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/eventhandler.h> 52 #include <sys/fcntl.h> 53 #include <sys/file.h> 54 #include <sys/kernel.h> 55 #include <sys/kthread.h> 56 #include <sys/malloc.h> 57 #include <sys/mbuf.h> 58 #include <sys/mount.h> 59 #include <sys/priv.h> 60 #include <sys/proc.h> 61 #include <sys/reboot.h> 62 #include <sys/socket.h> 63 #include <sys/stat.h> 64 #include <sys/sysctl.h> 65 #include <sys/syslog.h> 66 #include <sys/unistd.h> 67 #include <sys/vmmeter.h> 68 #include <sys/vnode.h> 69 70 #include <machine/limits.h> 71 72 #include <vm/vm.h> 73 #include <vm/vm_object.h> 74 #include <vm/vm_extern.h> 75 #include <vm/vm_kern.h> 76 #include <vm/pmap.h> 77 #include <vm/vm_map.h> 78 #include <vm/vm_page.h> 79 #include <vm/vm_pager.h> 80 #include <vm/vnode_pager.h> 81 #include <vm/vm_zone.h> 82 83 #include <sys/buf2.h> 84 #include <sys/mplock2.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 }; 210 211 __read_mostly static int timestamp_precision = TSP_USEC; 212 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 213 ×tamp_precision, 0, "Precision of file timestamps"); 214 215 /* 216 * Get a current timestamp. 217 * 218 * MPSAFE 219 */ 220 void 221 vfs_timestamp(struct timespec *tsp) 222 { 223 struct timeval tv; 224 225 switch (timestamp_precision) { 226 case TSP_SEC: 227 tsp->tv_sec = time_second; 228 tsp->tv_nsec = 0; 229 break; 230 case TSP_HZ: 231 getnanotime(tsp); 232 break; 233 case TSP_USEC: 234 microtime(&tv); 235 TIMEVAL_TO_TIMESPEC(&tv, tsp); 236 break; 237 case TSP_NSEC: 238 default: 239 nanotime(tsp); 240 break; 241 } 242 } 243 244 /* 245 * Set vnode attributes to VNOVAL 246 */ 247 void 248 vattr_null(struct vattr *vap) 249 { 250 vap->va_type = VNON; 251 vap->va_size = VNOVAL; 252 vap->va_bytes = VNOVAL; 253 vap->va_mode = VNOVAL; 254 vap->va_nlink = VNOVAL; 255 vap->va_uid = VNOVAL; 256 vap->va_gid = VNOVAL; 257 vap->va_fsid = VNOVAL; 258 vap->va_fileid = VNOVAL; 259 vap->va_blocksize = VNOVAL; 260 vap->va_rmajor = VNOVAL; 261 vap->va_rminor = VNOVAL; 262 vap->va_atime.tv_sec = VNOVAL; 263 vap->va_atime.tv_nsec = VNOVAL; 264 vap->va_mtime.tv_sec = VNOVAL; 265 vap->va_mtime.tv_nsec = VNOVAL; 266 vap->va_ctime.tv_sec = VNOVAL; 267 vap->va_ctime.tv_nsec = VNOVAL; 268 vap->va_flags = VNOVAL; 269 vap->va_gen = VNOVAL; 270 vap->va_vaflags = 0; 271 /* va_*_uuid fields are only valid if related flags are set */ 272 } 273 274 /* 275 * Flush out and invalidate all buffers associated with a vnode. 276 * 277 * vp must be locked. 278 */ 279 static int vinvalbuf_bp(struct buf *bp, void *data); 280 281 struct vinvalbuf_bp_info { 282 struct vnode *vp; 283 int slptimeo; 284 int lkflags; 285 int flags; 286 int clean; 287 }; 288 289 int 290 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 291 { 292 struct vinvalbuf_bp_info info; 293 vm_object_t object; 294 int error; 295 296 lwkt_gettoken(&vp->v_token); 297 298 /* 299 * If we are being asked to save, call fsync to ensure that the inode 300 * is updated. 301 */ 302 if (flags & V_SAVE) { 303 error = bio_track_wait(&vp->v_track_write, slpflag, slptimeo); 304 if (error) 305 goto done; 306 if (!RB_EMPTY(&vp->v_rbdirty_tree)) { 307 if ((error = VOP_FSYNC(vp, MNT_WAIT, 0)) != 0) 308 goto done; 309 #if 0 310 /* 311 * Dirty bufs may be left or generated via races 312 * in circumstances where vinvalbuf() is called on 313 * a vnode not undergoing reclamation. Only 314 * panic if we are trying to reclaim the vnode. 315 */ 316 if ((vp->v_flag & VRECLAIMED) && 317 (bio_track_active(&vp->v_track_write) || 318 !RB_EMPTY(&vp->v_rbdirty_tree))) { 319 panic("vinvalbuf: dirty bufs"); 320 } 321 #endif 322 } 323 } 324 info.slptimeo = slptimeo; 325 info.lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 326 if (slpflag & PCATCH) 327 info.lkflags |= LK_PCATCH; 328 info.flags = flags; 329 info.vp = vp; 330 331 /* 332 * Flush the buffer cache until nothing is left, wait for all I/O 333 * to complete. At least one pass is required. We might block 334 * in the pip code so we have to re-check. Order is important. 335 */ 336 do { 337 /* 338 * Flush buffer cache 339 */ 340 if (!RB_EMPTY(&vp->v_rbclean_tree)) { 341 info.clean = 1; 342 error = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 343 NULL, vinvalbuf_bp, &info); 344 } 345 if (!RB_EMPTY(&vp->v_rbdirty_tree)) { 346 info.clean = 0; 347 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 348 NULL, vinvalbuf_bp, &info); 349 } 350 351 /* 352 * Wait for I/O completion. 353 */ 354 bio_track_wait(&vp->v_track_write, 0, 0); 355 if ((object = vp->v_object) != NULL) 356 refcount_wait(&object->paging_in_progress, "vnvlbx"); 357 } while (bio_track_active(&vp->v_track_write) || 358 !RB_EMPTY(&vp->v_rbclean_tree) || 359 !RB_EMPTY(&vp->v_rbdirty_tree)); 360 361 /* 362 * Destroy the copy in the VM cache, too. 363 */ 364 if ((object = vp->v_object) != NULL) { 365 vm_object_page_remove(object, 0, 0, 366 (flags & V_SAVE) ? TRUE : FALSE); 367 } 368 369 if (!RB_EMPTY(&vp->v_rbdirty_tree) || !RB_EMPTY(&vp->v_rbclean_tree)) 370 panic("vinvalbuf: flush failed"); 371 if (!RB_EMPTY(&vp->v_rbhash_tree)) 372 panic("vinvalbuf: flush failed, buffers still present"); 373 error = 0; 374 done: 375 lwkt_reltoken(&vp->v_token); 376 return (error); 377 } 378 379 static int 380 vinvalbuf_bp(struct buf *bp, void *data) 381 { 382 struct vinvalbuf_bp_info *info = data; 383 int error; 384 385 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 386 atomic_add_int(&bp->b_refs, 1); 387 error = BUF_TIMELOCK(bp, info->lkflags, 388 "vinvalbuf", info->slptimeo); 389 atomic_subtract_int(&bp->b_refs, 1); 390 if (error == 0) { 391 BUF_UNLOCK(bp); 392 error = ENOLCK; 393 } 394 if (error == ENOLCK) 395 return(0); 396 return (-error); 397 } 398 KKASSERT(bp->b_vp == info->vp); 399 400 /* 401 * Must check clean/dirty status after successfully locking as 402 * it may race. 403 */ 404 if ((info->clean && (bp->b_flags & B_DELWRI)) || 405 (info->clean == 0 && (bp->b_flags & B_DELWRI) == 0)) { 406 BUF_UNLOCK(bp); 407 return(0); 408 } 409 410 /* 411 * NOTE: NO B_LOCKED CHECK. Also no buf_checkwrite() 412 * check. This code will write out the buffer, period. 413 */ 414 bremfree(bp); 415 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 416 (info->flags & V_SAVE)) { 417 cluster_awrite(bp); 418 } else if (info->flags & V_SAVE) { 419 /* 420 * Cannot set B_NOCACHE on a clean buffer as this will 421 * destroy the VM backing store which might actually 422 * be dirty (and unsynchronized). 423 */ 424 bp->b_flags |= (B_INVAL | B_RELBUF); 425 brelse(bp); 426 } else { 427 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 428 brelse(bp); 429 } 430 return(0); 431 } 432 433 /* 434 * Truncate a file's buffer and pages to a specified length. This 435 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 436 * sync activity. 437 * 438 * The vnode must be locked. 439 */ 440 static int vtruncbuf_bp_trunc_cmp(struct buf *bp, void *data); 441 static int vtruncbuf_bp_trunc(struct buf *bp, void *data); 442 static int vtruncbuf_bp_metasync_cmp(struct buf *bp, void *data); 443 static int vtruncbuf_bp_metasync(struct buf *bp, void *data); 444 445 struct vtruncbuf_info { 446 struct vnode *vp; 447 off_t truncloffset; 448 int clean; 449 }; 450 451 int 452 vtruncbuf(struct vnode *vp, off_t length, int blksize) 453 { 454 struct vtruncbuf_info info; 455 const char *filename; 456 int count; 457 458 /* 459 * Round up to the *next* block, then destroy the buffers in question. 460 * Since we are only removing some of the buffers we must rely on the 461 * scan count to determine whether a loop is necessary. 462 */ 463 if ((count = (int)(length % blksize)) != 0) 464 info.truncloffset = length + (blksize - count); 465 else 466 info.truncloffset = length; 467 info.vp = vp; 468 469 lwkt_gettoken(&vp->v_token); 470 do { 471 info.clean = 1; 472 count = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 473 vtruncbuf_bp_trunc_cmp, 474 vtruncbuf_bp_trunc, &info); 475 info.clean = 0; 476 count += RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 477 vtruncbuf_bp_trunc_cmp, 478 vtruncbuf_bp_trunc, &info); 479 } while(count); 480 481 /* 482 * For safety, fsync any remaining metadata if the file is not being 483 * truncated to 0. Since the metadata does not represent the entire 484 * dirty list we have to rely on the hit count to ensure that we get 485 * all of it. 486 */ 487 if (length > 0) { 488 do { 489 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 490 vtruncbuf_bp_metasync_cmp, 491 vtruncbuf_bp_metasync, &info); 492 } while (count); 493 } 494 495 /* 496 * Clean out any left over VM backing store. 497 * 498 * It is possible to have in-progress I/O from buffers that were 499 * not part of the truncation. This should not happen if we 500 * are truncating to 0-length. 501 */ 502 vnode_pager_setsize(vp, length); 503 bio_track_wait(&vp->v_track_write, 0, 0); 504 505 /* 506 * Debugging only 507 */ 508 spin_lock(&vp->v_spin); 509 filename = TAILQ_FIRST(&vp->v_namecache) ? 510 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"; 511 spin_unlock(&vp->v_spin); 512 513 /* 514 * Make sure no buffers were instantiated while we were trying 515 * to clean out the remaining VM pages. This could occur due 516 * to busy dirty VM pages being flushed out to disk. 517 */ 518 do { 519 info.clean = 1; 520 count = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 521 vtruncbuf_bp_trunc_cmp, 522 vtruncbuf_bp_trunc, &info); 523 info.clean = 0; 524 count += RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 525 vtruncbuf_bp_trunc_cmp, 526 vtruncbuf_bp_trunc, &info); 527 if (count) { 528 kprintf("Warning: vtruncbuf(): Had to re-clean %d " 529 "left over buffers in %s\n", count, filename); 530 } 531 } while(count); 532 533 lwkt_reltoken(&vp->v_token); 534 535 return (0); 536 } 537 538 /* 539 * The callback buffer is beyond the new file EOF and must be destroyed. 540 * Note that the compare function must conform to the RB_SCAN's requirements. 541 */ 542 static 543 int 544 vtruncbuf_bp_trunc_cmp(struct buf *bp, void *data) 545 { 546 struct vtruncbuf_info *info = data; 547 548 if (bp->b_loffset >= info->truncloffset) 549 return(0); 550 return(-1); 551 } 552 553 static 554 int 555 vtruncbuf_bp_trunc(struct buf *bp, void *data) 556 { 557 struct vtruncbuf_info *info = data; 558 559 /* 560 * Do not try to use a buffer we cannot immediately lock, but sleep 561 * anyway to prevent a livelock. The code will loop until all buffers 562 * can be acted upon. 563 * 564 * We must always revalidate the buffer after locking it to deal 565 * with MP races. 566 */ 567 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 568 atomic_add_int(&bp->b_refs, 1); 569 if (BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL) == 0) 570 BUF_UNLOCK(bp); 571 atomic_subtract_int(&bp->b_refs, 1); 572 } else if ((info->clean && (bp->b_flags & B_DELWRI)) || 573 (info->clean == 0 && (bp->b_flags & B_DELWRI) == 0) || 574 bp->b_vp != info->vp || 575 vtruncbuf_bp_trunc_cmp(bp, data)) { 576 BUF_UNLOCK(bp); 577 } else { 578 bremfree(bp); 579 bp->b_flags |= (B_INVAL | B_RELBUF | B_NOCACHE); 580 brelse(bp); 581 } 582 return(1); 583 } 584 585 /* 586 * Fsync all meta-data after truncating a file to be non-zero. Only metadata 587 * blocks (with a negative loffset) are scanned. 588 * Note that the compare function must conform to the RB_SCAN's requirements. 589 */ 590 static int 591 vtruncbuf_bp_metasync_cmp(struct buf *bp, void *data __unused) 592 { 593 if (bp->b_loffset < 0) 594 return(0); 595 return(1); 596 } 597 598 static int 599 vtruncbuf_bp_metasync(struct buf *bp, void *data) 600 { 601 struct vtruncbuf_info *info = data; 602 603 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 604 atomic_add_int(&bp->b_refs, 1); 605 if (BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL) == 0) 606 BUF_UNLOCK(bp); 607 atomic_subtract_int(&bp->b_refs, 1); 608 } else if ((bp->b_flags & B_DELWRI) == 0 || 609 bp->b_vp != info->vp || 610 vtruncbuf_bp_metasync_cmp(bp, data)) { 611 BUF_UNLOCK(bp); 612 } else { 613 bremfree(bp); 614 if (bp->b_vp == info->vp) 615 bawrite(bp); 616 else 617 bwrite(bp); 618 } 619 return(1); 620 } 621 622 /* 623 * vfsync - implements a multipass fsync on a file which understands 624 * dependancies and meta-data. The passed vnode must be locked. The 625 * waitfor argument may be MNT_WAIT or MNT_NOWAIT, or MNT_LAZY. 626 * 627 * When fsyncing data asynchronously just do one consolidated pass starting 628 * with the most negative block number. This may not get all the data due 629 * to dependancies. 630 * 631 * When fsyncing data synchronously do a data pass, then a metadata pass, 632 * then do additional data+metadata passes to try to get all the data out. 633 * 634 * Caller must ref the vnode but does not have to lock it. 635 */ 636 static int vfsync_wait_output(struct vnode *vp, 637 int (*waitoutput)(struct vnode *, struct thread *)); 638 static int vfsync_dummy_cmp(struct buf *bp __unused, void *data __unused); 639 static int vfsync_data_only_cmp(struct buf *bp, void *data); 640 static int vfsync_meta_only_cmp(struct buf *bp, void *data); 641 static int vfsync_lazy_range_cmp(struct buf *bp, void *data); 642 static int vfsync_bp(struct buf *bp, void *data); 643 644 struct vfsync_info { 645 struct vnode *vp; 646 int fastpass; 647 int synchronous; 648 int syncdeps; 649 int lazycount; 650 int lazylimit; 651 int skippedbufs; 652 int (*checkdef)(struct buf *); 653 int (*cmpfunc)(struct buf *, void *); 654 }; 655 656 int 657 vfsync(struct vnode *vp, int waitfor, int passes, 658 int (*checkdef)(struct buf *), 659 int (*waitoutput)(struct vnode *, struct thread *)) 660 { 661 struct vfsync_info info; 662 int error; 663 664 bzero(&info, sizeof(info)); 665 info.vp = vp; 666 if ((info.checkdef = checkdef) == NULL) 667 info.syncdeps = 1; 668 669 lwkt_gettoken(&vp->v_token); 670 671 switch(waitfor) { 672 case MNT_LAZY | MNT_NOWAIT: 673 case MNT_LAZY: 674 /* 675 * Lazy (filesystem syncer typ) Asynchronous plus limit the 676 * number of data (not meta) pages we try to flush to 1MB. 677 * A non-zero return means that lazy limit was reached. 678 */ 679 info.lazylimit = 1024 * 1024; 680 info.syncdeps = 1; 681 info.cmpfunc = vfsync_lazy_range_cmp; 682 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 683 vfsync_lazy_range_cmp, vfsync_bp, &info); 684 info.cmpfunc = vfsync_meta_only_cmp; 685 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 686 vfsync_meta_only_cmp, vfsync_bp, &info); 687 if (error == 0) 688 vp->v_lazyw = 0; 689 else if (!RB_EMPTY(&vp->v_rbdirty_tree)) 690 vn_syncer_add(vp, 1); 691 error = 0; 692 break; 693 case MNT_NOWAIT: 694 /* 695 * Asynchronous. Do a data-only pass and a meta-only pass. 696 */ 697 info.syncdeps = 1; 698 info.cmpfunc = vfsync_data_only_cmp; 699 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, vfsync_data_only_cmp, 700 vfsync_bp, &info); 701 info.cmpfunc = vfsync_meta_only_cmp; 702 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, vfsync_meta_only_cmp, 703 vfsync_bp, &info); 704 error = 0; 705 break; 706 default: 707 /* 708 * Synchronous. Do a data-only pass, then a meta-data+data 709 * pass, then additional integrated passes to try to get 710 * all the dependancies flushed. 711 */ 712 info.cmpfunc = vfsync_data_only_cmp; 713 info.fastpass = 1; 714 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, vfsync_data_only_cmp, 715 vfsync_bp, &info); 716 info.fastpass = 0; 717 error = vfsync_wait_output(vp, waitoutput); 718 if (error == 0) { 719 info.skippedbufs = 0; 720 info.cmpfunc = vfsync_dummy_cmp; 721 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 722 vfsync_bp, &info); 723 error = vfsync_wait_output(vp, waitoutput); 724 if (info.skippedbufs) { 725 kprintf("Warning: vfsync skipped %d dirty " 726 "buf%s in pass2!\n", 727 info.skippedbufs, 728 ((info.skippedbufs > 1) ? "s" : "")); 729 } 730 } 731 while (error == 0 && passes > 0 && 732 !RB_EMPTY(&vp->v_rbdirty_tree) 733 ) { 734 info.skippedbufs = 0; 735 if (--passes == 0) { 736 info.synchronous = 1; 737 info.syncdeps = 1; 738 } 739 info.cmpfunc = vfsync_dummy_cmp; 740 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 741 vfsync_bp, &info); 742 if (error < 0) 743 error = -error; 744 info.syncdeps = 1; 745 if (error == 0) 746 error = vfsync_wait_output(vp, waitoutput); 747 if (info.skippedbufs && passes == 0) { 748 kprintf("Warning: vfsync skipped %d dirty " 749 "buf%s in final pass!\n", 750 info.skippedbufs, 751 ((info.skippedbufs > 1) ? "s" : "")); 752 } 753 } 754 #if 0 755 /* 756 * This case can occur normally because vnode lock might 757 * not be held. 758 */ 759 if (!RB_EMPTY(&vp->v_rbdirty_tree)) 760 kprintf("dirty bufs left after final pass\n"); 761 #endif 762 break; 763 } 764 lwkt_reltoken(&vp->v_token); 765 766 return(error); 767 } 768 769 static int 770 vfsync_wait_output(struct vnode *vp, 771 int (*waitoutput)(struct vnode *, struct thread *)) 772 { 773 int error; 774 775 error = bio_track_wait(&vp->v_track_write, 0, 0); 776 if (waitoutput) 777 error = waitoutput(vp, curthread); 778 return(error); 779 } 780 781 static int 782 vfsync_dummy_cmp(struct buf *bp __unused, void *data __unused) 783 { 784 return(0); 785 } 786 787 static int 788 vfsync_data_only_cmp(struct buf *bp, void *data) 789 { 790 if (bp->b_loffset < 0) 791 return(-1); 792 return(0); 793 } 794 795 static int 796 vfsync_meta_only_cmp(struct buf *bp, void *data) 797 { 798 if (bp->b_loffset < 0) 799 return(0); 800 return(1); 801 } 802 803 static int 804 vfsync_lazy_range_cmp(struct buf *bp, void *data) 805 { 806 struct vfsync_info *info = data; 807 808 if (bp->b_loffset < info->vp->v_lazyw) 809 return(-1); 810 return(0); 811 } 812 813 static int 814 vfsync_bp(struct buf *bp, void *data) 815 { 816 struct vfsync_info *info = data; 817 struct vnode *vp = info->vp; 818 int error; 819 820 if (info->fastpass) { 821 /* 822 * Ignore buffers that we cannot immediately lock. 823 */ 824 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 825 /* 826 * Removed BUF_TIMELOCK(..., 1), even a 1-tick 827 * delay can mess up performance 828 * 829 * Another reason is that during a dirty-buffer 830 * scan a clustered write can start I/O on buffers 831 * ahead of the scan, causing the scan to not 832 * get a lock here. Usually this means the write 833 * is already in progress so, in fact, we *want* 834 * to skip the buffer. 835 */ 836 ++info->skippedbufs; 837 return(0); 838 } 839 } else if (info->synchronous == 0) { 840 /* 841 * Normal pass, give the buffer a little time to become 842 * available to us. 843 */ 844 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE, "bflst2", hz / 10)) { 845 ++info->skippedbufs; 846 return(0); 847 } 848 } else { 849 /* 850 * Synchronous pass, give the buffer a lot of time before 851 * giving up. 852 */ 853 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE, "bflst3", hz * 10)) { 854 ++info->skippedbufs; 855 return(0); 856 } 857 } 858 859 /* 860 * We must revalidate the buffer after locking. 861 */ 862 if ((bp->b_flags & B_DELWRI) == 0 || 863 bp->b_vp != info->vp || 864 info->cmpfunc(bp, data)) { 865 BUF_UNLOCK(bp); 866 return(0); 867 } 868 869 /* 870 * If syncdeps is not set we do not try to write buffers which have 871 * dependancies. 872 */ 873 if (!info->synchronous && info->syncdeps == 0 && info->checkdef(bp)) { 874 BUF_UNLOCK(bp); 875 return(0); 876 } 877 878 /* 879 * B_NEEDCOMMIT (primarily used by NFS) is a state where the buffer 880 * has been written but an additional handshake with the device 881 * is required before we can dispose of the buffer. We have no idea 882 * how to do this so we have to skip these buffers. 883 */ 884 if (bp->b_flags & B_NEEDCOMMIT) { 885 BUF_UNLOCK(bp); 886 return(0); 887 } 888 889 /* 890 * Ask bioops if it is ok to sync. If not the VFS may have 891 * set B_LOCKED so we have to cycle the buffer. 892 */ 893 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 894 bremfree(bp); 895 brelse(bp); 896 return(0); 897 } 898 899 if (info->synchronous) { 900 /* 901 * Synchronous flush. An error may be returned and will 902 * stop the scan. 903 */ 904 bremfree(bp); 905 error = bwrite(bp); 906 } else { 907 /* 908 * Asynchronous flush. We use the error return to support 909 * MNT_LAZY flushes. 910 * 911 * In low-memory situations we revert to synchronous 912 * operation. This should theoretically prevent the I/O 913 * path from exhausting memory in a non-recoverable way. 914 */ 915 vp->v_lazyw = bp->b_loffset; 916 bremfree(bp); 917 if (vm_page_count_min(0)) { 918 /* low memory */ 919 info->lazycount += bp->b_bufsize; 920 bwrite(bp); 921 } else { 922 /* normal */ 923 info->lazycount += cluster_awrite(bp); 924 waitrunningbufspace(); 925 /*vm_wait_nominal();*/ 926 } 927 if (info->lazylimit && info->lazycount >= info->lazylimit) 928 error = 1; 929 else 930 error = 0; 931 } 932 return(-error); 933 } 934 935 /* 936 * Associate a buffer with a vnode. 937 * 938 * MPSAFE 939 */ 940 int 941 bgetvp(struct vnode *vp, struct buf *bp, int testsize) 942 { 943 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 944 KKASSERT((bp->b_flags & (B_HASHED|B_DELWRI|B_VNCLEAN|B_VNDIRTY)) == 0); 945 946 /* 947 * Insert onto list for new vnode. 948 */ 949 lwkt_gettoken(&vp->v_token); 950 951 if (buf_rb_hash_RB_INSERT(&vp->v_rbhash_tree, bp)) { 952 lwkt_reltoken(&vp->v_token); 953 return (EEXIST); 954 } 955 956 /* 957 * Diagnostics (mainly for HAMMER debugging). Check for 958 * overlapping buffers. 959 */ 960 if (check_buf_overlap) { 961 struct buf *bx; 962 bx = buf_rb_hash_RB_PREV(bp); 963 if (bx) { 964 if (bx->b_loffset + bx->b_bufsize > bp->b_loffset) { 965 kprintf("bgetvp: overlapl %016jx/%d %016jx " 966 "bx %p bp %p\n", 967 (intmax_t)bx->b_loffset, 968 bx->b_bufsize, 969 (intmax_t)bp->b_loffset, 970 bx, bp); 971 if (check_buf_overlap > 1) 972 panic("bgetvp - overlapping buffer"); 973 } 974 } 975 bx = buf_rb_hash_RB_NEXT(bp); 976 if (bx) { 977 if (bp->b_loffset + testsize > bx->b_loffset) { 978 kprintf("bgetvp: overlapr %016jx/%d %016jx " 979 "bp %p bx %p\n", 980 (intmax_t)bp->b_loffset, 981 testsize, 982 (intmax_t)bx->b_loffset, 983 bp, bx); 984 if (check_buf_overlap > 1) 985 panic("bgetvp - overlapping buffer"); 986 } 987 } 988 } 989 bp->b_vp = vp; 990 bp->b_flags |= B_HASHED; 991 bp->b_flags |= B_VNCLEAN; 992 if (buf_rb_tree_RB_INSERT(&vp->v_rbclean_tree, bp)) 993 panic("reassignbuf: dup lblk/clean vp %p bp %p", vp, bp); 994 /*vhold(vp);*/ 995 lwkt_reltoken(&vp->v_token); 996 return(0); 997 } 998 999 /* 1000 * Disassociate a buffer from a vnode. 1001 * 1002 * MPSAFE 1003 */ 1004 void 1005 brelvp(struct buf *bp) 1006 { 1007 struct vnode *vp; 1008 1009 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1010 1011 /* 1012 * Delete from old vnode list, if on one. 1013 */ 1014 vp = bp->b_vp; 1015 lwkt_gettoken(&vp->v_token); 1016 if (bp->b_flags & (B_VNDIRTY | B_VNCLEAN)) { 1017 if (bp->b_flags & B_VNDIRTY) 1018 buf_rb_tree_RB_REMOVE(&vp->v_rbdirty_tree, bp); 1019 else 1020 buf_rb_tree_RB_REMOVE(&vp->v_rbclean_tree, bp); 1021 bp->b_flags &= ~(B_VNDIRTY | B_VNCLEAN); 1022 } 1023 if (bp->b_flags & B_HASHED) { 1024 buf_rb_hash_RB_REMOVE(&vp->v_rbhash_tree, bp); 1025 bp->b_flags &= ~B_HASHED; 1026 } 1027 1028 /* 1029 * Only remove from synclist when no dirty buffers are left AND 1030 * the VFS has not flagged the vnode's inode as being dirty. 1031 */ 1032 if ((vp->v_flag & (VONWORKLST | VISDIRTY | VOBJDIRTY)) == VONWORKLST && 1033 RB_EMPTY(&vp->v_rbdirty_tree)) { 1034 vn_syncer_remove(vp, 0); 1035 } 1036 bp->b_vp = NULL; 1037 1038 lwkt_reltoken(&vp->v_token); 1039 1040 /*vdrop(vp);*/ 1041 } 1042 1043 /* 1044 * Reassign the buffer to the proper clean/dirty list based on B_DELWRI. 1045 * This routine is called when the state of the B_DELWRI bit is changed. 1046 * 1047 * Must be called with vp->v_token held. 1048 * MPSAFE 1049 */ 1050 void 1051 reassignbuf(struct buf *bp) 1052 { 1053 struct vnode *vp = bp->b_vp; 1054 int delay; 1055 1056 ASSERT_LWKT_TOKEN_HELD(&vp->v_token); 1057 ++reassignbufcalls; 1058 1059 /* 1060 * B_PAGING flagged buffers cannot be reassigned because their vp 1061 * is not fully linked in. 1062 */ 1063 if (bp->b_flags & B_PAGING) 1064 panic("cannot reassign paging buffer"); 1065 1066 if (bp->b_flags & B_DELWRI) { 1067 /* 1068 * Move to the dirty list, add the vnode to the worklist 1069 */ 1070 if (bp->b_flags & B_VNCLEAN) { 1071 buf_rb_tree_RB_REMOVE(&vp->v_rbclean_tree, bp); 1072 bp->b_flags &= ~B_VNCLEAN; 1073 } 1074 if ((bp->b_flags & B_VNDIRTY) == 0) { 1075 if (buf_rb_tree_RB_INSERT(&vp->v_rbdirty_tree, bp)) { 1076 panic("reassignbuf: dup lblk vp %p bp %p", 1077 vp, bp); 1078 } 1079 bp->b_flags |= B_VNDIRTY; 1080 } 1081 if ((vp->v_flag & VONWORKLST) == 0) { 1082 switch (vp->v_type) { 1083 case VDIR: 1084 delay = dirdelay; 1085 break; 1086 case VCHR: 1087 case VBLK: 1088 if (vp->v_rdev && 1089 vp->v_rdev->si_mountpoint != NULL) { 1090 delay = metadelay; 1091 break; 1092 } 1093 /* fall through */ 1094 default: 1095 delay = filedelay; 1096 } 1097 vn_syncer_add(vp, delay); 1098 } 1099 } else { 1100 /* 1101 * Move to the clean list, remove the vnode from the worklist 1102 * if no dirty blocks remain. 1103 */ 1104 if (bp->b_flags & B_VNDIRTY) { 1105 buf_rb_tree_RB_REMOVE(&vp->v_rbdirty_tree, bp); 1106 bp->b_flags &= ~B_VNDIRTY; 1107 } 1108 if ((bp->b_flags & B_VNCLEAN) == 0) { 1109 if (buf_rb_tree_RB_INSERT(&vp->v_rbclean_tree, bp)) { 1110 panic("reassignbuf: dup lblk vp %p bp %p", 1111 vp, bp); 1112 } 1113 bp->b_flags |= B_VNCLEAN; 1114 } 1115 1116 /* 1117 * Only remove from synclist when no dirty buffers are left 1118 * AND the VFS has not flagged the vnode's inode as being 1119 * dirty. 1120 */ 1121 if ((vp->v_flag & (VONWORKLST | VISDIRTY | VOBJDIRTY)) == 1122 VONWORKLST && 1123 RB_EMPTY(&vp->v_rbdirty_tree)) { 1124 vn_syncer_remove(vp, 0); 1125 } 1126 } 1127 } 1128 1129 /* 1130 * Create a vnode for a block device. Used for mounting the root file 1131 * system. 1132 * 1133 * A vref()'d vnode is returned. 1134 */ 1135 extern struct vop_ops *devfs_vnode_dev_vops_p; 1136 int 1137 bdevvp(cdev_t dev, struct vnode **vpp) 1138 { 1139 struct vnode *vp; 1140 struct vnode *nvp; 1141 int error; 1142 1143 if (dev == NULL) { 1144 *vpp = NULLVP; 1145 return (ENXIO); 1146 } 1147 error = getspecialvnode(VT_NON, NULL, &devfs_vnode_dev_vops_p, 1148 &nvp, 0, 0); 1149 if (error) { 1150 *vpp = NULLVP; 1151 return (error); 1152 } 1153 vp = nvp; 1154 vp->v_type = VCHR; 1155 #if 0 1156 vp->v_rdev = dev; 1157 #endif 1158 v_associate_rdev(vp, dev); 1159 vp->v_umajor = dev->si_umajor; 1160 vp->v_uminor = dev->si_uminor; 1161 vx_unlock(vp); 1162 *vpp = vp; 1163 return (0); 1164 } 1165 1166 int 1167 v_associate_rdev(struct vnode *vp, cdev_t dev) 1168 { 1169 if (dev == NULL) 1170 return(ENXIO); 1171 if (dev_is_good(dev) == 0) 1172 return(ENXIO); 1173 KKASSERT(vp->v_rdev == NULL); 1174 vp->v_rdev = reference_dev(dev); 1175 lwkt_gettoken(&spechash_token); 1176 SLIST_INSERT_HEAD(&dev->si_hlist, vp, v_cdevnext); 1177 lwkt_reltoken(&spechash_token); 1178 return(0); 1179 } 1180 1181 void 1182 v_release_rdev(struct vnode *vp) 1183 { 1184 cdev_t dev; 1185 1186 if ((dev = vp->v_rdev) != NULL) { 1187 lwkt_gettoken(&spechash_token); 1188 SLIST_REMOVE(&dev->si_hlist, vp, vnode, v_cdevnext); 1189 vp->v_rdev = NULL; 1190 release_dev(dev); 1191 lwkt_reltoken(&spechash_token); 1192 } 1193 } 1194 1195 /* 1196 * Add a vnode to the alias list hung off the cdev_t. We only associate 1197 * the device number with the vnode. The actual device is not associated 1198 * until the vnode is opened (usually in spec_open()), and will be 1199 * disassociated on last close. 1200 */ 1201 void 1202 addaliasu(struct vnode *nvp, int x, int y) 1203 { 1204 if (nvp->v_type != VBLK && nvp->v_type != VCHR) 1205 panic("addaliasu on non-special vnode"); 1206 nvp->v_umajor = x; 1207 nvp->v_uminor = y; 1208 } 1209 1210 /* 1211 * Simple call that a filesystem can make to try to get rid of a 1212 * vnode. It will fail if anyone is referencing the vnode (including 1213 * the caller). 1214 * 1215 * The filesystem can check whether its in-memory inode structure still 1216 * references the vp on return. 1217 * 1218 * May only be called if the vnode is in a known state (i.e. being prevented 1219 * from being deallocated by some other condition such as a vfs inode hold). 1220 * 1221 * This call might not succeed. 1222 */ 1223 void 1224 vclean_unlocked(struct vnode *vp) 1225 { 1226 vx_get(vp); 1227 if (VREFCNT(vp) <= 1) 1228 vgone_vxlocked(vp); 1229 vx_put(vp); 1230 } 1231 1232 /* 1233 * Disassociate a vnode from its underlying filesystem. 1234 * 1235 * The vnode must be VX locked and referenced. In all normal situations 1236 * there are no active references. If vclean_vxlocked() is called while 1237 * there are active references, the vnode is being ripped out and we have 1238 * to call VOP_CLOSE() as appropriate before we can reclaim it. 1239 */ 1240 static void 1241 vclean_vxlocked(struct vnode *vp, int flags) 1242 { 1243 int active; 1244 int n; 1245 vm_object_t object; 1246 struct namecache *ncp; 1247 1248 /* 1249 * If the vnode has already been reclaimed we have nothing to do. 1250 */ 1251 if (vp->v_flag & VRECLAIMED) 1252 return; 1253 1254 /* 1255 * Set flag to interlock operation, flag finalization to ensure 1256 * that the vnode winds up on the inactive list, and set v_act to 0. 1257 */ 1258 vsetflags(vp, VRECLAIMED); 1259 atomic_set_int(&vp->v_refcnt, VREF_FINALIZE); 1260 vp->v_act = 0; 1261 1262 if (verbose_reclaims) { 1263 if ((ncp = TAILQ_FIRST(&vp->v_namecache)) != NULL) 1264 kprintf("Debug: reclaim %p %s\n", vp, ncp->nc_name); 1265 } 1266 1267 /* 1268 * Scrap the vfs cache 1269 */ 1270 while (cache_inval_vp(vp, 0) != 0) { 1271 kprintf("Warning: vnode %p clean/cache_resolution " 1272 "race detected\n", vp); 1273 tsleep(vp, 0, "vclninv", 2); 1274 } 1275 1276 /* 1277 * Check to see if the vnode is in use. If so we have to reference it 1278 * before we clean it out so that its count cannot fall to zero and 1279 * generate a race against ourselves to recycle it. 1280 */ 1281 active = (VREFCNT(vp) > 0); 1282 1283 /* 1284 * Clean out any buffers associated with the vnode and destroy its 1285 * object, if it has one. 1286 */ 1287 vinvalbuf(vp, V_SAVE, 0, 0); 1288 1289 /* 1290 * If purging an active vnode (typically during a forced unmount 1291 * or reboot), it must be closed and deactivated before being 1292 * reclaimed. This isn't really all that safe, but what can 1293 * we do? XXX. 1294 * 1295 * Note that neither of these routines unlocks the vnode. 1296 */ 1297 if (active && (flags & DOCLOSE)) { 1298 while ((n = vp->v_opencount) != 0) { 1299 if (vp->v_writecount) 1300 VOP_CLOSE(vp, FWRITE|FNONBLOCK, NULL); 1301 else 1302 VOP_CLOSE(vp, FNONBLOCK, NULL); 1303 if (vp->v_opencount == n) { 1304 kprintf("Warning: unable to force-close" 1305 " vnode %p\n", vp); 1306 break; 1307 } 1308 } 1309 } 1310 1311 /* 1312 * If the vnode has not been deactivated, deactivated it. Deactivation 1313 * can create new buffers and VM pages so we have to call vinvalbuf() 1314 * again to make sure they all get flushed. 1315 * 1316 * This can occur if a file with a link count of 0 needs to be 1317 * truncated. 1318 * 1319 * If the vnode is already dead don't try to deactivate it. 1320 */ 1321 if ((vp->v_flag & VINACTIVE) == 0) { 1322 vsetflags(vp, VINACTIVE); 1323 if (vp->v_mount) 1324 VOP_INACTIVE(vp); 1325 vinvalbuf(vp, V_SAVE, 0, 0); 1326 } 1327 1328 /* 1329 * If the vnode has an object, destroy it. 1330 */ 1331 while ((object = vp->v_object) != NULL) { 1332 vm_object_hold(object); 1333 if (object == vp->v_object) 1334 break; 1335 vm_object_drop(object); 1336 } 1337 1338 if (object != NULL) { 1339 if (object->ref_count == 0) { 1340 if ((object->flags & OBJ_DEAD) == 0) 1341 vm_object_terminate(object); 1342 vm_object_drop(object); 1343 vclrflags(vp, VOBJBUF); 1344 } else { 1345 vm_pager_deallocate(object); 1346 vclrflags(vp, VOBJBUF); 1347 vm_object_drop(object); 1348 } 1349 } 1350 KKASSERT((vp->v_flag & VOBJBUF) == 0); 1351 1352 if (vp->v_flag & VOBJDIRTY) 1353 vclrobjdirty(vp); 1354 1355 /* 1356 * Reclaim the vnode if not already dead. 1357 */ 1358 if (vp->v_mount && VOP_RECLAIM(vp)) 1359 panic("vclean: cannot reclaim"); 1360 1361 /* 1362 * Done with purge, notify sleepers of the grim news. 1363 */ 1364 vp->v_ops = &dead_vnode_vops_p; 1365 vn_gone(vp); 1366 vp->v_tag = VT_NON; 1367 1368 /* 1369 * If we are destroying an active vnode, reactivate it now that 1370 * we have reassociated it with deadfs. This prevents the system 1371 * from crashing on the vnode due to it being unexpectedly marked 1372 * as inactive or reclaimed. 1373 */ 1374 if (active && (flags & DOCLOSE)) { 1375 vclrflags(vp, VINACTIVE | VRECLAIMED); 1376 } 1377 } 1378 1379 /* 1380 * Eliminate all activity associated with the requested vnode 1381 * and with all vnodes aliased to the requested vnode. 1382 * 1383 * The vnode must be referenced but should not be locked. 1384 */ 1385 int 1386 vrevoke(struct vnode *vp, struct ucred *cred) 1387 { 1388 struct vnode *vq; 1389 struct vnode *vqn; 1390 cdev_t dev; 1391 int error; 1392 1393 /* 1394 * If the vnode has a device association, scrap all vnodes associated 1395 * with the device. Don't let the device disappear on us while we 1396 * are scrapping the vnodes. 1397 * 1398 * The passed vp will probably show up in the list, do not VX lock 1399 * it twice! 1400 * 1401 * Releasing the vnode's rdev here can mess up specfs's call to 1402 * device close, so don't do it. The vnode has been disassociated 1403 * and the device will be closed after the last ref on the related 1404 * fp goes away (if not still open by e.g. the kernel). 1405 */ 1406 if (vp->v_type != VCHR) { 1407 error = fdrevoke(vp, DTYPE_VNODE, cred); 1408 return (error); 1409 } 1410 if ((dev = vp->v_rdev) == NULL) { 1411 return(0); 1412 } 1413 reference_dev(dev); 1414 lwkt_gettoken(&spechash_token); 1415 1416 restart: 1417 vqn = SLIST_FIRST(&dev->si_hlist); 1418 if (vqn) 1419 vhold(vqn); 1420 while ((vq = vqn) != NULL) { 1421 if (VREFCNT(vq) > 0) { 1422 vref(vq); 1423 fdrevoke(vq, DTYPE_VNODE, cred); 1424 /*v_release_rdev(vq);*/ 1425 vrele(vq); 1426 if (vq->v_rdev != dev) { 1427 vdrop(vq); 1428 goto restart; 1429 } 1430 } 1431 vqn = SLIST_NEXT(vq, v_cdevnext); 1432 if (vqn) 1433 vhold(vqn); 1434 vdrop(vq); 1435 } 1436 lwkt_reltoken(&spechash_token); 1437 dev_drevoke(dev); 1438 release_dev(dev); 1439 return (0); 1440 } 1441 1442 /* 1443 * This is called when the object underlying a vnode is being destroyed, 1444 * such as in a remove(). Try to recycle the vnode immediately if the 1445 * only active reference is our reference. 1446 * 1447 * Directory vnodes in the namecache with children cannot be immediately 1448 * recycled because numerous VOP_N*() ops require them to be stable. 1449 * 1450 * To avoid recursive recycling from VOP_INACTIVE implemenetations this 1451 * function is a NOP if VRECLAIMED is already set. 1452 */ 1453 int 1454 vrecycle(struct vnode *vp) 1455 { 1456 if (VREFCNT(vp) <= 1 && (vp->v_flag & VRECLAIMED) == 0) { 1457 if (cache_inval_vp_nonblock(vp)) 1458 return(0); 1459 vgone_vxlocked(vp); 1460 return (1); 1461 } 1462 return (0); 1463 } 1464 1465 /* 1466 * Return the maximum I/O size allowed for strategy calls on VP. 1467 * 1468 * If vp is VCHR or VBLK we dive the device, otherwise we use 1469 * the vp's mount info. 1470 * 1471 * The returned value is clamped at MAXPHYS as most callers cannot use 1472 * buffers larger than that size. 1473 */ 1474 int 1475 vmaxiosize(struct vnode *vp) 1476 { 1477 int maxiosize; 1478 1479 if (vp->v_type == VBLK || vp->v_type == VCHR) 1480 maxiosize = vp->v_rdev->si_iosize_max; 1481 else 1482 maxiosize = vp->v_mount->mnt_iosize_max; 1483 1484 if (maxiosize > MAXPHYS) 1485 maxiosize = MAXPHYS; 1486 return (maxiosize); 1487 } 1488 1489 /* 1490 * Eliminate all activity associated with a vnode in preparation for 1491 * destruction. 1492 * 1493 * The vnode must be VX locked and refd and will remain VX locked and refd 1494 * on return. This routine may be called with the vnode in any state, as 1495 * long as it is VX locked. The vnode will be cleaned out and marked 1496 * VRECLAIMED but will not actually be reused until all existing refs and 1497 * holds go away. 1498 * 1499 * NOTE: This routine may be called on a vnode which has not yet been 1500 * already been deactivated (VOP_INACTIVE), or on a vnode which has 1501 * already been reclaimed. 1502 * 1503 * This routine is not responsible for placing us back on the freelist. 1504 * Instead, it happens automatically when the caller releases the VX lock 1505 * (assuming there aren't any other references). 1506 */ 1507 void 1508 vgone_vxlocked(struct vnode *vp) 1509 { 1510 /* 1511 * assert that the VX lock is held. This is an absolute requirement 1512 * now for vgone_vxlocked() to be called. 1513 */ 1514 KKASSERT(lockinuse(&vp->v_lock)); 1515 1516 /* 1517 * Clean out the filesystem specific data and set the VRECLAIMED 1518 * bit. Also deactivate the vnode if necessary. 1519 * 1520 * The vnode should have automatically been removed from the syncer 1521 * list as syncer/dirty flags cleared during the cleaning. 1522 */ 1523 vclean_vxlocked(vp, DOCLOSE); 1524 1525 /* 1526 * Normally panic if the vnode is still dirty, unless we are doing 1527 * a forced unmount (tmpfs typically). 1528 */ 1529 if (vp->v_flag & VONWORKLST) { 1530 if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) { 1531 /* force removal */ 1532 vn_syncer_remove(vp, 1); 1533 } else { 1534 panic("vp %p still dirty in vgone after flush", vp); 1535 } 1536 } 1537 1538 /* 1539 * Delete from old mount point vnode list, if on one. 1540 */ 1541 if (vp->v_mount != NULL) { 1542 KKASSERT(vp->v_data == NULL); 1543 insmntque(vp, NULL); 1544 } 1545 1546 /* 1547 * If special device, remove it from special device alias list 1548 * if it is on one. This should normally only occur if a vnode is 1549 * being revoked as the device should otherwise have been released 1550 * naturally. 1551 */ 1552 if ((vp->v_type == VBLK || vp->v_type == VCHR) && vp->v_rdev != NULL) { 1553 v_release_rdev(vp); 1554 } 1555 1556 /* 1557 * Set us to VBAD 1558 */ 1559 vp->v_type = VBAD; 1560 } 1561 1562 /* 1563 * Calculate the total number of references to a special device. This 1564 * routine may only be called for VBLK and VCHR vnodes since v_rdev is 1565 * an overloaded field. Since dev_from_devid() can now return NULL, we 1566 * have to check for a NULL v_rdev. 1567 */ 1568 int 1569 count_dev(cdev_t dev) 1570 { 1571 struct vnode *vp; 1572 int count = 0; 1573 1574 if (SLIST_FIRST(&dev->si_hlist)) { 1575 lwkt_gettoken(&spechash_token); 1576 SLIST_FOREACH(vp, &dev->si_hlist, v_cdevnext) { 1577 count += vp->v_opencount; 1578 } 1579 lwkt_reltoken(&spechash_token); 1580 } 1581 return(count); 1582 } 1583 1584 int 1585 vcount(struct vnode *vp) 1586 { 1587 if (vp->v_rdev == NULL) 1588 return(0); 1589 return(count_dev(vp->v_rdev)); 1590 } 1591 1592 /* 1593 * Initialize VMIO for a vnode. This routine MUST be called before a 1594 * VFS can issue buffer cache ops on a vnode. It is typically called 1595 * when a vnode is initialized from its inode. 1596 */ 1597 int 1598 vinitvmio(struct vnode *vp, off_t filesize, int blksize, int boff) 1599 { 1600 vm_object_t object; 1601 int error = 0; 1602 1603 object = vp->v_object; 1604 if (object) { 1605 vm_object_hold(object); 1606 KKASSERT(vp->v_object == object); 1607 } 1608 1609 if (object == NULL) { 1610 object = vnode_pager_alloc(vp, filesize, 0, 0, blksize, boff); 1611 1612 /* 1613 * Dereference the reference we just created. This assumes 1614 * that the object is associated with the vp. Allow it to 1615 * have zero refs. It cannot be destroyed as long as it 1616 * is associated with the vnode. 1617 */ 1618 vm_object_hold(object); 1619 atomic_add_int(&object->ref_count, -1); 1620 vrele(vp); 1621 } else { 1622 KKASSERT((object->flags & OBJ_DEAD) == 0); 1623 } 1624 KASSERT(vp->v_object != NULL, ("vinitvmio: NULL object")); 1625 vsetflags(vp, VOBJBUF); 1626 vm_object_drop(object); 1627 1628 return (error); 1629 } 1630 1631 1632 /* 1633 * Print out a description of a vnode. 1634 */ 1635 static char *typename[] = 1636 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 1637 1638 void 1639 vprint(char *label, struct vnode *vp) 1640 { 1641 char buf[96]; 1642 1643 if (label != NULL) 1644 kprintf("%s: %p: ", label, (void *)vp); 1645 else 1646 kprintf("%p: ", (void *)vp); 1647 kprintf("type %s, refcnt %08x, writecount %d, holdcnt %d,", 1648 typename[vp->v_type], 1649 vp->v_refcnt, vp->v_writecount, vp->v_auxrefs); 1650 buf[0] = '\0'; 1651 if (vp->v_flag & VROOT) 1652 strcat(buf, "|VROOT"); 1653 if (vp->v_flag & VPFSROOT) 1654 strcat(buf, "|VPFSROOT"); 1655 if (vp->v_flag & VTEXT) 1656 strcat(buf, "|VTEXT"); 1657 if (vp->v_flag & VSYSTEM) 1658 strcat(buf, "|VSYSTEM"); 1659 if (vp->v_flag & VOBJBUF) 1660 strcat(buf, "|VOBJBUF"); 1661 if (buf[0] != '\0') 1662 kprintf(" flags (%s)", &buf[1]); 1663 if (vp->v_data == NULL) { 1664 kprintf("\n"); 1665 } else { 1666 kprintf("\n\t"); 1667 VOP_PRINT(vp); 1668 } 1669 } 1670 1671 /* 1672 * Do the usual access checking. 1673 * file_mode, uid and gid are from the vnode in question, 1674 * while acc_mode and cred are from the VOP_ACCESS parameter list 1675 */ 1676 int 1677 vaccess(enum vtype type, mode_t file_mode, uid_t uid, gid_t gid, 1678 mode_t acc_mode, struct ucred *cred) 1679 { 1680 mode_t mask; 1681 int ismember; 1682 1683 /* 1684 * Super-user always gets read/write access, but execute access depends 1685 * on at least one execute bit being set. 1686 */ 1687 if (priv_check_cred(cred, PRIV_ROOT, 0) == 0) { 1688 if ((acc_mode & VEXEC) && type != VDIR && 1689 (file_mode & (S_IXUSR|S_IXGRP|S_IXOTH)) == 0) 1690 return (EACCES); 1691 return (0); 1692 } 1693 1694 mask = 0; 1695 1696 /* Otherwise, check the owner. */ 1697 if (cred->cr_uid == uid) { 1698 if (acc_mode & VEXEC) 1699 mask |= S_IXUSR; 1700 if (acc_mode & VREAD) 1701 mask |= S_IRUSR; 1702 if (acc_mode & VWRITE) 1703 mask |= S_IWUSR; 1704 return ((file_mode & mask) == mask ? 0 : EACCES); 1705 } 1706 1707 /* Otherwise, check the groups. */ 1708 ismember = groupmember(gid, cred); 1709 if (cred->cr_svgid == gid || ismember) { 1710 if (acc_mode & VEXEC) 1711 mask |= S_IXGRP; 1712 if (acc_mode & VREAD) 1713 mask |= S_IRGRP; 1714 if (acc_mode & VWRITE) 1715 mask |= S_IWGRP; 1716 return ((file_mode & mask) == mask ? 0 : EACCES); 1717 } 1718 1719 /* Otherwise, check everyone else. */ 1720 if (acc_mode & VEXEC) 1721 mask |= S_IXOTH; 1722 if (acc_mode & VREAD) 1723 mask |= S_IROTH; 1724 if (acc_mode & VWRITE) 1725 mask |= S_IWOTH; 1726 return ((file_mode & mask) == mask ? 0 : EACCES); 1727 } 1728 1729 #ifdef DDB 1730 #include <ddb/ddb.h> 1731 1732 static int db_show_locked_vnodes(struct mount *mp, void *data); 1733 1734 /* 1735 * List all of the locked vnodes in the system. 1736 * Called when debugging the kernel. 1737 */ 1738 DB_SHOW_COMMAND(lockedvnodes, lockedvnodes) 1739 { 1740 kprintf("Locked vnodes\n"); 1741 mountlist_scan(db_show_locked_vnodes, NULL, 1742 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 1743 } 1744 1745 static int 1746 db_show_locked_vnodes(struct mount *mp, void *data __unused) 1747 { 1748 struct vnode *vp; 1749 1750 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 1751 if (vn_islocked(vp)) 1752 vprint(NULL, vp); 1753 } 1754 return(0); 1755 } 1756 #endif 1757 1758 /* 1759 * Top level filesystem related information gathering. 1760 */ 1761 static int sysctl_ovfs_conf (SYSCTL_HANDLER_ARGS); 1762 1763 static int 1764 vfs_sysctl(SYSCTL_HANDLER_ARGS) 1765 { 1766 int *name = (int *)arg1 - 1; /* XXX */ 1767 u_int namelen = arg2 + 1; /* XXX */ 1768 struct vfsconf *vfsp; 1769 int maxtypenum; 1770 1771 #if 1 || defined(COMPAT_PRELITE2) 1772 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 1773 if (namelen == 1) 1774 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 1775 #endif 1776 1777 #ifdef notyet 1778 /* all sysctl names at this level are at least name and field */ 1779 if (namelen < 2) 1780 return (ENOTDIR); /* overloaded */ 1781 if (name[0] != VFS_GENERIC) { 1782 vfsp = vfsconf_find_by_typenum(name[0]); 1783 if (vfsp == NULL) 1784 return (EOPNOTSUPP); 1785 return ((*vfsp->vfc_vfsops->vfs_sysctl)(&name[1], namelen - 1, 1786 oldp, oldlenp, newp, newlen, p)); 1787 } 1788 #endif 1789 switch (name[1]) { 1790 case VFS_MAXTYPENUM: 1791 if (namelen != 2) 1792 return (ENOTDIR); 1793 maxtypenum = vfsconf_get_maxtypenum(); 1794 return (SYSCTL_OUT(req, &maxtypenum, sizeof(maxtypenum))); 1795 case VFS_CONF: 1796 if (namelen != 3) 1797 return (ENOTDIR); /* overloaded */ 1798 vfsp = vfsconf_find_by_typenum(name[2]); 1799 if (vfsp == NULL) 1800 return (EOPNOTSUPP); 1801 return (SYSCTL_OUT(req, vfsp, sizeof *vfsp)); 1802 } 1803 return (EOPNOTSUPP); 1804 } 1805 1806 SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD, vfs_sysctl, 1807 "Generic filesystem"); 1808 1809 #if 1 || defined(COMPAT_PRELITE2) 1810 1811 static int 1812 sysctl_ovfs_conf_iter(struct vfsconf *vfsp, void *data) 1813 { 1814 int error; 1815 struct ovfsconf ovfs; 1816 struct sysctl_req *req = (struct sysctl_req*) data; 1817 1818 bzero(&ovfs, sizeof(ovfs)); 1819 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 1820 strcpy(ovfs.vfc_name, vfsp->vfc_name); 1821 ovfs.vfc_index = vfsp->vfc_typenum; 1822 ovfs.vfc_refcount = vfsp->vfc_refcount; 1823 ovfs.vfc_flags = vfsp->vfc_flags; 1824 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 1825 if (error) 1826 return error; /* abort iteration with error code */ 1827 else 1828 return 0; /* continue iterating with next element */ 1829 } 1830 1831 static int 1832 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 1833 { 1834 return vfsconf_each(sysctl_ovfs_conf_iter, (void*)req); 1835 } 1836 1837 #endif /* 1 || COMPAT_PRELITE2 */ 1838 1839 /* 1840 * Check to see if a filesystem is mounted on a block device. 1841 */ 1842 int 1843 vfs_mountedon(struct vnode *vp) 1844 { 1845 cdev_t dev; 1846 1847 if ((dev = vp->v_rdev) == NULL) { 1848 /* if (vp->v_type != VBLK) 1849 dev = get_dev(vp->v_uminor, vp->v_umajor); */ 1850 } 1851 if (dev != NULL && dev->si_mountpoint) 1852 return (EBUSY); 1853 return (0); 1854 } 1855 1856 /* 1857 * Unmount all filesystems. The list is traversed in reverse order 1858 * of mounting to avoid dependencies. 1859 * 1860 * We want the umountall to be able to break out of its loop if a 1861 * failure occurs, after scanning all possible mounts, so the callback 1862 * returns 0 on error. 1863 * 1864 * NOTE: Do not call mountlist_remove(mp) on error any more, this will 1865 * confuse mountlist_scan()'s unbusy check. 1866 */ 1867 static int vfs_umountall_callback(struct mount *mp, void *data); 1868 1869 void 1870 vfs_unmountall(int halting) 1871 { 1872 int count; 1873 1874 do { 1875 count = mountlist_scan(vfs_umountall_callback, &halting, 1876 MNTSCAN_REVERSE|MNTSCAN_NOBUSY); 1877 } while (count); 1878 } 1879 1880 static 1881 int 1882 vfs_umountall_callback(struct mount *mp, void *data) 1883 { 1884 int error; 1885 int halting = *(int *)data; 1886 1887 /* 1888 * NOTE: When halting, dounmount will disconnect but leave 1889 * certain mount points intact. e.g. devfs. 1890 */ 1891 error = dounmount(mp, MNT_FORCE, halting); 1892 if (error) { 1893 kprintf("unmount of filesystem mounted from %s failed (", 1894 mp->mnt_stat.f_mntfromname); 1895 if (error == EBUSY) 1896 kprintf("BUSY)\n"); 1897 else 1898 kprintf("%d)\n", error); 1899 return 0; 1900 } else { 1901 return 1; 1902 } 1903 } 1904 1905 /* 1906 * Checks the mount flags for parameter mp and put the names comma-separated 1907 * into a string buffer buf with a size limit specified by len. 1908 * 1909 * It returns the number of bytes written into buf, and (*errorp) will be 1910 * set to 0, EINVAL (if passed length is 0), or ENOSPC (supplied buffer was 1911 * not large enough). The buffer will be 0-terminated if len was not 0. 1912 */ 1913 size_t 1914 vfs_flagstostr(int flags, const struct mountctl_opt *optp, 1915 char *buf, size_t len, int *errorp) 1916 { 1917 static const struct mountctl_opt optnames[] = { 1918 { MNT_RDONLY, "read-only" }, 1919 { MNT_SYNCHRONOUS, "synchronous" }, 1920 { MNT_NOEXEC, "noexec" }, 1921 { MNT_NOSUID, "nosuid" }, 1922 { MNT_NODEV, "nodev" }, 1923 { MNT_AUTOMOUNTED, "automounted" }, 1924 { MNT_ASYNC, "asynchronous" }, 1925 { MNT_SUIDDIR, "suiddir" }, 1926 { MNT_SOFTDEP, "soft-updates" }, 1927 { MNT_NOSYMFOLLOW, "nosymfollow" }, 1928 { MNT_TRIM, "trim" }, 1929 { MNT_NOATIME, "noatime" }, 1930 { MNT_NOCLUSTERR, "noclusterr" }, 1931 { MNT_NOCLUSTERW, "noclusterw" }, 1932 { MNT_EXRDONLY, "NFS read-only" }, 1933 { MNT_EXPORTED, "NFS exported" }, 1934 /* Remaining NFS flags could come here */ 1935 { MNT_LOCAL, "local" }, 1936 { MNT_QUOTA, "with-quotas" }, 1937 /* { MNT_ROOTFS, "rootfs" }, */ 1938 /* { MNT_IGNORE, "ignore" }, */ 1939 { 0, NULL} 1940 }; 1941 int bwritten; 1942 int bleft; 1943 int optlen; 1944 int actsize; 1945 1946 *errorp = 0; 1947 bwritten = 0; 1948 bleft = len - 1; /* leave room for trailing \0 */ 1949 1950 /* 1951 * Checks the size of the string. If it contains 1952 * any data, then we will append the new flags to 1953 * it. 1954 */ 1955 actsize = strlen(buf); 1956 if (actsize > 0) 1957 buf += actsize; 1958 1959 /* Default flags if no flags passed */ 1960 if (optp == NULL) 1961 optp = optnames; 1962 1963 if (bleft < 0) { /* degenerate case, 0-length buffer */ 1964 *errorp = EINVAL; 1965 return(0); 1966 } 1967 1968 for (; flags && optp->o_opt; ++optp) { 1969 if ((flags & optp->o_opt) == 0) 1970 continue; 1971 optlen = strlen(optp->o_name); 1972 if (bwritten || actsize > 0) { 1973 if (bleft < 2) { 1974 *errorp = ENOSPC; 1975 break; 1976 } 1977 buf[bwritten++] = ','; 1978 buf[bwritten++] = ' '; 1979 bleft -= 2; 1980 } 1981 if (bleft < optlen) { 1982 *errorp = ENOSPC; 1983 break; 1984 } 1985 bcopy(optp->o_name, buf + bwritten, optlen); 1986 bwritten += optlen; 1987 bleft -= optlen; 1988 flags &= ~optp->o_opt; 1989 } 1990 1991 /* 1992 * Space already reserved for trailing \0 1993 */ 1994 buf[bwritten] = 0; 1995 return (bwritten); 1996 } 1997 1998 /* 1999 * Build hash lists of net addresses and hang them off the mount point. 2000 * Called by ufs_mount() to set up the lists of export addresses. 2001 */ 2002 static int 2003 vfs_hang_addrlist(struct mount *mp, struct netexport *nep, 2004 const struct export_args *argp) 2005 { 2006 struct netcred *np; 2007 struct radix_node_head *rnh; 2008 int i; 2009 struct radix_node *rn; 2010 struct sockaddr *saddr, *smask = NULL; 2011 int error; 2012 2013 if (argp->ex_addrlen == 0) { 2014 if (mp->mnt_flag & MNT_DEFEXPORTED) 2015 return (EPERM); 2016 np = &nep->ne_defexported; 2017 np->netc_exflags = argp->ex_flags; 2018 np->netc_anon = argp->ex_anon; 2019 np->netc_anon.cr_ref = 1; 2020 mp->mnt_flag |= MNT_DEFEXPORTED; 2021 return (0); 2022 } 2023 2024 if (argp->ex_addrlen < 0 || argp->ex_addrlen > MLEN) 2025 return (EINVAL); 2026 if (argp->ex_masklen < 0 || argp->ex_masklen > MLEN) 2027 return (EINVAL); 2028 2029 i = sizeof(struct netcred) + argp->ex_addrlen + argp->ex_masklen; 2030 np = (struct netcred *)kmalloc(i, M_NETCRED, M_WAITOK | M_ZERO); 2031 saddr = (struct sockaddr *) (np + 1); 2032 if ((error = copyin(argp->ex_addr, (caddr_t) saddr, argp->ex_addrlen))) 2033 goto out; 2034 if (saddr->sa_len > argp->ex_addrlen) 2035 saddr->sa_len = argp->ex_addrlen; 2036 if (argp->ex_masklen) { 2037 smask = (struct sockaddr *)((caddr_t)saddr + argp->ex_addrlen); 2038 error = copyin(argp->ex_mask, (caddr_t)smask, argp->ex_masklen); 2039 if (error) 2040 goto out; 2041 if (smask->sa_len > argp->ex_masklen) 2042 smask->sa_len = argp->ex_masklen; 2043 } 2044 NE_LOCK(nep); 2045 if (nep->ne_maskhead == NULL) { 2046 if (!rn_inithead((void **)&nep->ne_maskhead, NULL, 0)) { 2047 error = ENOBUFS; 2048 goto out; 2049 } 2050 } 2051 if ((rnh = vfs_create_addrlist_af(saddr->sa_family, nep)) == NULL) { 2052 error = ENOBUFS; 2053 goto out; 2054 } 2055 rn = (*rnh->rnh_addaddr)((char *)saddr, (char *)smask, rnh, 2056 np->netc_rnodes); 2057 NE_UNLOCK(nep); 2058 if (rn == NULL || np != (struct netcred *)rn) { /* already exists */ 2059 error = EPERM; 2060 goto out; 2061 } 2062 np->netc_exflags = argp->ex_flags; 2063 np->netc_anon = argp->ex_anon; 2064 np->netc_anon.cr_ref = 1; 2065 return (0); 2066 2067 out: 2068 kfree(np, M_NETCRED); 2069 return (error); 2070 } 2071 2072 /* 2073 * Free netcred structures installed in the netexport 2074 */ 2075 static int 2076 vfs_free_netcred(struct radix_node *rn, void *w) 2077 { 2078 struct radix_node_head *rnh = (struct radix_node_head *)w; 2079 2080 (*rnh->rnh_deladdr) (rn->rn_key, rn->rn_mask, rnh); 2081 kfree(rn, M_NETCRED); 2082 2083 return (0); 2084 } 2085 2086 /* 2087 * callback to free an element of the mask table installed in the 2088 * netexport. These may be created indirectly and are not netcred 2089 * structures. 2090 */ 2091 static int 2092 vfs_free_netcred_mask(struct radix_node *rn, void *w) 2093 { 2094 struct radix_node_head *rnh = (struct radix_node_head *)w; 2095 2096 (*rnh->rnh_deladdr) (rn->rn_key, rn->rn_mask, rnh); 2097 kfree(rn, M_RTABLE); 2098 2099 return (0); 2100 } 2101 2102 static struct radix_node_head * 2103 vfs_create_addrlist_af(int af, struct netexport *nep) 2104 { 2105 struct radix_node_head *rnh = NULL; 2106 #if defined(INET) || defined(INET6) 2107 struct radix_node_head *maskhead = nep->ne_maskhead; 2108 int off; 2109 #endif 2110 2111 NE_ASSERT_LOCKED(nep); 2112 #if defined(INET) || defined(INET6) 2113 KKASSERT(maskhead != NULL); 2114 #endif 2115 switch (af) { 2116 #ifdef INET 2117 case AF_INET: 2118 if ((rnh = nep->ne_inethead) == NULL) { 2119 off = offsetof(struct sockaddr_in, sin_addr) << 3; 2120 if (!rn_inithead((void **)&rnh, maskhead, off)) 2121 return (NULL); 2122 nep->ne_inethead = rnh; 2123 } 2124 break; 2125 #endif 2126 #ifdef INET6 2127 case AF_INET6: 2128 if ((rnh = nep->ne_inet6head) == NULL) { 2129 off = offsetof(struct sockaddr_in6, sin6_addr) << 3; 2130 if (!rn_inithead((void **)&rnh, maskhead, off)) 2131 return (NULL); 2132 nep->ne_inet6head = rnh; 2133 } 2134 break; 2135 #endif 2136 } 2137 return (rnh); 2138 } 2139 2140 /* 2141 * helper function for freeing netcred elements 2142 */ 2143 static void 2144 vfs_free_addrlist_af(struct radix_node_head **prnh) 2145 { 2146 struct radix_node_head *rnh = *prnh; 2147 2148 (*rnh->rnh_walktree) (rnh, vfs_free_netcred, rnh); 2149 kfree(rnh, M_RTABLE); 2150 *prnh = NULL; 2151 } 2152 2153 /* 2154 * helper function for freeing mask elements 2155 */ 2156 static void 2157 vfs_free_addrlist_masks(struct radix_node_head **prnh) 2158 { 2159 struct radix_node_head *rnh = *prnh; 2160 2161 (*rnh->rnh_walktree) (rnh, vfs_free_netcred_mask, rnh); 2162 kfree(rnh, M_RTABLE); 2163 *prnh = NULL; 2164 } 2165 2166 /* 2167 * Free the net address hash lists that are hanging off the mount points. 2168 */ 2169 static void 2170 vfs_free_addrlist(struct netexport *nep) 2171 { 2172 NE_LOCK(nep); 2173 if (nep->ne_inethead != NULL) 2174 vfs_free_addrlist_af(&nep->ne_inethead); 2175 if (nep->ne_inet6head != NULL) 2176 vfs_free_addrlist_af(&nep->ne_inet6head); 2177 if (nep->ne_maskhead) 2178 vfs_free_addrlist_masks(&nep->ne_maskhead); 2179 NE_UNLOCK(nep); 2180 } 2181 2182 int 2183 vfs_export(struct mount *mp, struct netexport *nep, 2184 const struct export_args *argp) 2185 { 2186 int error; 2187 2188 if (argp->ex_flags & MNT_DELEXPORT) { 2189 if (mp->mnt_flag & MNT_EXPUBLIC) { 2190 vfs_setpublicfs(NULL, NULL, NULL); 2191 mp->mnt_flag &= ~MNT_EXPUBLIC; 2192 } 2193 vfs_free_addrlist(nep); 2194 mp->mnt_flag &= ~(MNT_EXPORTED | MNT_DEFEXPORTED); 2195 } 2196 if (argp->ex_flags & MNT_EXPORTED) { 2197 if (argp->ex_flags & MNT_EXPUBLIC) { 2198 if ((error = vfs_setpublicfs(mp, nep, argp)) != 0) 2199 return (error); 2200 mp->mnt_flag |= MNT_EXPUBLIC; 2201 } 2202 if ((error = vfs_hang_addrlist(mp, nep, argp))) 2203 return (error); 2204 mp->mnt_flag |= MNT_EXPORTED; 2205 } 2206 return (0); 2207 } 2208 2209 2210 /* 2211 * Set the publicly exported filesystem (WebNFS). Currently, only 2212 * one public filesystem is possible in the spec (RFC 2054 and 2055) 2213 */ 2214 int 2215 vfs_setpublicfs(struct mount *mp, struct netexport *nep, 2216 const struct export_args *argp) 2217 { 2218 int error; 2219 struct vnode *rvp; 2220 char *cp; 2221 2222 /* 2223 * mp == NULL -> invalidate the current info, the FS is 2224 * no longer exported. May be called from either vfs_export 2225 * or unmount, so check if it hasn't already been done. 2226 */ 2227 if (mp == NULL) { 2228 if (nfs_pub.np_valid) { 2229 nfs_pub.np_valid = 0; 2230 if (nfs_pub.np_index != NULL) { 2231 kfree(nfs_pub.np_index, M_TEMP); 2232 nfs_pub.np_index = NULL; 2233 } 2234 } 2235 return (0); 2236 } 2237 2238 /* 2239 * Only one allowed at a time. 2240 */ 2241 if (nfs_pub.np_valid != 0 && mp != nfs_pub.np_mount) 2242 return (EBUSY); 2243 2244 /* 2245 * Get real filehandle for root of exported FS. 2246 */ 2247 bzero((caddr_t)&nfs_pub.np_handle, sizeof(nfs_pub.np_handle)); 2248 nfs_pub.np_handle.fh_fsid = mp->mnt_stat.f_fsid; 2249 2250 if ((error = VFS_ROOT(mp, &rvp))) 2251 return (error); 2252 2253 if ((error = VFS_VPTOFH(rvp, &nfs_pub.np_handle.fh_fid))) 2254 return (error); 2255 2256 vput(rvp); 2257 2258 /* 2259 * If an indexfile was specified, pull it in. 2260 */ 2261 if (argp->ex_indexfile != NULL) { 2262 int namelen; 2263 2264 error = vn_get_namelen(rvp, &namelen); 2265 if (error) 2266 return (error); 2267 nfs_pub.np_index = kmalloc(namelen, M_TEMP, M_WAITOK); 2268 error = copyinstr(argp->ex_indexfile, nfs_pub.np_index, 2269 namelen, NULL); 2270 if (!error) { 2271 /* 2272 * Check for illegal filenames. 2273 */ 2274 for (cp = nfs_pub.np_index; *cp; cp++) { 2275 if (*cp == '/') { 2276 error = EINVAL; 2277 break; 2278 } 2279 } 2280 } 2281 if (error) { 2282 kfree(nfs_pub.np_index, M_TEMP); 2283 return (error); 2284 } 2285 } 2286 2287 nfs_pub.np_mount = mp; 2288 nfs_pub.np_valid = 1; 2289 return (0); 2290 } 2291 2292 struct netcred * 2293 vfs_export_lookup(struct mount *mp, struct netexport *nep, 2294 struct sockaddr *nam) 2295 { 2296 struct netcred *np; 2297 struct radix_node_head *rnh; 2298 struct sockaddr *saddr; 2299 2300 np = NULL; 2301 if (mp->mnt_flag & MNT_EXPORTED) { 2302 /* 2303 * Lookup in the export list first. 2304 */ 2305 NE_LOCK(nep); 2306 if (nam != NULL) { 2307 saddr = nam; 2308 switch (saddr->sa_family) { 2309 #ifdef INET 2310 case AF_INET: 2311 rnh = nep->ne_inethead; 2312 break; 2313 #endif 2314 #ifdef INET6 2315 case AF_INET6: 2316 rnh = nep->ne_inet6head; 2317 break; 2318 #endif 2319 default: 2320 rnh = NULL; 2321 } 2322 if (rnh != NULL) { 2323 np = (struct netcred *) 2324 (*rnh->rnh_matchaddr)((char *)saddr, 2325 rnh); 2326 if (np && np->netc_rnodes->rn_flags & RNF_ROOT) 2327 np = NULL; 2328 } 2329 } 2330 NE_UNLOCK(nep); 2331 /* 2332 * If no address match, use the default if it exists. 2333 */ 2334 if (np == NULL && mp->mnt_flag & MNT_DEFEXPORTED) 2335 np = &nep->ne_defexported; 2336 } 2337 return (np); 2338 } 2339 2340 /* 2341 * perform msync on all vnodes under a mount point. The mount point must 2342 * be locked. This code is also responsible for lazy-freeing unreferenced 2343 * vnodes whos VM objects no longer contain pages. 2344 * 2345 * NOTE: MNT_WAIT still skips vnodes in the VXLOCK state. 2346 * 2347 * NOTE: XXX VOP_PUTPAGES and friends requires that the vnode be locked, 2348 * but vnode_pager_putpages() doesn't lock the vnode. We have to do it 2349 * way up in this high level function. 2350 */ 2351 static int vfs_msync_scan1(struct mount *mp, struct vnode *vp, void *data); 2352 static int vfs_msync_scan2(struct mount *mp, struct vnode *vp, void *data); 2353 2354 void 2355 vfs_msync(struct mount *mp, int flags) 2356 { 2357 int vmsc_flags; 2358 2359 /* 2360 * tmpfs sets this flag to prevent msync(), sync, and the 2361 * filesystem periodic syncer from trying to flush VM pages 2362 * to swap. Only pure memory pressure flushes tmpfs VM pages 2363 * to swap. 2364 */ 2365 if (mp->mnt_kern_flag & MNTK_NOMSYNC) 2366 return; 2367 2368 /* 2369 * Ok, scan the vnodes for work. If the filesystem is using the 2370 * syncer thread feature we can use vsyncscan() instead of 2371 * vmntvnodescan(), which is much faster. 2372 */ 2373 vmsc_flags = VMSC_GETVP; 2374 if (flags != MNT_WAIT) 2375 vmsc_flags |= VMSC_NOWAIT; 2376 2377 if (mp->mnt_kern_flag & MNTK_THR_SYNC) { 2378 vsyncscan(mp, vmsc_flags, vfs_msync_scan2, 2379 (void *)(intptr_t)flags); 2380 } else { 2381 vmntvnodescan(mp, vmsc_flags, 2382 vfs_msync_scan1, vfs_msync_scan2, 2383 (void *)(intptr_t)flags); 2384 } 2385 } 2386 2387 /* 2388 * scan1 is a fast pre-check. There could be hundreds of thousands of 2389 * vnodes, we cannot afford to do anything heavy weight until we have a 2390 * fairly good indication that there is work to do. 2391 * 2392 * The new namecache holds the vnode for each v_namecache association 2393 * so allow these refs. 2394 */ 2395 static 2396 int 2397 vfs_msync_scan1(struct mount *mp, struct vnode *vp, void *data) 2398 { 2399 int flags = (int)(intptr_t)data; 2400 2401 if ((vp->v_flag & VRECLAIMED) == 0) { 2402 if (vp->v_auxrefs == vp->v_namecache_count && 2403 VREFCNT(vp) <= 0 && vp->v_object) { 2404 return(0); /* call scan2 */ 2405 } 2406 if ((mp->mnt_flag & MNT_RDONLY) == 0 && 2407 (vp->v_flag & VOBJDIRTY) && 2408 (flags == MNT_WAIT || vn_islocked(vp) == 0)) { 2409 return(0); /* call scan2 */ 2410 } 2411 } 2412 2413 /* 2414 * do not call scan2, continue the loop 2415 */ 2416 return(-1); 2417 } 2418 2419 /* 2420 * This callback is handed a locked vnode. 2421 */ 2422 static 2423 int 2424 vfs_msync_scan2(struct mount *mp, struct vnode *vp, void *data) 2425 { 2426 vm_object_t obj; 2427 int flags = (int)(intptr_t)data; 2428 int opcflags; 2429 2430 if (vp->v_flag & VRECLAIMED) 2431 return(0); 2432 2433 if ((mp->mnt_flag & MNT_RDONLY) == 0 && (vp->v_flag & VOBJDIRTY)) { 2434 if ((obj = vp->v_object) != NULL) { 2435 if (flags == MNT_WAIT) { 2436 /* 2437 * VFS_MSYNC is called with MNT_WAIT when 2438 * unmounting. 2439 */ 2440 opcflags = OBJPC_SYNC; 2441 } else if (vp->v_writecount || obj->ref_count) { 2442 /* 2443 * VFS_MSYNC is otherwise called via the 2444 * periodic filesystem sync or the 'sync' 2445 * command. Honor MADV_NOSYNC / MAP_NOSYNC 2446 * if the file is open for writing or memory 2447 * mapped. Pages flagged PG_NOSYNC will not 2448 * be automatically flushed at this time. 2449 * 2450 * The obj->ref_count test is not perfect 2451 * since temporary refs may be present, but 2452 * the periodic filesystem sync will ultimately 2453 * catch it if the file is not open and not 2454 * mapped. 2455 */ 2456 opcflags = OBJPC_NOSYNC; 2457 } else { 2458 /* 2459 * If the file is no longer open for writing 2460 * and also no longer mapped, do not honor 2461 * MAP_NOSYNC. That is, fully synchronize 2462 * the file. 2463 * 2464 * This still occurs on the periodic fs sync, 2465 * so frontend programs which turn the file 2466 * over quickly enough can still avoid the 2467 * sync, but ultimately we do want to flush 2468 * even MADV_NOSYNC pages once it is no longer 2469 * mapped or open for writing. 2470 */ 2471 opcflags = 0; 2472 } 2473 vm_object_page_clean(obj, 0, 0, opcflags); 2474 } 2475 } 2476 return(0); 2477 } 2478 2479 /* 2480 * Wake up anyone interested in vp because it is being revoked. 2481 */ 2482 void 2483 vn_gone(struct vnode *vp) 2484 { 2485 lwkt_gettoken(&vp->v_token); 2486 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, NOTE_REVOKE); 2487 lwkt_reltoken(&vp->v_token); 2488 } 2489 2490 /* 2491 * extract the cdev_t from a VBLK or VCHR. The vnode must have been opened 2492 * (or v_rdev might be NULL). 2493 */ 2494 cdev_t 2495 vn_todev(struct vnode *vp) 2496 { 2497 if (vp->v_type != VBLK && vp->v_type != VCHR) 2498 return (NULL); 2499 KKASSERT(vp->v_rdev != NULL); 2500 return (vp->v_rdev); 2501 } 2502 2503 /* 2504 * Check if vnode represents a disk device. The vnode does not need to be 2505 * opened. 2506 * 2507 * MPALMOSTSAFE 2508 */ 2509 int 2510 vn_isdisk(struct vnode *vp, int *errp) 2511 { 2512 cdev_t dev; 2513 2514 if (vp->v_type != VCHR) { 2515 if (errp != NULL) 2516 *errp = ENOTBLK; 2517 return (0); 2518 } 2519 2520 dev = vp->v_rdev; 2521 2522 if (dev == NULL) { 2523 if (errp != NULL) 2524 *errp = ENXIO; 2525 return (0); 2526 } 2527 if (dev_is_good(dev) == 0) { 2528 if (errp != NULL) 2529 *errp = ENXIO; 2530 return (0); 2531 } 2532 if ((dev_dflags(dev) & D_DISK) == 0) { 2533 if (errp != NULL) 2534 *errp = ENOTBLK; 2535 return (0); 2536 } 2537 if (errp != NULL) 2538 *errp = 0; 2539 return (1); 2540 } 2541 2542 int 2543 vn_get_namelen(struct vnode *vp, int *namelen) 2544 { 2545 int error; 2546 register_t retval[2]; 2547 2548 error = VOP_PATHCONF(vp, _PC_NAME_MAX, retval); 2549 if (error) 2550 return (error); 2551 *namelen = (int)retval[0]; 2552 return (0); 2553 } 2554 2555 int 2556 vop_write_dirent(int *error, struct uio *uio, ino_t d_ino, uint8_t d_type, 2557 uint16_t d_namlen, const char *d_name) 2558 { 2559 struct dirent *dp; 2560 size_t len; 2561 2562 len = _DIRENT_RECLEN(d_namlen); 2563 if (len > uio->uio_resid) 2564 return(1); 2565 2566 dp = kmalloc(len, M_TEMP, M_WAITOK | M_ZERO); 2567 2568 dp->d_ino = d_ino; 2569 dp->d_namlen = d_namlen; 2570 dp->d_type = d_type; 2571 bcopy(d_name, dp->d_name, d_namlen); 2572 2573 *error = uiomove((caddr_t)dp, len, uio); 2574 2575 kfree(dp, M_TEMP); 2576 2577 return(0); 2578 } 2579 2580 void 2581 vn_mark_atime(struct vnode *vp, struct thread *td) 2582 { 2583 struct proc *p = td->td_proc; 2584 struct ucred *cred = p ? p->p_ucred : proc0.p_ucred; 2585 2586 if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) { 2587 VOP_MARKATIME(vp, cred); 2588 } 2589 } 2590 2591 /* 2592 * Calculate the number of entries in an inode-related chained hash table. 2593 * With today's memory sizes, maxvnodes can wind up being a very large 2594 * number. There is no reason to waste memory, so tolerate some stacking. 2595 */ 2596 int 2597 vfs_inodehashsize(void) 2598 { 2599 int hsize; 2600 2601 hsize = 32; 2602 while (hsize < maxvnodes) 2603 hsize <<= 1; 2604 while (hsize > maxvnodes * 2) 2605 hsize >>= 1; /* nominal 2x stacking */ 2606 2607 if (maxvnodes > 1024 * 1024) 2608 hsize >>= 1; /* nominal 8x stacking */ 2609 2610 if (maxvnodes > 128 * 1024) 2611 hsize >>= 1; /* nominal 4x stacking */ 2612 2613 if (hsize < 16) 2614 hsize = 16; 2615 2616 return hsize; 2617 } 2618 2619 union _qcvt { 2620 quad_t qcvt; 2621 int32_t val[2]; 2622 }; 2623 2624 #define SETHIGH(q, h) { \ 2625 union _qcvt tmp; \ 2626 tmp.qcvt = (q); \ 2627 tmp.val[_QUAD_HIGHWORD] = (h); \ 2628 (q) = tmp.qcvt; \ 2629 } 2630 #define SETLOW(q, l) { \ 2631 union _qcvt tmp; \ 2632 tmp.qcvt = (q); \ 2633 tmp.val[_QUAD_LOWWORD] = (l); \ 2634 (q) = tmp.qcvt; \ 2635 } 2636 2637 u_quad_t 2638 init_va_filerev(void) 2639 { 2640 struct timeval tv; 2641 u_quad_t ret = 0; 2642 2643 getmicrouptime(&tv); 2644 SETHIGH(ret, tv.tv_sec); 2645 SETLOW(ret, tv.tv_usec * 4294); 2646 2647 return ret; 2648 } 2649