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