rb_tree_RB_REMOVE(&vp->v_rbclean_tree, bp); 1079 bp->b_flags &= ~B_VNCLEAN; 1080 } 1081 if ((bp->b_flags & B_VNDIRTY) == 0) { 1082 if (buf_rb_tree_RB_INSERT(&vp->v_rbdirty_tree, bp)) { 1083 panic("reassignbuf: dup lblk vp %p bp %p", 1084 vp, bp); 1085 } 1086 bp->b_flags |= B_VNDIRTY; 1087 } 1088 if ((vp->v_flag & VONWORKLST) == 0) { 1089 switch (vp->v_type) { 1090 case VDIR: 1091 delay = dirdelay; 1092 break; 1093 case VCHR: 1094 case VBLK: 1095 if (vp->v_rdev && 1096 vp->v_rdev->si_mountpoint != NULL) { 1097 delay = metadelay; 1098 break; 1099 } 1100 /* fall through */ 1101 default: 1102 delay = filedelay; 1103 } 1104 vn_syncer_add(vp, delay); 1105 } 1106 } else { 1107 /* 1108 * Move to the clean list, remove the vnode from the worklist 1109 * if no dirty blocks remain. 1110 */ 1111 if (bp->b_flags & B_VNDIRTY) { 1112 buf_rb_tree_RB_REMOVE(&vp->v_rbdirty_tree, bp); 1113 bp->b_flags &= ~B_VNDIRTY; 1114 } 1115 if ((bp->b_flags & B_VNCLEAN) == 0) { 1116 if (buf_rb_tree_RB_INSERT(&vp->v_rbclean_tree, bp)) { 1117 panic("reassignbuf: dup lblk vp %p bp %p", 1118 vp, bp); 1119 } 1120 bp->b_flags |= B_VNCLEAN; 1121 } 1122 1123 /* 1124 * Only remove from synclist when no dirty buffers are left 1125 * AND the VFS has not flagged the vnode's inode as being 1126 * dirty. 1127 */ 1128 if ((vp->v_flag & (VONWORKLST | VISDIRTY | VOBJDIRTY)) == 1129 VONWORKLST && 1130 RB_EMPTY(&vp->v_rbdirty_tree)) { 1131 vn_syncer_remove(vp, 0); 1132 } 1133 } 1134 } 1135 1136 /* 1137 * Create a vnode for a block device. Used for mounting the root file 1138 * system. 1139 * 1140 * A vref()'d vnode is returned. 1141 */ 1142 extern struct vop_ops *devfs_vnode_dev_vops_p; 1143 int 1144 bdevvp(cdev_t dev, struct vnode **vpp) 1145 { 1146 struct vnode *vp; 1147 struct vnode *nvp; 1148 int error; 1149 1150 if (dev == NULL) { 1151 *vpp = NULLVP; 1152 return (ENXIO); 1153 } 1154 error = getspecialvnode(VT_NON, NULL, &devfs_vnode_dev_vops_p, 1155 &nvp, 0, 0); 1156 if (error) { 1157 *vpp = NULLVP; 1158 return (error); 1159 } 1160 vp = nvp; 1161 vp->v_type = VCHR; 1162 #if 0 1163 vp->v_rdev = dev; 1164 #endif 1165 v_associate_rdev(vp, dev); 1166 vp->v_umajor = dev->si_umajor; 1167 vp->v_uminor = dev->si_uminor; 1168 vx_unlock(vp); 1169 *vpp = vp; 1170 return (0); 1171 } 1172 1173 int 1174 v_associate_rdev(struct vnode *vp, cdev_t dev) 1175 { 1176 if (dev == NULL) 1177 return(ENXIO); 1178 if (dev_is_good(dev) == 0) 1179 return(ENXIO); 1180 KKASSERT(vp->v_rdev == NULL); 1181 vp->v_rdev = reference_dev(dev); 1182 lwkt_gettoken(&spechash_token); 1183 SLIST_INSERT_HEAD(&dev->si_hlist, vp, v_cdevnext); 1184 lwkt_reltoken(&spechash_token); 1185 return(0); 1186 } 1187 1188 void 1189 v_release_rdev(struct vnode *vp) 1190 { 1191 cdev_t dev; 1192 1193 if ((dev = vp->v_rdev) != NULL) { 1194 lwkt_gettoken(&spechash_token); 1195 SLIST_REMOVE(&dev->si_hlist, vp, vnode, v_cdevnext); 1196 vp->v_rdev = NULL; 1197 release_dev(dev); 1198 lwkt_reltoken(&spechash_token); 1199 } 1200 } 1201 1202 /* 1203 * Add a vnode to the alias list hung off the cdev_t. We only associate 1204 * the device number with the vnode. The actual device is not associated 1205 * until the vnode is opened (usually in spec_open()), and will be 1206 * disassociated on last close. 1207 */ 1208 void 1209 addaliasu(struct vnode *nvp, int x, int y) 1210 { 1211 if (nvp->v_type != VBLK && nvp->v_type != VCHR) 1212 panic("addaliasu on non-special vnode"); 1213 nvp->v_umajor = x; 1214 nvp->v_uminor = y; 1215 } 1216 1217 /* 1218 * Simple call that a filesystem can make to try to get rid of a 1219 * vnode. It will fail if anyone is referencing the vnode (including 1220 * the caller). 1221 * 1222 * The filesystem can check whether its in-memory inode structure still 1223 * references the vp on return. 1224 * 1225 * May only be called if the vnode is in a known state (i.e. being prevented 1226 * from being deallocated by some other condition such as a vfs inode hold). 1227 * 1228 * This call might not succeed. 1229 */ 1230 void 1231 vclean_unlocked(struct vnode *vp) 1232 { 1233 vx_get(vp); 1234 if (VREFCNT(vp) <= 1) 1235 vgone_vxlocked(vp); 1236 vx_put(vp); 1237 } 1238 1239 /* 1240 * Disassociate a vnode from its underlying filesystem. 1241 * 1242 * The vnode must be VX locked and referenced. In all normal situations 1243 * there are no active references. If vclean_vxlocked() is called while 1244 * there are active references, the vnode is being ripped out and we have 1245 * to call VOP_CLOSE() as appropriate before we can reclaim it. 1246 */ 1247 static void 1248 vclean_vxlocked(struct vnode *vp, int flags) 1249 { 1250 int active; 1251 int n; 1252 vm_object_t object; 1253 struct namecache *ncp; 1254 1255 /* 1256 * If the vnode has already been reclaimed we have nothing to do. 1257 */ 1258 if (vp->v_flag & VRECLAIMED) 1259 return; 1260 1261 /* 1262 * Set flag to interlock operation, flag finalization to ensure 1263 * that the vnode winds up on the inactive list, and set v_act to 0. 1264 */ 1265 vsetflags(vp, VRECLAIMED); 1266 atomic_set_int(&vp->v_refcnt, VREF_FINALIZE); 1267 vp->v_act = 0; 1268 1269 if (verbose_reclaims) { 1270 if ((ncp = TAILQ_FIRST(&vp->v_namecache)) != NULL) 1271 kprintf("Debug: reclaim %p %s\n", vp, ncp->nc_name); 1272 } 1273 1274 /* 1275 * Scrap the vfs cache 1276 */ 1277 while (cache_inval_vp(vp, 0) != 0) { 1278 kprintf("Warning: vnode %p clean/cache_resolution " 1279 "race detected\n", vp); 1280 tsleep(vp, 0, "vclninv", 2); 1281 } 1282 1283 /* 1284 * Check to see if the vnode is in use. If so we have to reference it 1285 * before we clean it out so that its count cannot fall to zero and 1286 * generate a race against ourselves to recycle it. 1287 */ 1288 active = (VREFCNT(vp) > 0); 1289 1290 /* 1291 * Clean out any buffers associated with the vnode and destroy its 1292 * object, if it has one. 1293 */ 1294 vinvalbuf(vp, V_SAVE, 0, 0); 1295 1296 /* 1297 * If purging an active vnode (typically during a forced unmount 1298 * or reboot), it must be closed and deactivated before being 1299 * reclaimed. This isn't really all that safe, but what can 1300 * we do? XXX. 1301 * 1302 * Note that neither of these routines unlocks the vnode. 1303 */ 1304 if (active && (flags & DOCLOSE)) { 1305 while ((n = vp->v_opencount) != 0) { 1306 if (vp->v_writecount) 1307 VOP_CLOSE(vp, FWRITE|FNONBLOCK, NULL); 1308 else 1309 VOP_CLOSE(vp, FNONBLOCK, NULL); 1310 if (vp->v_opencount == n) { 1311 kprintf("Warning: unable to force-close" 1312 " vnode %p\n", vp); 1313 break; 1314 } 1315 } 1316 } 1317 1318 /* 1319 * If the vnode has not been deactivated, deactivated it. Deactivation 1320 * can create new buffers and VM pages so we have to call vinvalbuf() 1321 * again to make sure they all get flushed. 1322 * 1323 * This can occur if a file with a link count of 0 needs to be 1324 * truncated. 1325 * 1326 * If the vnode is already dead don't try to deactivate it. 1327 */ 1328 if ((vp->v_flag & VINACTIVE) == 0) { 1329 vsetflags(vp, VINACTIVE); 1330 if (vp->v_mount) 1331 VOP_INACTIVE(vp); 1332 vinvalbuf(vp, V_SAVE, 0, 0); 1333 } 1334 1335 /* 1336 * If the vnode has an object, destroy it. 1337 */ 1338 while ((object = vp->v_object) != NULL) { 1339 vm_object_hold(object); 1340 if (object == vp->v_object) 1341 break; 1342 vm_object_drop(object); 1343 } 1344 1345 if (object != NULL) { 1346 if (object->ref_count == 0) { 1347 if ((object->flags & OBJ_DEAD) == 0) 1348 vm_object_terminate(object); 1349 vm_object_drop(object); 1350 vclrflags(vp, VOBJBUF); 1351 } else { 1352 vm_pager_deallocate(object); 1353 vclrflags(vp, VOBJBUF); 1354 vm_object_drop(object); 1355 } 1356 } 1357 KKASSERT((vp->v_flag & VOBJBUF) == 0); 1358 1359 if (vp->v_flag & VOBJDIRTY) 1360 vclrobjdirty(vp); 1361 1362 /* 1363 * Reclaim the vnode if not already dead. 1364 */ 1365 if (vp->v_mount && VOP_RECLAIM(vp)) 1366 panic("vclean: cannot reclaim"); 1367 1368 /* 1369 * Done with purge, notify sleepers of the grim news. 1370 */ 1371 vp->v_ops = &dead_vnode_vops_p; 1372 vn_gone(vp); 1373 vp->v_tag = VT_NON; 1374 1375 /* 1376 * If we are destroying an active vnode, reactivate it now that 1377 * we have reassociated it with deadfs. This prevents the system 1378 * from crashing on the vnode due to it being unexpectedly marked 1379 * as inactive or reclaimed. 1380 */ 1381 if (active && (flags & DOCLOSE)) { 1382 vclrflags(vp, VINACTIVE | VRECLAIMED); 1383 } 1384 } 1385 1386 /* 1387 * Eliminate all activity associated with the requested vnode 1388 * and with all vnodes aliased to the requested vnode. 1389 * 1390 * The vnode must be referenced but should not be locked. 1391 */ 1392 int 1393 vrevoke(struct vnode *vp, struct ucred *cred) 1394 { 1395 struct vnode *vq; 1396 struct vnode *vqn; 1397 cdev_t dev; 1398 int error; 1399 1400 /* 1401 * If the vnode has a device association, scrap all vnodes associated 1402 * with the device. Don't let the device disappear on us while we 1403 * are scrapping the vnodes. 1404 * 1405 * The passed vp will probably show up in the list, do not VX lock 1406 * it twice! 1407 * 1408 * Releasing the vnode's rdev here can mess up specfs's call to 1409 * device close, so don't do it. The vnode has been disassociated 1410 * and the device will be closed after the last ref on the related 1411 * fp goes away (if not still open by e.g. the kernel). 1412 */ 1413 if (vp->v_type != VCHR) { 1414 error = fdrevoke(vp, DTYPE_VNODE, cred); 1415 return (error); 1416 } 1417 if ((dev = vp->v_rdev) == NULL) { 1418 return(0); 1419 } 1420 reference_dev(dev); 1421 lwkt_gettoken(&spechash_token); 1422 1423 restart: 1424 vqn = SLIST_FIRST(&dev->si_hlist); 1425 if (vqn) 1426 vhold(vqn); 1427 while ((vq = vqn) != NULL) { 1428 if (VREFCNT(vq) > 0) { 1429 vref(vq); 1430 fdrevoke(vq, DTYPE_VNODE, cred); 1431 /*v_release_rdev(vq);*/ 1432 vrele(vq); 1433 if (vq->v_rdev != dev) { 1434 vdrop(vq); 1435 goto restart; 1436 } 1437 } 1438 vqn = SLIST_NEXT(vq, v_cdevnext); 1439 if (vqn) 1440 vhold(vqn); 1441 vdrop(vq); 1442 } 1443 lwkt_reltoken(&spechash_token); 1444 dev_drevoke(dev); 1445 release_dev(dev); 1446 return (0); 1447 } 1448 1449 /* 1450 * This is called when the object underlying a vnode is being destroyed, 1451 * such as in a remove(). Try to recycle the vnode immediately if the 1452 * only active reference is our reference. 1453 * 1454 * Directory vnodes in the namecache with children cannot be immediately 1455 * recycled because numerous VOP_N*() ops require them to be stable. 1456 * 1457 * To avoid recursive recycling from VOP_INACTIVE implemenetations this 1458 * function is a NOP if VRECLAIMED is already set. 1459 */ 1460 int 1461 vrecycle(struct vnode *vp) 1462 { 1463 if (VREFCNT(vp) <= 1 && (vp->v_flag & VRECLAIMED) == 0) { 1464 if (cache_inval_vp_nonblock(vp)) 1465 return(0); 1466 vgone_vxlocked(vp); 1467 return (1); 1468 } 1469 return (0); 1470 } 1471 1472 /* 1473 * Return the maximum I/O size allowed for strategy calls on VP. 1474 * 1475 * If vp is VCHR or VBLK we dive the device, otherwise we use 1476 * the vp's mount info. 1477 * 1478 * The returned value is clamped at MAXPHYS as most callers cannot use 1479 * buffers larger than that size. 1480 */ 1481 int 1482 vmaxiosize(struct vnode *vp) 1483 { 1484 int maxiosize; 1485 1486 if (vp->v_type == VBLK || vp->v_type == VCHR) 1487 maxiosize = vp->v_rdev->si_iosize_max; 1488 else 1489 maxiosize = vp->v_mount->mnt_iosize_max; 1490 1491 if (maxiosize > MAXPHYS) 1492 maxiosize = MAXPHYS; 1493 return (maxiosize); 1494 } 1495 1496 /* 1497 * Eliminate all activity associated with a vnode in preparation for 1498 * destruction. 1499 * 1500 * The vnode must be VX locked and refd and will remain VX locked and refd 1501 * on return. This routine may be called with the vnode in any state, as 1502 * long as it is VX locked. The vnode will be cleaned out and marked 1503 * VRECLAIMED but will not actually be reused until all existing refs and 1504 * holds go away. 1505 * 1506 * NOTE: This routine may be called on a vnode which has not yet been 1507 * already been deactivated (VOP_INACTIVE), or on a vnode which has 1508 * already been reclaimed. 1509 * 1510 * This routine is not responsible for placing us back on the freelist. 1511 * Instead, it happens automatically when the caller releases the VX lock 1512 * (assuming there aren't any other references). 1513 */ 1514 void 1515 vgone_vxlocked(struct vnode *vp) 1516 { 1517 /* 1518 * assert that the VX lock is held. This is an absolute requirement 1519 * now for vgone_vxlocked() to be called. 1520 */ 1521 KKASSERT(lockinuse(&vp->v_lock)); 1522 1523 /* 1524 * Clean out the filesystem specific data and set the VRECLAIMED 1525 * bit. Also deactivate the vnode if necessary. 1526 * 1527 * The vnode should have automatically been removed from the syncer 1528 * list as syncer/dirty flags cleared during the cleaning. 1529 */ 1530 vclean_vxlocked(vp, DOCLOSE); 1531 1532 /* 1533 * Normally panic if the vnode is still dirty, unless we are doing 1534 * a forced unmount (tmpfs typically). 1535 */ 1536 if (vp->v_flag & VONWORKLST) { 1537 if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) { 1538 /* force removal */ 1539 vn_syncer_remove(vp, 1); 1540 } else { 1541 panic("vp %p still dirty in vgone after flush", vp); 1542 } 1543 } 1544 1545 /* 1546 * Delete from old mount point vnode list, if on one. 1547 */ 1548 if (vp->v_mount != NULL) { 1549 KKASSERT(vp->v_data == NULL); 1550 insmntque(vp, NULL); 1551 } 1552 1553 /* 1554 * If special device, remove it from special device alias list 1555 * if it is on one. This should normally only occur if a vnode is 1556 * being revoked as the device should otherwise have been released 1557 * naturally. 1558 */ 1559 if ((vp->v_type == VBLK || vp->v_type == VCHR) && vp->v_rdev != NULL) { 1560 v_release_rdev(vp); 1561 } 1562 1563 /* 1564 * Set us to VBAD 1565 */ 1566 vp->v_type = VBAD; 1567 } 1568 1569 /* 1570 * Calculate the total number of references to a special device. This 1571 * routine may only be called for VBLK and VCHR vnodes since v_rdev is 1572 * an overloaded field. Since dev_from_devid() can now return NULL, we 1573 * have to check for a NULL v_rdev. 1574 */ 1575 int 1576 count_dev(cdev_t dev) 1577 { 1578 struct vnode *vp; 1579 int count = 0; 1580 1581 if (SLIST_FIRST(&dev->si_hlist)) { 1582 lwkt_gettoken(&spechash_token); 1583 SLIST_FOREACH(vp, &dev->si_hlist, v_cdevnext) { 1584 count += vp->v_opencount; 1585 } 1586 lwkt_reltoken(&spechash_token); 1587 } 1588 return(count); 1589 } 1590 1591 int 1592 vcount(struct vnode *vp) 1593 { 1594 if (vp->v_rdev == NULL) 1595 return(0); 1596 return(count_dev(vp->v_rdev)); 1597 } 1598 1599 /* 1600 * Initialize VMIO for a vnode. This routine MUST be called before a 1601 * VFS can issue buffer cache ops on a vnode. It is typically called 1602 * when a vnode is initialized from its inode. 1603 */ 1604 int 1605 vinitvmio(struct vnode *vp, off_t filesize, int blksize, int boff) 1606 { 1607 vm_object_t object; 1608 int error = 0; 1609 1610 object = vp->v_object; 1611 if (object) { 1612 vm_object_hold(object); 1613 KKASSERT(vp->v_object == object); 1614 } 1615 1616 if (object == NULL) { 1617 object = vnode_pager_alloc(vp, filesize, 0, 0, blksize, boff); 1618 1619 /* 1620 * Dereference the reference we just created. This assumes 1621 * that the object is associated with the vp. Allow it to 1622 * have zero refs. It cannot be destroyed as long as it 1623 * is associated with the vnode. 1624 */ 1625 vm_object_hold(object); 1626 atomic_add_int(&object->ref_count, -1); 1627 vrele(vp); 1628 } else { 1629 KKASSERT((object->flags & OBJ_DEAD) == 0); 1630 } 1631 KASSERT(vp->v_object != NULL, ("vinitvmio: NULL object")); 1632 vsetflags(vp, VOBJBUF); 1633 vm_object_drop(object); 1634 1635 return (error); 1636 } 1637 1638 1639 /* 1640 * Print out a description of a vnode. 1641 */ 1642 static char *typename[] = 1643 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 1644 1645 void 1646 vprint(char *label, struct vnode *vp) 1647 { 1648 char buf[96]; 1649 1650 if (label != NULL) 1651 kprintf("%s: %p: ", label, (void *)vp); 1652 else 1653 kprintf("%p: ", (void *)vp); 1654 kprintf("type %s, refcnt %08x, writecount %d, holdcnt %d,", 1655 typename[vp->v_type], 1656 vp->v_refcnt, vp->v_writecount, vp->v_auxrefs); 1657 buf[0] = '\0'; 1658 if (vp->v_flag & VROOT) 1659 strcat(buf, "|VROOT"); 1660 if (vp->v_flag & VPFSROOT) 1661 strcat(buf, "|VPFSROOT"); 1662 if (vp->v_flag & VTEXT) 1663 strcat(buf, "|VTEXT"); 1664 if (vp->v_flag & VSYSTEM) 1665 strcat(buf, "|VSYSTEM"); 1666 if (vp->v_flag & VOBJBUF) 1667 strcat(buf, "|VOBJBUF"); 1668 if (buf[0] != '\0') 1669 kprintf(" flags (%s)", &buf[1]); 1670 if (vp->v_data == NULL) { 1671 kprintf("\n"); 1672 } else { 1673 kprintf("\n\t"); 1674 VOP_PRINT(vp); 1675 } 1676 } 1677 1678 /* 1679 * Do the usual access checking. 1680 * file_mode, uid and gid are from the vnode in question, 1681 * while acc_mode and cred are from the VOP_ACCESS parameter list 1682 */ 1683 int 1684 vaccess(enum vtype type, mode_t file_mode, uid_t uid, gid_t gid, 1685 mode_t acc_mode, struct ucred *cred) 1686 { 1687 mode_t mask; 1688 int ismember; 1689 1690 /* 1691 * Super-user always gets read/write access, but execute access depends 1692 * on at least one execute bit being set. 1693 */ 1694 if (priv_check_cred(cred, PRIV_ROOT, 0) == 0) { 1695 if ((acc_mode & VEXEC) && type != VDIR && 1696 (file_mode & (S_IXUSR|S_IXGRP|S_IXOTH)) == 0) 1697 return (EACCES); 1698 return (0); 1699 } 1700 1701 mask = 0; 1702 1703 /* Otherwise, check the owner. */ 1704 if (cred->cr_uid == uid) { 1705 if (acc_mode & VEXEC) 1706 mask |= S_IXUSR; 1707 if (acc_mode & VREAD) 1708 mask |= S_IRUSR; 1709 if (acc_mode & VWRITE) 1710 mask |= S_IWUSR; 1711 return ((file_mode & mask) == mask ? 0 : EACCES); 1712 } 1713 1714 /* Otherwise, check the groups. */ 1715 ismember = groupmember(gid, cred); 1716 if (cred->cr_svgid == gid || ismember) { 1717 if (acc_mode & VEXEC) 1718 mask |= S_IXGRP; 1719 if (acc_mode & VREAD) 1720 mask |= S_IRGRP; 1721 if (acc_mode & VWRITE) 1722 mask |= S_IWGRP; 1723 return ((file_mode & mask) == mask ? 0 : EACCES); 1724 } 1725 1726 /* Otherwise, check everyone else. */ 1727 if (acc_mode & VEXEC) 1728 mask |= S_IXOTH; 1729 if (acc_mode & VREAD) 1730 mask |= S_IROTH; 1731 if (acc_mode & VWRITE) 1732 mask |= S_IWOTH; 1733 return ((file_mode & mask) == mask ? 0 : EACCES); 1734 } 1735 1736 #ifdef DDB 1737 #include <ddb/ddb.h> 1738 1739 static int db_show_locked_vnodes(struct mount *mp, void *data); 1740 1741 /* 1742 * List all of the locked vnodes in the system. 1743 * Called when debugging the kernel. 1744 */ 1745 DB_SHOW_COMMAND(lockedvnodes, lockedvnodes) 1746 { 1747 kprintf("Locked vnodes\n"); 1748 mountlist_scan(db_show_locked_vnodes, NULL, 1749 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 1750 } 1751 1752 static int 1753 db_show_locked_vnodes(struct mount *mp, void *data __unused) 1754 { 1755 struct vnode *vp; 1756 1757 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 1758 if (vn_islocked(vp)) 1759 vprint(NULL, vp); 1760 } 1761 return(0); 1762 } 1763 #endif 1764 1765 /* 1766 * Top level filesystem related information gathering. 1767 */ 1768 static int sysctl_ovfs_conf (SYSCTL_HANDLER_ARGS); 1769 1770 static int 1771 vfs_sysctl(SYSCTL_HANDLER_ARGS) 1772 { 1773 int *name = (int *)arg1 - 1; /* XXX */ 1774 u_int namelen = arg2 + 1; /* XXX */ 1775 struct vfsconf *vfsp; 1776 int maxtypenum; 1777 1778 #if 1 || defined(COMPAT_PRELITE2) 1779 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 1780 if (namelen == 1) 1781 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 1782 #endif 1783 1784 #ifdef notyet 1785 /* all sysctl names at this level are at least name and field */ 1786 if (namelen < 2) 1787 return (ENOTDIR); /* overloaded */ 1788 if (name[0] != VFS_GENERIC) { 1789 vfsp = vfsconf_find_by_typenum(name[0]); 1790 if (vfsp == NULL) 1791 return (EOPNOTSUPP); 1792 return ((*vfsp->vfc_vfsops->vfs_sysctl)(&name[1], namelen - 1, 1793 oldp, oldlenp, newp, newlen, p)); 1794 } 1795 #endif 1796 switch (name[1]) { 1797 case VFS_MAXTYPENUM: 1798 if (namelen != 2) 1799 return (ENOTDIR); 1800 maxtypenum = vfsconf_get_maxtypenum(); 1801 return (SYSCTL_OUT(req, &maxtypenum, sizeof(maxtypenum))); 1802 case VFS_CONF: 1803 if (namelen != 3) 1804 return (ENOTDIR); /* overloaded */ 1805 vfsp = vfsconf_find_by_typenum(name[2]); 1806 if (vfsp == NULL) 1807 return (EOPNOTSUPP); 1808 return (SYSCTL_OUT(req, vfsp, sizeof *vfsp)); 1809 } 1810 return (EOPNOTSUPP); 1811 } 1812 1813 SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD, vfs_sysctl, 1814 "Generic filesystem"); 1815 1816 #if 1 || defined(COMPAT_PRELITE2) 1817 1818 static int 1819 sysctl_ovfs_conf_iter(struct vfsconf *vfsp, void *data) 1820 { 1821 int error; 1822 struct ovfsconf ovfs; 1823 struct sysctl_req *req = (struct sysctl_req*) data; 1824 1825 bzero(&ovfs, sizeof(ovfs)); 1826 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 1827 strcpy(ovfs.vfc_name, vfsp->vfc_name); 1828 ovfs.vfc_index = vfsp->vfc_typenum; 1829 ovfs.vfc_refcount = vfsp->vfc_refcount; 1830 ovfs.vfc_flags = vfsp->vfc_flags; 1831 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 1832 if (error) 1833 return error; /* abort iteration with error code */ 1834 else 1835 return 0; /* continue iterating with next element */ 1836 } 1837 1838 static int 1839 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 1840 { 1841 return vfsconf_each(sysctl_ovfs_conf_iter, (void*)req); 1842 } 1843 1844 #endif /* 1 || COMPAT_PRELITE2 */ 1845 1846 /* 1847 * Check to see if a filesystem is mounted on a block device. 1848 */ 1849 int 1850 vfs_mountedon(struct vnode *vp) 1851 { 1852 cdev_t dev; 1853 1854 dev = vp->v_rdev; 1855 if (dev != NULL && dev->si_mountpoint) 1856 return (EBUSY); 1857 return (0); 1858 } 1859 1860 /* 1861 * Unmount all filesystems. The list is traversed in reverse order 1862 * of mounting to avoid dependencies. 1863 * 1864 * We want the umountall to be able to break out of its loop if a 1865 * failure occurs, after scanning all possible mounts, so the callback 1866 * returns 0 on error. 1867 * 1868 * NOTE: Do not call mountlist_remove(mp) on error any more, this will 1869 * confuse mountlist_scan()'s unbusy check. 1870 */ 1871 static int vfs_umountall_callback(struct mount *mp, void *data); 1872 1873 void 1874 vfs_unmountall(int halting) 1875 { 1876 int count; 1877 1878 do { 1879 count = mountlist_scan(vfs_umountall_callback, &halting, 1880 MNTSCAN_REVERSE|MNTSCAN_NOBUSY); 1881 } while (count); 1882 } 1883 1884 static 1885 int 1886 vfs_umountall_callback(struct mount *mp, void *data) 1887 { 1888 int error; 1889 int halting = *(int *)data; 1890 1891 /* 1892 * NOTE: When halting, dounmount will disconnect but leave 1893 * certain mount points intact. e.g. devfs. 1894 */ 1895 error = dounmount(mp, MNT_FORCE, halting); 1896 if (error) { 1897 kprintf("unmount of filesystem mounted from %s failed (", 1898 mp->mnt_stat.f_mntfromname); 1899 if (error == EBUSY) 1900 kprintf("BUSY)\n"); 1901 else 1902 kprintf("%d)\n", error); 1903 return 0; 1904 } else { 1905 return 1; 1906 } 1907 } 1908 1909 /* 1910 * Checks the mount flags for parameter mp and put the names comma-separated 1911 * into a string buffer buf with a size limit specified by len. 1912 * 1913 * It returns the number of bytes written into buf, and (*errorp) will be 1914 * set to 0, EINVAL (if passed length is 0), or ENOSPC (supplied buffer was 1915 * not large enough). The buffer will be 0-terminated if len was not 0. 1916 */ 1917 size_t 1918 vfs_flagstostr(int flags, const struct mountctl_opt *optp, 1919 char *buf, size_t len, int *errorp) 1920 { 1921 static const struct mountctl_opt optnames[] = { 1922 { MNT_RDONLY, "read-only" }, 1923 { MNT_SYNCHRONOUS, "synchronous" }, 1924 { MNT_NOEXEC, "noexec" }, 1925 { MNT_NOSUID, "nosuid" }, 1926 { MNT_NODEV, "nodev" }, 1927 { MNT_AUTOMOUNTED, "automounted" }, 1928 { MNT_ASYNC, "asynchronous" }, 1929 { MNT_SUIDDIR, "suiddir" }, 1930 { MNT_SOFTDEP, "soft-updates" }, 1931 { MNT_NOSYMFOLLOW, "nosymfollow" }, 1932 { MNT_TRIM, "trim" }, 1933 { MNT_NOATIME, "noatime" }, 1934 { MNT_NOCLUSTERR, "noclusterr" }, 1935 { MNT_NOCLUSTERW, "noclusterw" }, 1936 { MNT_EXRDONLY, "NFS read-only" }, 1937 { MNT_EXPORTED, "NFS exported" }, 1938 /* Remaining NFS flags could come here */ 1939 { MNT_LOCAL, "local" }, 1940 { MNT_QUOTA, "with-quotas" }, 1941 /* { MNT_ROOTFS, "rootfs" }, */ 1942 /* { MNT_IGNORE, "ignore" }, */ 1943 { 0, NULL} 1944 }; 1945 int bwritten; 1946 int bleft; 1947 int optlen; 1948 int actsize; 1949 1950 *errorp = 0; 1951 bwritten = 0; 1952 bleft = len - 1; /* leave room for trailing \0 */ 1953 1954 /* 1955 * Checks the size of the string. If it contains 1956 * any data, then we will append the new flags to 1957 * it. 1958 */ 1959 actsize = strlen(buf); 1960 if (actsize > 0) 1961 buf += actsize; 1962 1963 /* Default flags if no flags passed */ 1964 if (optp == NULL) 1965 optp = optnames; 1966 1967 if (bleft < 0) { /* degenerate case, 0-length buffer */ 1968 *errorp = EINVAL; 1969 return(0); 1970 } 1971 1972 for (; flags && optp->o_opt; ++optp) { 1973 if ((flags & optp->o_opt) == 0) 1974 continue; 1975 optlen = strlen(optp->o_name); 1976 if (bwritten || actsize > 0) { 1977 if (bleft < 2) { 1978 *errorp = ENOSPC; 1979 break; 1980 } 1981 buf[bwritten++] = ','; 1982 buf[bwritten++] = ' '; 1983 bleft -= 2; 1984 } 1985 if (bleft < optlen) { 1986 *errorp = ENOSPC; 1987 break; 1988 } 1989 bcopy(optp->o_name, buf + bwritten, optlen); 1990 bwritten += optlen; 1991 bleft -= optlen; 1992 flags &= ~optp->o_opt; 1993 } 1994 1995 /* 1996 * Space already reserved for trailing \0 1997 */ 1998 buf[bwritten] = 0; 1999 return (bwritten); 2000 } 2001 2002 /* 2003 * Build hash lists of net addresses and hang them off the mount point. 2004 * Called by ufs_mount() to set up the lists of export addresses. 2005 */ 2006 static int 2007 vfs_hang_addrlist(struct mount *mp, struct netexport *nep, 2008 const struct export_args *argp) 2009 { 2010 struct netcred *np; 2011 struct radix_node_head *rnh; 2012 int i; 2013 struct radix_node *rn; 2014 struct sockaddr *saddr, *smask = NULL; 2015 int error; 2016 2017 if (argp->ex_addrlen == 0) { 2018 if (mp->mnt_flag & MNT_DEFEXPORTED) 2019 return (EPERM); 2020 np = &nep->ne_defexported; 2021 np->netc_exflags = argp->ex_flags; 2022 np->netc_anon = argp->ex_anon; 2023 np->netc_anon.cr_ref = 1; 2024 mp->mnt_flag |= MNT_DEFEXPORTED; 2025 return (0); 2026 } 2027 2028 if (argp->ex_addrlen < 0 || argp->ex_addrlen > MLEN) 2029 return (EINVAL); 2030 if (argp->ex_masklen < 0 || argp->ex_masklen > MLEN) 2031 return (EINVAL); 2032 2033 i = sizeof(struct netcred) + argp->ex_addrlen + argp->ex_masklen; 2034 np = (struct netcred *)kmalloc(i, M_NETCRED, M_WAITOK | M_ZERO); 2035 saddr = (struct sockaddr *) (np + 1); 2036 if ((error = copyin(argp->ex_addr, (caddr_t) saddr, argp->ex_addrlen))) 2037 goto out; 2038 if (saddr->sa_len > argp->ex_addrlen) 2039 saddr->sa_len = argp->ex_addrlen; 2040 if (argp->ex_masklen) { 2041 smask = (struct sockaddr *)((caddr_t)saddr + argp->ex_addrlen); 2042 error = copyin(argp->ex_mask, (caddr_t)smask, argp->ex_masklen); 2043 if (error) 2044 goto out; 2045 if (smask->sa_len > argp->ex_masklen) 2046 smask->sa_len = argp->ex_masklen; 2047 } 2048 NE_LOCK(nep); 2049 if (nep->ne_maskhead == NULL) { 2050 if (!rn_inithead((void **)&nep->ne_maskhead, NULL, 0)) { 2051 error = ENOBUFS; 2052 goto out; 2053 } 2054 } 2055 if ((rnh = vfs_create_addrlist_af(saddr->sa_family, nep)) == NULL) { 2056 error = ENOBUFS; 2057 goto out; 2058 } 2059 rn = (*rnh->rnh_addaddr)((char *)saddr, (char *)smask, rnh, 2060 np->netc_rnodes); 2061 NE_UNLOCK(nep); 2062 if (rn == NULL || np != (struct netcred *)rn) { /* already exists */ 2063 error = EPERM; 2064 goto out; 2065 } 2066 np->netc_exflags = argp->ex_flags; 2067 np->netc_anon = argp->ex_anon; 2068 np->netc_anon.cr_ref = 1; 2069 return (0); 2070 2071 out: 2072 kfree(np, M_NETCRED); 2073 return (error); 2074 } 2075 2076 /* 2077 * Free netcred structures installed in the netexport 2078 */ 2079 static int 2080 vfs_free_netcred(struct radix_node *rn, void *w) 2081 { 2082 struct radix_node_head *rnh = (struct radix_node_head *)w; 2083 2084 (*rnh->rnh_deladdr) (rn->rn_key, rn->rn_mask, rnh); 2085 kfree(rn, M_NETCRED); 2086 2087 return (0); 2088 } 2089 2090 /* 2091 * callback to free an element of the mask table installed in the 2092 * netexport. These may be created indirectly and are not netcred 2093 * structures. 2094 */ 2095 static int 2096 vfs_free_netcred_mask(struct radix_node *rn, void *w) 2097 { 2098 struct radix_node_head *rnh = (struct radix_node_head *)w; 2099 2100 (*rnh->rnh_deladdr) (rn->rn_key, rn->rn_mask, rnh); 2101 kfree(rn, M_RTABLE); 2102 2103 return (0); 2104 } 2105 2106 static struct radix_node_head * 2107 vfs_create_addrlist_af(int af, struct netexport *nep) 2108 { 2109 struct radix_node_head *rnh = NULL; 2110 #if defined(INET) || defined(INET6) 2111 struct radix_node_head *maskhead = nep->ne_maskhead; 2112 int off; 2113 #endif 2114 2115 NE_ASSERT_LOCKED(nep); 2116 #if defined(INET) || defined(INET6) 2117 KKASSERT(maskhead != NULL); 2118 #endif 2119 switch (af) { 2120 #ifdef INET 2121 case AF_INET: 2122 if ((rnh = nep->ne_inethead) == NULL) { 2123 off = offsetof(struct sockaddr_in, sin_addr) << 3; 2124 if (!rn_inithead((void **)&rnh, maskhead, off)) 2125 return (NULL); 2126 nep->ne_inethead = rnh; 2127 } 2128 break; 2129 #endif 2130 #ifdef INET6 2131 case AF_INET6: 2132 if ((rnh = nep->ne_inet6head) == NULL) { 2133 off = offsetof(struct sockaddr_in6, sin6_addr) << 3; 2134 if (!rn_inithead((void **)&rnh, maskhead, off)) 2135 return (NULL); 2136 nep->ne_inet6head = rnh; 2137 } 2138 break; 2139 #endif 2140 } 2141 return (rnh); 2142 } 2143 2144 /* 2145 * helper function for freeing netcred elements 2146 */ 2147 static void 2148 vfs_free_addrlist_af(struct radix_node_head **prnh) 2149 { 2150 struct radix_node_head *rnh = *prnh; 2151 2152 (*rnh->rnh_walktree) (rnh, vfs_free_netcred, rnh); 2153 kfree(rnh, M_RTABLE); 2154 *prnh = NULL; 2155 } 2156 2157 /* 2158 * helper function for freeing mask elements 2159 */ 2160 static void 2161 vfs_free_addrlist_masks(struct radix_node_head **prnh) 2162 { 2163 struct radix_node_head *rnh = *prnh; 2164 2165 (*rnh->rnh_walktree) (rnh, vfs_free_netcred_mask, rnh); 2166 kfree(rnh, M_RTABLE); 2167 *prnh = NULL; 2168 } 2169 2170 /* 2171 * Free the net address hash lists that are hanging off the mount points. 2172 */ 2173 static void 2174 vfs_free_addrlist(struct netexport *nep) 2175 { 2176 NE_LOCK(nep); 2177 if (nep->ne_inethead != NULL) 2178 vfs_free_addrlist_af(&nep->ne_inethead); 2179 if (nep->ne_inet6head != NULL) 2180 vfs_free_addrlist_af(&nep->ne_inet6head); 2181 if (nep->ne_maskhead) 2182 vfs_free_addrlist_masks(&nep->ne_maskhead); 2183 NE_UNLOCK(nep); 2184 } 2185 2186 int 2187 vfs_export(struct mount *mp, struct netexport *nep, 2188 const struct export_args *argp) 2189 { 2190 int error; 2191 2192 if (argp->ex_flags & MNT_DELEXPORT) { 2193 if (mp->mnt_flag & MNT_EXPUBLIC) { 2194 vfs_setpublicfs(NULL, NULL, NULL); 2195 mp->mnt_flag &= ~MNT_EXPUBLIC; 2196 } 2197 vfs_free_addrlist(nep); 2198 mp->mnt_flag &= ~(MNT_EXPORTED | MNT_DEFEXPORTED); 2199 } 2200 if (argp->ex_flags & MNT_EXPORTED) { 2201 if (argp->ex_flags & MNT_EXPUBLIC) { 2202 if ((error = vfs_setpublicfs(mp, nep, argp)) != 0) 2203 return (error); 2204 mp->mnt_flag |= MNT_EXPUBLIC; 2205 } 2206 if ((error = vfs_hang_addrlist(mp, nep, argp))) 2207 return (error); 2208 mp->mnt_flag |= MNT_EXPORTED; 2209 } 2210 return (0); 2211 } 2212 2213 2214 /* 2215 * Set the publicly exported filesystem (WebNFS). Currently, only 2216 * one public filesystem is possible in the spec (RFC 2054 and 2055) 2217 */ 2218 int 2219 vfs_setpublicfs(struct mount *mp, struct netexport *nep, 2220 const struct export_args *argp) 2221 { 2222 int error; 2223 struct vnode *rvp; 2224 char *cp; 2225 2226 /* 2227 * mp == NULL -> invalidate the current info, the FS is 2228 * no longer exported. May be called from either vfs_export 2229 * or unmount, so check if it hasn't already been done. 2230 */ 2231 if (mp == NULL) { 2232 if (nfs_pub.np_valid) { 2233 nfs_pub.np_valid = 0; 2234 if (nfs_pub.np_index != NULL) { 2235 kfree(nfs_pub.np_index, M_TEMP); 2236 nfs_pub.np_index = NULL; 2237 } 2238 } 2239 return (0); 2240 } 2241 2242 /* 2243 * Only one allowed at a time. 2244 */ 2245 if (nfs_pub.np_valid != 0 && mp != nfs_pub.np_mount) 2246 return (EBUSY); 2247 2248 /* 2249 * Get real filehandle for root of exported FS. 2250 */ 2251 bzero((caddr_t)&nfs_pub.np_handle, sizeof(nfs_pub.np_handle)); 2252 nfs_pub.np_handle.fh_fsid = mp->mnt_stat.f_fsid; 2253 2254 if ((error = VFS_ROOT(mp, &rvp))) 2255 return (error); 2256 2257 if ((error = VFS_VPTOFH(rvp, &nfs_pub.np_handle.fh_fid))) 2258 return (error); 2259 2260 vput(rvp); 2261 2262 /* 2263 * If an indexfile was specified, pull it in. 2264 */ 2265 if (argp->ex_indexfile != NULL) { 2266 int namelen; 2267 2268 error = vn_get_namelen(rvp, &namelen); 2269 if (error) 2270 return (error); 2271 nfs_pub.np_index = kmalloc(namelen, M_TEMP, M_WAITOK); 2272 error = copyinstr(argp->ex_indexfile, nfs_pub.np_index, 2273 namelen, NULL); 2274 if (!error) { 2275 /* 2276 * Check for illegal filenames. 2277 */ 2278 for (cp = nfs_pub.np_index; *cp; cp++) { 2279 if (*cp == '/') { 2280 error = EINVAL; 2281 break; 2282 } 2283 } 2284 } 2285 if (error) { 2286 kfree(nfs_pub.np_index, M_TEMP); 2287 return (error); 2288 } 2289 } 2290 2291 nfs_pub.np_mount = mp; 2292 nfs_pub.np_valid = 1; 2293 return (0); 2294 } 2295 2296 struct netcred * 2297 vfs_export_lookup(struct mount *mp, struct netexport *nep, 2298 struct sockaddr *nam) 2299 { 2300 struct netcred *np; 2301 struct radix_node_head *rnh; 2302 struct sockaddr *saddr; 2303 2304 np = NULL; 2305 if (mp->mnt_flag & MNT_EXPORTED) { 2306 /* 2307 * Lookup in the export list first. 2308 */ 2309 NE_LOCK(nep); 2310 if (nam != NULL) { 2311 saddr = nam; 2312 switch (saddr->sa_family) { 2313 #ifdef INET 2314 case AF_INET: 2315 rnh = nep->ne_inethead; 2316 break; 2317 #endif 2318 #ifdef INET6 2319 case AF_INET6: 2320 rnh = nep->ne_inet6head; 2321 break; 2322 #endif 2323 default: 2324 rnh = NULL; 2325 } 2326 if (rnh != NULL) { 2327 np = (struct netcred *) 2328 (*rnh->rnh_matchaddr)((char *)saddr, 2329 rnh); 2330 if (np && np->netc_rnodes->rn_flags & RNF_ROOT) 2331 np = NULL; 2332 } 2333 } 2334 NE_UNLOCK(nep); 2335 /* 2336 * If no address match, use the default if it exists. 2337 */ 2338 if (np == NULL && mp->mnt_flag & MNT_DEFEXPORTED) 2339 np = &nep->ne_defexported; 2340 } 2341 return (np); 2342 } 2343 2344 /* 2345 * perform msync on all vnodes under a mount point. The mount point must 2346 * be locked. This code is also responsible for lazy-freeing unreferenced 2347 * vnodes whos VM objects no longer contain pages. 2348 * 2349 * NOTE: MNT_WAIT still skips vnodes in the VXLOCK state. 2350 * 2351 * NOTE: XXX VOP_PUTPAGES and friends requires that the vnode be locked, 2352 * but vnode_pager_putpages() doesn't lock the vnode. We have to do it 2353 * way up in this high level function. 2354 */ 2355 static int vfs_msync_scan1(struct mount *mp, struct vnode *vp, void *data); 2356 static int vfs_msync_scan2(struct mount *mp, struct vnode *vp, void *data); 2357 2358 void 2359 vfs_msync(struct mount *mp, int flags) 2360 { 2361 int vmsc_flags; 2362 2363 /* 2364 * tmpfs sets this flag to prevent msync(), sync, and the 2365 * filesystem periodic syncer from trying to flush VM pages 2366 * to swap. Only pure memory pressure flushes tmpfs VM pages 2367 * to swap. 2368 */ 2369 if (mp->mnt_kern_flag & MNTK_NOMSYNC) 2370 return; 2371 2372 /* 2373 * Ok, scan the vnodes for work. If the filesystem is using the 2374 * syncer thread feature we can use vsyncscan() instead of 2375 * vmntvnodescan(), which is much faster. 2376 */ 2377 vmsc_flags = VMSC_GETVP; 2378 if (flags != MNT_WAIT) 2379 vmsc_flags |= VMSC_NOWAIT; 2380 2381 if (mp->mnt_kern_flag & MNTK_THR_SYNC) { 2382 vsyncscan(mp, vmsc_flags, vfs_msync_scan2, 2383 (void *)(intptr_t)flags); 2384 } else { 2385 vmntvnodescan(mp, vmsc_flags, 2386 vfs_msync_scan1, vfs_msync_scan2, 2387 (void *)(intptr_t)flags); 2388 } 2389 } 2390 2391 /* 2392 * scan1 is a fast pre-check. There could be hundreds of thousands of 2393 * vnodes, we cannot afford to do anything heavy weight until we have a 2394 * fairly good indication that there is work to do. 2395 * 2396 * The new namecache holds the vnode for each v_namecache association 2397 * so allow these refs. 2398 */ 2399 static 2400 int 2401 vfs_msync_scan1(struct mount *mp, struct vnode *vp, void *data) 2402 { 2403 int flags = (int)(intptr_t)data; 2404 2405 if ((vp->v_flag & VRECLAIMED) == 0) { 2406 if (vp->v_auxrefs == vp->v_namecache_count && 2407 VREFCNT(vp) <= 0 && vp->v_object) { 2408 return(0); /* call scan2 */ 2409 } 2410 if ((mp->mnt_flag & MNT_RDONLY) == 0 && 2411 (vp->v_flag & VOBJDIRTY) && 2412 (flags == MNT_WAIT || vn_islocked(vp) == 0)) { 2413 return(0); /* call scan2 */ 2414 } 2415 } 2416 2417 /* 2418 * do not call scan2, continue the loop 2419 */ 2420 return(-1); 2421 } 2422 2423 /* 2424 * This callback is handed a locked vnode. 2425 */ 2426 static 2427 int 2428 vfs_msync_scan2(struct mount *mp, struct vnode *vp, void *data) 2429 { 2430 vm_object_t obj; 2431 int flags = (int)(intptr_t)data; 2432 int opcflags; 2433 2434 if (vp->v_flag & VRECLAIMED) 2435 return(0); 2436 2437 if ((mp->mnt_flag & MNT_RDONLY) == 0 && (vp->v_flag & VOBJDIRTY)) { 2438 if ((obj = vp->v_object) != NULL) { 2439 if (flags == MNT_WAIT) { 2440 /* 2441 * VFS_MSYNC is called with MNT_WAIT when 2442 * unmounting. 2443 */ 2444 opcflags = OBJPC_SYNC; 2445 } else if (vp->v_writecount || obj->ref_count) { 2446 /* 2447 * VFS_MSYNC is otherwise called via the 2448 * periodic filesystem sync or the 'sync' 2449 * command. Honor MADV_NOSYNC / MAP_NOSYNC 2450 * if the file is open for writing or memory 2451 * mapped. Pages flagged PG_NOSYNC will not 2452 * be automatically flushed at this time. 2453 * 2454 * The obj->ref_count test is not perfect 2455 * since temporary refs may be present, but 2456 * the periodic filesystem sync will ultimately 2457 * catch it if the file is not open and not 2458 * mapped. 2459 */ 2460 opcflags = OBJPC_NOSYNC; 2461 } else { 2462 /* 2463 * If the file is no longer open for writing 2464 * and also no longer mapped, do not honor 2465 * MAP_NOSYNC. That is, fully synchronize 2466 * the file. 2467 * 2468 * This still occurs on the periodic fs sync, 2469 * so frontend programs which turn the file 2470 * over quickly enough can still avoid the 2471 * sync, but ultimately we do want to flush 2472 * even MADV_NOSYNC pages once it is no longer 2473 * mapped or open for writing. 2474 */ 2475 opcflags = 0; 2476 } 2477 vm_object_page_clean(obj, 0, 0, opcflags); 2478 } 2479 } 2480 return(0); 2481 } 2482 2483 /* 2484 * Wake up anyone interested in vp because it is being revoked. 2485 */ 2486 void 2487 vn_gone(struct vnode *vp) 2488 { 2489 lwkt_gettoken(&vp->v_token); 2490 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, NOTE_REVOKE); 2491 lwkt_reltoken(&vp->v_token); 2492 } 2493 2494 /* 2495 * extract the cdev_t from a VBLK or VCHR. The vnode must have been opened 2496 * (or v_rdev might be NULL). 2497 */ 2498 cdev_t 2499 vn_todev(struct vnode *vp) 2500 { 2501 if (vp->v_type != VBLK && vp->v_type != VCHR) 2502 return (NULL); 2503 KKASSERT(vp->v_rdev != NULL); 2504 return (vp->v_rdev); 2505 } 2506 2507 /* 2508 * Check if vnode represents a disk device. The vnode does not need to be 2509 * opened. 2510 * 2511 * MPALMOSTSAFE 2512 */ 2513 int 2514 vn_isdisk(struct vnode *vp, int *errp) 2515 { 2516 cdev_t dev; 2517 2518 if (vp->v_type != VCHR) { 2519 if (errp != NULL) 2520 *errp = ENOTBLK; 2521 return (0); 2522 } 2523 2524 dev = vp->v_rdev; 2525 2526 if (dev == NULL) { 2527 if (errp != NULL) 2528 *errp = ENXIO; 2529 return (0); 2530 } 2531 if (dev_is_good(dev) == 0) { 2532 if (errp != NULL) 2533 *errp = ENXIO; 2534 return (0); 2535 } 2536 if ((dev_dflags(dev) & D_DISK) == 0) { 2537 if (errp != NULL) 2538 *errp = ENOTBLK; 2539 return (0); 2540 } 2541 if (errp != NULL) 2542 *errp = 0; 2543 return (1); 2544 } 2545 2546 int 2547 vn_get_namelen(struct vnode *vp, int *namelen) 2548 { 2549 int error; 2550 register_t retval[2]; 2551 2552 error = VOP_PATHCONF(vp, _PC_NAME_MAX, retval); 2553 if (error) 2554 return (error); 2555 *namelen = (int)retval[0]; 2556 return (0); 2557 } 2558 2559 int 2560 vop_write_dirent(int *error, struct uio *uio, ino_t d_ino, uint8_t d_type, 2561 uint16_t d_namlen, const char *d_name) 2562 { 2563 struct dirent *dp; 2564 size_t len; 2565 2566 len = _DIRENT_RECLEN(d_namlen); 2567 if (len > uio->uio_resid) 2568 return(1); 2569 2570 dp = kmalloc(len, M_TEMP, M_WAITOK | M_ZERO); 2571 2572 dp->d_ino = d_ino; 2573 dp->d_namlen = d_namlen; 2574 dp->d_type = d_type; 2575 bcopy(d_name, dp->d_name, d_namlen); 2576 2577 *error = uiomove((caddr_t)dp, len, uio); 2578 2579 kfree(dp, M_TEMP); 2580 2581 return(0); 2582 } 2583 2584 void 2585 vn_mark_atime(struct vnode *vp, struct thread *td) 2586 { 2587 struct proc *p = td->td_proc; 2588 struct ucred *cred = p ? p->p_ucred : proc0.p_ucred; 2589 2590 if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) { 2591 VOP_MARKATIME(vp, cred); 2592 } 2593 } 2594 2595 /* 2596 * Calculate the number of entries in an inode-related chained hash table. 2597 * With today's memory sizes, maxvnodes can wind up being a very large 2598 * number. There is no reason to waste memory, so tolerate some stacking. 2599 */ 2600 int 2601 vfs_inodehashsize(void) 2602 { 2603 int hsize; 2604 2605 hsize = 32; 2606 while (hsize < maxvnodes) 2607 hsize <<= 1; 2608 while (hsize > maxvnodes * 2) 2609 hsize >>= 1; /* nominal 2x stacking */ 2610 2611 if (maxvnodes > 1024 * 1024) 2612 hsize >>= 1; /* nominal 8x stacking */ 2613 2614 if (maxvnodes > 128 * 1024) 2615 hsize >>= 1; /* nominal 4x stacking */ 2616 2617 if (hsize < 16) 2618 hsize = 16; 2619 2620 return hsize; 2621 } 2622 2623 union _qcvt { 2624 quad_t qcvt; 2625 int32_t val[2]; 2626 }; 2627 2628 #define SETHIGH(q, h) { \ 2629 union _qcvt tmp; \ 2630 tmp.qcvt = (q); \ 2631 tmp.val[_QUAD_HIGHWORD] = (h); \ 2632 (q) = tmp.qcvt; \ 2633 } 2634 #define SETLOW(q, l) { \ 2635 union _qcvt tmp; \ 2636 tmp.qcvt = (q); \ 2637 tmp.val[_QUAD_LOWWORD] = (l); \ 2638 (q) = tmp.qcvt; \ 2639 } 2640 2641 u_quad_t 2642 init_va_filerev(void) 2643 { 2644 struct timeval tv; 2645 u_quad_t ret = 0; 2646 2647 getmicrouptime(&tv); 2648 SETHIGH(ret, tv.tv_sec); 2649 SETLOW(ret, tv.tv_usec * 4294); 2650 2651 return ret; 2652 } 2653