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/buf.h> 48 #include <sys/conf.h> 49 #include <sys/dirent.h> 50 #include <sys/eventhandler.h> 51 #include <sys/fcntl.h> 52 #include <sys/file.h> 53 #include <sys/kernel.h> 54 #include <sys/kthread.h> 55 #include <sys/malloc.h> 56 #include <sys/mbuf.h> 57 #include <sys/mount.h> 58 #include <sys/priv.h> 59 #include <sys/proc.h> 60 #include <sys/reboot.h> 61 #include <sys/socket.h> 62 #include <sys/stat.h> 63 #include <sys/sysctl.h> 64 #include <sys/syslog.h> 65 #include <sys/unistd.h> 66 #include <sys/vmmeter.h> 67 #include <sys/vnode.h> 68 69 #include <machine/limits.h> 70 71 #include <vm/vm.h> 72 #include <vm/vm_object.h> 73 #include <vm/vm_extern.h> 74 #include <vm/vm_kern.h> 75 #include <vm/pmap.h> 76 #include <vm/vm_map.h> 77 #include <vm/vm_page.h> 78 #include <vm/vm_pager.h> 79 #include <vm/vnode_pager.h> 80 #include <vm/vm_zone.h> 81 82 #include <sys/buf2.h> 83 #include <sys/mplock2.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); 129 130 __read_mostly int prtactive = 0; /* 1 => print out reclaim of active vnodes */ 131 132 /* 133 * Red black tree functions 134 */ 135 static int rb_buf_compare(struct buf *b1, struct buf *b2); 136 RB_GENERATE2(buf_rb_tree, buf, b_rbnode, rb_buf_compare, off_t, b_loffset); 137 RB_GENERATE2(buf_rb_hash, buf, b_rbhash, rb_buf_compare, off_t, b_loffset); 138 139 static int 140 rb_buf_compare(struct buf *b1, struct buf *b2) 141 { 142 if (b1->b_loffset < b2->b_loffset) 143 return(-1); 144 if (b1->b_loffset > b2->b_loffset) 145 return(1); 146 return(0); 147 } 148 149 /* 150 * Initialize the vnode management data structures. 151 * 152 * Called from vfsinit() 153 */ 154 #define MAXVNBREAKMEM (1L * 1024 * 1024 * 1024) 155 #define MINVNODES 2000 156 #define MAXVNODES 4000000 157 158 void 159 vfs_subr_init(void) 160 { 161 int factor1; /* Limit based on ram (x 2 above 1GB) */ 162 int factor2; /* Limit based on available KVM */ 163 size_t freemem; 164 165 /* 166 * Size maxvnodes to available memory. Size significantly 167 * smaller on low-memory systems (calculations for the first 168 * 1GB of ram), and pump it up a bit when free memory is 169 * above 1GB. 170 * 171 * The general minimum is maxproc * 8 (we want someone pushing 172 * up maxproc a lot to also get more vnodes). Usually maxproc 173 * does not affect this calculation. 174 * 175 * There isn't much of a point allowing maxvnodes to exceed a 176 * few million as our modern filesystems cache pages in the 177 * underlying block device and not so much hanging off of VM 178 * objects. 179 */ 180 factor1 = 50 * (sizeof(struct vm_object) + sizeof(struct vnode)); 181 factor2 = 30 * (sizeof(struct vm_object) + sizeof(struct vnode)); 182 183 freemem = (int64_t)vmstats.v_page_count * PAGE_SIZE; 184 185 maxvnodes = freemem / factor1; 186 if (freemem > MAXVNBREAKMEM) 187 maxvnodes += (freemem - MAXVNBREAKMEM) / factor1; 188 maxvnodes = imax(maxvnodes, maxproc * 8); 189 maxvnodes = imin(maxvnodes, KvaSize / factor2); 190 maxvnodes = imin(maxvnodes, MAXVNODES); 191 maxvnodes = imax(maxvnodes, MINVNODES); 192 193 lwkt_token_init(&spechash_token, "spechash"); 194 } 195 196 /* 197 * Knob to control the precision of file timestamps: 198 * 199 * 0 = seconds only; nanoseconds zeroed. 200 * 1 = seconds and nanoseconds, accurate within 1/HZ. 201 * 2 = seconds and nanoseconds, truncated to microseconds. 202 * >=3 = seconds and nanoseconds, maximum precision. 203 */ 204 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 205 206 __read_mostly static int timestamp_precision = TSP_SEC; 207 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 208 ×tamp_precision, 0, "Precision of file timestamps"); 209 210 /* 211 * Get a current timestamp. 212 * 213 * MPSAFE 214 */ 215 void 216 vfs_timestamp(struct timespec *tsp) 217 { 218 struct timeval tv; 219 220 switch (timestamp_precision) { 221 case TSP_SEC: 222 tsp->tv_sec = time_second; 223 tsp->tv_nsec = 0; 224 break; 225 case TSP_HZ: 226 getnanotime(tsp); 227 break; 228 case TSP_USEC: 229 microtime(&tv); 230 TIMEVAL_TO_TIMESPEC(&tv, tsp); 231 break; 232 case TSP_NSEC: 233 default: 234 nanotime(tsp); 235 break; 236 } 237 } 238 239 /* 240 * Set vnode attributes to VNOVAL 241 */ 242 void 243 vattr_null(struct vattr *vap) 244 { 245 vap->va_type = VNON; 246 vap->va_size = VNOVAL; 247 vap->va_bytes = VNOVAL; 248 vap->va_mode = VNOVAL; 249 vap->va_nlink = VNOVAL; 250 vap->va_uid = VNOVAL; 251 vap->va_gid = VNOVAL; 252 vap->va_fsid = VNOVAL; 253 vap->va_fileid = VNOVAL; 254 vap->va_blocksize = VNOVAL; 255 vap->va_rmajor = VNOVAL; 256 vap->va_rminor = VNOVAL; 257 vap->va_atime.tv_sec = VNOVAL; 258 vap->va_atime.tv_nsec = VNOVAL; 259 vap->va_mtime.tv_sec = VNOVAL; 260 vap->va_mtime.tv_nsec = VNOVAL; 261 vap->va_ctime.tv_sec = VNOVAL; 262 vap->va_ctime.tv_nsec = VNOVAL; 263 vap->va_flags = VNOVAL; 264 vap->va_gen = VNOVAL; 265 vap->va_vaflags = 0; 266 /* va_*_uuid fields are only valid if related flags are set */ 267 } 268 269 /* 270 * Flush out and invalidate all buffers associated with a vnode. 271 * 272 * vp must be locked. 273 */ 274 static int vinvalbuf_bp(struct buf *bp, void *data); 275 276 struct vinvalbuf_bp_info { 277 struct vnode *vp; 278 int slptimeo; 279 int lkflags; 280 int flags; 281 int clean; 282 }; 283 284 int 285 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 286 { 287 struct vinvalbuf_bp_info info; 288 vm_object_t object; 289 int error; 290 291 lwkt_gettoken(&vp->v_token); 292 293 /* 294 * If we are being asked to save, call fsync to ensure that the inode 295 * is updated. 296 */ 297 if (flags & V_SAVE) { 298 error = bio_track_wait(&vp->v_track_write, slpflag, slptimeo); 299 if (error) 300 goto done; 301 if (!RB_EMPTY(&vp->v_rbdirty_tree)) { 302 if ((error = VOP_FSYNC(vp, MNT_WAIT, 0)) != 0) 303 goto done; 304 #if 0 305 /* 306 * Dirty bufs may be left or generated via races 307 * in circumstances where vinvalbuf() is called on 308 * a vnode not undergoing reclamation. Only 309 * panic if we are trying to reclaim the vnode. 310 */ 311 if ((vp->v_flag & VRECLAIMED) && 312 (bio_track_active(&vp->v_track_write) || 313 !RB_EMPTY(&vp->v_rbdirty_tree))) { 314 panic("vinvalbuf: dirty bufs"); 315 } 316 #endif 317 } 318 } 319 info.slptimeo = slptimeo; 320 info.lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 321 if (slpflag & PCATCH) 322 info.lkflags |= LK_PCATCH; 323 info.flags = flags; 324 info.vp = vp; 325 326 /* 327 * Flush the buffer cache until nothing is left, wait for all I/O 328 * to complete. At least one pass is required. We might block 329 * in the pip code so we have to re-check. Order is important. 330 */ 331 do { 332 /* 333 * Flush buffer cache 334 */ 335 if (!RB_EMPTY(&vp->v_rbclean_tree)) { 336 info.clean = 1; 337 error = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 338 NULL, vinvalbuf_bp, &info); 339 } 340 if (!RB_EMPTY(&vp->v_rbdirty_tree)) { 341 info.clean = 0; 342 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 343 NULL, vinvalbuf_bp, &info); 344 } 345 346 /* 347 * Wait for I/O completion. 348 */ 349 bio_track_wait(&vp->v_track_write, 0, 0); 350 if ((object = vp->v_object) != NULL) 351 refcount_wait(&object->paging_in_progress, "vnvlbx"); 352 } while (bio_track_active(&vp->v_track_write) || 353 !RB_EMPTY(&vp->v_rbclean_tree) || 354 !RB_EMPTY(&vp->v_rbdirty_tree)); 355 356 /* 357 * Destroy the copy in the VM cache, too. 358 */ 359 if ((object = vp->v_object) != NULL) { 360 vm_object_page_remove(object, 0, 0, 361 (flags & V_SAVE) ? TRUE : FALSE); 362 } 363 364 if (!RB_EMPTY(&vp->v_rbdirty_tree) || !RB_EMPTY(&vp->v_rbclean_tree)) 365 panic("vinvalbuf: flush failed"); 366 if (!RB_EMPTY(&vp->v_rbhash_tree)) 367 panic("vinvalbuf: flush failed, buffers still present"); 368 error = 0; 369 done: 370 lwkt_reltoken(&vp->v_token); 371 return (error); 372 } 373 374 static int 375 vinvalbuf_bp(struct buf *bp, void *data) 376 { 377 struct vinvalbuf_bp_info *info = data; 378 int error; 379 380 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 381 atomic_add_int(&bp->b_refs, 1); 382 error = BUF_TIMELOCK(bp, info->lkflags, 383 "vinvalbuf", info->slptimeo); 384 atomic_subtract_int(&bp->b_refs, 1); 385 if (error == 0) { 386 BUF_UNLOCK(bp); 387 error = ENOLCK; 388 } 389 if (error == ENOLCK) 390 return(0); 391 return (-error); 392 } 393 KKASSERT(bp->b_vp == info->vp); 394 395 /* 396 * Must check clean/dirty status after successfully locking as 397 * it may race. 398 */ 399 if ((info->clean && (bp->b_flags & B_DELWRI)) || 400 (info->clean == 0 && (bp->b_flags & B_DELWRI) == 0)) { 401 BUF_UNLOCK(bp); 402 return(0); 403 } 404 405 /* 406 * NOTE: NO B_LOCKED CHECK. Also no buf_checkwrite() 407 * check. This code will write out the buffer, period. 408 */ 409 bremfree(bp); 410 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 411 (info->flags & V_SAVE)) { 412 cluster_awrite(bp); 413 } else if (info->flags & V_SAVE) { 414 /* 415 * Cannot set B_NOCACHE on a clean buffer as this will 416 * destroy the VM backing store which might actually 417 * be dirty (and unsynchronized). 418 */ 419 bp->b_flags |= (B_INVAL | B_RELBUF); 420 brelse(bp); 421 } else { 422 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 423 brelse(bp); 424 } 425 return(0); 426 } 427 428 /* 429 * Truncate a file's buffer and pages to a specified length. This 430 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 431 * sync activity. 432 * 433 * The vnode must be locked. 434 */ 435 static int vtruncbuf_bp_trunc_cmp(struct buf *bp, void *data); 436 static int vtruncbuf_bp_trunc(struct buf *bp, void *data); 437 static int vtruncbuf_bp_metasync_cmp(struct buf *bp, void *data); 438 static int vtruncbuf_bp_metasync(struct buf *bp, void *data); 439 440 struct vtruncbuf_info { 441 struct vnode *vp; 442 off_t truncloffset; 443 int clean; 444 }; 445 446 int 447 vtruncbuf(struct vnode *vp, off_t length, int blksize) 448 { 449 struct vtruncbuf_info info; 450 const char *filename; 451 int count; 452 453 /* 454 * Round up to the *next* block, then destroy the buffers in question. 455 * Since we are only removing some of the buffers we must rely on the 456 * scan count to determine whether a loop is necessary. 457 */ 458 if ((count = (int)(length % blksize)) != 0) 459 info.truncloffset = length + (blksize - count); 460 else 461 info.truncloffset = length; 462 info.vp = vp; 463 464 lwkt_gettoken(&vp->v_token); 465 do { 466 info.clean = 1; 467 count = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 468 vtruncbuf_bp_trunc_cmp, 469 vtruncbuf_bp_trunc, &info); 470 info.clean = 0; 471 count += RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 472 vtruncbuf_bp_trunc_cmp, 473 vtruncbuf_bp_trunc, &info); 474 } while(count); 475 476 /* 477 * For safety, fsync any remaining metadata if the file is not being 478 * truncated to 0. Since the metadata does not represent the entire 479 * dirty list we have to rely on the hit count to ensure that we get 480 * all of it. 481 */ 482 if (length > 0) { 483 do { 484 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 485 vtruncbuf_bp_metasync_cmp, 486 vtruncbuf_bp_metasync, &info); 487 } while (count); 488 } 489 490 /* 491 * Clean out any left over VM backing store. 492 * 493 * It is possible to have in-progress I/O from buffers that were 494 * not part of the truncation. This should not happen if we 495 * are truncating to 0-length. 496 */ 497 vnode_pager_setsize(vp, length); 498 bio_track_wait(&vp->v_track_write, 0, 0); 499 500 /* 501 * Debugging only 502 */ 503 spin_lock(&vp->v_spin); 504 filename = TAILQ_FIRST(&vp->v_namecache) ? 505 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"; 506 spin_unlock(&vp->v_spin); 507 508 /* 509 * Make sure no buffers were instantiated while we were trying 510 * to clean out the remaining VM pages. This could occur due 511 * to busy dirty VM pages being flushed out to disk. 512 */ 513 do { 514 info.clean = 1; 515 count = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 516 vtruncbuf_bp_trunc_cmp, 517 vtruncbuf_bp_trunc, &info); 518 info.clean = 0; 519 count += RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 520 vtruncbuf_bp_trunc_cmp, 521 vtruncbuf_bp_trunc, &info); 522 if (count) { 523 kprintf("Warning: vtruncbuf(): Had to re-clean %d " 524 "left over buffers in %s\n", count, filename); 525 } 526 } while(count); 527 528 lwkt_reltoken(&vp->v_token); 529 530 return (0); 531 } 532 533 /* 534 * The callback buffer is beyond the new file EOF and must be destroyed. 535 * Note that the compare function must conform to the RB_SCAN's requirements. 536 */ 537 static 538 int 539 vtruncbuf_bp_trunc_cmp(struct buf *bp, void *data) 540 { 541 struct vtruncbuf_info *info = data; 542 543 if (bp->b_loffset >= info->truncloffset) 544 return(0); 545 return(-1); 546 } 547 548 static 549 int 550 vtruncbuf_bp_trunc(struct buf *bp, void *data) 551 { 552 struct vtruncbuf_info *info = data; 553 554 /* 555 * Do not try to use a buffer we cannot immediately lock, but sleep 556 * anyway to prevent a livelock. The code will loop until all buffers 557 * can be acted upon. 558 * 559 * We must always revalidate the buffer after locking it to deal 560 * with MP races. 561 */ 562 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 563 atomic_add_int(&bp->b_refs, 1); 564 if (BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL) == 0) 565 BUF_UNLOCK(bp); 566 atomic_subtract_int(&bp->b_refs, 1); 567 } else if ((info->clean && (bp->b_flags & B_DELWRI)) || 568 (info->clean == 0 && (bp->b_flags & B_DELWRI) == 0) || 569 bp->b_vp != info->vp || 570 vtruncbuf_bp_trunc_cmp(bp, data)) { 571 BUF_UNLOCK(bp); 572 } else { 573 bremfree(bp); 574 bp->b_flags |= (B_INVAL | B_RELBUF | B_NOCACHE); 575 brelse(bp); 576 } 577 return(1); 578 } 579 580 /* 581 * Fsync all meta-data after truncating a file to be non-zero. Only metadata 582 * blocks (with a negative loffset) are scanned. 583 * Note that the compare function must conform to the RB_SCAN's requirements. 584 */ 585 static int 586 vtruncbuf_bp_metasync_cmp(struct buf *bp, void *data __unused) 587 { 588 if (bp->b_loffset < 0) 589 return(0); 590 return(1); 591 } 592 593 static int 594 vtruncbuf_bp_metasync(struct buf *bp, void *data) 595 { 596 struct vtruncbuf_info *info = data; 597 598 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 599 atomic_add_int(&bp->b_refs, 1); 600 if (BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL) == 0) 601 BUF_UNLOCK(bp); 602 atomic_subtract_int(&bp->b_refs, 1); 603 } else if ((bp->b_flags & B_DELWRI) == 0 || 604 bp->b_vp != info->vp || 605 vtruncbuf_bp_metasync_cmp(bp, data)) { 606 BUF_UNLOCK(bp); 607 } else { 608 bremfree(bp); 609 if (bp->b_vp == info->vp) 610 bawrite(bp); 611 else 612 bwrite(bp); 613 } 614 return(1); 615 } 616 617 /* 618 * vfsync - implements a multipass fsync on a file which understands 619 * dependancies and meta-data. The passed vnode must be locked. The 620 * waitfor argument may be MNT_WAIT or MNT_NOWAIT, or MNT_LAZY. 621 * 622 * When fsyncing data asynchronously just do one consolidated pass starting 623 * with the most negative block number. This may not get all the data due 624 * to dependancies. 625 * 626 * When fsyncing data synchronously do a data pass, then a metadata pass, 627 * then do additional data+metadata passes to try to get all the data out. 628 * 629 * Caller must ref the vnode but does not have to lock it. 630 */ 631 static int vfsync_wait_output(struct vnode *vp, 632 int (*waitoutput)(struct vnode *, struct thread *)); 633 static int vfsync_dummy_cmp(struct buf *bp __unused, void *data __unused); 634 static int vfsync_data_only_cmp(struct buf *bp, void *data); 635 static int vfsync_meta_only_cmp(struct buf *bp, void *data); 636 static int vfsync_lazy_range_cmp(struct buf *bp, void *data); 637 static int vfsync_bp(struct buf *bp, void *data); 638 639 struct vfsync_info { 640 struct vnode *vp; 641 int fastpass; 642 int synchronous; 643 int syncdeps; 644 int lazycount; 645 int lazylimit; 646 int skippedbufs; 647 int (*checkdef)(struct buf *); 648 int (*cmpfunc)(struct buf *, void *); 649 }; 650 651 int 652 vfsync(struct vnode *vp, int waitfor, int passes, 653 int (*checkdef)(struct buf *), 654 int (*waitoutput)(struct vnode *, struct thread *)) 655 { 656 struct vfsync_info info; 657 int error; 658 659 bzero(&info, sizeof(info)); 660 info.vp = vp; 661 if ((info.checkdef = checkdef) == NULL) 662 info.syncdeps = 1; 663 664 lwkt_gettoken(&vp->v_token); 665 666 switch(waitfor) { 667 case MNT_LAZY | MNT_NOWAIT: 668 case MNT_LAZY: 669 /* 670 * Lazy (filesystem syncer typ) Asynchronous plus limit the 671 * number of data (not meta) pages we try to flush to 1MB. 672 * A non-zero return means that lazy limit was reached. 673 */ 674 info.lazylimit = 1024 * 1024; 675 info.syncdeps = 1; 676 info.cmpfunc = vfsync_lazy_range_cmp; 677 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 678 vfsync_lazy_range_cmp, vfsync_bp, &info); 679 info.cmpfunc = vfsync_meta_only_cmp; 680 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 681 vfsync_meta_only_cmp, vfsync_bp, &info); 682 if (error == 0) 683 vp->v_lazyw = 0; 684 else if (!RB_EMPTY(&vp->v_rbdirty_tree)) 685 vn_syncer_add(vp, 1); 686 error = 0; 687 break; 688 case MNT_NOWAIT: 689 /* 690 * Asynchronous. Do a data-only pass and a meta-only pass. 691 */ 692 info.syncdeps = 1; 693 info.cmpfunc = vfsync_data_only_cmp; 694 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, vfsync_data_only_cmp, 695 vfsync_bp, &info); 696 info.cmpfunc = vfsync_meta_only_cmp; 697 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, vfsync_meta_only_cmp, 698 vfsync_bp, &info); 699 error = 0; 700 break; 701 default: 702 /* 703 * Synchronous. Do a data-only pass, then a meta-data+data 704 * pass, then additional integrated passes to try to get 705 * all the dependancies flushed. 706 */ 707 info.cmpfunc = vfsync_data_only_cmp; 708 info.fastpass = 1; 709 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, vfsync_data_only_cmp, 710 vfsync_bp, &info); 711 info.fastpass = 0; 712 error = vfsync_wait_output(vp, waitoutput); 713 if (error == 0) { 714 info.skippedbufs = 0; 715 info.cmpfunc = vfsync_dummy_cmp; 716 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 717 vfsync_bp, &info); 718 error = vfsync_wait_output(vp, waitoutput); 719 if (info.skippedbufs) { 720 kprintf("Warning: vfsync skipped %d dirty " 721 "buf%s in pass2!\n", 722 info.skippedbufs, 723 ((info.skippedbufs > 1) ? "s" : "")); 724 } 725 } 726 while (error == 0 && passes > 0 && 727 !RB_EMPTY(&vp->v_rbdirty_tree) 728 ) { 729 info.skippedbufs = 0; 730 if (--passes == 0) { 731 info.synchronous = 1; 732 info.syncdeps = 1; 733 } 734 info.cmpfunc = vfsync_dummy_cmp; 735 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 736 vfsync_bp, &info); 737 if (error < 0) 738 error = -error; 739 info.syncdeps = 1; 740 if (error == 0) 741 error = vfsync_wait_output(vp, waitoutput); 742 if (info.skippedbufs && passes == 0) { 743 kprintf("Warning: vfsync skipped %d dirty " 744 "buf%s in final pass!\n", 745 info.skippedbufs, 746 ((info.skippedbufs > 1) ? "s" : "")); 747 } 748 } 749 #if 0 750 /* 751 * This case can occur normally because vnode lock might 752 * not be held. 753 */ 754 if (!RB_EMPTY(&vp->v_rbdirty_tree)) 755 kprintf("dirty bufs left after final pass\n"); 756 #endif 757 break; 758 } 759 lwkt_reltoken(&vp->v_token); 760 761 return(error); 762 } 763 764 static int 765 vfsync_wait_output(struct vnode *vp, 766 int (*waitoutput)(struct vnode *, struct thread *)) 767 { 768 int error; 769 770 error = bio_track_wait(&vp->v_track_write, 0, 0); 771 if (waitoutput) 772 error = waitoutput(vp, curthread); 773 return(error); 774 } 775 776 static int 777 vfsync_dummy_cmp(struct buf *bp __unused, void *data __unused) 778 { 779 return(0); 780 } 781 782 static int 783 vfsync_data_only_cmp(struct buf *bp, void *data) 784 { 785 if (bp->b_loffset < 0) 786 return(-1); 787 return(0); 788 } 789 790 static int 791 vfsync_meta_only_cmp(struct buf *bp, void *data) 792 { 793 if (bp->b_loffset < 0) 794 return(0); 795 return(1); 796 } 797 798 static int 799 vfsync_lazy_range_cmp(struct buf *bp, void *data) 800 { 801 struct vfsync_info *info = data; 802 803 if (bp->b_loffset < info->vp->v_lazyw) 804 return(-1); 805 return(0); 806 } 807 808 static int 809 vfsync_bp(struct buf *bp, void *data) 810 { 811 struct vfsync_info *info = data; 812 struct vnode *vp = info->vp; 813 int error; 814 815 if (info->fastpass) { 816 /* 817 * Ignore buffers that we cannot immediately lock. 818 */ 819 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 820 /* 821 * Removed BUF_TIMELOCK(..., 1), even a 1-tick 822 * delay can mess up performance 823 * 824 * Another reason is that during a dirty-buffer 825 * scan a clustered write can start I/O on buffers 826 * ahead of the scan, causing the scan to not 827 * get a lock here. Usually this means the write 828 * is already in progress so, in fact, we *want* 829 * to skip the buffer. 830 */ 831 ++info->skippedbufs; 832 return(0); 833 } 834 } else if (info->synchronous == 0) { 835 /* 836 * Normal pass, give the buffer a little time to become 837 * available to us. 838 */ 839 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE, "bflst2", hz / 10)) { 840 ++info->skippedbufs; 841 return(0); 842 } 843 } else { 844 /* 845 * Synchronous pass, give the buffer a lot of time before 846 * giving up. 847 */ 848 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE, "bflst3", hz * 10)) { 849 ++info->skippedbufs; 850 return(0); 851 } 852 } 853 854 /* 855 * We must revalidate the buffer after locking. 856 */ 857 if ((bp->b_flags & B_DELWRI) == 0 || 858 bp->b_vp != info->vp || 859 info->cmpfunc(bp, data)) { 860 BUF_UNLOCK(bp); 861 return(0); 862 } 863 864 /* 865 * If syncdeps is not set we do not try to write buffers which have 866 * dependancies. 867 */ 868 if (!info->synchronous && info->syncdeps == 0 && info->checkdef(bp)) { 869 BUF_UNLOCK(bp); 870 return(0); 871 } 872 873 /* 874 * B_NEEDCOMMIT (primarily used by NFS) is a state where the buffer 875 * has been written but an additional handshake with the device 876 * is required before we can dispose of the buffer. We have no idea 877 * how to do this so we have to skip these buffers. 878 */ 879 if (bp->b_flags & B_NEEDCOMMIT) { 880 BUF_UNLOCK(bp); 881 return(0); 882 } 883 884 /* 885 * Ask bioops if it is ok to sync. If not the VFS may have 886 * set B_LOCKED so we have to cycle the buffer. 887 */ 888 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 889 bremfree(bp); 890 brelse(bp); 891 return(0); 892 } 893 894 if (info->synchronous) { 895 /* 896 * Synchronous flush. An error may be returned and will 897 * stop the scan. 898 */ 899 bremfree(bp); 900 error = bwrite(bp); 901 } else { 902 /* 903 * Asynchronous flush. We use the error return to support 904 * MNT_LAZY flushes. 905 * 906 * In low-memory situations we revert to synchronous 907 * operation. This should theoretically prevent the I/O 908 * path from exhausting memory in a non-recoverable way. 909 */ 910 vp->v_lazyw = bp->b_loffset; 911 bremfree(bp); 912 if (vm_page_count_min(0)) { 913 /* low memory */ 914 info->lazycount += bp->b_bufsize; 915 bwrite(bp); 916 } else { 917 /* normal */ 918 info->lazycount += cluster_awrite(bp); 919 waitrunningbufspace(); 920 /*vm_wait_nominal();*/ 921 } 922 if (info->lazylimit && info->lazycount >= info->lazylimit) 923 error = 1; 924 else 925 error = 0; 926 } 927 return(-error); 928 } 929 930 /* 931 * Associate a buffer with a vnode. 932 * 933 * MPSAFE 934 */ 935 int 936 bgetvp(struct vnode *vp, struct buf *bp, int testsize) 937 { 938 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 939 KKASSERT((bp->b_flags & (B_HASHED|B_DELWRI|B_VNCLEAN|B_VNDIRTY)) == 0); 940 941 /* 942 * Insert onto list for new vnode. 943 */ 944 lwkt_gettoken(&vp->v_token); 945 946 if (buf_rb_hash_RB_INSERT(&vp->v_rbhash_tree, bp)) { 947 lwkt_reltoken(&vp->v_token); 948 return (EEXIST); 949 } 950 951 /* 952 * Diagnostics (mainly for HAMMER debugging). Check for 953 * overlapping buffers. 954 */ 955 if (check_buf_overlap) { 956 struct buf *bx; 957 bx = buf_rb_hash_RB_PREV(bp); 958 if (bx) { 959 if (bx->b_loffset + bx->b_bufsize > bp->b_loffset) { 960 kprintf("bgetvp: overlapl %016jx/%d %016jx " 961 "bx %p bp %p\n", 962 (intmax_t)bx->b_loffset, 963 bx->b_bufsize, 964 (intmax_t)bp->b_loffset, 965 bx, bp); 966 if (check_buf_overlap > 1) 967 panic("bgetvp - overlapping buffer"); 968 } 969 } 970 bx = buf_rb_hash_RB_NEXT(bp); 971 if (bx) { 972 if (bp->b_loffset + testsize > bx->b_loffset) { 973 kprintf("bgetvp: overlapr %016jx/%d %016jx " 974 "bp %p bx %p\n", 975 (intmax_t)bp->b_loffset, 976 testsize, 977 (intmax_t)bx->b_loffset, 978 bp, bx); 979 if (check_buf_overlap > 1) 980 panic("bgetvp - overlapping buffer"); 981 } 982 } 983 } 984 bp->b_vp = vp; 985 bp->b_flags |= B_HASHED; 986 bp->b_flags |= B_VNCLEAN; 987 if (buf_rb_tree_RB_INSERT(&vp->v_rbclean_tree, bp)) 988 panic("reassignbuf: dup lblk/clean vp %p bp %p", vp, bp); 989 /*vhold(vp);*/ 990 lwkt_reltoken(&vp->v_token); 991 return(0); 992 } 993 994 /* 995 * Disassociate a buffer from a vnode. 996 * 997 * MPSAFE 998 */ 999 void 1000 brelvp(struct buf *bp) 1001 { 1002 struct vnode *vp; 1003 1004 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1005 1006 /* 1007 * Delete from old vnode list, if on one. 1008 */ 1009 vp = bp->b_vp; 1010 lwkt_gettoken(&vp->v_token); 1011 if (bp->b_flags & (B_VNDIRTY | B_VNCLEAN)) { 1012 if (bp->b_flags & B_VNDIRTY) 1013 buf_rb_tree_RB_REMOVE(&vp->v_rbdirty_tree, bp); 1014 else 1015 buf_rb_tree_RB_REMOVE(&vp->v_rbclean_tree, bp); 1016 bp->b_flags &= ~(B_VNDIRTY | B_VNCLEAN); 1017 } 1018 if (bp->b_flags & B_HASHED) { 1019 buf_rb_hash_RB_REMOVE(&vp->v_rbhash_tree, bp); 1020 bp->b_flags &= ~B_HASHED; 1021 } 1022 1023 /* 1024 * Only remove from synclist when no dirty buffers are left AND 1025 * the VFS has not flagged the vnode's inode as being dirty. 1026 */ 1027 if ((vp->v_flag & (VONWORKLST | VISDIRTY | VOBJDIRTY)) == VONWORKLST && 1028 RB_EMPTY(&vp->v_rbdirty_tree)) { 1029 vn_syncer_remove(vp, 0); 1030 } 1031 bp->b_vp = NULL; 1032 1033 lwkt_reltoken(&vp->v_token); 1034 1035 /*vdrop(vp);*/ 1036 } 1037 1038 /* 1039 * Reassign the buffer to the proper clean/dirty list based on B_DELWRI. 1040 * This routine is called when the state of the B_DELWRI bit is changed. 1041 * 1042 * Must be called with vp->v_token held. 1043 * MPSAFE 1044 */ 1045 void 1046 reassignbuf(struct buf *bp) 1047 { 1048 struct vnode *vp = bp->b_vp; 1049 int delay; 1050 1051 ASSERT_LWKT_TOKEN_HELD(&vp->v_token); 1052 ++reassignbufcalls; 1053 1054 /* 1055 * B_PAGING flagged buffers cannot be reassigned because their vp 1056 * is not fully linked in. 1057 */ 1058 if (bp->b_flags & B_PAGING) 1059 panic("cannot reassign paging buffer"); 1060 1061 if (bp->b_flags & B_DELWRI) { 1062 /* 1063 * Move to the dirty list, add the vnode to the worklist 1064 */ 1065 if (bp->b_flags & B_VNCLEAN) { 1066 buf_rb_tree_RB_REMOVE(&vp->v_rbclean_tree, bp); 1067 bp->b_flags &= ~B_VNCLEAN; 1068 } 1069 if ((bp->b_flags & B_VNDIRTY) == 0) { 1070 if (buf_rb_tree_RB_INSERT(&vp->v_rbdirty_tree, bp)) { 1071 panic("reassignbuf: dup lblk vp %p bp %p", 1072 vp, bp); 1073 } 1074 bp->b_flags |= B_VNDIRTY; 1075 } 1076 if ((vp->v_flag & VONWORKLST) == 0) { 1077 switch (vp->v_type) { 1078 case VDIR: 1079 delay = dirdelay; 1080 break; 1081 case VCHR: 1082 case VBLK: 1083 if (vp->v_rdev && 1084 vp->v_rdev->si_mountpoint != NULL) { 1085 delay = metadelay; 1086 break; 1087 } 1088 /* fall through */ 1089 default: 1090 delay = filedelay; 1091 } 1092 vn_syncer_add(vp, delay); 1093 } 1094 } else { 1095 /* 1096 * Move to the clean list, remove the vnode from the worklist 1097 * if no dirty blocks remain. 1098 */ 1099 if (bp->b_flags & B_VNDIRTY) { 1100 buf_rb_tree_RB_REMOVE(&vp->v_rbdirty_tree, bp); 1101 bp->b_flags &= ~B_VNDIRTY; 1102 } 1103 if ((bp->b_flags & B_VNCLEAN) == 0) { 1104 if (buf_rb_tree_RB_INSERT(&vp->v_rbclean_tree, bp)) { 1105 panic("reassignbuf: dup lblk vp %p bp %p", 1106 vp, bp); 1107 } 1108 bp->b_flags |= B_VNCLEAN; 1109 } 1110 1111 /* 1112 * Only remove from synclist when no dirty buffers are left 1113 * AND the VFS has not flagged the vnode's inode as being 1114 * dirty. 1115 */ 1116 if ((vp->v_flag & (VONWORKLST | VISDIRTY | VOBJDIRTY)) == 1117 VONWORKLST && 1118 RB_EMPTY(&vp->v_rbdirty_tree)) { 1119 vn_syncer_remove(vp, 0); 1120 } 1121 } 1122 } 1123 1124 /* 1125 * Create a vnode for a block device. Used for mounting the root file 1126 * system. 1127 * 1128 * A vref()'d vnode is returned. 1129 */ 1130 extern struct vop_ops *devfs_vnode_dev_vops_p; 1131 int 1132 bdevvp(cdev_t dev, struct vnode **vpp) 1133 { 1134 struct vnode *vp; 1135 struct vnode *nvp; 1136 int error; 1137 1138 if (dev == NULL) { 1139 *vpp = NULLVP; 1140 return (ENXIO); 1141 } 1142 error = getspecialvnode(VT_NON, NULL, &devfs_vnode_dev_vops_p, 1143 &nvp, 0, 0); 1144 if (error) { 1145 *vpp = NULLVP; 1146 return (error); 1147 } 1148 vp = nvp; 1149 vp->v_type = VCHR; 1150 #if 0 1151 vp->v_rdev = dev; 1152 #endif 1153 v_associate_rdev(vp, dev); 1154 vp->v_umajor = dev->si_umajor; 1155 vp->v_uminor = dev->si_uminor; 1156 vx_unlock(vp); 1157 *vpp = vp; 1158 return (0); 1159 } 1160 1161 int 1162 v_associate_rdev(struct vnode *vp, cdev_t dev) 1163 { 1164 if (dev == NULL) 1165 return(ENXIO); 1166 if (dev_is_good(dev) == 0) 1167 return(ENXIO); 1168 KKASSERT(vp->v_rdev == NULL); 1169 vp->v_rdev = reference_dev(dev); 1170 lwkt_gettoken(&spechash_token); 1171 SLIST_INSERT_HEAD(&dev->si_hlist, vp, v_cdevnext); 1172 lwkt_reltoken(&spechash_token); 1173 return(0); 1174 } 1175 1176 void 1177 v_release_rdev(struct vnode *vp) 1178 { 1179 cdev_t dev; 1180 1181 if ((dev = vp->v_rdev) != NULL) { 1182 lwkt_gettoken(&spechash_token); 1183 SLIST_REMOVE(&dev->si_hlist, vp, vnode, v_cdevnext); 1184 vp->v_rdev = NULL; 1185 release_dev(dev); 1186 lwkt_reltoken(&spechash_token); 1187 } 1188 } 1189 1190 /* 1191 * Add a vnode to the alias list hung off the cdev_t. We only associate 1192 * the device number with the vnode. The actual device is not associated 1193 * until the vnode is opened (usually in spec_open()), and will be 1194 * disassociated on last close. 1195 */ 1196 void 1197 addaliasu(struct vnode *nvp, int x, int y) 1198 { 1199 if (nvp->v_type != VBLK && nvp->v_type != VCHR) 1200 panic("addaliasu on non-special vnode"); 1201 nvp->v_umajor = x; 1202 nvp->v_uminor = y; 1203 } 1204 1205 /* 1206 * Simple call that a filesystem can make to try to get rid of a 1207 * vnode. It will fail if anyone is referencing the vnode (including 1208 * the caller). 1209 * 1210 * The filesystem can check whether its in-memory inode structure still 1211 * references the vp on return. 1212 * 1213 * May only be called if the vnode is in a known state (i.e. being prevented 1214 * from being deallocated by some other condition such as a vfs inode hold). 1215 */ 1216 void 1217 vclean_unlocked(struct vnode *vp) 1218 { 1219 vx_get(vp); 1220 if (VREFCNT(vp) <= 1) 1221 vgone_vxlocked(vp); 1222 vx_put(vp); 1223 } 1224 1225 /* 1226 * Disassociate a vnode from its underlying filesystem. 1227 * 1228 * The vnode must be VX locked and referenced. In all normal situations 1229 * there are no active references. If vclean_vxlocked() is called while 1230 * there are active references, the vnode is being ripped out and we have 1231 * to call VOP_CLOSE() as appropriate before we can reclaim it. 1232 */ 1233 void 1234 vclean_vxlocked(struct vnode *vp, int flags) 1235 { 1236 int active; 1237 int n; 1238 vm_object_t object; 1239 struct namecache *ncp; 1240 1241 /* 1242 * If the vnode has already been reclaimed we have nothing to do. 1243 */ 1244 if (vp->v_flag & VRECLAIMED) 1245 return; 1246 1247 /* 1248 * Set flag to interlock operation, flag finalization to ensure 1249 * that the vnode winds up on the inactive list, and set v_act to 0. 1250 */ 1251 vsetflags(vp, VRECLAIMED); 1252 atomic_set_int(&vp->v_refcnt, VREF_FINALIZE); 1253 vp->v_act = 0; 1254 1255 if (verbose_reclaims) { 1256 if ((ncp = TAILQ_FIRST(&vp->v_namecache)) != NULL) 1257 kprintf("Debug: reclaim %p %s\n", vp, ncp->nc_name); 1258 } 1259 1260 /* 1261 * Scrap the vfs cache 1262 */ 1263 while (cache_inval_vp(vp, 0) != 0) { 1264 kprintf("Warning: vnode %p clean/cache_resolution " 1265 "race detected\n", vp); 1266 tsleep(vp, 0, "vclninv", 2); 1267 } 1268 1269 /* 1270 * Check to see if the vnode is in use. If so we have to reference it 1271 * before we clean it out so that its count cannot fall to zero and 1272 * generate a race against ourselves to recycle it. 1273 */ 1274 active = (VREFCNT(vp) > 0); 1275 1276 /* 1277 * Clean out any buffers associated with the vnode and destroy its 1278 * object, if it has one. 1279 */ 1280 vinvalbuf(vp, V_SAVE, 0, 0); 1281 1282 /* 1283 * If purging an active vnode (typically during a forced unmount 1284 * or reboot), it must be closed and deactivated before being 1285 * reclaimed. This isn't really all that safe, but what can 1286 * we do? XXX. 1287 * 1288 * Note that neither of these routines unlocks the vnode. 1289 */ 1290 if (active && (flags & DOCLOSE)) { 1291 while ((n = vp->v_opencount) != 0) { 1292 if (vp->v_writecount) 1293 VOP_CLOSE(vp, FWRITE|FNONBLOCK, NULL); 1294 else 1295 VOP_CLOSE(vp, FNONBLOCK, NULL); 1296 if (vp->v_opencount == n) { 1297 kprintf("Warning: unable to force-close" 1298 " vnode %p\n", vp); 1299 break; 1300 } 1301 } 1302 } 1303 1304 /* 1305 * If the vnode has not been deactivated, deactivated it. Deactivation 1306 * can create new buffers and VM pages so we have to call vinvalbuf() 1307 * again to make sure they all get flushed. 1308 * 1309 * This can occur if a file with a link count of 0 needs to be 1310 * truncated. 1311 * 1312 * If the vnode is already dead don't try to deactivate it. 1313 */ 1314 if ((vp->v_flag & VINACTIVE) == 0) { 1315 vsetflags(vp, VINACTIVE); 1316 if (vp->v_mount) 1317 VOP_INACTIVE(vp); 1318 vinvalbuf(vp, V_SAVE, 0, 0); 1319 } 1320 1321 /* 1322 * If the vnode has an object, destroy it. 1323 */ 1324 while ((object = vp->v_object) != NULL) { 1325 vm_object_hold(object); 1326 if (object == vp->v_object) 1327 break; 1328 vm_object_drop(object); 1329 } 1330 1331 if (object != NULL) { 1332 if (object->ref_count == 0) { 1333 if ((object->flags & OBJ_DEAD) == 0) 1334 vm_object_terminate(object); 1335 vm_object_drop(object); 1336 vclrflags(vp, VOBJBUF); 1337 } else { 1338 vm_pager_deallocate(object); 1339 vclrflags(vp, VOBJBUF); 1340 vm_object_drop(object); 1341 } 1342 } 1343 KKASSERT((vp->v_flag & VOBJBUF) == 0); 1344 1345 if (vp->v_flag & VOBJDIRTY) 1346 vclrobjdirty(vp); 1347 1348 /* 1349 * Reclaim the vnode if not already dead. 1350 */ 1351 if (vp->v_mount && VOP_RECLAIM(vp)) 1352 panic("vclean: cannot reclaim"); 1353 1354 /* 1355 * Done with purge, notify sleepers of the grim news. 1356 */ 1357 vp->v_ops = &dead_vnode_vops_p; 1358 vn_gone(vp); 1359 vp->v_tag = VT_NON; 1360 1361 /* 1362 * If we are destroying an active vnode, reactivate it now that 1363 * we have reassociated it with deadfs. This prevents the system 1364 * from crashing on the vnode due to it being unexpectedly marked 1365 * as inactive or reclaimed. 1366 */ 1367 if (active && (flags & DOCLOSE)) { 1368 vclrflags(vp, VINACTIVE | VRECLAIMED); 1369 } 1370 } 1371 1372 /* 1373 * Eliminate all activity associated with the requested vnode 1374 * and with all vnodes aliased to the requested vnode. 1375 * 1376 * The vnode must be referenced but should not be locked. 1377 */ 1378 int 1379 vrevoke(struct vnode *vp, struct ucred *cred) 1380 { 1381 struct vnode *vq; 1382 struct vnode *vqn; 1383 cdev_t dev; 1384 int error; 1385 1386 /* 1387 * If the vnode has a device association, scrap all vnodes associated 1388 * with the device. Don't let the device disappear on us while we 1389 * are scrapping the vnodes. 1390 * 1391 * The passed vp will probably show up in the list, do not VX lock 1392 * it twice! 1393 * 1394 * Releasing the vnode's rdev here can mess up specfs's call to 1395 * device close, so don't do it. The vnode has been disassociated 1396 * and the device will be closed after the last ref on the related 1397 * fp goes away (if not still open by e.g. the kernel). 1398 */ 1399 if (vp->v_type != VCHR) { 1400 error = fdrevoke(vp, DTYPE_VNODE, cred); 1401 return (error); 1402 } 1403 if ((dev = vp->v_rdev) == NULL) { 1404 return(0); 1405 } 1406 reference_dev(dev); 1407 lwkt_gettoken(&spechash_token); 1408 1409 restart: 1410 vqn = SLIST_FIRST(&dev->si_hlist); 1411 if (vqn) 1412 vhold(vqn); 1413 while ((vq = vqn) != NULL) { 1414 if (VREFCNT(vq) > 0) { 1415 vref(vq); 1416 fdrevoke(vq, DTYPE_VNODE, cred); 1417 /*v_release_rdev(vq);*/ 1418 vrele(vq); 1419 if (vq->v_rdev != dev) { 1420 vdrop(vq); 1421 goto restart; 1422 } 1423 } 1424 vqn = SLIST_NEXT(vq, v_cdevnext); 1425 if (vqn) 1426 vhold(vqn); 1427 vdrop(vq); 1428 } 1429 lwkt_reltoken(&spechash_token); 1430 dev_drevoke(dev); 1431 release_dev(dev); 1432 return (0); 1433 } 1434 1435 /* 1436 * This is called when the object underlying a vnode is being destroyed, 1437 * such as in a remove(). Try to recycle the vnode immediately if the 1438 * only active reference is our reference. 1439 * 1440 * Directory vnodes in the namecache with children cannot be immediately 1441 * recycled because numerous VOP_N*() ops require them to be stable. 1442 * 1443 * To avoid recursive recycling from VOP_INACTIVE implemenetations this 1444 * function is a NOP if VRECLAIMED is already set. 1445 */ 1446 int 1447 vrecycle(struct vnode *vp) 1448 { 1449 if (VREFCNT(vp) <= 1 && (vp->v_flag & VRECLAIMED) == 0) { 1450 if (cache_inval_vp_nonblock(vp)) 1451 return(0); 1452 vgone_vxlocked(vp); 1453 return (1); 1454 } 1455 return (0); 1456 } 1457 1458 /* 1459 * Return the maximum I/O size allowed for strategy calls on VP. 1460 * 1461 * If vp is VCHR or VBLK we dive the device, otherwise we use 1462 * the vp's mount info. 1463 * 1464 * The returned value is clamped at MAXPHYS as most callers cannot use 1465 * buffers larger than that size. 1466 */ 1467 int 1468 vmaxiosize(struct vnode *vp) 1469 { 1470 int maxiosize; 1471 1472 if (vp->v_type == VBLK || vp->v_type == VCHR) 1473 maxiosize = vp->v_rdev->si_iosize_max; 1474 else 1475 maxiosize = vp->v_mount->mnt_iosize_max; 1476 1477 if (maxiosize > MAXPHYS) 1478 maxiosize = MAXPHYS; 1479 return (maxiosize); 1480 } 1481 1482 /* 1483 * Eliminate all activity associated with a vnode in preparation for 1484 * destruction. 1485 * 1486 * The vnode must be VX locked and refd and will remain VX locked and refd 1487 * on return. This routine may be called with the vnode in any state, as 1488 * long as it is VX locked. The vnode will be cleaned out and marked 1489 * VRECLAIMED but will not actually be reused until all existing refs and 1490 * holds go away. 1491 * 1492 * NOTE: This routine may be called on a vnode which has not yet been 1493 * already been deactivated (VOP_INACTIVE), or on a vnode which has 1494 * already been reclaimed. 1495 * 1496 * This routine is not responsible for placing us back on the freelist. 1497 * Instead, it happens automatically when the caller releases the VX lock 1498 * (assuming there aren't any other references). 1499 */ 1500 void 1501 vgone_vxlocked(struct vnode *vp) 1502 { 1503 /* 1504 * assert that the VX lock is held. This is an absolute requirement 1505 * now for vgone_vxlocked() to be called. 1506 */ 1507 KKASSERT(lockinuse(&vp->v_lock)); 1508 1509 /* 1510 * Clean out the filesystem specific data and set the VRECLAIMED 1511 * bit. Also deactivate the vnode if necessary. 1512 * 1513 * The vnode should have automatically been removed from the syncer 1514 * list as syncer/dirty flags cleared during the cleaning. 1515 */ 1516 vclean_vxlocked(vp, DOCLOSE); 1517 1518 /* 1519 * Normally panic if the vnode is still dirty, unless we are doing 1520 * a forced unmount (tmpfs typically). 1521 */ 1522 if (vp->v_flag & VONWORKLST) { 1523 if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) { 1524 /* force removal */ 1525 vn_syncer_remove(vp, 1); 1526 } else { 1527 panic("vp %p still dirty in vgone after flush", vp); 1528 } 1529 } 1530 1531 /* 1532 * Delete from old mount point vnode list, if on one. 1533 */ 1534 if (vp->v_mount != NULL) { 1535 KKASSERT(vp->v_data == NULL); 1536 insmntque(vp, NULL); 1537 } 1538 1539 /* 1540 * If special device, remove it from special device alias list 1541 * if it is on one. This should normally only occur if a vnode is 1542 * being revoked as the device should otherwise have been released 1543 * naturally. 1544 */ 1545 if ((vp->v_type == VBLK || vp->v_type == VCHR) && vp->v_rdev != NULL) { 1546 v_release_rdev(vp); 1547 } 1548 1549 /* 1550 * Set us to VBAD 1551 */ 1552 vp->v_type = VBAD; 1553 } 1554 1555 /* 1556 * Lookup a vnode by device number. 1557 * 1558 * Returns non-zero and *vpp set to a vref'd vnode on success. 1559 * Returns zero on failure. 1560 */ 1561 int 1562 vfinddev(cdev_t dev, enum vtype type, struct vnode **vpp) 1563 { 1564 struct vnode *vp; 1565 1566 lwkt_gettoken(&spechash_token); 1567 SLIST_FOREACH(vp, &dev->si_hlist, v_cdevnext) { 1568 if (type == vp->v_type) { 1569 *vpp = vp; 1570 vref(vp); 1571 lwkt_reltoken(&spechash_token); 1572 return (1); 1573 } 1574 } 1575 lwkt_reltoken(&spechash_token); 1576 return (0); 1577 } 1578 1579 /* 1580 * Calculate the total number of references to a special device. This 1581 * routine may only be called for VBLK and VCHR vnodes since v_rdev is 1582 * an overloaded field. Since udev2dev can now return NULL, we have 1583 * to check for a NULL v_rdev. 1584 */ 1585 int 1586 count_dev(cdev_t dev) 1587 { 1588 struct vnode *vp; 1589 int count = 0; 1590 1591 if (SLIST_FIRST(&dev->si_hlist)) { 1592 lwkt_gettoken(&spechash_token); 1593 SLIST_FOREACH(vp, &dev->si_hlist, v_cdevnext) { 1594 count += vp->v_opencount; 1595 } 1596 lwkt_reltoken(&spechash_token); 1597 } 1598 return(count); 1599 } 1600 1601 int 1602 vcount(struct vnode *vp) 1603 { 1604 if (vp->v_rdev == NULL) 1605 return(0); 1606 return(count_dev(vp->v_rdev)); 1607 } 1608 1609 /* 1610 * Initialize VMIO for a vnode. This routine MUST be called before a 1611 * VFS can issue buffer cache ops on a vnode. It is typically called 1612 * when a vnode is initialized from its inode. 1613 */ 1614 int 1615 vinitvmio(struct vnode *vp, off_t filesize, int blksize, int boff) 1616 { 1617 vm_object_t object; 1618 int error = 0; 1619 1620 object = vp->v_object; 1621 if (object) { 1622 vm_object_hold(object); 1623 KKASSERT(vp->v_object == object); 1624 } 1625 1626 if (object == NULL) { 1627 object = vnode_pager_alloc(vp, filesize, 0, 0, blksize, boff); 1628 1629 /* 1630 * Dereference the reference we just created. This assumes 1631 * that the object is associated with the vp. Allow it to 1632 * have zero refs. It cannot be destroyed as long as it 1633 * is associated with the vnode. 1634 */ 1635 vm_object_hold(object); 1636 atomic_add_int(&object->ref_count, -1); 1637 vrele(vp); 1638 } else { 1639 KKASSERT((object->flags & OBJ_DEAD) == 0); 1640 } 1641 KASSERT(vp->v_object != NULL, ("vinitvmio: NULL object")); 1642 vsetflags(vp, VOBJBUF); 1643 vm_object_drop(object); 1644 1645 return (error); 1646 } 1647 1648 1649 /* 1650 * Print out a description of a vnode. 1651 */ 1652 static char *typename[] = 1653 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 1654 1655 void 1656 vprint(char *label, struct vnode *vp) 1657 { 1658 char buf[96]; 1659 1660 if (label != NULL) 1661 kprintf("%s: %p: ", label, (void *)vp); 1662 else 1663 kprintf("%p: ", (void *)vp); 1664 kprintf("type %s, refcnt %08x, writecount %d, holdcnt %d,", 1665 typename[vp->v_type], 1666 vp->v_refcnt, vp->v_writecount, vp->v_auxrefs); 1667 buf[0] = '\0'; 1668 if (vp->v_flag & VROOT) 1669 strcat(buf, "|VROOT"); 1670 if (vp->v_flag & VPFSROOT) 1671 strcat(buf, "|VPFSROOT"); 1672 if (vp->v_flag & VTEXT) 1673 strcat(buf, "|VTEXT"); 1674 if (vp->v_flag & VSYSTEM) 1675 strcat(buf, "|VSYSTEM"); 1676 if (vp->v_flag & VOBJBUF) 1677 strcat(buf, "|VOBJBUF"); 1678 if (buf[0] != '\0') 1679 kprintf(" flags (%s)", &buf[1]); 1680 if (vp->v_data == NULL) { 1681 kprintf("\n"); 1682 } else { 1683 kprintf("\n\t"); 1684 VOP_PRINT(vp); 1685 } 1686 } 1687 1688 /* 1689 * Do the usual access checking. 1690 * file_mode, uid and gid are from the vnode in question, 1691 * while acc_mode and cred are from the VOP_ACCESS parameter list 1692 */ 1693 int 1694 vaccess(enum vtype type, mode_t file_mode, uid_t uid, gid_t gid, 1695 mode_t acc_mode, struct ucred *cred) 1696 { 1697 mode_t mask; 1698 int ismember; 1699 1700 /* 1701 * Super-user always gets read/write access, but execute access depends 1702 * on at least one execute bit being set. 1703 */ 1704 if (priv_check_cred(cred, PRIV_ROOT, 0) == 0) { 1705 if ((acc_mode & VEXEC) && type != VDIR && 1706 (file_mode & (S_IXUSR|S_IXGRP|S_IXOTH)) == 0) 1707 return (EACCES); 1708 return (0); 1709 } 1710 1711 mask = 0; 1712 1713 /* Otherwise, check the owner. */ 1714 if (cred->cr_uid == uid) { 1715 if (acc_mode & VEXEC) 1716 mask |= S_IXUSR; 1717 if (acc_mode & VREAD) 1718 mask |= S_IRUSR; 1719 if (acc_mode & VWRITE) 1720 mask |= S_IWUSR; 1721 return ((file_mode & mask) == mask ? 0 : EACCES); 1722 } 1723 1724 /* Otherwise, check the groups. */ 1725 ismember = groupmember(gid, cred); 1726 if (cred->cr_svgid == gid || ismember) { 1727 if (acc_mode & VEXEC) 1728 mask |= S_IXGRP; 1729 if (acc_mode & VREAD) 1730 mask |= S_IRGRP; 1731 if (acc_mode & VWRITE) 1732 mask |= S_IWGRP; 1733 return ((file_mode & mask) == mask ? 0 : EACCES); 1734 } 1735 1736 /* Otherwise, check everyone else. */ 1737 if (acc_mode & VEXEC) 1738 mask |= S_IXOTH; 1739 if (acc_mode & VREAD) 1740 mask |= S_IROTH; 1741 if (acc_mode & VWRITE) 1742 mask |= S_IWOTH; 1743 return ((file_mode & mask) == mask ? 0 : EACCES); 1744 } 1745 1746 #ifdef DDB 1747 #include <ddb/ddb.h> 1748 1749 static int db_show_locked_vnodes(struct mount *mp, void *data); 1750 1751 /* 1752 * List all of the locked vnodes in the system. 1753 * Called when debugging the kernel. 1754 */ 1755 DB_SHOW_COMMAND(lockedvnodes, lockedvnodes) 1756 { 1757 kprintf("Locked vnodes\n"); 1758 mountlist_scan(db_show_locked_vnodes, NULL, 1759 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 1760 } 1761 1762 static int 1763 db_show_locked_vnodes(struct mount *mp, void *data __unused) 1764 { 1765 struct vnode *vp; 1766 1767 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 1768 if (vn_islocked(vp)) 1769 vprint(NULL, vp); 1770 } 1771 return(0); 1772 } 1773 #endif 1774 1775 /* 1776 * Top level filesystem related information gathering. 1777 */ 1778 static int sysctl_ovfs_conf (SYSCTL_HANDLER_ARGS); 1779 1780 static int 1781 vfs_sysctl(SYSCTL_HANDLER_ARGS) 1782 { 1783 int *name = (int *)arg1 - 1; /* XXX */ 1784 u_int namelen = arg2 + 1; /* XXX */ 1785 struct vfsconf *vfsp; 1786 int maxtypenum; 1787 1788 #if 1 || defined(COMPAT_PRELITE2) 1789 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 1790 if (namelen == 1) 1791 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 1792 #endif 1793 1794 #ifdef notyet 1795 /* all sysctl names at this level are at least name and field */ 1796 if (namelen < 2) 1797 return (ENOTDIR); /* overloaded */ 1798 if (name[0] != VFS_GENERIC) { 1799 vfsp = vfsconf_find_by_typenum(name[0]); 1800 if (vfsp == NULL) 1801 return (EOPNOTSUPP); 1802 return ((*vfsp->vfc_vfsops->vfs_sysctl)(&name[1], namelen - 1, 1803 oldp, oldlenp, newp, newlen, p)); 1804 } 1805 #endif 1806 switch (name[1]) { 1807 case VFS_MAXTYPENUM: 1808 if (namelen != 2) 1809 return (ENOTDIR); 1810 maxtypenum = vfsconf_get_maxtypenum(); 1811 return (SYSCTL_OUT(req, &maxtypenum, sizeof(maxtypenum))); 1812 case VFS_CONF: 1813 if (namelen != 3) 1814 return (ENOTDIR); /* overloaded */ 1815 vfsp = vfsconf_find_by_typenum(name[2]); 1816 if (vfsp == NULL) 1817 return (EOPNOTSUPP); 1818 return (SYSCTL_OUT(req, vfsp, sizeof *vfsp)); 1819 } 1820 return (EOPNOTSUPP); 1821 } 1822 1823 SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD, vfs_sysctl, 1824 "Generic filesystem"); 1825 1826 #if 1 || defined(COMPAT_PRELITE2) 1827 1828 static int 1829 sysctl_ovfs_conf_iter(struct vfsconf *vfsp, void *data) 1830 { 1831 int error; 1832 struct ovfsconf ovfs; 1833 struct sysctl_req *req = (struct sysctl_req*) data; 1834 1835 bzero(&ovfs, sizeof(ovfs)); 1836 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 1837 strcpy(ovfs.vfc_name, vfsp->vfc_name); 1838 ovfs.vfc_index = vfsp->vfc_typenum; 1839 ovfs.vfc_refcount = vfsp->vfc_refcount; 1840 ovfs.vfc_flags = vfsp->vfc_flags; 1841 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 1842 if (error) 1843 return error; /* abort iteration with error code */ 1844 else 1845 return 0; /* continue iterating with next element */ 1846 } 1847 1848 static int 1849 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 1850 { 1851 return vfsconf_each(sysctl_ovfs_conf_iter, (void*)req); 1852 } 1853 1854 #endif /* 1 || COMPAT_PRELITE2 */ 1855 1856 /* 1857 * Check to see if a filesystem is mounted on a block device. 1858 */ 1859 int 1860 vfs_mountedon(struct vnode *vp) 1861 { 1862 cdev_t dev; 1863 1864 if ((dev = vp->v_rdev) == NULL) { 1865 /* if (vp->v_type != VBLK) 1866 dev = get_dev(vp->v_uminor, vp->v_umajor); */ 1867 } 1868 if (dev != NULL && dev->si_mountpoint) 1869 return (EBUSY); 1870 return (0); 1871 } 1872 1873 /* 1874 * Unmount all filesystems. The list is traversed in reverse order 1875 * of mounting to avoid dependencies. 1876 * 1877 * We want the umountall to be able to break out of its loop if a 1878 * failure occurs, after scanning all possible mounts, so the callback 1879 * returns 0 on error. 1880 * 1881 * NOTE: Do not call mountlist_remove(mp) on error any more, this will 1882 * confuse mountlist_scan()'s unbusy check. 1883 */ 1884 static int vfs_umountall_callback(struct mount *mp, void *data); 1885 1886 void 1887 vfs_unmountall(int halting) 1888 { 1889 int count; 1890 1891 do { 1892 count = mountlist_scan(vfs_umountall_callback, &halting, 1893 MNTSCAN_REVERSE|MNTSCAN_NOBUSY); 1894 } while (count); 1895 } 1896 1897 static 1898 int 1899 vfs_umountall_callback(struct mount *mp, void *data) 1900 { 1901 int error; 1902 int halting = *(int *)data; 1903 1904 /* 1905 * NOTE: When halting, dounmount will disconnect but leave 1906 * certain mount points intact. e.g. devfs. 1907 */ 1908 error = dounmount(mp, MNT_FORCE, halting); 1909 if (error) { 1910 kprintf("unmount of filesystem mounted from %s failed (", 1911 mp->mnt_stat.f_mntfromname); 1912 if (error == EBUSY) 1913 kprintf("BUSY)\n"); 1914 else 1915 kprintf("%d)\n", error); 1916 return 0; 1917 } else { 1918 return 1; 1919 } 1920 } 1921 1922 /* 1923 * Checks the mount flags for parameter mp and put the names comma-separated 1924 * into a string buffer buf with a size limit specified by len. 1925 * 1926 * It returns the number of bytes written into buf, and (*errorp) will be 1927 * set to 0, EINVAL (if passed length is 0), or ENOSPC (supplied buffer was 1928 * not large enough). The buffer will be 0-terminated if len was not 0. 1929 */ 1930 size_t 1931 vfs_flagstostr(int flags, const struct mountctl_opt *optp, 1932 char *buf, size_t len, int *errorp) 1933 { 1934 static const struct mountctl_opt optnames[] = { 1935 { MNT_RDONLY, "read-only" }, 1936 { MNT_SYNCHRONOUS, "synchronous" }, 1937 { MNT_NOEXEC, "noexec" }, 1938 { MNT_NOSUID, "nosuid" }, 1939 { MNT_NODEV, "nodev" }, 1940 { MNT_AUTOMOUNTED, "automounted" }, 1941 { MNT_ASYNC, "asynchronous" }, 1942 { MNT_SUIDDIR, "suiddir" }, 1943 { MNT_SOFTDEP, "soft-updates" }, 1944 { MNT_NOSYMFOLLOW, "nosymfollow" }, 1945 { MNT_TRIM, "trim" }, 1946 { MNT_NOATIME, "noatime" }, 1947 { MNT_NOCLUSTERR, "noclusterr" }, 1948 { MNT_NOCLUSTERW, "noclusterw" }, 1949 { MNT_EXRDONLY, "NFS read-only" }, 1950 { MNT_EXPORTED, "NFS exported" }, 1951 /* Remaining NFS flags could come here */ 1952 { MNT_LOCAL, "local" }, 1953 { MNT_QUOTA, "with-quotas" }, 1954 /* { MNT_ROOTFS, "rootfs" }, */ 1955 /* { MNT_IGNORE, "ignore" }, */ 1956 { 0, NULL} 1957 }; 1958 int bwritten; 1959 int bleft; 1960 int optlen; 1961 int actsize; 1962 1963 *errorp = 0; 1964 bwritten = 0; 1965 bleft = len - 1; /* leave room for trailing \0 */ 1966 1967 /* 1968 * Checks the size of the string. If it contains 1969 * any data, then we will append the new flags to 1970 * it. 1971 */ 1972 actsize = strlen(buf); 1973 if (actsize > 0) 1974 buf += actsize; 1975 1976 /* Default flags if no flags passed */ 1977 if (optp == NULL) 1978 optp = optnames; 1979 1980 if (bleft < 0) { /* degenerate case, 0-length buffer */ 1981 *errorp = EINVAL; 1982 return(0); 1983 } 1984 1985 for (; flags && optp->o_opt; ++optp) { 1986 if ((flags & optp->o_opt) == 0) 1987 continue; 1988 optlen = strlen(optp->o_name); 1989 if (bwritten || actsize > 0) { 1990 if (bleft < 2) { 1991 *errorp = ENOSPC; 1992 break; 1993 } 1994 buf[bwritten++] = ','; 1995 buf[bwritten++] = ' '; 1996 bleft -= 2; 1997 } 1998 if (bleft < optlen) { 1999 *errorp = ENOSPC; 2000 break; 2001 } 2002 bcopy(optp->o_name, buf + bwritten, optlen); 2003 bwritten += optlen; 2004 bleft -= optlen; 2005 flags &= ~optp->o_opt; 2006 } 2007 2008 /* 2009 * Space already reserved for trailing \0 2010 */ 2011 buf[bwritten] = 0; 2012 return (bwritten); 2013 } 2014 2015 /* 2016 * Build hash lists of net addresses and hang them off the mount point. 2017 * Called by ufs_mount() to set up the lists of export addresses. 2018 */ 2019 static int 2020 vfs_hang_addrlist(struct mount *mp, struct netexport *nep, 2021 const struct export_args *argp) 2022 { 2023 struct netcred *np; 2024 struct radix_node_head *rnh; 2025 int i; 2026 struct radix_node *rn; 2027 struct sockaddr *saddr, *smask = NULL; 2028 int error; 2029 2030 if (argp->ex_addrlen == 0) { 2031 if (mp->mnt_flag & MNT_DEFEXPORTED) 2032 return (EPERM); 2033 np = &nep->ne_defexported; 2034 np->netc_exflags = argp->ex_flags; 2035 np->netc_anon = argp->ex_anon; 2036 np->netc_anon.cr_ref = 1; 2037 mp->mnt_flag |= MNT_DEFEXPORTED; 2038 return (0); 2039 } 2040 2041 if (argp->ex_addrlen < 0 || argp->ex_addrlen > MLEN) 2042 return (EINVAL); 2043 if (argp->ex_masklen < 0 || argp->ex_masklen > MLEN) 2044 return (EINVAL); 2045 2046 i = sizeof(struct netcred) + argp->ex_addrlen + argp->ex_masklen; 2047 np = (struct netcred *)kmalloc(i, M_NETCRED, M_WAITOK | M_ZERO); 2048 saddr = (struct sockaddr *) (np + 1); 2049 if ((error = copyin(argp->ex_addr, (caddr_t) saddr, argp->ex_addrlen))) 2050 goto out; 2051 if (saddr->sa_len > argp->ex_addrlen) 2052 saddr->sa_len = argp->ex_addrlen; 2053 if (argp->ex_masklen) { 2054 smask = (struct sockaddr *)((caddr_t)saddr + argp->ex_addrlen); 2055 error = copyin(argp->ex_mask, (caddr_t)smask, argp->ex_masklen); 2056 if (error) 2057 goto out; 2058 if (smask->sa_len > argp->ex_masklen) 2059 smask->sa_len = argp->ex_masklen; 2060 } 2061 NE_LOCK(nep); 2062 if (nep->ne_maskhead == NULL) { 2063 if (!rn_inithead((void **)&nep->ne_maskhead, NULL, 0)) { 2064 error = ENOBUFS; 2065 goto out; 2066 } 2067 } 2068 if ((rnh = vfs_create_addrlist_af(saddr->sa_family, nep)) == NULL) { 2069 error = ENOBUFS; 2070 goto out; 2071 } 2072 rn = (*rnh->rnh_addaddr)((char *)saddr, (char *)smask, rnh, 2073 np->netc_rnodes); 2074 NE_UNLOCK(nep); 2075 if (rn == NULL || np != (struct netcred *)rn) { /* already exists */ 2076 error = EPERM; 2077 goto out; 2078 } 2079 np->netc_exflags = argp->ex_flags; 2080 np->netc_anon = argp->ex_anon; 2081 np->netc_anon.cr_ref = 1; 2082 return (0); 2083 2084 out: 2085 kfree(np, M_NETCRED); 2086 return (error); 2087 } 2088 2089 /* 2090 * Free netcred structures installed in the netexport 2091 */ 2092 static int 2093 vfs_free_netcred(struct radix_node *rn, void *w) 2094 { 2095 struct radix_node_head *rnh = (struct radix_node_head *)w; 2096 2097 (*rnh->rnh_deladdr) (rn->rn_key, rn->rn_mask, rnh); 2098 kfree(rn, M_NETCRED); 2099 2100 return (0); 2101 } 2102 2103 /* 2104 * callback to free an element of the mask table installed in the 2105 * netexport. These may be created indirectly and are not netcred 2106 * structures. 2107 */ 2108 static int 2109 vfs_free_netcred_mask(struct radix_node *rn, void *w) 2110 { 2111 struct radix_node_head *rnh = (struct radix_node_head *)w; 2112 2113 (*rnh->rnh_deladdr) (rn->rn_key, rn->rn_mask, rnh); 2114 kfree(rn, M_RTABLE); 2115 2116 return (0); 2117 } 2118 2119 static struct radix_node_head * 2120 vfs_create_addrlist_af(int af, struct netexport *nep) 2121 { 2122 struct radix_node_head *rnh = NULL; 2123 #if defined(INET) || defined(INET6) 2124 struct radix_node_head *maskhead = nep->ne_maskhead; 2125 int off; 2126 #endif 2127 2128 NE_ASSERT_LOCKED(nep); 2129 #if defined(INET) || defined(INET6) 2130 KKASSERT(maskhead != NULL); 2131 #endif 2132 switch (af) { 2133 #ifdef INET 2134 case AF_INET: 2135 if ((rnh = nep->ne_inethead) == NULL) { 2136 off = offsetof(struct sockaddr_in, sin_addr) << 3; 2137 if (!rn_inithead((void **)&rnh, maskhead, off)) 2138 return (NULL); 2139 nep->ne_inethead = rnh; 2140 } 2141 break; 2142 #endif 2143 #ifdef INET6 2144 case AF_INET6: 2145 if ((rnh = nep->ne_inet6head) == NULL) { 2146 off = offsetof(struct sockaddr_in6, sin6_addr) << 3; 2147 if (!rn_inithead((void **)&rnh, maskhead, off)) 2148 return (NULL); 2149 nep->ne_inet6head = rnh; 2150 } 2151 break; 2152 #endif 2153 } 2154 return (rnh); 2155 } 2156 2157 /* 2158 * helper function for freeing netcred elements 2159 */ 2160 static void 2161 vfs_free_addrlist_af(struct radix_node_head **prnh) 2162 { 2163 struct radix_node_head *rnh = *prnh; 2164 2165 (*rnh->rnh_walktree) (rnh, vfs_free_netcred, rnh); 2166 kfree(rnh, M_RTABLE); 2167 *prnh = NULL; 2168 } 2169 2170 /* 2171 * helper function for freeing mask elements 2172 */ 2173 static void 2174 vfs_free_addrlist_masks(struct radix_node_head **prnh) 2175 { 2176 struct radix_node_head *rnh = *prnh; 2177 2178 (*rnh->rnh_walktree) (rnh, vfs_free_netcred_mask, rnh); 2179 kfree(rnh, M_RTABLE); 2180 *prnh = NULL; 2181 } 2182 2183 /* 2184 * Free the net address hash lists that are hanging off the mount points. 2185 */ 2186 static void 2187 vfs_free_addrlist(struct netexport *nep) 2188 { 2189 NE_LOCK(nep); 2190 if (nep->ne_inethead != NULL) 2191 vfs_free_addrlist_af(&nep->ne_inethead); 2192 if (nep->ne_inet6head != NULL) 2193 vfs_free_addrlist_af(&nep->ne_inet6head); 2194 if (nep->ne_maskhead) 2195 vfs_free_addrlist_masks(&nep->ne_maskhead); 2196 NE_UNLOCK(nep); 2197 } 2198 2199 int 2200 vfs_export(struct mount *mp, struct netexport *nep, 2201 const struct export_args *argp) 2202 { 2203 int error; 2204 2205 if (argp->ex_flags & MNT_DELEXPORT) { 2206 if (mp->mnt_flag & MNT_EXPUBLIC) { 2207 vfs_setpublicfs(NULL, NULL, NULL); 2208 mp->mnt_flag &= ~MNT_EXPUBLIC; 2209 } 2210 vfs_free_addrlist(nep); 2211 mp->mnt_flag &= ~(MNT_EXPORTED | MNT_DEFEXPORTED); 2212 } 2213 if (argp->ex_flags & MNT_EXPORTED) { 2214 if (argp->ex_flags & MNT_EXPUBLIC) { 2215 if ((error = vfs_setpublicfs(mp, nep, argp)) != 0) 2216 return (error); 2217 mp->mnt_flag |= MNT_EXPUBLIC; 2218 } 2219 if ((error = vfs_hang_addrlist(mp, nep, argp))) 2220 return (error); 2221 mp->mnt_flag |= MNT_EXPORTED; 2222 } 2223 return (0); 2224 } 2225 2226 2227 /* 2228 * Set the publicly exported filesystem (WebNFS). Currently, only 2229 * one public filesystem is possible in the spec (RFC 2054 and 2055) 2230 */ 2231 int 2232 vfs_setpublicfs(struct mount *mp, struct netexport *nep, 2233 const struct export_args *argp) 2234 { 2235 int error; 2236 struct vnode *rvp; 2237 char *cp; 2238 2239 /* 2240 * mp == NULL -> invalidate the current info, the FS is 2241 * no longer exported. May be called from either vfs_export 2242 * or unmount, so check if it hasn't already been done. 2243 */ 2244 if (mp == NULL) { 2245 if (nfs_pub.np_valid) { 2246 nfs_pub.np_valid = 0; 2247 if (nfs_pub.np_index != NULL) { 2248 kfree(nfs_pub.np_index, M_TEMP); 2249 nfs_pub.np_index = NULL; 2250 } 2251 } 2252 return (0); 2253 } 2254 2255 /* 2256 * Only one allowed at a time. 2257 */ 2258 if (nfs_pub.np_valid != 0 && mp != nfs_pub.np_mount) 2259 return (EBUSY); 2260 2261 /* 2262 * Get real filehandle for root of exported FS. 2263 */ 2264 bzero((caddr_t)&nfs_pub.np_handle, sizeof(nfs_pub.np_handle)); 2265 nfs_pub.np_handle.fh_fsid = mp->mnt_stat.f_fsid; 2266 2267 if ((error = VFS_ROOT(mp, &rvp))) 2268 return (error); 2269 2270 if ((error = VFS_VPTOFH(rvp, &nfs_pub.np_handle.fh_fid))) 2271 return (error); 2272 2273 vput(rvp); 2274 2275 /* 2276 * If an indexfile was specified, pull it in. 2277 */ 2278 if (argp->ex_indexfile != NULL) { 2279 int namelen; 2280 2281 error = vn_get_namelen(rvp, &namelen); 2282 if (error) 2283 return (error); 2284 nfs_pub.np_index = kmalloc(namelen, M_TEMP, M_WAITOK); 2285 error = copyinstr(argp->ex_indexfile, nfs_pub.np_index, 2286 namelen, NULL); 2287 if (!error) { 2288 /* 2289 * Check for illegal filenames. 2290 */ 2291 for (cp = nfs_pub.np_index; *cp; cp++) { 2292 if (*cp == '/') { 2293 error = EINVAL; 2294 break; 2295 } 2296 } 2297 } 2298 if (error) { 2299 kfree(nfs_pub.np_index, M_TEMP); 2300 return (error); 2301 } 2302 } 2303 2304 nfs_pub.np_mount = mp; 2305 nfs_pub.np_valid = 1; 2306 return (0); 2307 } 2308 2309 struct netcred * 2310 vfs_export_lookup(struct mount *mp, struct netexport *nep, 2311 struct sockaddr *nam) 2312 { 2313 struct netcred *np; 2314 struct radix_node_head *rnh; 2315 struct sockaddr *saddr; 2316 2317 np = NULL; 2318 if (mp->mnt_flag & MNT_EXPORTED) { 2319 /* 2320 * Lookup in the export list first. 2321 */ 2322 NE_LOCK(nep); 2323 if (nam != NULL) { 2324 saddr = nam; 2325 switch (saddr->sa_family) { 2326 #ifdef INET 2327 case AF_INET: 2328 rnh = nep->ne_inethead; 2329 break; 2330 #endif 2331 #ifdef INET6 2332 case AF_INET6: 2333 rnh = nep->ne_inet6head; 2334 break; 2335 #endif 2336 default: 2337 rnh = NULL; 2338 } 2339 if (rnh != NULL) { 2340 np = (struct netcred *) 2341 (*rnh->rnh_matchaddr)((char *)saddr, 2342 rnh); 2343 if (np && np->netc_rnodes->rn_flags & RNF_ROOT) 2344 np = NULL; 2345 } 2346 } 2347 NE_UNLOCK(nep); 2348 /* 2349 * If no address match, use the default if it exists. 2350 */ 2351 if (np == NULL && mp->mnt_flag & MNT_DEFEXPORTED) 2352 np = &nep->ne_defexported; 2353 } 2354 return (np); 2355 } 2356 2357 /* 2358 * perform msync on all vnodes under a mount point. The mount point must 2359 * be locked. This code is also responsible for lazy-freeing unreferenced 2360 * vnodes whos VM objects no longer contain pages. 2361 * 2362 * NOTE: MNT_WAIT still skips vnodes in the VXLOCK state. 2363 * 2364 * NOTE: XXX VOP_PUTPAGES and friends requires that the vnode be locked, 2365 * but vnode_pager_putpages() doesn't lock the vnode. We have to do it 2366 * way up in this high level function. 2367 */ 2368 static int vfs_msync_scan1(struct mount *mp, struct vnode *vp, void *data); 2369 static int vfs_msync_scan2(struct mount *mp, struct vnode *vp, void *data); 2370 2371 void 2372 vfs_msync(struct mount *mp, int flags) 2373 { 2374 int vmsc_flags; 2375 2376 /* 2377 * tmpfs sets this flag to prevent msync(), sync, and the 2378 * filesystem periodic syncer from trying to flush VM pages 2379 * to swap. Only pure memory pressure flushes tmpfs VM pages 2380 * to swap. 2381 */ 2382 if (mp->mnt_kern_flag & MNTK_NOMSYNC) 2383 return; 2384 2385 /* 2386 * Ok, scan the vnodes for work. If the filesystem is using the 2387 * syncer thread feature we can use vsyncscan() instead of 2388 * vmntvnodescan(), which is much faster. 2389 */ 2390 vmsc_flags = VMSC_GETVP; 2391 if (flags != MNT_WAIT) 2392 vmsc_flags |= VMSC_NOWAIT; 2393 2394 if (mp->mnt_kern_flag & MNTK_THR_SYNC) { 2395 vsyncscan(mp, vmsc_flags, vfs_msync_scan2, 2396 (void *)(intptr_t)flags); 2397 } else { 2398 vmntvnodescan(mp, vmsc_flags, 2399 vfs_msync_scan1, vfs_msync_scan2, 2400 (void *)(intptr_t)flags); 2401 } 2402 } 2403 2404 /* 2405 * scan1 is a fast pre-check. There could be hundreds of thousands of 2406 * vnodes, we cannot afford to do anything heavy weight until we have a 2407 * fairly good indication that there is work to do. 2408 */ 2409 static 2410 int 2411 vfs_msync_scan1(struct mount *mp, struct vnode *vp, void *data) 2412 { 2413 int flags = (int)(intptr_t)data; 2414 2415 if ((vp->v_flag & VRECLAIMED) == 0) { 2416 if (vp->v_auxrefs == 0 && VREFCNT(vp) <= 0 && 2417 vp->v_object) { 2418 return(0); /* call scan2 */ 2419 } 2420 if ((mp->mnt_flag & MNT_RDONLY) == 0 && 2421 (vp->v_flag & VOBJDIRTY) && 2422 (flags == MNT_WAIT || vn_islocked(vp) == 0)) { 2423 return(0); /* call scan2 */ 2424 } 2425 } 2426 2427 /* 2428 * do not call scan2, continue the loop 2429 */ 2430 return(-1); 2431 } 2432 2433 /* 2434 * This callback is handed a locked vnode. 2435 */ 2436 static 2437 int 2438 vfs_msync_scan2(struct mount *mp, struct vnode *vp, void *data) 2439 { 2440 vm_object_t obj; 2441 int flags = (int)(intptr_t)data; 2442 int opcflags; 2443 2444 if (vp->v_flag & VRECLAIMED) 2445 return(0); 2446 2447 if ((mp->mnt_flag & MNT_RDONLY) == 0 && (vp->v_flag & VOBJDIRTY)) { 2448 if ((obj = vp->v_object) != NULL) { 2449 if (flags == MNT_WAIT) { 2450 /* 2451 * VFS_MSYNC is called with MNT_WAIT when 2452 * unmounting. 2453 */ 2454 opcflags = OBJPC_SYNC; 2455 } else if (vp->v_writecount || obj->ref_count) { 2456 /* 2457 * VFS_MSYNC is otherwise called via the 2458 * periodic filesystem sync or the 'sync' 2459 * command. Honor MADV_NOSYNC / MAP_NOSYNC 2460 * if the file is open for writing or memory 2461 * mapped. Pages flagged PG_NOSYNC will not 2462 * be automatically flushed at this time. 2463 * 2464 * The obj->ref_count test is not perfect 2465 * since temporary refs may be present, but 2466 * the periodic filesystem sync will ultimately 2467 * catch it if the file is not open and not 2468 * mapped. 2469 */ 2470 opcflags = OBJPC_NOSYNC; 2471 } else { 2472 /* 2473 * If the file is no longer open for writing 2474 * and also no longer mapped, do not honor 2475 * MAP_NOSYNC. That is, fully synchronize 2476 * the file. 2477 * 2478 * This still occurs on the periodic fs sync, 2479 * so frontend programs which turn the file 2480 * over quickly enough can still avoid the 2481 * sync, but ultimately we do want to flush 2482 * even MADV_NOSYNC pages once it is no longer 2483 * mapped or open for writing. 2484 */ 2485 opcflags = 0; 2486 } 2487 vm_object_page_clean(obj, 0, 0, opcflags); 2488 } 2489 } 2490 return(0); 2491 } 2492 2493 /* 2494 * Wake up anyone interested in vp because it is being revoked. 2495 */ 2496 void 2497 vn_gone(struct vnode *vp) 2498 { 2499 lwkt_gettoken(&vp->v_token); 2500 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, NOTE_REVOKE); 2501 lwkt_reltoken(&vp->v_token); 2502 } 2503 2504 /* 2505 * extract the cdev_t from a VBLK or VCHR. The vnode must have been opened 2506 * (or v_rdev might be NULL). 2507 */ 2508 cdev_t 2509 vn_todev(struct vnode *vp) 2510 { 2511 if (vp->v_type != VBLK && vp->v_type != VCHR) 2512 return (NULL); 2513 KKASSERT(vp->v_rdev != NULL); 2514 return (vp->v_rdev); 2515 } 2516 2517 /* 2518 * Check if vnode represents a disk device. The vnode does not need to be 2519 * opened. 2520 * 2521 * MPALMOSTSAFE 2522 */ 2523 int 2524 vn_isdisk(struct vnode *vp, int *errp) 2525 { 2526 cdev_t dev; 2527 2528 if (vp->v_type != VCHR) { 2529 if (errp != NULL) 2530 *errp = ENOTBLK; 2531 return (0); 2532 } 2533 2534 dev = vp->v_rdev; 2535 2536 if (dev == NULL) { 2537 if (errp != NULL) 2538 *errp = ENXIO; 2539 return (0); 2540 } 2541 if (dev_is_good(dev) == 0) { 2542 if (errp != NULL) 2543 *errp = ENXIO; 2544 return (0); 2545 } 2546 if ((dev_dflags(dev) & D_DISK) == 0) { 2547 if (errp != NULL) 2548 *errp = ENOTBLK; 2549 return (0); 2550 } 2551 if (errp != NULL) 2552 *errp = 0; 2553 return (1); 2554 } 2555 2556 int 2557 vn_get_namelen(struct vnode *vp, int *namelen) 2558 { 2559 int error; 2560 register_t retval[2]; 2561 2562 error = VOP_PATHCONF(vp, _PC_NAME_MAX, retval); 2563 if (error) 2564 return (error); 2565 *namelen = (int)retval[0]; 2566 return (0); 2567 } 2568 2569 int 2570 vop_write_dirent(int *error, struct uio *uio, ino_t d_ino, uint8_t d_type, 2571 uint16_t d_namlen, const char *d_name) 2572 { 2573 struct dirent *dp; 2574 size_t len; 2575 2576 len = _DIRENT_RECLEN(d_namlen); 2577 if (len > uio->uio_resid) 2578 return(1); 2579 2580 dp = kmalloc(len, M_TEMP, M_WAITOK | M_ZERO); 2581 2582 dp->d_ino = d_ino; 2583 dp->d_namlen = d_namlen; 2584 dp->d_type = d_type; 2585 bcopy(d_name, dp->d_name, d_namlen); 2586 2587 *error = uiomove((caddr_t)dp, len, uio); 2588 2589 kfree(dp, M_TEMP); 2590 2591 return(0); 2592 } 2593 2594 void 2595 vn_mark_atime(struct vnode *vp, struct thread *td) 2596 { 2597 struct proc *p = td->td_proc; 2598 struct ucred *cred = p ? p->p_ucred : proc0.p_ucred; 2599 2600 if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) { 2601 VOP_MARKATIME(vp, cred); 2602 } 2603 } 2604 2605 /* 2606 * Calculate the number of entries in an inode-related chained hash table. 2607 * With today's memory sizes, maxvnodes can wind up being a very large 2608 * number. There is no reason to waste memory, so tolerate some stacking. 2609 */ 2610 int 2611 vfs_inodehashsize(void) 2612 { 2613 int hsize; 2614 2615 hsize = 32; 2616 while (hsize < maxvnodes) 2617 hsize <<= 1; 2618 while (hsize > maxvnodes * 2) 2619 hsize >>= 1; /* nominal 2x stacking */ 2620 2621 if (maxvnodes > 1024 * 1024) 2622 hsize >>= 1; /* nominal 8x stacking */ 2623 2624 if (maxvnodes > 128 * 1024) 2625 hsize >>= 1; /* nominal 4x stacking */ 2626 2627 if (hsize < 16) 2628 hsize = 16; 2629 2630 return hsize; 2631 } 2632