1 /* 2 * Copyright (c) 1989, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 35 * $FreeBSD: src/sys/kern/vfs_subr.c,v 1.249.2.30 2003/04/04 20:35:57 tegge Exp $ 36 */ 37 38 /* 39 * External virtual filesystem routines 40 */ 41 #include "opt_ddb.h" 42 #include "opt_inet.h" 43 #include "opt_inet6.h" 44 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/uio.h> 48 #include <sys/buf.h> 49 #include <sys/conf.h> 50 #include <sys/dirent.h> 51 #include <sys/eventhandler.h> 52 #include <sys/fcntl.h> 53 #include <sys/file.h> 54 #include <sys/kernel.h> 55 #include <sys/kthread.h> 56 #include <sys/malloc.h> 57 #include <sys/mbuf.h> 58 #include <sys/mount.h> 59 #include <sys/priv.h> 60 #include <sys/proc.h> 61 #include <sys/reboot.h> 62 #include <sys/socket.h> 63 #include <sys/stat.h> 64 #include <sys/sysctl.h> 65 #include <sys/syslog.h> 66 #include <sys/unistd.h> 67 #include <sys/vmmeter.h> 68 #include <sys/vnode.h> 69 70 #include <machine/limits.h> 71 72 #include <vm/vm.h> 73 #include <vm/vm_object.h> 74 #include <vm/vm_extern.h> 75 #include <vm/vm_kern.h> 76 #include <vm/pmap.h> 77 #include <vm/vm_map.h> 78 #include <vm/vm_page.h> 79 #include <vm/vm_pager.h> 80 #include <vm/vnode_pager.h> 81 #include <vm/vm_zone.h> 82 83 #include <sys/buf2.h> 84 #include <sys/mplock2.h> 85 #include <vm/vm_page2.h> 86 87 #include <netinet/in.h> 88 89 static MALLOC_DEFINE(M_NETCRED, "Export Host", "Export host address structure"); 90 91 __read_mostly int numvnodes; 92 SYSCTL_INT(_debug, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, 93 "Number of vnodes allocated"); 94 __read_mostly int verbose_reclaims; 95 SYSCTL_INT(_debug, OID_AUTO, verbose_reclaims, CTLFLAG_RD, &verbose_reclaims, 0, 96 "Output filename of reclaimed vnode(s)"); 97 98 __read_mostly enum vtype iftovt_tab[16] = { 99 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 100 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 101 }; 102 __read_mostly int vttoif_tab[9] = { 103 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 104 S_IFSOCK, S_IFIFO, S_IFMT, 105 }; 106 107 static int reassignbufcalls; 108 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 109 0, "Number of times buffers have been reassigned to the proper list"); 110 111 __read_mostly static int check_buf_overlap = 2; /* invasive check */ 112 SYSCTL_INT(_vfs, OID_AUTO, check_buf_overlap, CTLFLAG_RW, &check_buf_overlap, 113 0, "Enable overlapping buffer checks"); 114 115 int nfs_mount_type = -1; 116 static struct lwkt_token spechash_token; 117 struct nfs_public nfs_pub; /* publicly exported FS */ 118 119 __read_mostly int maxvnodes; 120 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 121 &maxvnodes, 0, "Maximum number of vnodes"); 122 123 static struct radix_node_head *vfs_create_addrlist_af(int af, 124 struct netexport *nep); 125 static void vfs_free_addrlist (struct netexport *nep); 126 static int vfs_free_netcred (struct radix_node *rn, void *w); 127 static void vfs_free_addrlist_af (struct radix_node_head **prnh); 128 static int vfs_hang_addrlist (struct mount *mp, struct netexport *nep, 129 const struct export_args *argp); 130 131 __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 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 206 207 __read_mostly static int timestamp_precision = TSP_SEC; 208 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 209 ×tamp_precision, 0, "Precision of file timestamps"); 210 211 /* 212 * Get a current timestamp. 213 * 214 * MPSAFE 215 */ 216 void 217 vfs_timestamp(struct timespec *tsp) 218 { 219 struct timeval tv; 220 221 switch (timestamp_precision) { 222 case TSP_SEC: 223 tsp->tv_sec = time_second; 224 tsp->tv_nsec = 0; 225 break; 226 case TSP_HZ: 227 getnanotime(tsp); 228 break; 229 case TSP_USEC: 230 microtime(&tv); 231 TIMEVAL_TO_TIMESPEC(&tv, tsp); 232 break; 233 case TSP_NSEC: 234 default: 235 nanotime(tsp); 236 break; 237 } 238 } 239 240 /* 241 * Set vnode attributes to VNOVAL 242 */ 243 void 244 vattr_null(struct vattr *vap) 245 { 246 vap->va_type = VNON; 247 vap->va_size = VNOVAL; 248 vap->va_bytes = VNOVAL; 249 vap->va_mode = VNOVAL; 250 vap->va_nlink = VNOVAL; 251 vap->va_uid = VNOVAL; 252 vap->va_gid = VNOVAL; 253 vap->va_fsid = VNOVAL; 254 vap->va_fileid = VNOVAL; 255 vap->va_blocksize = VNOVAL; 256 vap->va_rmajor = VNOVAL; 257 vap->va_rminor = VNOVAL; 258 vap->va_atime.tv_sec = VNOVAL; 259 vap->va_atime.tv_nsec = VNOVAL; 260 vap->va_mtime.tv_sec = VNOVAL; 261 vap->va_mtime.tv_nsec = VNOVAL; 262 vap->va_ctime.tv_sec = VNOVAL; 263 vap->va_ctime.tv_nsec = VNOVAL; 264 vap->va_flags = VNOVAL; 265 vap->va_gen = VNOVAL; 266 vap->va_vaflags = 0; 267 /* va_*_uuid fields are only valid if related flags are set */ 268 } 269 270 /* 271 * Flush out and invalidate all buffers associated with a vnode. 272 * 273 * vp must be locked. 274 */ 275 static int vinvalbuf_bp(struct buf *bp, void *data); 276 277 struct vinvalbuf_bp_info { 278 struct vnode *vp; 279 int slptimeo; 280 int lkflags; 281 int flags; 282 int clean; 283 }; 284 285 int 286 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 287 { 288 struct vinvalbuf_bp_info info; 289 vm_object_t object; 290 int error; 291 292 lwkt_gettoken(&vp->v_token); 293 294 /* 295 * If we are being asked to save, call fsync to ensure that the inode 296 * is updated. 297 */ 298 if (flags & V_SAVE) { 299 error = bio_track_wait(&vp->v_track_write, slpflag, slptimeo); 300 if (error) 301 goto done; 302 if (!RB_EMPTY(&vp->v_rbdirty_tree)) { 303 if ((error = VOP_FSYNC(vp, MNT_WAIT, 0)) != 0) 304 goto done; 305 #if 0 306 /* 307 * Dirty bufs may be left or generated via races 308 * in circumstances where vinvalbuf() is called on 309 * a vnode not undergoing reclamation. Only 310 * panic if we are trying to reclaim the vnode. 311 */ 312 if ((vp->v_flag & VRECLAIMED) && 313 (bio_track_active(&vp->v_track_write) || 314 !RB_EMPTY(&vp->v_rbdirty_tree))) { 315 panic("vinvalbuf: dirty bufs"); 316 } 317 #endif 318 } 319 } 320 info.slptimeo = slptimeo; 321 info.lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 322 if (slpflag & PCATCH) 323 info.lkflags |= LK_PCATCH; 324 info.flags = flags; 325 info.vp = vp; 326 327 /* 328 * Flush the buffer cache until nothing is left, wait for all I/O 329 * to complete. At least one pass is required. We might block 330 * in the pip code so we have to re-check. Order is important. 331 */ 332 do { 333 /* 334 * Flush buffer cache 335 */ 336 if (!RB_EMPTY(&vp->v_rbclean_tree)) { 337 info.clean = 1; 338 error = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 339 NULL, vinvalbuf_bp, &info); 340 } 341 if (!RB_EMPTY(&vp->v_rbdirty_tree)) { 342 info.clean = 0; 343 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 344 NULL, vinvalbuf_bp, &info); 345 } 346 347 /* 348 * Wait for I/O completion. 349 */ 350 bio_track_wait(&vp->v_track_write, 0, 0); 351 if ((object = vp->v_object) != NULL) 352 refcount_wait(&object->paging_in_progress, "vnvlbx"); 353 } while (bio_track_active(&vp->v_track_write) || 354 !RB_EMPTY(&vp->v_rbclean_tree) || 355 !RB_EMPTY(&vp->v_rbdirty_tree)); 356 357 /* 358 * Destroy the copy in the VM cache, too. 359 */ 360 if ((object = vp->v_object) != NULL) { 361 vm_object_page_remove(object, 0, 0, 362 (flags & V_SAVE) ? TRUE : FALSE); 363 } 364 365 if (!RB_EMPTY(&vp->v_rbdirty_tree) || !RB_EMPTY(&vp->v_rbclean_tree)) 366 panic("vinvalbuf: flush failed"); 367 if (!RB_EMPTY(&vp->v_rbhash_tree)) 368 panic("vinvalbuf: flush failed, buffers still present"); 369 error = 0; 370 done: 371 lwkt_reltoken(&vp->v_token); 372 return (error); 373 } 374 375 static int 376 vinvalbuf_bp(struct buf *bp, void *data) 377 { 378 struct vinvalbuf_bp_info *info = data; 379 int error; 380 381 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 382 atomic_add_int(&bp->b_refs, 1); 383 error = BUF_TIMELOCK(bp, info->lkflags, 384 "vinvalbuf", info->slptimeo); 385 atomic_subtract_int(&bp->b_refs, 1); 386 if (error == 0) { 387 BUF_UNLOCK(bp); 388 error = ENOLCK; 389 } 390 if (error == ENOLCK) 391 return(0); 392 return (-error); 393 } 394 KKASSERT(bp->b_vp == info->vp); 395 396 /* 397 * Must check clean/dirty status after successfully locking as 398 * it may race. 399 */ 400 if ((info->clean && (bp->b_flags & B_DELWRI)) || 401 (info->clean == 0 && (bp->b_flags & B_DELWRI) == 0)) { 402 BUF_UNLOCK(bp); 403 return(0); 404 } 405 406 /* 407 * NOTE: NO B_LOCKED CHECK. Also no buf_checkwrite() 408 * check. This code will write out the buffer, period. 409 */ 410 bremfree(bp); 411 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 412 (info->flags & V_SAVE)) { 413 cluster_awrite(bp); 414 } else if (info->flags & V_SAVE) { 415 /* 416 * Cannot set B_NOCACHE on a clean buffer as this will 417 * destroy the VM backing store which might actually 418 * be dirty (and unsynchronized). 419 */ 420 bp->b_flags |= (B_INVAL | B_RELBUF); 421 brelse(bp); 422 } else { 423 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 424 brelse(bp); 425 } 426 return(0); 427 } 428 429 /* 430 * Truncate a file's buffer and pages to a specified length. This 431 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 432 * sync activity. 433 * 434 * The vnode must be locked. 435 */ 436 static int vtruncbuf_bp_trunc_cmp(struct buf *bp, void *data); 437 static int vtruncbuf_bp_trunc(struct buf *bp, void *data); 438 static int vtruncbuf_bp_metasync_cmp(struct buf *bp, void *data); 439 static int vtruncbuf_bp_metasync(struct buf *bp, void *data); 440 441 struct vtruncbuf_info { 442 struct vnode *vp; 443 off_t truncloffset; 444 int clean; 445 }; 446 447 int 448 vtruncbuf(struct vnode *vp, off_t length, int blksize) 449 { 450 struct vtruncbuf_info info; 451 const char *filename; 452 int count; 453 454 /* 455 * Round up to the *next* block, then destroy the buffers in question. 456 * Since we are only removing some of the buffers we must rely on the 457 * scan count to determine whether a loop is necessary. 458 */ 459 if ((count = (int)(length % blksize)) != 0) 460 info.truncloffset = length + (blksize - count); 461 else 462 info.truncloffset = length; 463 info.vp = vp; 464 465 lwkt_gettoken(&vp->v_token); 466 do { 467 info.clean = 1; 468 count = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 469 vtruncbuf_bp_trunc_cmp, 470 vtruncbuf_bp_trunc, &info); 471 info.clean = 0; 472 count += RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 473 vtruncbuf_bp_trunc_cmp, 474 vtruncbuf_bp_trunc, &info); 475 } while(count); 476 477 /* 478 * For safety, fsync any remaining metadata if the file is not being 479 * truncated to 0. Since the metadata does not represent the entire 480 * dirty list we have to rely on the hit count to ensure that we get 481 * all of it. 482 */ 483 if (length > 0) { 484 do { 485 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 486 vtruncbuf_bp_metasync_cmp, 487 vtruncbuf_bp_metasync, &info); 488 } while (count); 489 } 490 491 /* 492 * Clean out any left over VM backing store. 493 * 494 * It is possible to have in-progress I/O from buffers that were 495 * not part of the truncation. This should not happen if we 496 * are truncating to 0-length. 497 */ 498 vnode_pager_setsize(vp, length); 499 bio_track_wait(&vp->v_track_write, 0, 0); 500 501 /* 502 * Debugging only 503 */ 504 spin_lock(&vp->v_spin); 505 filename = TAILQ_FIRST(&vp->v_namecache) ? 506 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"; 507 spin_unlock(&vp->v_spin); 508 509 /* 510 * Make sure no buffers were instantiated while we were trying 511 * to clean out the remaining VM pages. This could occur due 512 * to busy dirty VM pages being flushed out to disk. 513 */ 514 do { 515 info.clean = 1; 516 count = RB_SCAN(buf_rb_tree, &vp->v_rbclean_tree, 517 vtruncbuf_bp_trunc_cmp, 518 vtruncbuf_bp_trunc, &info); 519 info.clean = 0; 520 count += RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 521 vtruncbuf_bp_trunc_cmp, 522 vtruncbuf_bp_trunc, &info); 523 if (count) { 524 kprintf("Warning: vtruncbuf(): Had to re-clean %d " 525 "left over buffers in %s\n", count, filename); 526 } 527 } while(count); 528 529 lwkt_reltoken(&vp->v_token); 530 531 return (0); 532 } 533 534 /* 535 * The callback buffer is beyond the new file EOF and must be destroyed. 536 * Note that the compare function must conform to the RB_SCAN's requirements. 537 */ 538 static 539 int 540 vtruncbuf_bp_trunc_cmp(struct buf *bp, void *data) 541 { 542 struct vtruncbuf_info *info = data; 543 544 if (bp->b_loffset >= info->truncloffset) 545 return(0); 546 return(-1); 547 } 548 549 static 550 int 551 vtruncbuf_bp_trunc(struct buf *bp, void *data) 552 { 553 struct vtruncbuf_info *info = data; 554 555 /* 556 * Do not try to use a buffer we cannot immediately lock, but sleep 557 * anyway to prevent a livelock. The code will loop until all buffers 558 * can be acted upon. 559 * 560 * We must always revalidate the buffer after locking it to deal 561 * with MP races. 562 */ 563 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 564 atomic_add_int(&bp->b_refs, 1); 565 if (BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL) == 0) 566 BUF_UNLOCK(bp); 567 atomic_subtract_int(&bp->b_refs, 1); 568 } else if ((info->clean && (bp->b_flags & B_DELWRI)) || 569 (info->clean == 0 && (bp->b_flags & B_DELWRI) == 0) || 570 bp->b_vp != info->vp || 571 vtruncbuf_bp_trunc_cmp(bp, data)) { 572 BUF_UNLOCK(bp); 573 } else { 574 bremfree(bp); 575 bp->b_flags |= (B_INVAL | B_RELBUF | B_NOCACHE); 576 brelse(bp); 577 } 578 return(1); 579 } 580 581 /* 582 * Fsync all meta-data after truncating a file to be non-zero. Only metadata 583 * blocks (with a negative loffset) are scanned. 584 * Note that the compare function must conform to the RB_SCAN's requirements. 585 */ 586 static int 587 vtruncbuf_bp_metasync_cmp(struct buf *bp, void *data __unused) 588 { 589 if (bp->b_loffset < 0) 590 return(0); 591 return(1); 592 } 593 594 static int 595 vtruncbuf_bp_metasync(struct buf *bp, void *data) 596 { 597 struct vtruncbuf_info *info = data; 598 599 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 600 atomic_add_int(&bp->b_refs, 1); 601 if (BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL) == 0) 602 BUF_UNLOCK(bp); 603 atomic_subtract_int(&bp->b_refs, 1); 604 } else if ((bp->b_flags & B_DELWRI) == 0 || 605 bp->b_vp != info->vp || 606 vtruncbuf_bp_metasync_cmp(bp, data)) { 607 BUF_UNLOCK(bp); 608 } else { 609 bremfree(bp); 610 if (bp->b_vp == info->vp) 611 bawrite(bp); 612 else 613 bwrite(bp); 614 } 615 return(1); 616 } 617 618 /* 619 * vfsync - implements a multipass fsync on a file which understands 620 * dependancies and meta-data. The passed vnode must be locked. The 621 * waitfor argument may be MNT_WAIT or MNT_NOWAIT, or MNT_LAZY. 622 * 623 * When fsyncing data asynchronously just do one consolidated pass starting 624 * with the most negative block number. This may not get all the data due 625 * to dependancies. 626 * 627 * When fsyncing data synchronously do a data pass, then a metadata pass, 628 * then do additional data+metadata passes to try to get all the data out. 629 * 630 * Caller must ref the vnode but does not have to lock it. 631 */ 632 static int vfsync_wait_output(struct vnode *vp, 633 int (*waitoutput)(struct vnode *, struct thread *)); 634 static int vfsync_dummy_cmp(struct buf *bp __unused, void *data __unused); 635 static int vfsync_data_only_cmp(struct buf *bp, void *data); 636 static int vfsync_meta_only_cmp(struct buf *bp, void *data); 637 static int vfsync_lazy_range_cmp(struct buf *bp, void *data); 638 static int vfsync_bp(struct buf *bp, void *data); 639 640 struct vfsync_info { 641 struct vnode *vp; 642 int fastpass; 643 int synchronous; 644 int syncdeps; 645 int lazycount; 646 int lazylimit; 647 int skippedbufs; 648 int (*checkdef)(struct buf *); 649 int (*cmpfunc)(struct buf *, void *); 650 }; 651 652 int 653 vfsync(struct vnode *vp, int waitfor, int passes, 654 int (*checkdef)(struct buf *), 655 int (*waitoutput)(struct vnode *, struct thread *)) 656 { 657 struct vfsync_info info; 658 int error; 659 660 bzero(&info, sizeof(info)); 661 info.vp = vp; 662 if ((info.checkdef = checkdef) == NULL) 663 info.syncdeps = 1; 664 665 lwkt_gettoken(&vp->v_token); 666 667 switch(waitfor) { 668 case MNT_LAZY | MNT_NOWAIT: 669 case MNT_LAZY: 670 /* 671 * Lazy (filesystem syncer typ) Asynchronous plus limit the 672 * number of data (not meta) pages we try to flush to 1MB. 673 * A non-zero return means that lazy limit was reached. 674 */ 675 info.lazylimit = 1024 * 1024; 676 info.syncdeps = 1; 677 info.cmpfunc = vfsync_lazy_range_cmp; 678 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 679 vfsync_lazy_range_cmp, vfsync_bp, &info); 680 info.cmpfunc = vfsync_meta_only_cmp; 681 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, 682 vfsync_meta_only_cmp, vfsync_bp, &info); 683 if (error == 0) 684 vp->v_lazyw = 0; 685 else if (!RB_EMPTY(&vp->v_rbdirty_tree)) 686 vn_syncer_add(vp, 1); 687 error = 0; 688 break; 689 case MNT_NOWAIT: 690 /* 691 * Asynchronous. Do a data-only pass and a meta-only pass. 692 */ 693 info.syncdeps = 1; 694 info.cmpfunc = vfsync_data_only_cmp; 695 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, vfsync_data_only_cmp, 696 vfsync_bp, &info); 697 info.cmpfunc = vfsync_meta_only_cmp; 698 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, vfsync_meta_only_cmp, 699 vfsync_bp, &info); 700 error = 0; 701 break; 702 default: 703 /* 704 * Synchronous. Do a data-only pass, then a meta-data+data 705 * pass, then additional integrated passes to try to get 706 * all the dependancies flushed. 707 */ 708 info.cmpfunc = vfsync_data_only_cmp; 709 info.fastpass = 1; 710 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, vfsync_data_only_cmp, 711 vfsync_bp, &info); 712 info.fastpass = 0; 713 error = vfsync_wait_output(vp, waitoutput); 714 if (error == 0) { 715 info.skippedbufs = 0; 716 info.cmpfunc = vfsync_dummy_cmp; 717 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 718 vfsync_bp, &info); 719 error = vfsync_wait_output(vp, waitoutput); 720 if (info.skippedbufs) { 721 kprintf("Warning: vfsync skipped %d dirty " 722 "buf%s in pass2!\n", 723 info.skippedbufs, 724 ((info.skippedbufs > 1) ? "s" : "")); 725 } 726 } 727 while (error == 0 && passes > 0 && 728 !RB_EMPTY(&vp->v_rbdirty_tree) 729 ) { 730 info.skippedbufs = 0; 731 if (--passes == 0) { 732 info.synchronous = 1; 733 info.syncdeps = 1; 734 } 735 info.cmpfunc = vfsync_dummy_cmp; 736 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 737 vfsync_bp, &info); 738 if (error < 0) 739 error = -error; 740 info.syncdeps = 1; 741 if (error == 0) 742 error = vfsync_wait_output(vp, waitoutput); 743 if (info.skippedbufs && passes == 0) { 744 kprintf("Warning: vfsync skipped %d dirty " 745 "buf%s in final pass!\n", 746 info.skippedbufs, 747 ((info.skippedbufs > 1) ? "s" : "")); 748 } 749 } 750 #if 0 751 /* 752 * This case can occur normally because vnode lock might 753 * not be held. 754 */ 755 if (!RB_EMPTY(&vp->v_rbdirty_tree)) 756 kprintf("dirty bufs left after final pass\n"); 757 #endif 758 break; 759 } 760 lwkt_reltoken(&vp->v_token); 761 762 return(error); 763 } 764 765 static int 766 vfsync_wait_output(struct vnode *vp, 767 int (*waitoutput)(struct vnode *, struct thread *)) 768 { 769 int error; 770 771 error = bio_track_wait(&vp->v_track_write, 0, 0); 772 if (waitoutput) 773 error = waitoutput(vp, curthread); 774 return(error); 775 } 776 777 static int 778 vfsync_dummy_cmp(struct buf *bp __unused, void *data __unused) 779 { 780 return(0); 781 } 782 783 static int 784 vfsync_data_only_cmp(struct buf *bp, void *data) 785 { 786 if (bp->b_loffset < 0) 787 return(-1); 788 return(0); 789 } 790 791 static int 792 vfsync_meta_only_cmp(struct buf *bp, void *data) 793 { 794 if (bp->b_loffset < 0) 795 return(0); 796 return(1); 797 } 798 799 static int 800 vfsync_lazy_range_cmp(struct buf *bp, void *data) 801 { 802 struct vfsync_info *info = data; 803 804 if (bp->b_loffset < info->vp->v_lazyw) 805 return(-1); 806 return(0); 807 } 808 809 static int 810 vfsync_bp(struct buf *bp, void *data) 811 { 812 struct vfsync_info *info = data; 813 struct vnode *vp = info->vp; 814 int error; 815 816 if (info->fastpass) { 817 /* 818 * Ignore buffers that we cannot immediately lock. 819 */ 820 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 821 /* 822 * Removed BUF_TIMELOCK(..., 1), even a 1-tick 823 * delay can mess up performance 824 * 825 * Another reason is that during a dirty-buffer 826 * scan a clustered write can start I/O on buffers 827 * ahead of the scan, causing the scan to not 828 * get a lock here. Usually this means the write 829 * is already in progress so, in fact, we *want* 830 * to skip the buffer. 831 */ 832 ++info->skippedbufs; 833 return(0); 834 } 835 } else if (info->synchronous == 0) { 836 /* 837 * Normal pass, give the buffer a little time to become 838 * available to us. 839 */ 840 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE, "bflst2", hz / 10)) { 841 ++info->skippedbufs; 842 return(0); 843 } 844 } else { 845 /* 846 * Synchronous pass, give the buffer a lot of time before 847 * giving up. 848 */ 849 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE, "bflst3", hz * 10)) { 850 ++info->skippedbufs; 851 return(0); 852 } 853 } 854 855 /* 856 * We must revalidate the buffer after locking. 857 */ 858 if ((bp->b_flags & B_DELWRI) == 0 || 859 bp->b_vp != info->vp || 860 info->cmpfunc(bp, data)) { 861 BUF_UNLOCK(bp); 862 return(0); 863 } 864 865 /* 866 * If syncdeps is not set we do not try to write buffers which have 867 * dependancies. 868 */ 869 if (!info->synchronous && info->syncdeps == 0 && info->checkdef(bp)) { 870 BUF_UNLOCK(bp); 871 return(0); 872 } 873 874 /* 875 * B_NEEDCOMMIT (primarily used by NFS) is a state where the buffer 876 * has been written but an additional handshake with the device 877 * is required before we can dispose of the buffer. We have no idea 878 * how to do this so we have to skip these buffers. 879 */ 880 if (bp->b_flags & B_NEEDCOMMIT) { 881 BUF_UNLOCK(bp); 882 return(0); 883 } 884 885 /* 886 * Ask bioops if it is ok to sync. If not the VFS may have 887 * set B_LOCKED so we have to cycle the buffer. 888 */ 889 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) { 890 bremfree(bp); 891 brelse(bp); 892 return(0); 893 } 894 895 if (info->synchronous) { 896 /* 897 * Synchronous flush. An error may be returned and will 898 * stop the scan. 899 */ 900 bremfree(bp); 901 error = bwrite(bp); 902 } else { 903 /* 904 * Asynchronous flush. We use the error return to support 905 * MNT_LAZY flushes. 906 * 907 * In low-memory situations we revert to synchronous 908 * operation. This should theoretically prevent the I/O 909 * path from exhausting memory in a non-recoverable way. 910 */ 911 vp->v_lazyw = bp->b_loffset; 912 bremfree(bp); 913 if (vm_page_count_min(0)) { 914 /* low memory */ 915 info->lazycount += bp->b_bufsize; 916 bwrite(bp); 917 } else { 918 /* normal */ 919 info->lazycount += cluster_awrite(bp); 920 waitrunningbufspace(); 921 /*vm_wait_nominal();*/ 922 } 923 if (info->lazylimit && info->lazycount >= info->lazylimit) 924 error = 1; 925 else 926 error = 0; 927 } 928 return(-error); 929 } 930 931 /* 932 * Associate a buffer with a vnode. 933 * 934 * MPSAFE 935 */ 936 int 937 bgetvp(struct vnode *vp, struct buf *bp, int testsize) 938 { 939 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 940 KKASSERT((bp->b_flags & (B_HASHED|B_DELWRI|B_VNCLEAN|B_VNDIRTY)) == 0); 941 942 /* 943 * Insert onto list for new vnode. 944 */ 945 lwkt_gettoken(&vp->v_token); 946 947 if (buf_rb_hash_RB_INSERT(&vp->v_rbhash_tree, bp)) { 948 lwkt_reltoken(&vp->v_token); 949 return (EEXIST); 950 } 951 952 /* 953 * Diagnostics (mainly for HAMMER debugging). Check for 954 * overlapping buffers. 955 */ 956 if (check_buf_overlap) { 957 struct buf *bx; 958 bx = buf_rb_hash_RB_PREV(bp); 959 if (bx) { 960 if (bx->b_loffset + bx->b_bufsize > bp->b_loffset) { 961 kprintf("bgetvp: overlapl %016jx/%d %016jx " 962 "bx %p bp %p\n", 963 (intmax_t)bx->b_loffset, 964 bx->b_bufsize, 965 (intmax_t)bp->b_loffset, 966 bx, bp); 967 if (check_buf_overlap > 1) 968 panic("bgetvp - overlapping buffer"); 969 } 970 } 971 bx = buf_rb_hash_RB_NEXT(bp); 972 if (bx) { 973 if (bp->b_loffset + testsize > bx->b_loffset) { 974 kprintf("bgetvp: overlapr %016jx/%d %016jx " 975 "bp %p bx %p\n", 976 (intmax_t)bp->b_loffset, 977 testsize, 978 (intmax_t)bx->b_loffset, 979 bp, bx); 980 if (check_buf_overlap > 1) 981 panic("bgetvp - overlapping buffer"); 982 } 983 } 984 } 985 bp->b_vp = vp; 986 bp->b_flags |= B_HASHED; 987 bp->b_flags |= B_VNCLEAN; 988 if (buf_rb_tree_RB_INSERT(&vp->v_rbclean_tree, bp)) 989 panic("reassignbuf: dup lblk/clean vp %p bp %p", vp, bp); 990 /*vhold(vp);*/ 991 lwkt_reltoken(&vp->v_token); 992 return(0); 993 } 994 995 /* 996 * Disassociate a buffer from a vnode. 997 * 998 * MPSAFE 999 */ 1000 void 1001 brelvp(struct buf *bp) 1002 { 1003 struct vnode *vp; 1004 1005 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1006 1007 /* 1008 * Delete from old vnode list, if on one. 1009 */ 1010 vp = bp->b_vp; 1011 lwkt_gettoken(&vp->v_token); 1012 if (bp->b_flags & (B_VNDIRTY | B_VNCLEAN)) { 1013 if (bp->b_flags & B_VNDIRTY) 1014 buf_rb_tree_RB_REMOVE(&vp->v_rbdirty_tree, bp); 1015 else 1016 buf_rb_tree_RB_REMOVE(&vp->v_rbclean_tree, bp); 1017 bp->b_flags &= ~(B_VNDIRTY | B_VNCLEAN); 1018 } 1019 if (bp->b_flags & B_HASHED) { 1020 buf_rb_hash_RB_REMOVE(&vp->v_rbhash_tree, bp); 1021 bp->b_flags &= ~B_HASHED; 1022 } 1023 1024 /* 1025 * Only remove from synclist when no dirty buffers are left AND 1026 * the VFS has not flagged the vnode's inode as being dirty. 1027 */ 1028 if ((vp->v_flag & (VONWORKLST | VISDIRTY | VOBJDIRTY)) == VONWORKLST && 1029 RB_EMPTY(&vp->v_rbdirty_tree)) { 1030 vn_syncer_remove(vp, 0); 1031 } 1032 bp->b_vp = NULL; 1033 1034 lwkt_reltoken(&vp->v_token); 1035 1036 /*vdrop(vp);*/ 1037 } 1038 1039 /* 1040 * Reassign the buffer to the proper clean/dirty list based on B_DELWRI. 1041 * This routine is called when the state of the B_DELWRI bit is changed. 1042 * 1043 * Must be called with vp->v_token held. 1044 * MPSAFE 1045 */ 1046 void 1047 reassignbuf(struct buf *bp) 1048 { 1049 struct vnode *vp = bp->b_vp; 1050 int delay; 1051 1052 ASSERT_LWKT_TOKEN_HELD(&vp->v_token); 1053 ++reassignbufcalls; 1054 1055 /* 1056 * B_PAGING flagged buffers cannot be reassigned because their vp 1057 * is not fully linked in. 1058 */ 1059 if (bp->b_flags & B_PAGING) 1060 panic("cannot reassign paging buffer"); 1061 1062 if (bp->b_flags & B_DELWRI) { 1063 /* 1064 * Move to the dirty list, add the vnode to the worklist 1065 */ 1066 if (bp->b_flags & B_VNCLEAN) { 1067 buf_rb_tree_RB_REMOVE(&vp->v_rbclean_tree, bp); 1068 bp->b_flags &= ~B_VNCLEAN; 1069 } 1070 if ((bp->b_flags & B_VNDIRTY) == 0) { 1071 if (buf_rb_tree_RB_INSERT(&vp->v_rbdirty_tree, bp)) { 1072 panic("reassignbuf: dup lblk vp %p bp %p", 1073 vp, bp); 1074 } 1075 bp->b_flags |= B_VNDIRTY; 1076 } 1077 if ((vp->v_flag & VONWORKLST) == 0) { 1078 switch (vp->v_type) { 1079 case VDIR: 1080 delay = dirdelay; 1081 break; 1082 case VCHR: 1083 case VBLK: 1084 if (vp->v_rdev && 1085 vp->v_rdev->si_mountpoint != NULL) { 1086 delay = metadelay; 1087 break; 1088 } 1089 /* fall through */ 1090 default: 1091 delay = filedelay; 1092 } 1093 vn_syncer_add(vp, delay); 1094 } 1095 } else { 1096 /* 1097 * Move to the clean list, remove the vnode from the worklist 1098 * if no dirty blocks remain. 1099 */ 1100 if (bp->b_flags & B_VNDIRTY) { 1101 buf_rb_tree_RB_REMOVE(&vp->v_rbdirty_tree, bp); 1102 bp->b_flags &= ~B_VNDIRTY; 1103 } 1104 if ((bp->b_flags & B_VNCLEAN) == 0) { 1105 if (buf_rb_tree_RB_INSERT(&vp->v_rbclean_tree, bp)) { 1106 panic("reassignbuf: dup lblk vp %p bp %p", 1107 vp, bp); 1108 } 1109 bp->b_flags |= B_VNCLEAN; 1110 } 1111 1112 /* 1113 * Only remove from synclist when no dirty buffers are left 1114 * AND the VFS has not flagged the vnode's inode as being 1115 * dirty. 1116 */ 1117 if ((vp->v_flag & (VONWORKLST | VISDIRTY | VOBJDIRTY)) == 1118 VONWORKLST && 1119 RB_EMPTY(&vp->v_rbdirty_tree)) { 1120 vn_syncer_remove(vp, 0); 1121 } 1122 } 1123 } 1124 1125 /* 1126 * Create a vnode for a block device. Used for mounting the root file 1127 * system. 1128 * 1129 * A vref()'d vnode is returned. 1130 */ 1131 extern struct vop_ops *devfs_vnode_dev_vops_p; 1132 int 1133 bdevvp(cdev_t dev, struct vnode **vpp) 1134 { 1135 struct vnode *vp; 1136 struct vnode *nvp; 1137 int error; 1138 1139 if (dev == NULL) { 1140 *vpp = NULLVP; 1141 return (ENXIO); 1142 } 1143 error = getspecialvnode(VT_NON, NULL, &devfs_vnode_dev_vops_p, 1144 &nvp, 0, 0); 1145 if (error) { 1146 *vpp = NULLVP; 1147 return (error); 1148 } 1149 vp = nvp; 1150 vp->v_type = VCHR; 1151 #if 0 1152 vp->v_rdev = dev; 1153 #endif 1154 v_associate_rdev(vp, dev); 1155 vp->v_umajor = dev->si_umajor; 1156 vp->v_uminor = dev->si_uminor; 1157 vx_unlock(vp); 1158 *vpp = vp; 1159 return (0); 1160 } 1161 1162 int 1163 v_associate_rdev(struct vnode *vp, cdev_t dev) 1164 { 1165 if (dev == NULL) 1166 return(ENXIO); 1167 if (dev_is_good(dev) == 0) 1168 return(ENXIO); 1169 KKASSERT(vp->v_rdev == NULL); 1170 vp->v_rdev = reference_dev(dev); 1171 lwkt_gettoken(&spechash_token); 1172 SLIST_INSERT_HEAD(&dev->si_hlist, vp, v_cdevnext); 1173 lwkt_reltoken(&spechash_token); 1174 return(0); 1175 } 1176 1177 void 1178 v_release_rdev(struct vnode *vp) 1179 { 1180 cdev_t dev; 1181 1182 if ((dev = vp->v_rdev) != NULL) { 1183 lwkt_gettoken(&spechash_token); 1184 SLIST_REMOVE(&dev->si_hlist, vp, vnode, v_cdevnext); 1185 vp->v_rdev = NULL; 1186 release_dev(dev); 1187 lwkt_reltoken(&spechash_token); 1188 } 1189 } 1190 1191 /* 1192 * Add a vnode to the alias list hung off the cdev_t. We only associate 1193 * the device number with the vnode. The actual device is not associated 1194 * until the vnode is opened (usually in spec_open()), and will be 1195 * disassociated on last close. 1196 */ 1197 void 1198 addaliasu(struct vnode *nvp, int x, int y) 1199 { 1200 if (nvp->v_type != VBLK && nvp->v_type != VCHR) 1201 panic("addaliasu on non-special vnode"); 1202 nvp->v_umajor = x; 1203 nvp->v_uminor = y; 1204 } 1205 1206 /* 1207 * Simple call that a filesystem can make to try to get rid of a 1208 * vnode. It will fail if anyone is referencing the vnode (including 1209 * the caller). 1210 * 1211 * The filesystem can check whether its in-memory inode structure still 1212 * references the vp on return. 1213 * 1214 * May only be called if the vnode is in a known state (i.e. being prevented 1215 * from being deallocated by some other condition such as a vfs inode hold). 1216 */ 1217 void 1218 vclean_unlocked(struct vnode *vp) 1219 { 1220 vx_get(vp); 1221 if (VREFCNT(vp) <= 1) 1222 vgone_vxlocked(vp); 1223 vx_put(vp); 1224 } 1225 1226 /* 1227 * Disassociate a vnode from its underlying filesystem. 1228 * 1229 * The vnode must be VX locked and referenced. In all normal situations 1230 * there are no active references. If vclean_vxlocked() is called while 1231 * there are active references, the vnode is being ripped out and we have 1232 * to call VOP_CLOSE() as appropriate before we can reclaim it. 1233 */ 1234 void 1235 vclean_vxlocked(struct vnode *vp, int flags) 1236 { 1237 int active; 1238 int n; 1239 vm_object_t object; 1240 struct namecache *ncp; 1241 1242 /* 1243 * If the vnode has already been reclaimed we have nothing to do. 1244 */ 1245 if (vp->v_flag & VRECLAIMED) 1246 return; 1247 1248 /* 1249 * Set flag to interlock operation, flag finalization to ensure 1250 * that the vnode winds up on the inactive list, and set v_act to 0. 1251 */ 1252 vsetflags(vp, VRECLAIMED); 1253 atomic_set_int(&vp->v_refcnt, VREF_FINALIZE); 1254 vp->v_act = 0; 1255 1256 if (verbose_reclaims) { 1257 if ((ncp = TAILQ_FIRST(&vp->v_namecache)) != NULL) 1258 kprintf("Debug: reclaim %p %s\n", vp, ncp->nc_name); 1259 } 1260 1261 /* 1262 * Scrap the vfs cache 1263 */ 1264 while (cache_inval_vp(vp, 0) != 0) { 1265 kprintf("Warning: vnode %p clean/cache_resolution " 1266 "race detected\n", vp); 1267 tsleep(vp, 0, "vclninv", 2); 1268 } 1269 1270 /* 1271 * Check to see if the vnode is in use. If so we have to reference it 1272 * before we clean it out so that its count cannot fall to zero and 1273 * generate a race against ourselves to recycle it. 1274 */ 1275 active = (VREFCNT(vp) > 0); 1276 1277 /* 1278 * Clean out any buffers associated with the vnode and destroy its 1279 * object, if it has one. 1280 */ 1281 vinvalbuf(vp, V_SAVE, 0, 0); 1282 1283 /* 1284 * If purging an active vnode (typically during a forced unmount 1285 * or reboot), it must be closed and deactivated before being 1286 * reclaimed. This isn't really all that safe, but what can 1287 * we do? XXX. 1288 * 1289 * Note that neither of these routines unlocks the vnode. 1290 */ 1291 if (active && (flags & DOCLOSE)) { 1292 while ((n = vp->v_opencount) != 0) { 1293 if (vp->v_writecount) 1294 VOP_CLOSE(vp, FWRITE|FNONBLOCK, NULL); 1295 else 1296 VOP_CLOSE(vp, FNONBLOCK, NULL); 1297 if (vp->v_opencount == n) { 1298 kprintf("Warning: unable to force-close" 1299 " vnode %p\n", vp); 1300 break; 1301 } 1302 } 1303 } 1304 1305 /* 1306 * If the vnode has not been deactivated, deactivated it. Deactivation 1307 * can create new buffers and VM pages so we have to call vinvalbuf() 1308 * again to make sure they all get flushed. 1309 * 1310 * This can occur if a file with a link count of 0 needs to be 1311 * truncated. 1312 * 1313 * If the vnode is already dead don't try to deactivate it. 1314 */ 1315 if ((vp->v_flag & VINACTIVE) == 0) { 1316 vsetflags(vp, VINACTIVE); 1317 if (vp->v_mount) 1318 VOP_INACTIVE(vp); 1319 vinvalbuf(vp, V_SAVE, 0, 0); 1320 } 1321 1322 /* 1323 * If the vnode has an object, destroy it. 1324 */ 1325 while ((object = vp->v_object) != NULL) { 1326 vm_object_hold(object); 1327 if (object == vp->v_object) 1328 break; 1329 vm_object_drop(object); 1330 } 1331 1332 if (object != NULL) { 1333 if (object->ref_count == 0) { 1334 if ((object->flags & OBJ_DEAD) == 0) 1335 vm_object_terminate(object); 1336 vm_object_drop(object); 1337 vclrflags(vp, VOBJBUF); 1338 } else { 1339 vm_pager_deallocate(object); 1340 vclrflags(vp, VOBJBUF); 1341 vm_object_drop(object); 1342 } 1343 } 1344 KKASSERT((vp->v_flag & VOBJBUF) == 0); 1345 1346 if (vp->v_flag & VOBJDIRTY) 1347 vclrobjdirty(vp); 1348 1349 /* 1350 * Reclaim the vnode if not already dead. 1351 */ 1352 if (vp->v_mount && VOP_RECLAIM(vp)) 1353 panic("vclean: cannot reclaim"); 1354 1355 /* 1356 * Done with purge, notify sleepers of the grim news. 1357 */ 1358 vp->v_ops = &dead_vnode_vops_p; 1359 vn_gone(vp); 1360 vp->v_tag = VT_NON; 1361 1362 /* 1363 * If we are destroying an active vnode, reactivate it now that 1364 * we have reassociated it with deadfs. This prevents the system 1365 * from crashing on the vnode due to it being unexpectedly marked 1366 * as inactive or reclaimed. 1367 */ 1368 if (active && (flags & DOCLOSE)) { 1369 vclrflags(vp, VINACTIVE | VRECLAIMED); 1370 } 1371 } 1372 1373 /* 1374 * Eliminate all activity associated with the requested vnode 1375 * and with all vnodes aliased to the requested vnode. 1376 * 1377 * The vnode must be referenced but should not be locked. 1378 */ 1379 int 1380 vrevoke(struct vnode *vp, struct ucred *cred) 1381 { 1382 struct vnode *vq; 1383 struct vnode *vqn; 1384 cdev_t dev; 1385 int error; 1386 1387 /* 1388 * If the vnode has a device association, scrap all vnodes associated 1389 * with the device. Don't let the device disappear on us while we 1390 * are scrapping the vnodes. 1391 * 1392 * The passed vp will probably show up in the list, do not VX lock 1393 * it twice! 1394 * 1395 * Releasing the vnode's rdev here can mess up specfs's call to 1396 * device close, so don't do it. The vnode has been disassociated 1397 * and the device will be closed after the last ref on the related 1398 * fp goes away (if not still open by e.g. the kernel). 1399 */ 1400 if (vp->v_type != VCHR) { 1401 error = fdrevoke(vp, DTYPE_VNODE, cred); 1402 return (error); 1403 } 1404 if ((dev = vp->v_rdev) == NULL) { 1405 return(0); 1406 } 1407 reference_dev(dev); 1408 lwkt_gettoken(&spechash_token); 1409 1410 restart: 1411 vqn = SLIST_FIRST(&dev->si_hlist); 1412 if (vqn) 1413 vhold(vqn); 1414 while ((vq = vqn) != NULL) { 1415 if (VREFCNT(vq) > 0) { 1416 vref(vq); 1417 fdrevoke(vq, DTYPE_VNODE, cred); 1418 /*v_release_rdev(vq);*/ 1419 vrele(vq); 1420 if (vq->v_rdev != dev) { 1421 vdrop(vq); 1422 goto restart; 1423 } 1424 } 1425 vqn = SLIST_NEXT(vq, v_cdevnext); 1426 if (vqn) 1427 vhold(vqn); 1428 vdrop(vq); 1429 } 1430 lwkt_reltoken(&spechash_token); 1431 dev_drevoke(dev); 1432 release_dev(dev); 1433 return (0); 1434 } 1435 1436 /* 1437 * This is called when the object underlying a vnode is being destroyed, 1438 * such as in a remove(). Try to recycle the vnode immediately if the 1439 * only active reference is our reference. 1440 * 1441 * Directory vnodes in the namecache with children cannot be immediately 1442 * recycled because numerous VOP_N*() ops require them to be stable. 1443 * 1444 * To avoid recursive recycling from VOP_INACTIVE implemenetations this 1445 * function is a NOP if VRECLAIMED is already set. 1446 */ 1447 int 1448 vrecycle(struct vnode *vp) 1449 { 1450 if (VREFCNT(vp) <= 1 && (vp->v_flag & VRECLAIMED) == 0) { 1451 if (cache_inval_vp_nonblock(vp)) 1452 return(0); 1453 vgone_vxlocked(vp); 1454 return (1); 1455 } 1456 return (0); 1457 } 1458 1459 /* 1460 * Return the maximum I/O size allowed for strategy calls on VP. 1461 * 1462 * If vp is VCHR or VBLK we dive the device, otherwise we use 1463 * the vp's mount info. 1464 * 1465 * The returned value is clamped at MAXPHYS as most callers cannot use 1466 * buffers larger than that size. 1467 */ 1468 int 1469 vmaxiosize(struct vnode *vp) 1470 { 1471 int maxiosize; 1472 1473 if (vp->v_type == VBLK || vp->v_type == VCHR) 1474 maxiosize = vp->v_rdev->si_iosize_max; 1475 else 1476 maxiosize = vp->v_mount->mnt_iosize_max; 1477 1478 if (maxiosize > MAXPHYS) 1479 maxiosize = MAXPHYS; 1480 return (maxiosize); 1481 } 1482 1483 /* 1484 * Eliminate all activity associated with a vnode in preparation for 1485 * destruction. 1486 * 1487 * The vnode must be VX locked and refd and will remain VX locked and refd 1488 * on return. This routine may be called with the vnode in any state, as 1489 * long as it is VX locked. The vnode will be cleaned out and marked 1490 * VRECLAIMED but will not actually be reused until all existing refs and 1491 * holds go away. 1492 * 1493 * NOTE: This routine may be called on a vnode which has not yet been 1494 * already been deactivated (VOP_INACTIVE), or on a vnode which has 1495 * already been reclaimed. 1496 * 1497 * This routine is not responsible for placing us back on the freelist. 1498 * Instead, it happens automatically when the caller releases the VX lock 1499 * (assuming there aren't any other references). 1500 */ 1501 void 1502 vgone_vxlocked(struct vnode *vp) 1503 { 1504 /* 1505 * assert that the VX lock is held. This is an absolute requirement 1506 * now for vgone_vxlocked() to be called. 1507 */ 1508 KKASSERT(lockinuse(&vp->v_lock)); 1509 1510 /* 1511 * Clean out the filesystem specific data and set the VRECLAIMED 1512 * bit. Also deactivate the vnode if necessary. 1513 * 1514 * The vnode should have automatically been removed from the syncer 1515 * list as syncer/dirty flags cleared during the cleaning. 1516 */ 1517 vclean_vxlocked(vp, DOCLOSE); 1518 1519 /* 1520 * Normally panic if the vnode is still dirty, unless we are doing 1521 * a forced unmount (tmpfs typically). 1522 */ 1523 if (vp->v_flag & VONWORKLST) { 1524 if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) { 1525 /* force removal */ 1526 vn_syncer_remove(vp, 1); 1527 } else { 1528 panic("vp %p still dirty in vgone after flush", vp); 1529 } 1530 } 1531 1532 /* 1533 * Delete from old mount point vnode list, if on one. 1534 */ 1535 if (vp->v_mount != NULL) { 1536 KKASSERT(vp->v_data == NULL); 1537 insmntque(vp, NULL); 1538 } 1539 1540 /* 1541 * If special device, remove it from special device alias list 1542 * if it is on one. This should normally only occur if a vnode is 1543 * being revoked as the device should otherwise have been released 1544 * naturally. 1545 */ 1546 if ((vp->v_type == VBLK || vp->v_type == VCHR) && vp->v_rdev != NULL) { 1547 v_release_rdev(vp); 1548 } 1549 1550 /* 1551 * Set us to VBAD 1552 */ 1553 vp->v_type = VBAD; 1554 } 1555 1556 /* 1557 * Lookup a vnode by device number. 1558 * 1559 * Returns non-zero and *vpp set to a vref'd vnode on success. 1560 * Returns zero on failure. 1561 */ 1562 int 1563 vfinddev(cdev_t dev, enum vtype type, struct vnode **vpp) 1564 { 1565 struct vnode *vp; 1566 1567 lwkt_gettoken(&spechash_token); 1568 SLIST_FOREACH(vp, &dev->si_hlist, v_cdevnext) { 1569 if (type == vp->v_type) { 1570 *vpp = vp; 1571 vref(vp); 1572 lwkt_reltoken(&spechash_token); 1573 return (1); 1574 } 1575 } 1576 lwkt_reltoken(&spechash_token); 1577 return (0); 1578 } 1579 1580 /* 1581 * Calculate the total number of references to a special device. This 1582 * routine may only be called for VBLK and VCHR vnodes since v_rdev is 1583 * an overloaded field. Since udev2dev can now return NULL, we have 1584 * to check for a NULL v_rdev. 1585 */ 1586 int 1587 count_dev(cdev_t dev) 1588 { 1589 struct vnode *vp; 1590 int count = 0; 1591 1592 if (SLIST_FIRST(&dev->si_hlist)) { 1593 lwkt_gettoken(&spechash_token); 1594 SLIST_FOREACH(vp, &dev->si_hlist, v_cdevnext) { 1595 count += vp->v_opencount; 1596 } 1597 lwkt_reltoken(&spechash_token); 1598 } 1599 return(count); 1600 } 1601 1602 int 1603 vcount(struct vnode *vp) 1604 { 1605 if (vp->v_rdev == NULL) 1606 return(0); 1607 return(count_dev(vp->v_rdev)); 1608 } 1609 1610 /* 1611 * Initialize VMIO for a vnode. This routine MUST be called before a 1612 * VFS can issue buffer cache ops on a vnode. It is typically called 1613 * when a vnode is initialized from its inode. 1614 */ 1615 int 1616 vinitvmio(struct vnode *vp, off_t filesize, int blksize, int boff) 1617 { 1618 vm_object_t object; 1619 int error = 0; 1620 1621 object = vp->v_object; 1622 if (object) { 1623 vm_object_hold(object); 1624 KKASSERT(vp->v_object == object); 1625 } 1626 1627 if (object == NULL) { 1628 object = vnode_pager_alloc(vp, filesize, 0, 0, blksize, boff); 1629 1630 /* 1631 * Dereference the reference we just created. This assumes 1632 * that the object is associated with the vp. Allow it to 1633 * have zero refs. It cannot be destroyed as long as it 1634 * is associated with the vnode. 1635 */ 1636 vm_object_hold(object); 1637 atomic_add_int(&object->ref_count, -1); 1638 vrele(vp); 1639 } else { 1640 KKASSERT((object->flags & OBJ_DEAD) == 0); 1641 } 1642 KASSERT(vp->v_object != NULL, ("vinitvmio: NULL object")); 1643 vsetflags(vp, VOBJBUF); 1644 vm_object_drop(object); 1645 1646 return (error); 1647 } 1648 1649 1650 /* 1651 * Print out a description of a vnode. 1652 */ 1653 static char *typename[] = 1654 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 1655 1656 void 1657 vprint(char *label, struct vnode *vp) 1658 { 1659 char buf[96]; 1660 1661 if (label != NULL) 1662 kprintf("%s: %p: ", label, (void *)vp); 1663 else 1664 kprintf("%p: ", (void *)vp); 1665 kprintf("type %s, refcnt %08x, writecount %d, holdcnt %d,", 1666 typename[vp->v_type], 1667 vp->v_refcnt, vp->v_writecount, vp->v_auxrefs); 1668 buf[0] = '\0'; 1669 if (vp->v_flag & VROOT) 1670 strcat(buf, "|VROOT"); 1671 if (vp->v_flag & VPFSROOT) 1672 strcat(buf, "|VPFSROOT"); 1673 if (vp->v_flag & VTEXT) 1674 strcat(buf, "|VTEXT"); 1675 if (vp->v_flag & VSYSTEM) 1676 strcat(buf, "|VSYSTEM"); 1677 if (vp->v_flag & VOBJBUF) 1678 strcat(buf, "|VOBJBUF"); 1679 if (buf[0] != '\0') 1680 kprintf(" flags (%s)", &buf[1]); 1681 if (vp->v_data == NULL) { 1682 kprintf("\n"); 1683 } else { 1684 kprintf("\n\t"); 1685 VOP_PRINT(vp); 1686 } 1687 } 1688 1689 /* 1690 * Do the usual access checking. 1691 * file_mode, uid and gid are from the vnode in question, 1692 * while acc_mode and cred are from the VOP_ACCESS parameter list 1693 */ 1694 int 1695 vaccess(enum vtype type, mode_t file_mode, uid_t uid, gid_t gid, 1696 mode_t acc_mode, struct ucred *cred) 1697 { 1698 mode_t mask; 1699 int ismember; 1700 1701 /* 1702 * Super-user always gets read/write access, but execute access depends 1703 * on at least one execute bit being set. 1704 */ 1705 if (priv_check_cred(cred, PRIV_ROOT, 0) == 0) { 1706 if ((acc_mode & VEXEC) && type != VDIR && 1707 (file_mode & (S_IXUSR|S_IXGRP|S_IXOTH)) == 0) 1708 return (EACCES); 1709 return (0); 1710 } 1711 1712 mask = 0; 1713 1714 /* Otherwise, check the owner. */ 1715 if (cred->cr_uid == uid) { 1716 if (acc_mode & VEXEC) 1717 mask |= S_IXUSR; 1718 if (acc_mode & VREAD) 1719 mask |= S_IRUSR; 1720 if (acc_mode & VWRITE) 1721 mask |= S_IWUSR; 1722 return ((file_mode & mask) == mask ? 0 : EACCES); 1723 } 1724 1725 /* Otherwise, check the groups. */ 1726 ismember = groupmember(gid, cred); 1727 if (cred->cr_svgid == gid || ismember) { 1728 if (acc_mode & VEXEC) 1729 mask |= S_IXGRP; 1730 if (acc_mode & VREAD) 1731 mask |= S_IRGRP; 1732 if (acc_mode & VWRITE) 1733 mask |= S_IWGRP; 1734 return ((file_mode & mask) == mask ? 0 : EACCES); 1735 } 1736 1737 /* Otherwise, check everyone else. */ 1738 if (acc_mode & VEXEC) 1739 mask |= S_IXOTH; 1740 if (acc_mode & VREAD) 1741 mask |= S_IROTH; 1742 if (acc_mode & VWRITE) 1743 mask |= S_IWOTH; 1744 return ((file_mode & mask) == mask ? 0 : EACCES); 1745 } 1746 1747 #ifdef DDB 1748 #include <ddb/ddb.h> 1749 1750 static int db_show_locked_vnodes(struct mount *mp, void *data); 1751 1752 /* 1753 * List all of the locked vnodes in the system. 1754 * Called when debugging the kernel. 1755 */ 1756 DB_SHOW_COMMAND(lockedvnodes, lockedvnodes) 1757 { 1758 kprintf("Locked vnodes\n"); 1759 mountlist_scan(db_show_locked_vnodes, NULL, 1760 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 1761 } 1762 1763 static int 1764 db_show_locked_vnodes(struct mount *mp, void *data __unused) 1765 { 1766 struct vnode *vp; 1767 1768 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 1769 if (vn_islocked(vp)) 1770 vprint(NULL, vp); 1771 } 1772 return(0); 1773 } 1774 #endif 1775 1776 /* 1777 * Top level filesystem related information gathering. 1778 */ 1779 static int sysctl_ovfs_conf (SYSCTL_HANDLER_ARGS); 1780 1781 static int 1782 vfs_sysctl(SYSCTL_HANDLER_ARGS) 1783 { 1784 int *name = (int *)arg1 - 1; /* XXX */ 1785 u_int namelen = arg2 + 1; /* XXX */ 1786 struct vfsconf *vfsp; 1787 int maxtypenum; 1788 1789 #if 1 || defined(COMPAT_PRELITE2) 1790 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 1791 if (namelen == 1) 1792 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 1793 #endif 1794 1795 #ifdef notyet 1796 /* all sysctl names at this level are at least name and field */ 1797 if (namelen < 2) 1798 return (ENOTDIR); /* overloaded */ 1799 if (name[0] != VFS_GENERIC) { 1800 vfsp = vfsconf_find_by_typenum(name[0]); 1801 if (vfsp == NULL) 1802 return (EOPNOTSUPP); 1803 return ((*vfsp->vfc_vfsops->vfs_sysctl)(&name[1], namelen - 1, 1804 oldp, oldlenp, newp, newlen, p)); 1805 } 1806 #endif 1807 switch (name[1]) { 1808 case VFS_MAXTYPENUM: 1809 if (namelen != 2) 1810 return (ENOTDIR); 1811 maxtypenum = vfsconf_get_maxtypenum(); 1812 return (SYSCTL_OUT(req, &maxtypenum, sizeof(maxtypenum))); 1813 case VFS_CONF: 1814 if (namelen != 3) 1815 return (ENOTDIR); /* overloaded */ 1816 vfsp = vfsconf_find_by_typenum(name[2]); 1817 if (vfsp == NULL) 1818 return (EOPNOTSUPP); 1819 return (SYSCTL_OUT(req, vfsp, sizeof *vfsp)); 1820 } 1821 return (EOPNOTSUPP); 1822 } 1823 1824 SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD, vfs_sysctl, 1825 "Generic filesystem"); 1826 1827 #if 1 || defined(COMPAT_PRELITE2) 1828 1829 static int 1830 sysctl_ovfs_conf_iter(struct vfsconf *vfsp, void *data) 1831 { 1832 int error; 1833 struct ovfsconf ovfs; 1834 struct sysctl_req *req = (struct sysctl_req*) data; 1835 1836 bzero(&ovfs, sizeof(ovfs)); 1837 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 1838 strcpy(ovfs.vfc_name, vfsp->vfc_name); 1839 ovfs.vfc_index = vfsp->vfc_typenum; 1840 ovfs.vfc_refcount = vfsp->vfc_refcount; 1841 ovfs.vfc_flags = vfsp->vfc_flags; 1842 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 1843 if (error) 1844 return error; /* abort iteration with error code */ 1845 else 1846 return 0; /* continue iterating with next element */ 1847 } 1848 1849 static int 1850 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 1851 { 1852 return vfsconf_each(sysctl_ovfs_conf_iter, (void*)req); 1853 } 1854 1855 #endif /* 1 || COMPAT_PRELITE2 */ 1856 1857 /* 1858 * Check to see if a filesystem is mounted on a block device. 1859 */ 1860 int 1861 vfs_mountedon(struct vnode *vp) 1862 { 1863 cdev_t dev; 1864 1865 if ((dev = vp->v_rdev) == NULL) { 1866 /* if (vp->v_type != VBLK) 1867 dev = get_dev(vp->v_uminor, vp->v_umajor); */ 1868 } 1869 if (dev != NULL && dev->si_mountpoint) 1870 return (EBUSY); 1871 return (0); 1872 } 1873 1874 /* 1875 * Unmount all filesystems. The list is traversed in reverse order 1876 * of mounting to avoid dependencies. 1877 * 1878 * We want the umountall to be able to break out of its loop if a 1879 * failure occurs, after scanning all possible mounts, so the callback 1880 * returns 0 on error. 1881 * 1882 * NOTE: Do not call mountlist_remove(mp) on error any more, this will 1883 * confuse mountlist_scan()'s unbusy check. 1884 */ 1885 static int vfs_umountall_callback(struct mount *mp, void *data); 1886 1887 void 1888 vfs_unmountall(int halting) 1889 { 1890 int count; 1891 1892 do { 1893 count = mountlist_scan(vfs_umountall_callback, &halting, 1894 MNTSCAN_REVERSE|MNTSCAN_NOBUSY); 1895 } while (count); 1896 } 1897 1898 static 1899 int 1900 vfs_umountall_callback(struct mount *mp, void *data) 1901 { 1902 int error; 1903 int halting = *(int *)data; 1904 1905 /* 1906 * NOTE: When halting, dounmount will disconnect but leave 1907 * certain mount points intact. e.g. devfs. 1908 */ 1909 error = dounmount(mp, MNT_FORCE, halting); 1910 if (error) { 1911 kprintf("unmount of filesystem mounted from %s failed (", 1912 mp->mnt_stat.f_mntfromname); 1913 if (error == EBUSY) 1914 kprintf("BUSY)\n"); 1915 else 1916 kprintf("%d)\n", error); 1917 return 0; 1918 } else { 1919 return 1; 1920 } 1921 } 1922 1923 /* 1924 * Checks the mount flags for parameter mp and put the names comma-separated 1925 * into a string buffer buf with a size limit specified by len. 1926 * 1927 * It returns the number of bytes written into buf, and (*errorp) will be 1928 * set to 0, EINVAL (if passed length is 0), or ENOSPC (supplied buffer was 1929 * not large enough). The buffer will be 0-terminated if len was not 0. 1930 */ 1931 size_t 1932 vfs_flagstostr(int flags, const struct mountctl_opt *optp, 1933 char *buf, size_t len, int *errorp) 1934 { 1935 static const struct mountctl_opt optnames[] = { 1936 { MNT_RDONLY, "read-only" }, 1937 { MNT_SYNCHRONOUS, "synchronous" }, 1938 { MNT_NOEXEC, "noexec" }, 1939 { MNT_NOSUID, "nosuid" }, 1940 { MNT_NODEV, "nodev" }, 1941 { MNT_AUTOMOUNTED, "automounted" }, 1942 { MNT_ASYNC, "asynchronous" }, 1943 { MNT_SUIDDIR, "suiddir" }, 1944 { MNT_SOFTDEP, "soft-updates" }, 1945 { MNT_NOSYMFOLLOW, "nosymfollow" }, 1946 { MNT_TRIM, "trim" }, 1947 { MNT_NOATIME, "noatime" }, 1948 { MNT_NOCLUSTERR, "noclusterr" }, 1949 { MNT_NOCLUSTERW, "noclusterw" }, 1950 { MNT_EXRDONLY, "NFS read-only" }, 1951 { MNT_EXPORTED, "NFS exported" }, 1952 /* Remaining NFS flags could come here */ 1953 { MNT_LOCAL, "local" }, 1954 { MNT_QUOTA, "with-quotas" }, 1955 /* { MNT_ROOTFS, "rootfs" }, */ 1956 /* { MNT_IGNORE, "ignore" }, */ 1957 { 0, NULL} 1958 }; 1959 int bwritten; 1960 int bleft; 1961 int optlen; 1962 int actsize; 1963 1964 *errorp = 0; 1965 bwritten = 0; 1966 bleft = len - 1; /* leave room for trailing \0 */ 1967 1968 /* 1969 * Checks the size of the string. If it contains 1970 * any data, then we will append the new flags to 1971 * it. 1972 */ 1973 actsize = strlen(buf); 1974 if (actsize > 0) 1975 buf += actsize; 1976 1977 /* Default flags if no flags passed */ 1978 if (optp == NULL) 1979 optp = optnames; 1980 1981 if (bleft < 0) { /* degenerate case, 0-length buffer */ 1982 *errorp = EINVAL; 1983 return(0); 1984 } 1985 1986 for (; flags && optp->o_opt; ++optp) { 1987 if ((flags & optp->o_opt) == 0) 1988 continue; 1989 optlen = strlen(optp->o_name); 1990 if (bwritten || actsize > 0) { 1991 if (bleft < 2) { 1992 *errorp = ENOSPC; 1993 break; 1994 } 1995 buf[bwritten++] = ','; 1996 buf[bwritten++] = ' '; 1997 bleft -= 2; 1998 } 1999 if (bleft < optlen) { 2000 *errorp = ENOSPC; 2001 break; 2002 } 2003 bcopy(optp->o_name, buf + bwritten, optlen); 2004 bwritten += optlen; 2005 bleft -= optlen; 2006 flags &= ~optp->o_opt; 2007 } 2008 2009 /* 2010 * Space already reserved for trailing \0 2011 */ 2012 buf[bwritten] = 0; 2013 return (bwritten); 2014 } 2015 2016 /* 2017 * Build hash lists of net addresses and hang them off the mount point. 2018 * Called by ufs_mount() to set up the lists of export addresses. 2019 */ 2020 static int 2021 vfs_hang_addrlist(struct mount *mp, struct netexport *nep, 2022 const struct export_args *argp) 2023 { 2024 struct netcred *np; 2025 struct radix_node_head *rnh; 2026 int i; 2027 struct radix_node *rn; 2028 struct sockaddr *saddr, *smask = NULL; 2029 int error; 2030 2031 if (argp->ex_addrlen == 0) { 2032 if (mp->mnt_flag & MNT_DEFEXPORTED) 2033 return (EPERM); 2034 np = &nep->ne_defexported; 2035 np->netc_exflags = argp->ex_flags; 2036 np->netc_anon = argp->ex_anon; 2037 np->netc_anon.cr_ref = 1; 2038 mp->mnt_flag |= MNT_DEFEXPORTED; 2039 return (0); 2040 } 2041 2042 if (argp->ex_addrlen < 0 || argp->ex_addrlen > MLEN) 2043 return (EINVAL); 2044 if (argp->ex_masklen < 0 || argp->ex_masklen > MLEN) 2045 return (EINVAL); 2046 2047 i = sizeof(struct netcred) + argp->ex_addrlen + argp->ex_masklen; 2048 np = (struct netcred *)kmalloc(i, M_NETCRED, M_WAITOK | M_ZERO); 2049 saddr = (struct sockaddr *) (np + 1); 2050 if ((error = copyin(argp->ex_addr, (caddr_t) saddr, argp->ex_addrlen))) 2051 goto out; 2052 if (saddr->sa_len > argp->ex_addrlen) 2053 saddr->sa_len = argp->ex_addrlen; 2054 if (argp->ex_masklen) { 2055 smask = (struct sockaddr *)((caddr_t)saddr + argp->ex_addrlen); 2056 error = copyin(argp->ex_mask, (caddr_t)smask, argp->ex_masklen); 2057 if (error) 2058 goto out; 2059 if (smask->sa_len > argp->ex_masklen) 2060 smask->sa_len = argp->ex_masklen; 2061 } 2062 NE_LOCK(nep); 2063 if (nep->ne_maskhead == NULL) { 2064 if (!rn_inithead((void **)&nep->ne_maskhead, NULL, 0)) { 2065 error = ENOBUFS; 2066 goto out; 2067 } 2068 } 2069 if ((rnh = vfs_create_addrlist_af(saddr->sa_family, nep)) == NULL) { 2070 error = ENOBUFS; 2071 goto out; 2072 } 2073 rn = (*rnh->rnh_addaddr)((char *)saddr, (char *)smask, rnh, 2074 np->netc_rnodes); 2075 NE_UNLOCK(nep); 2076 if (rn == NULL || np != (struct netcred *)rn) { /* already exists */ 2077 error = EPERM; 2078 goto out; 2079 } 2080 np->netc_exflags = argp->ex_flags; 2081 np->netc_anon = argp->ex_anon; 2082 np->netc_anon.cr_ref = 1; 2083 return (0); 2084 2085 out: 2086 kfree(np, M_NETCRED); 2087 return (error); 2088 } 2089 2090 /* 2091 * Free netcred structures installed in the netexport 2092 */ 2093 static int 2094 vfs_free_netcred(struct radix_node *rn, void *w) 2095 { 2096 struct radix_node_head *rnh = (struct radix_node_head *)w; 2097 2098 (*rnh->rnh_deladdr) (rn->rn_key, rn->rn_mask, rnh); 2099 kfree(rn, M_NETCRED); 2100 2101 return (0); 2102 } 2103 2104 /* 2105 * callback to free an element of the mask table installed in the 2106 * netexport. These may be created indirectly and are not netcred 2107 * structures. 2108 */ 2109 static int 2110 vfs_free_netcred_mask(struct radix_node *rn, void *w) 2111 { 2112 struct radix_node_head *rnh = (struct radix_node_head *)w; 2113 2114 (*rnh->rnh_deladdr) (rn->rn_key, rn->rn_mask, rnh); 2115 kfree(rn, M_RTABLE); 2116 2117 return (0); 2118 } 2119 2120 static struct radix_node_head * 2121 vfs_create_addrlist_af(int af, struct netexport *nep) 2122 { 2123 struct radix_node_head *rnh = NULL; 2124 #if defined(INET) || defined(INET6) 2125 struct radix_node_head *maskhead = nep->ne_maskhead; 2126 int off; 2127 #endif 2128 2129 NE_ASSERT_LOCKED(nep); 2130 #if defined(INET) || defined(INET6) 2131 KKASSERT(maskhead != NULL); 2132 #endif 2133 switch (af) { 2134 #ifdef INET 2135 case AF_INET: 2136 if ((rnh = nep->ne_inethead) == NULL) { 2137 off = offsetof(struct sockaddr_in, sin_addr) << 3; 2138 if (!rn_inithead((void **)&rnh, maskhead, off)) 2139 return (NULL); 2140 nep->ne_inethead = rnh; 2141 } 2142 break; 2143 #endif 2144 #ifdef INET6 2145 case AF_INET6: 2146 if ((rnh = nep->ne_inet6head) == NULL) { 2147 off = offsetof(struct sockaddr_in6, sin6_addr) << 3; 2148 if (!rn_inithead((void **)&rnh, maskhead, off)) 2149 return (NULL); 2150 nep->ne_inet6head = rnh; 2151 } 2152 break; 2153 #endif 2154 } 2155 return (rnh); 2156 } 2157 2158 /* 2159 * helper function for freeing netcred elements 2160 */ 2161 static void 2162 vfs_free_addrlist_af(struct radix_node_head **prnh) 2163 { 2164 struct radix_node_head *rnh = *prnh; 2165 2166 (*rnh->rnh_walktree) (rnh, vfs_free_netcred, rnh); 2167 kfree(rnh, M_RTABLE); 2168 *prnh = NULL; 2169 } 2170 2171 /* 2172 * helper function for freeing mask elements 2173 */ 2174 static void 2175 vfs_free_addrlist_masks(struct radix_node_head **prnh) 2176 { 2177 struct radix_node_head *rnh = *prnh; 2178 2179 (*rnh->rnh_walktree) (rnh, vfs_free_netcred_mask, rnh); 2180 kfree(rnh, M_RTABLE); 2181 *prnh = NULL; 2182 } 2183 2184 /* 2185 * Free the net address hash lists that are hanging off the mount points. 2186 */ 2187 static void 2188 vfs_free_addrlist(struct netexport *nep) 2189 { 2190 NE_LOCK(nep); 2191 if (nep->ne_inethead != NULL) 2192 vfs_free_addrlist_af(&nep->ne_inethead); 2193 if (nep->ne_inet6head != NULL) 2194 vfs_free_addrlist_af(&nep->ne_inet6head); 2195 if (nep->ne_maskhead) 2196 vfs_free_addrlist_masks(&nep->ne_maskhead); 2197 NE_UNLOCK(nep); 2198 } 2199 2200 int 2201 vfs_export(struct mount *mp, struct netexport *nep, 2202 const struct export_args *argp) 2203 { 2204 int error; 2205 2206 if (argp->ex_flags & MNT_DELEXPORT) { 2207 if (mp->mnt_flag & MNT_EXPUBLIC) { 2208 vfs_setpublicfs(NULL, NULL, NULL); 2209 mp->mnt_flag &= ~MNT_EXPUBLIC; 2210 } 2211 vfs_free_addrlist(nep); 2212 mp->mnt_flag &= ~(MNT_EXPORTED | MNT_DEFEXPORTED); 2213 } 2214 if (argp->ex_flags & MNT_EXPORTED) { 2215 if (argp->ex_flags & MNT_EXPUBLIC) { 2216 if ((error = vfs_setpublicfs(mp, nep, argp)) != 0) 2217 return (error); 2218 mp->mnt_flag |= MNT_EXPUBLIC; 2219 } 2220 if ((error = vfs_hang_addrlist(mp, nep, argp))) 2221 return (error); 2222 mp->mnt_flag |= MNT_EXPORTED; 2223 } 2224 return (0); 2225 } 2226 2227 2228 /* 2229 * Set the publicly exported filesystem (WebNFS). Currently, only 2230 * one public filesystem is possible in the spec (RFC 2054 and 2055) 2231 */ 2232 int 2233 vfs_setpublicfs(struct mount *mp, struct netexport *nep, 2234 const struct export_args *argp) 2235 { 2236 int error; 2237 struct vnode *rvp; 2238 char *cp; 2239 2240 /* 2241 * mp == NULL -> invalidate the current info, the FS is 2242 * no longer exported. May be called from either vfs_export 2243 * or unmount, so check if it hasn't already been done. 2244 */ 2245 if (mp == NULL) { 2246 if (nfs_pub.np_valid) { 2247 nfs_pub.np_valid = 0; 2248 if (nfs_pub.np_index != NULL) { 2249 kfree(nfs_pub.np_index, M_TEMP); 2250 nfs_pub.np_index = NULL; 2251 } 2252 } 2253 return (0); 2254 } 2255 2256 /* 2257 * Only one allowed at a time. 2258 */ 2259 if (nfs_pub.np_valid != 0 && mp != nfs_pub.np_mount) 2260 return (EBUSY); 2261 2262 /* 2263 * Get real filehandle for root of exported FS. 2264 */ 2265 bzero((caddr_t)&nfs_pub.np_handle, sizeof(nfs_pub.np_handle)); 2266 nfs_pub.np_handle.fh_fsid = mp->mnt_stat.f_fsid; 2267 2268 if ((error = VFS_ROOT(mp, &rvp))) 2269 return (error); 2270 2271 if ((error = VFS_VPTOFH(rvp, &nfs_pub.np_handle.fh_fid))) 2272 return (error); 2273 2274 vput(rvp); 2275 2276 /* 2277 * If an indexfile was specified, pull it in. 2278 */ 2279 if (argp->ex_indexfile != NULL) { 2280 int namelen; 2281 2282 error = vn_get_namelen(rvp, &namelen); 2283 if (error) 2284 return (error); 2285 nfs_pub.np_index = kmalloc(namelen, M_TEMP, M_WAITOK); 2286 error = copyinstr(argp->ex_indexfile, nfs_pub.np_index, 2287 namelen, NULL); 2288 if (!error) { 2289 /* 2290 * Check for illegal filenames. 2291 */ 2292 for (cp = nfs_pub.np_index; *cp; cp++) { 2293 if (*cp == '/') { 2294 error = EINVAL; 2295 break; 2296 } 2297 } 2298 } 2299 if (error) { 2300 kfree(nfs_pub.np_index, M_TEMP); 2301 return (error); 2302 } 2303 } 2304 2305 nfs_pub.np_mount = mp; 2306 nfs_pub.np_valid = 1; 2307 return (0); 2308 } 2309 2310 struct netcred * 2311 vfs_export_lookup(struct mount *mp, struct netexport *nep, 2312 struct sockaddr *nam) 2313 { 2314 struct netcred *np; 2315 struct radix_node_head *rnh; 2316 struct sockaddr *saddr; 2317 2318 np = NULL; 2319 if (mp->mnt_flag & MNT_EXPORTED) { 2320 /* 2321 * Lookup in the export list first. 2322 */ 2323 NE_LOCK(nep); 2324 if (nam != NULL) { 2325 saddr = nam; 2326 switch (saddr->sa_family) { 2327 #ifdef INET 2328 case AF_INET: 2329 rnh = nep->ne_inethead; 2330 break; 2331 #endif 2332 #ifdef INET6 2333 case AF_INET6: 2334 rnh = nep->ne_inet6head; 2335 break; 2336 #endif 2337 default: 2338 rnh = NULL; 2339 } 2340 if (rnh != NULL) { 2341 np = (struct netcred *) 2342 (*rnh->rnh_matchaddr)((char *)saddr, 2343 rnh); 2344 if (np && np->netc_rnodes->rn_flags & RNF_ROOT) 2345 np = NULL; 2346 } 2347 } 2348 NE_UNLOCK(nep); 2349 /* 2350 * If no address match, use the default if it exists. 2351 */ 2352 if (np == NULL && mp->mnt_flag & MNT_DEFEXPORTED) 2353 np = &nep->ne_defexported; 2354 } 2355 return (np); 2356 } 2357 2358 /* 2359 * perform msync on all vnodes under a mount point. The mount point must 2360 * be locked. This code is also responsible for lazy-freeing unreferenced 2361 * vnodes whos VM objects no longer contain pages. 2362 * 2363 * NOTE: MNT_WAIT still skips vnodes in the VXLOCK state. 2364 * 2365 * NOTE: XXX VOP_PUTPAGES and friends requires that the vnode be locked, 2366 * but vnode_pager_putpages() doesn't lock the vnode. We have to do it 2367 * way up in this high level function. 2368 */ 2369 static int vfs_msync_scan1(struct mount *mp, struct vnode *vp, void *data); 2370 static int vfs_msync_scan2(struct mount *mp, struct vnode *vp, void *data); 2371 2372 void 2373 vfs_msync(struct mount *mp, int flags) 2374 { 2375 int vmsc_flags; 2376 2377 /* 2378 * tmpfs sets this flag to prevent msync(), sync, and the 2379 * filesystem periodic syncer from trying to flush VM pages 2380 * to swap. Only pure memory pressure flushes tmpfs VM pages 2381 * to swap. 2382 */ 2383 if (mp->mnt_kern_flag & MNTK_NOMSYNC) 2384 return; 2385 2386 /* 2387 * Ok, scan the vnodes for work. If the filesystem is using the 2388 * syncer thread feature we can use vsyncscan() instead of 2389 * vmntvnodescan(), which is much faster. 2390 */ 2391 vmsc_flags = VMSC_GETVP; 2392 if (flags != MNT_WAIT) 2393 vmsc_flags |= VMSC_NOWAIT; 2394 2395 if (mp->mnt_kern_flag & MNTK_THR_SYNC) { 2396 vsyncscan(mp, vmsc_flags, vfs_msync_scan2, 2397 (void *)(intptr_t)flags); 2398 } else { 2399 vmntvnodescan(mp, vmsc_flags, 2400 vfs_msync_scan1, vfs_msync_scan2, 2401 (void *)(intptr_t)flags); 2402 } 2403 } 2404 2405 /* 2406 * scan1 is a fast pre-check. There could be hundreds of thousands of 2407 * vnodes, we cannot afford to do anything heavy weight until we have a 2408 * fairly good indication that there is work to do. 2409 */ 2410 static 2411 int 2412 vfs_msync_scan1(struct mount *mp, struct vnode *vp, void *data) 2413 { 2414 int flags = (int)(intptr_t)data; 2415 2416 if ((vp->v_flag & VRECLAIMED) == 0) { 2417 if (vp->v_auxrefs == 0 && VREFCNT(vp) <= 0 && 2418 vp->v_object) { 2419 return(0); /* call scan2 */ 2420 } 2421 if ((mp->mnt_flag & MNT_RDONLY) == 0 && 2422 (vp->v_flag & VOBJDIRTY) && 2423 (flags == MNT_WAIT || vn_islocked(vp) == 0)) { 2424 return(0); /* call scan2 */ 2425 } 2426 } 2427 2428 /* 2429 * do not call scan2, continue the loop 2430 */ 2431 return(-1); 2432 } 2433 2434 /* 2435 * This callback is handed a locked vnode. 2436 */ 2437 static 2438 int 2439 vfs_msync_scan2(struct mount *mp, struct vnode *vp, void *data) 2440 { 2441 vm_object_t obj; 2442 int flags = (int)(intptr_t)data; 2443 int opcflags; 2444 2445 if (vp->v_flag & VRECLAIMED) 2446 return(0); 2447 2448 if ((mp->mnt_flag & MNT_RDONLY) == 0 && (vp->v_flag & VOBJDIRTY)) { 2449 if ((obj = vp->v_object) != NULL) { 2450 if (flags == MNT_WAIT) { 2451 /* 2452 * VFS_MSYNC is called with MNT_WAIT when 2453 * unmounting. 2454 */ 2455 opcflags = OBJPC_SYNC; 2456 } else if (vp->v_writecount || obj->ref_count) { 2457 /* 2458 * VFS_MSYNC is otherwise called via the 2459 * periodic filesystem sync or the 'sync' 2460 * command. Honor MADV_NOSYNC / MAP_NOSYNC 2461 * if the file is open for writing or memory 2462 * mapped. Pages flagged PG_NOSYNC will not 2463 * be automatically flushed at this time. 2464 * 2465 * The obj->ref_count test is not perfect 2466 * since temporary refs may be present, but 2467 * the periodic filesystem sync will ultimately 2468 * catch it if the file is not open and not 2469 * mapped. 2470 */ 2471 opcflags = OBJPC_NOSYNC; 2472 } else { 2473 /* 2474 * If the file is no longer open for writing 2475 * and also no longer mapped, do not honor 2476 * MAP_NOSYNC. That is, fully synchronize 2477 * the file. 2478 * 2479 * This still occurs on the periodic fs sync, 2480 * so frontend programs which turn the file 2481 * over quickly enough can still avoid the 2482 * sync, but ultimately we do want to flush 2483 * even MADV_NOSYNC pages once it is no longer 2484 * mapped or open for writing. 2485 */ 2486 opcflags = 0; 2487 } 2488 vm_object_page_clean(obj, 0, 0, opcflags); 2489 } 2490 } 2491 return(0); 2492 } 2493 2494 /* 2495 * Wake up anyone interested in vp because it is being revoked. 2496 */ 2497 void 2498 vn_gone(struct vnode *vp) 2499 { 2500 lwkt_gettoken(&vp->v_token); 2501 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, NOTE_REVOKE); 2502 lwkt_reltoken(&vp->v_token); 2503 } 2504 2505 /* 2506 * extract the cdev_t from a VBLK or VCHR. The vnode must have been opened 2507 * (or v_rdev might be NULL). 2508 */ 2509 cdev_t 2510 vn_todev(struct vnode *vp) 2511 { 2512 if (vp->v_type != VBLK && vp->v_type != VCHR) 2513 return (NULL); 2514 KKASSERT(vp->v_rdev != NULL); 2515 return (vp->v_rdev); 2516 } 2517 2518 /* 2519 * Check if vnode represents a disk device. The vnode does not need to be 2520 * opened. 2521 * 2522 * MPALMOSTSAFE 2523 */ 2524 int 2525 vn_isdisk(struct vnode *vp, int *errp) 2526 { 2527 cdev_t dev; 2528 2529 if (vp->v_type != VCHR) { 2530 if (errp != NULL) 2531 *errp = ENOTBLK; 2532 return (0); 2533 } 2534 2535 dev = vp->v_rdev; 2536 2537 if (dev == NULL) { 2538 if (errp != NULL) 2539 *errp = ENXIO; 2540 return (0); 2541 } 2542 if (dev_is_good(dev) == 0) { 2543 if (errp != NULL) 2544 *errp = ENXIO; 2545 return (0); 2546 } 2547 if ((dev_dflags(dev) & D_DISK) == 0) { 2548 if (errp != NULL) 2549 *errp = ENOTBLK; 2550 return (0); 2551 } 2552 if (errp != NULL) 2553 *errp = 0; 2554 return (1); 2555 } 2556 2557 int 2558 vn_get_namelen(struct vnode *vp, int *namelen) 2559 { 2560 int error; 2561 register_t retval[2]; 2562 2563 error = VOP_PATHCONF(vp, _PC_NAME_MAX, retval); 2564 if (error) 2565 return (error); 2566 *namelen = (int)retval[0]; 2567 return (0); 2568 } 2569 2570 int 2571 vop_write_dirent(int *error, struct uio *uio, ino_t d_ino, uint8_t d_type, 2572 uint16_t d_namlen, const char *d_name) 2573 { 2574 struct dirent *dp; 2575 size_t len; 2576 2577 len = _DIRENT_RECLEN(d_namlen); 2578 if (len > uio->uio_resid) 2579 return(1); 2580 2581 dp = kmalloc(len, M_TEMP, M_WAITOK | M_ZERO); 2582 2583 dp->d_ino = d_ino; 2584 dp->d_namlen = d_namlen; 2585 dp->d_type = d_type; 2586 bcopy(d_name, dp->d_name, d_namlen); 2587 2588 *error = uiomove((caddr_t)dp, len, uio); 2589 2590 kfree(dp, M_TEMP); 2591 2592 return(0); 2593 } 2594 2595 void 2596 vn_mark_atime(struct vnode *vp, struct thread *td) 2597 { 2598 struct proc *p = td->td_proc; 2599 struct ucred *cred = p ? p->p_ucred : proc0.p_ucred; 2600 2601 if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) { 2602 VOP_MARKATIME(vp, cred); 2603 } 2604 } 2605 2606 /* 2607 * Calculate the number of entries in an inode-related chained hash table. 2608 * With today's memory sizes, maxvnodes can wind up being a very large 2609 * number. There is no reason to waste memory, so tolerate some stacking. 2610 */ 2611 int 2612 vfs_inodehashsize(void) 2613 { 2614 int hsize; 2615 2616 hsize = 32; 2617 while (hsize < maxvnodes) 2618 hsize <<= 1; 2619 while (hsize > maxvnodes * 2) 2620 hsize >>= 1; /* nominal 2x stacking */ 2621 2622 if (maxvnodes > 1024 * 1024) 2623 hsize >>= 1; /* nominal 8x stacking */ 2624 2625 if (maxvnodes > 128 * 1024) 2626 hsize >>= 1; /* nominal 4x stacking */ 2627 2628 if (hsize < 16) 2629 hsize = 16; 2630 2631 return hsize; 2632 } 2633 2634 union _qcvt { 2635 quad_t qcvt; 2636 int32_t val[2]; 2637 }; 2638 2639 #define SETHIGH(q, h) { \ 2640 union _qcvt tmp; \ 2641 tmp.qcvt = (q); \ 2642 tmp.val[_QUAD_HIGHWORD] = (h); \ 2643 (q) = tmp.qcvt; \ 2644 } 2645 #define SETLOW(q, l) { \ 2646 union _qcvt tmp; \ 2647 tmp.qcvt = (q); \ 2648 tmp.val[_QUAD_LOWWORD] = (l); \ 2649 (q) = tmp.qcvt; \ 2650 } 2651 2652 u_quad_t 2653 init_va_filerev(void) 2654 { 2655 struct timeval tv; 2656 u_quad_t ret = 0; 2657 2658 getmicrouptime(&tv); 2659 SETHIGH(ret, tv.tv_sec); 2660 SETLOW(ret, tv.tv_usec * 4294); 2661 2662 return ret; 2663 } 2664