vfs_subr.c revision 112181
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. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 39 * $FreeBSD: head/sys/kern/vfs_subr.c 112181 2003-03-13 07:19:23Z jeff $ 40 */ 41 42/* 43 * External virtual filesystem routines 44 */ 45#include "opt_ddb.h" 46#include "opt_mac.h" 47 48#include <sys/param.h> 49#include <sys/systm.h> 50#include <sys/bio.h> 51#include <sys/buf.h> 52#include <sys/conf.h> 53#include <sys/eventhandler.h> 54#include <sys/extattr.h> 55#include <sys/fcntl.h> 56#include <sys/kernel.h> 57#include <sys/kthread.h> 58#include <sys/mac.h> 59#include <sys/malloc.h> 60#include <sys/mount.h> 61#include <sys/namei.h> 62#include <sys/stat.h> 63#include <sys/sysctl.h> 64#include <sys/syslog.h> 65#include <sys/vmmeter.h> 66#include <sys/vnode.h> 67 68#include <vm/vm.h> 69#include <vm/vm_object.h> 70#include <vm/vm_extern.h> 71#include <vm/pmap.h> 72#include <vm/vm_map.h> 73#include <vm/vm_page.h> 74#include <vm/uma.h> 75 76static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure"); 77 78static void addalias(struct vnode *vp, dev_t nvp_rdev); 79static void insmntque(struct vnode *vp, struct mount *mp); 80static void vclean(struct vnode *vp, int flags, struct thread *td); 81static void vlruvp(struct vnode *vp); 82static int flushbuflist(struct buf *blist, int flags, struct vnode *vp, 83 int slpflag, int slptimeo, int *errorp); 84static int vcanrecycle(struct vnode *vp, struct mount **vnmpp); 85 86 87/* 88 * Number of vnodes in existence. Increased whenever getnewvnode() 89 * allocates a new vnode, never decreased. 90 */ 91static unsigned long numvnodes; 92 93SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); 94 95/* 96 * Conversion tables for conversion from vnode types to inode formats 97 * and back. 98 */ 99enum vtype iftovt_tab[16] = { 100 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 101 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 102}; 103int vttoif_tab[9] = { 104 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 105 S_IFSOCK, S_IFIFO, S_IFMT, 106}; 107 108/* 109 * List of vnodes that are ready for recycling. 110 */ 111static TAILQ_HEAD(freelst, vnode) vnode_free_list; 112 113/* 114 * Minimum number of free vnodes. If there are fewer than this free vnodes, 115 * getnewvnode() will return a newly allocated vnode. 116 */ 117static u_long wantfreevnodes = 25; 118SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 119/* Number of vnodes in the free list. */ 120static u_long freevnodes; 121SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); 122 123/* 124 * Various variables used for debugging the new implementation of 125 * reassignbuf(). 126 * XXX these are probably of (very) limited utility now. 127 */ 128static int reassignbufcalls; 129SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); 130static int nameileafonly; 131SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, ""); 132 133/* 134 * Cache for the mount type id assigned to NFS. This is used for 135 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 136 */ 137int nfs_mount_type = -1; 138 139/* To keep more than one thread at a time from running vfs_getnewfsid */ 140static struct mtx mntid_mtx; 141 142/* 143 * Lock for any access to the following: 144 * vnode_free_list 145 * numvnodes 146 * freevnodes 147 */ 148static struct mtx vnode_free_list_mtx; 149 150/* 151 * For any iteration/modification of dev->si_hlist (linked through 152 * v_specnext) 153 */ 154static struct mtx spechash_mtx; 155 156/* Publicly exported FS */ 157struct nfs_public nfs_pub; 158 159/* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 160static uma_zone_t vnode_zone; 161static uma_zone_t vnodepoll_zone; 162 163/* Set to 1 to print out reclaim of active vnodes */ 164int prtactive; 165 166/* 167 * The workitem queue. 168 * 169 * It is useful to delay writes of file data and filesystem metadata 170 * for tens of seconds so that quickly created and deleted files need 171 * not waste disk bandwidth being created and removed. To realize this, 172 * we append vnodes to a "workitem" queue. When running with a soft 173 * updates implementation, most pending metadata dependencies should 174 * not wait for more than a few seconds. Thus, mounted on block devices 175 * are delayed only about a half the time that file data is delayed. 176 * Similarly, directory updates are more critical, so are only delayed 177 * about a third the time that file data is delayed. Thus, there are 178 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 179 * one each second (driven off the filesystem syncer process). The 180 * syncer_delayno variable indicates the next queue that is to be processed. 181 * Items that need to be processed soon are placed in this queue: 182 * 183 * syncer_workitem_pending[syncer_delayno] 184 * 185 * A delay of fifteen seconds is done by placing the request fifteen 186 * entries later in the queue: 187 * 188 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 189 * 190 */ 191static int syncer_delayno; 192static long syncer_mask; 193LIST_HEAD(synclist, vnode); 194static struct synclist *syncer_workitem_pending; 195/* 196 * The sync_mtx protects: 197 * vp->v_synclist 198 * syncer_delayno 199 * syncer_workitem_pending 200 * rushjob 201 */ 202static struct mtx sync_mtx; 203 204#define SYNCER_MAXDELAY 32 205static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 206static int syncdelay = 30; /* max time to delay syncing data */ 207static int filedelay = 30; /* time to delay syncing files */ 208SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); 209static int dirdelay = 29; /* time to delay syncing directories */ 210SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); 211static int metadelay = 28; /* time to delay syncing metadata */ 212SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); 213static int rushjob; /* number of slots to run ASAP */ 214static int stat_rush_requests; /* number of times I/O speeded up */ 215SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); 216 217/* 218 * Number of vnodes we want to exist at any one time. This is mostly used 219 * to size hash tables in vnode-related code. It is normally not used in 220 * getnewvnode(), as wantfreevnodes is normally nonzero.) 221 * 222 * XXX desiredvnodes is historical cruft and should not exist. 223 */ 224int desiredvnodes; 225SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 226 &desiredvnodes, 0, "Maximum number of vnodes"); 227static int minvnodes; 228SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 229 &minvnodes, 0, "Minimum number of vnodes"); 230static int vnlru_nowhere; 231SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, &vnlru_nowhere, 0, 232 "Number of times the vnlru process ran without success"); 233 234/* Hook for calling soft updates */ 235int (*softdep_process_worklist_hook)(struct mount *); 236 237/* 238 * This only exists to supress warnings from unlocked specfs accesses. It is 239 * no longer ok to have an unlocked VFS. 240 */ 241#define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD) 242 243/* Print lock violations */ 244int vfs_badlock_print = 1; 245 246/* Panic on violation */ 247int vfs_badlock_panic = 1; 248 249/* Check for interlock across VOPs */ 250int vfs_badlock_mutex = 1; 251 252static void 253vfs_badlock(char *msg, char *str, struct vnode *vp) 254{ 255 if (vfs_badlock_print) 256 printf("%s: %p %s\n", str, vp, msg); 257 if (vfs_badlock_panic) 258 Debugger("Lock violation.\n"); 259} 260 261void 262assert_vi_unlocked(struct vnode *vp, char *str) 263{ 264 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 265 vfs_badlock("interlock is locked but should not be", str, vp); 266} 267 268void 269assert_vi_locked(struct vnode *vp, char *str) 270{ 271 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 272 vfs_badlock("interlock is not locked but should be", str, vp); 273} 274 275void 276assert_vop_locked(struct vnode *vp, char *str) 277{ 278 if (vp && !IGNORE_LOCK(vp) && !VOP_ISLOCKED(vp, NULL)) 279 vfs_badlock("is not locked but should be", str, vp); 280} 281 282void 283assert_vop_unlocked(struct vnode *vp, char *str) 284{ 285 if (vp && !IGNORE_LOCK(vp) && 286 VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE) 287 vfs_badlock("is locked but should not be", str, vp); 288} 289 290void 291assert_vop_elocked(struct vnode *vp, char *str) 292{ 293 if (vp && !IGNORE_LOCK(vp) && 294 VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE) 295 vfs_badlock("is not exclusive locked but should be", str, vp); 296} 297 298void 299assert_vop_elocked_other(struct vnode *vp, char *str) 300{ 301 if (vp && !IGNORE_LOCK(vp) && 302 VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER) 303 vfs_badlock("is not exclusive locked by another thread", 304 str, vp); 305} 306 307void 308assert_vop_slocked(struct vnode *vp, char *str) 309{ 310 if (vp && !IGNORE_LOCK(vp) && 311 VOP_ISLOCKED(vp, curthread) != LK_SHARED) 312 vfs_badlock("is not locked shared but should be", str, vp); 313} 314 315void 316vop_rename_pre(void *ap) 317{ 318 struct vop_rename_args *a = ap; 319 320 if (a->a_tvp) 321 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 322 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 323 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 324 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 325 326 /* Check the source (from) */ 327 if (a->a_tdvp != a->a_fdvp) 328 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked.\n"); 329 if (a->a_tvp != a->a_fvp) 330 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked.\n"); 331 332 /* Check the target */ 333 if (a->a_tvp) 334 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked.\n"); 335 336 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked.\n"); 337} 338 339void 340vop_strategy_pre(void *ap) 341{ 342 struct vop_strategy_args *a = ap; 343 struct buf *bp; 344 345 bp = a->a_bp; 346 347 /* 348 * Cluster ops lock their component buffers but not the IO container. 349 */ 350 if ((bp->b_flags & B_CLUSTER) != 0) 351 return; 352 353 if (BUF_REFCNT(bp) < 1) { 354 if (vfs_badlock_print) 355 printf("VOP_STRATEGY: bp is not locked but should be.\n"); 356 if (vfs_badlock_panic) 357 Debugger("Lock violation.\n"); 358 } 359} 360 361void 362vop_lookup_pre(void *ap) 363{ 364 struct vop_lookup_args *a = ap; 365 struct vnode *dvp; 366 367 dvp = a->a_dvp; 368 369 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 370 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 371} 372 373void 374vop_lookup_post(void *ap, int rc) 375{ 376 struct vop_lookup_args *a = ap; 377 struct componentname *cnp; 378 struct vnode *dvp; 379 struct vnode *vp; 380 int flags; 381 382 dvp = a->a_dvp; 383 cnp = a->a_cnp; 384 vp = *(a->a_vpp); 385 flags = cnp->cn_flags; 386 387 388 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 389 /* 390 * If this is the last path component for this lookup and LOCPARENT 391 * is set, OR if there is an error the directory has to be locked. 392 */ 393 if ((flags & LOCKPARENT) && (flags & ISLASTCN)) 394 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)"); 395 else if (rc != 0) 396 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)"); 397 else if (dvp != vp) 398 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)"); 399 400 if (flags & PDIRUNLOCK) 401 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)"); 402} 403 404void 405vop_unlock_pre(void *ap) 406{ 407 struct vop_unlock_args *a = ap; 408 409 if (a->a_flags & LK_INTERLOCK) 410 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 411 412 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 413} 414 415void 416vop_unlock_post(void *ap, int rc) 417{ 418 struct vop_unlock_args *a = ap; 419 420 if (a->a_flags & LK_INTERLOCK) 421 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 422} 423 424void 425vop_lock_pre(void *ap) 426{ 427 struct vop_lock_args *a = ap; 428 429 if ((a->a_flags & LK_INTERLOCK) == 0) 430 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 431 else 432 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 433} 434 435void 436vop_lock_post(void *ap, int rc) 437{ 438 struct vop_lock_args *a; 439 440 a = ap; 441 442 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 443 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 444} 445 446void 447v_addpollinfo(struct vnode *vp) 448{ 449 vp->v_pollinfo = uma_zalloc(vnodepoll_zone, M_WAITOK); 450 mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 451} 452 453/* 454 * Initialize the vnode management data structures. 455 */ 456static void 457vntblinit(void *dummy __unused) 458{ 459 460 desiredvnodes = maxproc + cnt.v_page_count / 4; 461 minvnodes = desiredvnodes / 4; 462 mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF); 463 mtx_init(&mntvnode_mtx, "mntvnode", NULL, MTX_DEF); 464 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 465 mtx_init(&spechash_mtx, "spechash", NULL, MTX_DEF); 466 TAILQ_INIT(&vnode_free_list); 467 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 468 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 469 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 470 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 471 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 472 /* 473 * Initialize the filesystem syncer. 474 */ 475 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 476 &syncer_mask); 477 syncer_maxdelay = syncer_mask + 1; 478 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 479} 480SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) 481 482 483/* 484 * Mark a mount point as busy. Used to synchronize access and to delay 485 * unmounting. Interlock is not released on failure. 486 */ 487int 488vfs_busy(mp, flags, interlkp, td) 489 struct mount *mp; 490 int flags; 491 struct mtx *interlkp; 492 struct thread *td; 493{ 494 int lkflags; 495 496 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 497 if (flags & LK_NOWAIT) 498 return (ENOENT); 499 mp->mnt_kern_flag |= MNTK_MWAIT; 500 /* 501 * Since all busy locks are shared except the exclusive 502 * lock granted when unmounting, the only place that a 503 * wakeup needs to be done is at the release of the 504 * exclusive lock at the end of dounmount. 505 */ 506 msleep(mp, interlkp, PVFS, "vfs_busy", 0); 507 return (ENOENT); 508 } 509 lkflags = LK_SHARED | LK_NOPAUSE; 510 if (interlkp) 511 lkflags |= LK_INTERLOCK; 512 if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td)) 513 panic("vfs_busy: unexpected lock failure"); 514 return (0); 515} 516 517/* 518 * Free a busy filesystem. 519 */ 520void 521vfs_unbusy(mp, td) 522 struct mount *mp; 523 struct thread *td; 524{ 525 526 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); 527} 528 529/* 530 * Lookup a mount point by filesystem identifier. 531 */ 532struct mount * 533vfs_getvfs(fsid) 534 fsid_t *fsid; 535{ 536 register struct mount *mp; 537 538 mtx_lock(&mountlist_mtx); 539 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 540 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 541 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 542 mtx_unlock(&mountlist_mtx); 543 return (mp); 544 } 545 } 546 mtx_unlock(&mountlist_mtx); 547 return ((struct mount *) 0); 548} 549 550/* 551 * Get a new unique fsid. Try to make its val[0] unique, since this value 552 * will be used to create fake device numbers for stat(). Also try (but 553 * not so hard) make its val[0] unique mod 2^16, since some emulators only 554 * support 16-bit device numbers. We end up with unique val[0]'s for the 555 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 556 * 557 * Keep in mind that several mounts may be running in parallel. Starting 558 * the search one past where the previous search terminated is both a 559 * micro-optimization and a defense against returning the same fsid to 560 * different mounts. 561 */ 562void 563vfs_getnewfsid(mp) 564 struct mount *mp; 565{ 566 static u_int16_t mntid_base; 567 fsid_t tfsid; 568 int mtype; 569 570 mtx_lock(&mntid_mtx); 571 mtype = mp->mnt_vfc->vfc_typenum; 572 tfsid.val[1] = mtype; 573 mtype = (mtype & 0xFF) << 24; 574 for (;;) { 575 tfsid.val[0] = makeudev(255, 576 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 577 mntid_base++; 578 if (vfs_getvfs(&tfsid) == NULL) 579 break; 580 } 581 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 582 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 583 mtx_unlock(&mntid_mtx); 584} 585 586/* 587 * Knob to control the precision of file timestamps: 588 * 589 * 0 = seconds only; nanoseconds zeroed. 590 * 1 = seconds and nanoseconds, accurate within 1/HZ. 591 * 2 = seconds and nanoseconds, truncated to microseconds. 592 * >=3 = seconds and nanoseconds, maximum precision. 593 */ 594enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 595 596static int timestamp_precision = TSP_SEC; 597SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 598 ×tamp_precision, 0, ""); 599 600/* 601 * Get a current timestamp. 602 */ 603void 604vfs_timestamp(tsp) 605 struct timespec *tsp; 606{ 607 struct timeval tv; 608 609 switch (timestamp_precision) { 610 case TSP_SEC: 611 tsp->tv_sec = time_second; 612 tsp->tv_nsec = 0; 613 break; 614 case TSP_HZ: 615 getnanotime(tsp); 616 break; 617 case TSP_USEC: 618 microtime(&tv); 619 TIMEVAL_TO_TIMESPEC(&tv, tsp); 620 break; 621 case TSP_NSEC: 622 default: 623 nanotime(tsp); 624 break; 625 } 626} 627 628/* 629 * Set vnode attributes to VNOVAL 630 */ 631void 632vattr_null(vap) 633 register struct vattr *vap; 634{ 635 636 vap->va_type = VNON; 637 vap->va_size = VNOVAL; 638 vap->va_bytes = VNOVAL; 639 vap->va_mode = VNOVAL; 640 vap->va_nlink = VNOVAL; 641 vap->va_uid = VNOVAL; 642 vap->va_gid = VNOVAL; 643 vap->va_fsid = VNOVAL; 644 vap->va_fileid = VNOVAL; 645 vap->va_blocksize = VNOVAL; 646 vap->va_rdev = VNOVAL; 647 vap->va_atime.tv_sec = VNOVAL; 648 vap->va_atime.tv_nsec = VNOVAL; 649 vap->va_mtime.tv_sec = VNOVAL; 650 vap->va_mtime.tv_nsec = VNOVAL; 651 vap->va_ctime.tv_sec = VNOVAL; 652 vap->va_ctime.tv_nsec = VNOVAL; 653 vap->va_birthtime.tv_sec = VNOVAL; 654 vap->va_birthtime.tv_nsec = VNOVAL; 655 vap->va_flags = VNOVAL; 656 vap->va_gen = VNOVAL; 657 vap->va_vaflags = 0; 658} 659 660/* 661 * This routine is called when we have too many vnodes. It attempts 662 * to free <count> vnodes and will potentially free vnodes that still 663 * have VM backing store (VM backing store is typically the cause 664 * of a vnode blowout so we want to do this). Therefore, this operation 665 * is not considered cheap. 666 * 667 * A number of conditions may prevent a vnode from being reclaimed. 668 * the buffer cache may have references on the vnode, a directory 669 * vnode may still have references due to the namei cache representing 670 * underlying files, or the vnode may be in active use. It is not 671 * desireable to reuse such vnodes. These conditions may cause the 672 * number of vnodes to reach some minimum value regardless of what 673 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 674 */ 675static int 676vlrureclaim(struct mount *mp, int count) 677{ 678 struct vnode *vp; 679 int done; 680 int trigger; 681 int usevnodes; 682 683 /* 684 * Calculate the trigger point, don't allow user 685 * screwups to blow us up. This prevents us from 686 * recycling vnodes with lots of resident pages. We 687 * aren't trying to free memory, we are trying to 688 * free vnodes. 689 */ 690 usevnodes = desiredvnodes; 691 if (usevnodes <= 0) 692 usevnodes = 1; 693 trigger = cnt.v_page_count * 2 / usevnodes; 694 695 done = 0; 696 mtx_lock(&mntvnode_mtx); 697 while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) { 698 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 699 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 700 701 if (vp->v_type != VNON && 702 vp->v_type != VBAD && 703 VI_TRYLOCK(vp)) { 704 if (VMIGHTFREE(vp) && /* critical path opt */ 705 (vp->v_object == NULL || 706 vp->v_object->resident_page_count < trigger)) { 707 mtx_unlock(&mntvnode_mtx); 708 vgonel(vp, curthread); 709 done++; 710 mtx_lock(&mntvnode_mtx); 711 } else 712 VI_UNLOCK(vp); 713 } 714 --count; 715 } 716 mtx_unlock(&mntvnode_mtx); 717 return done; 718} 719 720/* 721 * Attempt to recycle vnodes in a context that is always safe to block. 722 * Calling vlrurecycle() from the bowels of filesystem code has some 723 * interesting deadlock problems. 724 */ 725static struct proc *vnlruproc; 726static int vnlruproc_sig; 727 728static void 729vnlru_proc(void) 730{ 731 struct mount *mp, *nmp; 732 int s; 733 int done, take; 734 struct proc *p = vnlruproc; 735 struct thread *td = FIRST_THREAD_IN_PROC(p); /* XXXKSE */ 736 737 mtx_lock(&Giant); 738 739 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 740 SHUTDOWN_PRI_FIRST); 741 742 s = splbio(); 743 for (;;) { 744 kthread_suspend_check(p); 745 mtx_lock(&vnode_free_list_mtx); 746 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { 747 mtx_unlock(&vnode_free_list_mtx); 748 vnlruproc_sig = 0; 749 wakeup(&vnlruproc_sig); 750 tsleep(vnlruproc, PVFS, "vlruwt", hz); 751 continue; 752 } 753 mtx_unlock(&vnode_free_list_mtx); 754 done = 0; 755 mtx_lock(&mountlist_mtx); 756 take = 0; 757 TAILQ_FOREACH(mp, &mountlist, mnt_list) 758 take++; 759 take = desiredvnodes / (take * 10); 760 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 761 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 762 nmp = TAILQ_NEXT(mp, mnt_list); 763 continue; 764 } 765 done += vlrureclaim(mp, take); 766 mtx_lock(&mountlist_mtx); 767 nmp = TAILQ_NEXT(mp, mnt_list); 768 vfs_unbusy(mp, td); 769 } 770 mtx_unlock(&mountlist_mtx); 771 if (done == 0) { 772#if 0 773 /* These messages are temporary debugging aids */ 774 if (vnlru_nowhere < 5) 775 printf("vnlru process getting nowhere..\n"); 776 else if (vnlru_nowhere == 5) 777 printf("vnlru process messages stopped.\n"); 778#endif 779 vnlru_nowhere++; 780 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 781 } 782 } 783 splx(s); 784} 785 786static struct kproc_desc vnlru_kp = { 787 "vnlru", 788 vnlru_proc, 789 &vnlruproc 790}; 791SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 792 793 794/* 795 * Routines having to do with the management of the vnode table. 796 */ 797 798/* 799 * Check to see if a free vnode can be recycled. If it can, 800 * return it locked with the vn lock, but not interlock. Also 801 * get the vn_start_write lock. Otherwise indicate the error. 802 */ 803static int 804vcanrecycle(struct vnode *vp, struct mount **vnmpp) 805{ 806 struct thread *td = curthread; 807 vm_object_t object; 808 int error; 809 810 /* Don't recycle if we can't get the interlock */ 811 if (!VI_TRYLOCK(vp)) 812 return (EWOULDBLOCK); 813 814 /* We should be able to immediately acquire this */ 815 /* XXX This looks like it should panic if it fails */ 816 if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td) != 0) { 817 if (VOP_ISLOCKED(vp, td)) 818 panic("vcanrecycle: locked vnode"); 819 return (EWOULDBLOCK); 820 } 821 822 /* 823 * Don't recycle if its filesystem is being suspended. 824 */ 825 if (vn_start_write(vp, vnmpp, V_NOWAIT) != 0) { 826 error = EBUSY; 827 goto done; 828 } 829 830 /* 831 * Don't recycle if we still have cached pages. 832 */ 833 if (VOP_GETVOBJECT(vp, &object) == 0 && 834 (object->resident_page_count || 835 object->ref_count)) { 836 error = EBUSY; 837 goto done; 838 } 839 if (LIST_FIRST(&vp->v_cache_src)) { 840 /* 841 * note: nameileafonly sysctl is temporary, 842 * for debugging only, and will eventually be 843 * removed. 844 */ 845 if (nameileafonly > 0) { 846 /* 847 * Do not reuse namei-cached directory 848 * vnodes that have cached 849 * subdirectories. 850 */ 851 if (cache_leaf_test(vp) < 0) { 852 error = EISDIR; 853 goto done; 854 } 855 } else if (nameileafonly < 0 || 856 vmiodirenable == 0) { 857 /* 858 * Do not reuse namei-cached directory 859 * vnodes if nameileafonly is -1 or 860 * if VMIO backing for directories is 861 * turned off (otherwise we reuse them 862 * too quickly). 863 */ 864 error = EBUSY; 865 goto done; 866 } 867 } 868 return (0); 869done: 870 VOP_UNLOCK(vp, 0, td); 871 return (error); 872} 873 874/* 875 * Return the next vnode from the free list. 876 */ 877int 878getnewvnode(tag, mp, vops, vpp) 879 const char *tag; 880 struct mount *mp; 881 vop_t **vops; 882 struct vnode **vpp; 883{ 884 int s; 885 struct thread *td = curthread; /* XXX */ 886 struct vnode *vp = NULL; 887 struct vpollinfo *pollinfo = NULL; 888 struct mount *vnmp; 889 890 s = splbio(); 891 mtx_lock(&vnode_free_list_mtx); 892 893 /* 894 * Try to reuse vnodes if we hit the max. This situation only 895 * occurs in certain large-memory (2G+) situations. We cannot 896 * attempt to directly reclaim vnodes due to nasty recursion 897 * problems. 898 */ 899 while (numvnodes - freevnodes > desiredvnodes) { 900 if (vnlruproc_sig == 0) { 901 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 902 wakeup(vnlruproc); 903 } 904 mtx_unlock(&vnode_free_list_mtx); 905 tsleep(&vnlruproc_sig, PVFS, "vlruwk", hz); 906 mtx_lock(&vnode_free_list_mtx); 907 } 908 909 /* 910 * Attempt to reuse a vnode already on the free list, allocating 911 * a new vnode if we can't find one or if we have not reached a 912 * good minimum for good LRU performance. 913 */ 914 915 if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) { 916 int error; 917 int count; 918 919 for (count = 0; count < freevnodes; count++) { 920 vp = TAILQ_FIRST(&vnode_free_list); 921 922 KASSERT(vp->v_usecount == 0 && 923 (vp->v_iflag & VI_DOINGINACT) == 0, 924 ("getnewvnode: free vnode isn't")); 925 926 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 927 /* 928 * We have to drop the free list mtx to avoid lock 929 * order reversals with interlock. 930 */ 931 mtx_unlock(&vnode_free_list_mtx); 932 error = vcanrecycle(vp, &vnmp); 933 mtx_lock(&vnode_free_list_mtx); 934 if (error == 0) 935 break; 936 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 937 vp = NULL; 938 } 939 } 940 if (vp) { 941 freevnodes--; 942 mtx_unlock(&vnode_free_list_mtx); 943 944 cache_purge(vp); 945 VI_LOCK(vp); 946 vp->v_iflag |= VI_DOOMED; 947 vp->v_iflag &= ~VI_FREE; 948 if (vp->v_type != VBAD) { 949 VOP_UNLOCK(vp, 0, td); 950 vgonel(vp, td); 951 VI_LOCK(vp); 952 } else { 953 VOP_UNLOCK(vp, 0, td); 954 } 955 vn_finished_write(vnmp); 956 957#ifdef INVARIANTS 958 { 959 if (vp->v_data) 960 panic("cleaned vnode isn't"); 961 if (vp->v_numoutput) 962 panic("Clean vnode has pending I/O's"); 963 if (vp->v_writecount != 0) 964 panic("Non-zero write count"); 965 } 966#endif 967 if ((pollinfo = vp->v_pollinfo) != NULL) { 968 /* 969 * To avoid lock order reversals, the call to 970 * uma_zfree() must be delayed until the vnode 971 * interlock is released. 972 */ 973 vp->v_pollinfo = NULL; 974 } 975#ifdef MAC 976 mac_destroy_vnode(vp); 977#endif 978 vp->v_iflag = 0; 979 vp->v_vflag = 0; 980 vp->v_lastw = 0; 981 vp->v_lasta = 0; 982 vp->v_cstart = 0; 983 vp->v_clen = 0; 984 vp->v_socket = 0; 985 lockdestroy(vp->v_vnlock); 986 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 987 KASSERT(vp->v_cleanbufcnt == 0, ("cleanbufcnt not 0")); 988 KASSERT(vp->v_cleanblkroot == NULL, ("cleanblkroot not NULL")); 989 KASSERT(vp->v_dirtybufcnt == 0, ("dirtybufcnt not 0")); 990 KASSERT(vp->v_dirtyblkroot == NULL, ("dirtyblkroot not NULL")); 991 } else { 992 numvnodes++; 993 mtx_unlock(&vnode_free_list_mtx); 994 995 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); 996 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 997 VI_LOCK(vp); 998 vp->v_dd = vp; 999 vp->v_vnlock = &vp->v_lock; 1000 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 1001 cache_purge(vp); 1002 LIST_INIT(&vp->v_cache_src); 1003 TAILQ_INIT(&vp->v_cache_dst); 1004 } 1005 1006 TAILQ_INIT(&vp->v_cleanblkhd); 1007 TAILQ_INIT(&vp->v_dirtyblkhd); 1008 vp->v_type = VNON; 1009 vp->v_tag = tag; 1010 vp->v_op = vops; 1011 *vpp = vp; 1012 vp->v_usecount = 1; 1013 vp->v_data = 0; 1014 vp->v_cachedid = -1; 1015 VI_UNLOCK(vp); 1016 if (pollinfo != NULL) { 1017 mtx_destroy(&pollinfo->vpi_lock); 1018 uma_zfree(vnodepoll_zone, pollinfo); 1019 } 1020#ifdef MAC 1021 mac_init_vnode(vp); 1022 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 1023 mac_associate_vnode_singlelabel(mp, vp); 1024#endif 1025 insmntque(vp, mp); 1026 1027 return (0); 1028} 1029 1030/* 1031 * Move a vnode from one mount queue to another. 1032 */ 1033static void 1034insmntque(vp, mp) 1035 register struct vnode *vp; 1036 register struct mount *mp; 1037{ 1038 1039 mtx_lock(&mntvnode_mtx); 1040 /* 1041 * Delete from old mount point vnode list, if on one. 1042 */ 1043 if (vp->v_mount != NULL) 1044 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes); 1045 /* 1046 * Insert into list of vnodes for the new mount point, if available. 1047 */ 1048 if ((vp->v_mount = mp) == NULL) { 1049 mtx_unlock(&mntvnode_mtx); 1050 return; 1051 } 1052 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1053 mtx_unlock(&mntvnode_mtx); 1054} 1055 1056/* 1057 * Update outstanding I/O count and do wakeup if requested. 1058 */ 1059void 1060vwakeup(bp) 1061 register struct buf *bp; 1062{ 1063 register struct vnode *vp; 1064 1065 bp->b_flags &= ~B_WRITEINPROG; 1066 if ((vp = bp->b_vp)) { 1067 VI_LOCK(vp); 1068 vp->v_numoutput--; 1069 if (vp->v_numoutput < 0) 1070 panic("vwakeup: neg numoutput"); 1071 if ((vp->v_numoutput == 0) && (vp->v_iflag & VI_BWAIT)) { 1072 vp->v_iflag &= ~VI_BWAIT; 1073 wakeup(&vp->v_numoutput); 1074 } 1075 VI_UNLOCK(vp); 1076 } 1077} 1078 1079/* 1080 * Flush out and invalidate all buffers associated with a vnode. 1081 * Called with the underlying object locked. 1082 */ 1083int 1084vinvalbuf(vp, flags, cred, td, slpflag, slptimeo) 1085 struct vnode *vp; 1086 int flags; 1087 struct ucred *cred; 1088 struct thread *td; 1089 int slpflag, slptimeo; 1090{ 1091 struct buf *blist; 1092 int s, error; 1093 vm_object_t object; 1094 1095 GIANT_REQUIRED; 1096 1097 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1098 1099 VI_LOCK(vp); 1100 if (flags & V_SAVE) { 1101 s = splbio(); 1102 while (vp->v_numoutput) { 1103 vp->v_iflag |= VI_BWAIT; 1104 error = msleep(&vp->v_numoutput, VI_MTX(vp), 1105 slpflag | (PRIBIO + 1), "vinvlbuf", slptimeo); 1106 if (error) { 1107 VI_UNLOCK(vp); 1108 splx(s); 1109 return (error); 1110 } 1111 } 1112 if (!TAILQ_EMPTY(&vp->v_dirtyblkhd)) { 1113 splx(s); 1114 VI_UNLOCK(vp); 1115 if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0) 1116 return (error); 1117 /* 1118 * XXX We could save a lock/unlock if this was only 1119 * enabled under INVARIANTS 1120 */ 1121 VI_LOCK(vp); 1122 s = splbio(); 1123 if (vp->v_numoutput > 0 || 1124 !TAILQ_EMPTY(&vp->v_dirtyblkhd)) 1125 panic("vinvalbuf: dirty bufs"); 1126 } 1127 splx(s); 1128 } 1129 s = splbio(); 1130 /* 1131 * If you alter this loop please notice that interlock is dropped and 1132 * reacquired in flushbuflist. Special care is needed to ensure that 1133 * no race conditions occur from this. 1134 */ 1135 for (error = 0;;) { 1136 if ((blist = TAILQ_FIRST(&vp->v_cleanblkhd)) != 0 && 1137 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 1138 if (error) 1139 break; 1140 continue; 1141 } 1142 if ((blist = TAILQ_FIRST(&vp->v_dirtyblkhd)) != 0 && 1143 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 1144 if (error) 1145 break; 1146 continue; 1147 } 1148 break; 1149 } 1150 if (error) { 1151 splx(s); 1152 VI_UNLOCK(vp); 1153 return (error); 1154 } 1155 1156 /* 1157 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1158 * have write I/O in-progress but if there is a VM object then the 1159 * VM object can also have read-I/O in-progress. 1160 */ 1161 do { 1162 while (vp->v_numoutput > 0) { 1163 vp->v_iflag |= VI_BWAIT; 1164 msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vnvlbv", 0); 1165 } 1166 VI_UNLOCK(vp); 1167 if (VOP_GETVOBJECT(vp, &object) == 0) { 1168 while (object->paging_in_progress) 1169 vm_object_pip_sleep(object, "vnvlbx"); 1170 } 1171 VI_LOCK(vp); 1172 } while (vp->v_numoutput > 0); 1173 VI_UNLOCK(vp); 1174 1175 splx(s); 1176 1177 /* 1178 * Destroy the copy in the VM cache, too. 1179 */ 1180 if (VOP_GETVOBJECT(vp, &object) == 0) { 1181 vm_object_lock(object); 1182 vm_object_page_remove(object, 0, 0, 1183 (flags & V_SAVE) ? TRUE : FALSE); 1184 vm_object_unlock(object); 1185 } 1186 1187#ifdef INVARIANTS 1188 VI_LOCK(vp); 1189 if ((flags & (V_ALT | V_NORMAL)) == 0 && 1190 (!TAILQ_EMPTY(&vp->v_dirtyblkhd) || 1191 !TAILQ_EMPTY(&vp->v_cleanblkhd))) 1192 panic("vinvalbuf: flush failed"); 1193 VI_UNLOCK(vp); 1194#endif 1195 return (0); 1196} 1197 1198/* 1199 * Flush out buffers on the specified list. 1200 * 1201 */ 1202static int 1203flushbuflist(blist, flags, vp, slpflag, slptimeo, errorp) 1204 struct buf *blist; 1205 int flags; 1206 struct vnode *vp; 1207 int slpflag, slptimeo; 1208 int *errorp; 1209{ 1210 struct buf *bp, *nbp; 1211 int found, error; 1212 1213 ASSERT_VI_LOCKED(vp, "flushbuflist"); 1214 1215 for (found = 0, bp = blist; bp; bp = nbp) { 1216 nbp = TAILQ_NEXT(bp, b_vnbufs); 1217 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1218 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1219 continue; 1220 } 1221 found += 1; 1222 error = BUF_TIMELOCK(bp, 1223 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp), 1224 "flushbuf", slpflag, slptimeo); 1225 if (error) { 1226 if (error != ENOLCK) 1227 *errorp = error; 1228 goto done; 1229 } 1230 /* 1231 * XXX Since there are no node locks for NFS, I 1232 * believe there is a slight chance that a delayed 1233 * write will occur while sleeping just above, so 1234 * check for it. Note that vfs_bio_awrite expects 1235 * buffers to reside on a queue, while BUF_WRITE and 1236 * brelse do not. 1237 */ 1238 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1239 (flags & V_SAVE)) { 1240 1241 if (bp->b_vp == vp) { 1242 if (bp->b_flags & B_CLUSTEROK) { 1243 vfs_bio_awrite(bp); 1244 } else { 1245 bremfree(bp); 1246 bp->b_flags |= B_ASYNC; 1247 BUF_WRITE(bp); 1248 } 1249 } else { 1250 bremfree(bp); 1251 (void) BUF_WRITE(bp); 1252 } 1253 goto done; 1254 } 1255 bremfree(bp); 1256 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 1257 bp->b_flags &= ~B_ASYNC; 1258 brelse(bp); 1259 VI_LOCK(vp); 1260 } 1261 return (found); 1262done: 1263 VI_LOCK(vp); 1264 return (found); 1265} 1266 1267/* 1268 * Truncate a file's buffer and pages to a specified length. This 1269 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1270 * sync activity. 1271 */ 1272int 1273vtruncbuf(vp, cred, td, length, blksize) 1274 register struct vnode *vp; 1275 struct ucred *cred; 1276 struct thread *td; 1277 off_t length; 1278 int blksize; 1279{ 1280 register struct buf *bp; 1281 struct buf *nbp; 1282 int s, anyfreed; 1283 int trunclbn; 1284 1285 /* 1286 * Round up to the *next* lbn. 1287 */ 1288 trunclbn = (length + blksize - 1) / blksize; 1289 1290 s = splbio(); 1291 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1292restart: 1293 VI_LOCK(vp); 1294 anyfreed = 1; 1295 for (;anyfreed;) { 1296 anyfreed = 0; 1297 for (bp = TAILQ_FIRST(&vp->v_cleanblkhd); bp; bp = nbp) { 1298 nbp = TAILQ_NEXT(bp, b_vnbufs); 1299 if (bp->b_lblkno >= trunclbn) { 1300 if (BUF_LOCK(bp, 1301 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1302 VI_MTX(vp)) == ENOLCK) 1303 goto restart; 1304 1305 bremfree(bp); 1306 bp->b_flags |= (B_INVAL | B_RELBUF); 1307 bp->b_flags &= ~B_ASYNC; 1308 brelse(bp); 1309 anyfreed = 1; 1310 1311 if (nbp && 1312 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1313 (nbp->b_vp != vp) || 1314 (nbp->b_flags & B_DELWRI))) { 1315 goto restart; 1316 } 1317 VI_LOCK(vp); 1318 } 1319 } 1320 1321 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1322 nbp = TAILQ_NEXT(bp, b_vnbufs); 1323 if (bp->b_lblkno >= trunclbn) { 1324 if (BUF_LOCK(bp, 1325 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1326 VI_MTX(vp)) == ENOLCK) 1327 goto restart; 1328 bremfree(bp); 1329 bp->b_flags |= (B_INVAL | B_RELBUF); 1330 bp->b_flags &= ~B_ASYNC; 1331 brelse(bp); 1332 anyfreed = 1; 1333 if (nbp && 1334 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1335 (nbp->b_vp != vp) || 1336 (nbp->b_flags & B_DELWRI) == 0)) { 1337 goto restart; 1338 } 1339 VI_LOCK(vp); 1340 } 1341 } 1342 } 1343 1344 if (length > 0) { 1345restartsync: 1346 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1347 nbp = TAILQ_NEXT(bp, b_vnbufs); 1348 if ((bp->b_flags & B_DELWRI) && (bp->b_lblkno < 0)) { 1349 if (BUF_LOCK(bp, 1350 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1351 VI_MTX(vp)) == ENOLCK) 1352 goto restart; 1353 bremfree(bp); 1354 if (bp->b_vp == vp) 1355 bp->b_flags |= B_ASYNC; 1356 else 1357 bp->b_flags &= ~B_ASYNC; 1358 1359 BUF_WRITE(bp); 1360 VI_LOCK(vp); 1361 goto restartsync; 1362 } 1363 } 1364 } 1365 1366 while (vp->v_numoutput > 0) { 1367 vp->v_iflag |= VI_BWAIT; 1368 msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vbtrunc", 0); 1369 } 1370 VI_UNLOCK(vp); 1371 splx(s); 1372 1373 vnode_pager_setsize(vp, length); 1374 1375 return (0); 1376} 1377 1378/* 1379 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1380 * a vnode. 1381 * 1382 * NOTE: We have to deal with the special case of a background bitmap 1383 * buffer, a situation where two buffers will have the same logical 1384 * block offset. We want (1) only the foreground buffer to be accessed 1385 * in a lookup and (2) must differentiate between the foreground and 1386 * background buffer in the splay tree algorithm because the splay 1387 * tree cannot normally handle multiple entities with the same 'index'. 1388 * We accomplish this by adding differentiating flags to the splay tree's 1389 * numerical domain. 1390 */ 1391static 1392struct buf * 1393buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1394{ 1395 struct buf dummy; 1396 struct buf *lefttreemax, *righttreemin, *y; 1397 1398 if (root == NULL) 1399 return (NULL); 1400 lefttreemax = righttreemin = &dummy; 1401 for (;;) { 1402 if (lblkno < root->b_lblkno || 1403 (lblkno == root->b_lblkno && 1404 (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1405 if ((y = root->b_left) == NULL) 1406 break; 1407 if (lblkno < y->b_lblkno) { 1408 /* Rotate right. */ 1409 root->b_left = y->b_right; 1410 y->b_right = root; 1411 root = y; 1412 if ((y = root->b_left) == NULL) 1413 break; 1414 } 1415 /* Link into the new root's right tree. */ 1416 righttreemin->b_left = root; 1417 righttreemin = root; 1418 } else if (lblkno > root->b_lblkno || 1419 (lblkno == root->b_lblkno && 1420 (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { 1421 if ((y = root->b_right) == NULL) 1422 break; 1423 if (lblkno > y->b_lblkno) { 1424 /* Rotate left. */ 1425 root->b_right = y->b_left; 1426 y->b_left = root; 1427 root = y; 1428 if ((y = root->b_right) == NULL) 1429 break; 1430 } 1431 /* Link into the new root's left tree. */ 1432 lefttreemax->b_right = root; 1433 lefttreemax = root; 1434 } else { 1435 break; 1436 } 1437 root = y; 1438 } 1439 /* Assemble the new root. */ 1440 lefttreemax->b_right = root->b_left; 1441 righttreemin->b_left = root->b_right; 1442 root->b_left = dummy.b_right; 1443 root->b_right = dummy.b_left; 1444 return (root); 1445} 1446 1447static 1448void 1449buf_vlist_remove(struct buf *bp) 1450{ 1451 struct vnode *vp = bp->b_vp; 1452 struct buf *root; 1453 1454 ASSERT_VI_LOCKED(vp, "buf_vlist_remove"); 1455 if (bp->b_xflags & BX_VNDIRTY) { 1456 if (bp != vp->v_dirtyblkroot) { 1457 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot); 1458 KASSERT(root == bp, ("splay lookup failed during dirty remove")); 1459 } 1460 if (bp->b_left == NULL) { 1461 root = bp->b_right; 1462 } else { 1463 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1464 root->b_right = bp->b_right; 1465 } 1466 vp->v_dirtyblkroot = root; 1467 TAILQ_REMOVE(&vp->v_dirtyblkhd, bp, b_vnbufs); 1468 vp->v_dirtybufcnt--; 1469 } else { 1470 /* KASSERT(bp->b_xflags & BX_VNCLEAN, ("bp wasn't clean")); */ 1471 if (bp != vp->v_cleanblkroot) { 1472 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot); 1473 KASSERT(root == bp, ("splay lookup failed during clean remove")); 1474 } 1475 if (bp->b_left == NULL) { 1476 root = bp->b_right; 1477 } else { 1478 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1479 root->b_right = bp->b_right; 1480 } 1481 vp->v_cleanblkroot = root; 1482 TAILQ_REMOVE(&vp->v_cleanblkhd, bp, b_vnbufs); 1483 vp->v_cleanbufcnt--; 1484 } 1485 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1486} 1487 1488/* 1489 * Add the buffer to the sorted clean or dirty block list using a 1490 * splay tree algorithm. 1491 * 1492 * NOTE: xflags is passed as a constant, optimizing this inline function! 1493 */ 1494static 1495void 1496buf_vlist_add(struct buf *bp, struct vnode *vp, b_xflags_t xflags) 1497{ 1498 struct buf *root; 1499 1500 ASSERT_VI_LOCKED(vp, "buf_vlist_add"); 1501 bp->b_xflags |= xflags; 1502 if (xflags & BX_VNDIRTY) { 1503 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot); 1504 if (root == NULL) { 1505 bp->b_left = NULL; 1506 bp->b_right = NULL; 1507 TAILQ_INSERT_TAIL(&vp->v_dirtyblkhd, bp, b_vnbufs); 1508 } else if (bp->b_lblkno < root->b_lblkno || 1509 (bp->b_lblkno == root->b_lblkno && 1510 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1511 bp->b_left = root->b_left; 1512 bp->b_right = root; 1513 root->b_left = NULL; 1514 TAILQ_INSERT_BEFORE(root, bp, b_vnbufs); 1515 } else { 1516 bp->b_right = root->b_right; 1517 bp->b_left = root; 1518 root->b_right = NULL; 1519 TAILQ_INSERT_AFTER(&vp->v_dirtyblkhd, 1520 root, bp, b_vnbufs); 1521 } 1522 vp->v_dirtybufcnt++; 1523 vp->v_dirtyblkroot = bp; 1524 } else { 1525 /* KASSERT(xflags & BX_VNCLEAN, ("xflags not clean")); */ 1526 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot); 1527 if (root == NULL) { 1528 bp->b_left = NULL; 1529 bp->b_right = NULL; 1530 TAILQ_INSERT_TAIL(&vp->v_cleanblkhd, bp, b_vnbufs); 1531 } else if (bp->b_lblkno < root->b_lblkno || 1532 (bp->b_lblkno == root->b_lblkno && 1533 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1534 bp->b_left = root->b_left; 1535 bp->b_right = root; 1536 root->b_left = NULL; 1537 TAILQ_INSERT_BEFORE(root, bp, b_vnbufs); 1538 } else { 1539 bp->b_right = root->b_right; 1540 bp->b_left = root; 1541 root->b_right = NULL; 1542 TAILQ_INSERT_AFTER(&vp->v_cleanblkhd, 1543 root, bp, b_vnbufs); 1544 } 1545 vp->v_cleanbufcnt++; 1546 vp->v_cleanblkroot = bp; 1547 } 1548} 1549 1550/* 1551 * Lookup a buffer using the splay tree. Note that we specifically avoid 1552 * shadow buffers used in background bitmap writes. 1553 * 1554 * This code isn't quite efficient as it could be because we are maintaining 1555 * two sorted lists and do not know which list the block resides in. 1556 */ 1557struct buf * 1558gbincore(struct vnode *vp, daddr_t lblkno) 1559{ 1560 struct buf *bp; 1561 1562 GIANT_REQUIRED; 1563 1564 ASSERT_VI_LOCKED(vp, "gbincore"); 1565 bp = vp->v_cleanblkroot = buf_splay(lblkno, 0, vp->v_cleanblkroot); 1566 if (bp && bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1567 return(bp); 1568 bp = vp->v_dirtyblkroot = buf_splay(lblkno, 0, vp->v_dirtyblkroot); 1569 if (bp && bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1570 return(bp); 1571 return(NULL); 1572} 1573 1574/* 1575 * Associate a buffer with a vnode. 1576 */ 1577void 1578bgetvp(vp, bp) 1579 register struct vnode *vp; 1580 register struct buf *bp; 1581{ 1582 int s; 1583 1584 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 1585 1586 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1587 ("bgetvp: bp already attached! %p", bp)); 1588 1589 ASSERT_VI_LOCKED(vp, "bgetvp"); 1590 vholdl(vp); 1591 bp->b_vp = vp; 1592 bp->b_dev = vn_todev(vp); 1593 /* 1594 * Insert onto list for new vnode. 1595 */ 1596 s = splbio(); 1597 buf_vlist_add(bp, vp, BX_VNCLEAN); 1598 splx(s); 1599} 1600 1601/* 1602 * Disassociate a buffer from a vnode. 1603 */ 1604void 1605brelvp(bp) 1606 register struct buf *bp; 1607{ 1608 struct vnode *vp; 1609 int s; 1610 1611 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1612 1613 /* 1614 * Delete from old vnode list, if on one. 1615 */ 1616 vp = bp->b_vp; 1617 s = splbio(); 1618 VI_LOCK(vp); 1619 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1620 buf_vlist_remove(bp); 1621 if ((vp->v_iflag & VI_ONWORKLST) && TAILQ_EMPTY(&vp->v_dirtyblkhd)) { 1622 vp->v_iflag &= ~VI_ONWORKLST; 1623 mtx_lock(&sync_mtx); 1624 LIST_REMOVE(vp, v_synclist); 1625 mtx_unlock(&sync_mtx); 1626 } 1627 vdropl(vp); 1628 VI_UNLOCK(vp); 1629 bp->b_vp = (struct vnode *) 0; 1630 if (bp->b_object) 1631 bp->b_object = NULL; 1632 splx(s); 1633} 1634 1635/* 1636 * Add an item to the syncer work queue. 1637 */ 1638static void 1639vn_syncer_add_to_worklist(struct vnode *vp, int delay) 1640{ 1641 int s, slot; 1642 1643 s = splbio(); 1644 ASSERT_VI_LOCKED(vp, "vn_syncer_add_to_worklist"); 1645 1646 mtx_lock(&sync_mtx); 1647 if (vp->v_iflag & VI_ONWORKLST) 1648 LIST_REMOVE(vp, v_synclist); 1649 else 1650 vp->v_iflag |= VI_ONWORKLST; 1651 1652 if (delay > syncer_maxdelay - 2) 1653 delay = syncer_maxdelay - 2; 1654 slot = (syncer_delayno + delay) & syncer_mask; 1655 1656 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist); 1657 mtx_unlock(&sync_mtx); 1658 1659 splx(s); 1660} 1661 1662struct proc *updateproc; 1663static void sched_sync(void); 1664static struct kproc_desc up_kp = { 1665 "syncer", 1666 sched_sync, 1667 &updateproc 1668}; 1669SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 1670 1671/* 1672 * System filesystem synchronizer daemon. 1673 */ 1674static void 1675sched_sync(void) 1676{ 1677 struct synclist *slp; 1678 struct vnode *vp; 1679 struct mount *mp; 1680 long starttime; 1681 int s; 1682 struct thread *td = FIRST_THREAD_IN_PROC(updateproc); /* XXXKSE */ 1683 1684 mtx_lock(&Giant); 1685 1686 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, td->td_proc, 1687 SHUTDOWN_PRI_LAST); 1688 1689 for (;;) { 1690 kthread_suspend_check(td->td_proc); 1691 1692 starttime = time_second; 1693 1694 /* 1695 * Push files whose dirty time has expired. Be careful 1696 * of interrupt race on slp queue. 1697 */ 1698 s = splbio(); 1699 mtx_lock(&sync_mtx); 1700 slp = &syncer_workitem_pending[syncer_delayno]; 1701 syncer_delayno += 1; 1702 if (syncer_delayno == syncer_maxdelay) 1703 syncer_delayno = 0; 1704 splx(s); 1705 1706 while ((vp = LIST_FIRST(slp)) != NULL) { 1707 mtx_unlock(&sync_mtx); 1708 if (VOP_ISLOCKED(vp, NULL) == 0 && 1709 vn_start_write(vp, &mp, V_NOWAIT) == 0) { 1710 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 1711 (void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td); 1712 VOP_UNLOCK(vp, 0, td); 1713 vn_finished_write(mp); 1714 } 1715 s = splbio(); 1716 mtx_lock(&sync_mtx); 1717 if (LIST_FIRST(slp) == vp) { 1718 mtx_unlock(&sync_mtx); 1719 /* 1720 * Note: VFS vnodes can remain on the 1721 * worklist too with no dirty blocks, but 1722 * since sync_fsync() moves it to a different 1723 * slot we are safe. 1724 */ 1725 VI_LOCK(vp); 1726 if (TAILQ_EMPTY(&vp->v_dirtyblkhd) && 1727 !vn_isdisk(vp, NULL)) { 1728 panic("sched_sync: fsync failed " 1729 "vp %p tag %s", vp, vp->v_tag); 1730 } 1731 /* 1732 * Put us back on the worklist. The worklist 1733 * routine will remove us from our current 1734 * position and then add us back in at a later 1735 * position. 1736 */ 1737 vn_syncer_add_to_worklist(vp, syncdelay); 1738 VI_UNLOCK(vp); 1739 mtx_lock(&sync_mtx); 1740 } 1741 splx(s); 1742 } 1743 mtx_unlock(&sync_mtx); 1744 1745 /* 1746 * Do soft update processing. 1747 */ 1748 if (softdep_process_worklist_hook != NULL) 1749 (*softdep_process_worklist_hook)(NULL); 1750 1751 /* 1752 * The variable rushjob allows the kernel to speed up the 1753 * processing of the filesystem syncer process. A rushjob 1754 * value of N tells the filesystem syncer to process the next 1755 * N seconds worth of work on its queue ASAP. Currently rushjob 1756 * is used by the soft update code to speed up the filesystem 1757 * syncer process when the incore state is getting so far 1758 * ahead of the disk that the kernel memory pool is being 1759 * threatened with exhaustion. 1760 */ 1761 mtx_lock(&sync_mtx); 1762 if (rushjob > 0) { 1763 rushjob -= 1; 1764 mtx_unlock(&sync_mtx); 1765 continue; 1766 } 1767 mtx_unlock(&sync_mtx); 1768 /* 1769 * If it has taken us less than a second to process the 1770 * current work, then wait. Otherwise start right over 1771 * again. We can still lose time if any single round 1772 * takes more than two seconds, but it does not really 1773 * matter as we are just trying to generally pace the 1774 * filesystem activity. 1775 */ 1776 if (time_second == starttime) 1777 tsleep(&lbolt, PPAUSE, "syncer", 0); 1778 } 1779} 1780 1781/* 1782 * Request the syncer daemon to speed up its work. 1783 * We never push it to speed up more than half of its 1784 * normal turn time, otherwise it could take over the cpu. 1785 * XXXKSE only one update? 1786 */ 1787int 1788speedup_syncer() 1789{ 1790 struct thread *td; 1791 int ret = 0; 1792 1793 td = FIRST_THREAD_IN_PROC(updateproc); 1794 mtx_lock_spin(&sched_lock); 1795 if (td->td_wchan == &lbolt) { 1796 unsleep(td); 1797 TD_CLR_SLEEPING(td); 1798 setrunnable(td); 1799 } 1800 mtx_unlock_spin(&sched_lock); 1801 mtx_lock(&sync_mtx); 1802 if (rushjob < syncdelay / 2) { 1803 rushjob += 1; 1804 stat_rush_requests += 1; 1805 ret = 1; 1806 } 1807 mtx_unlock(&sync_mtx); 1808 return (ret); 1809} 1810 1811/* 1812 * Associate a p-buffer with a vnode. 1813 * 1814 * Also sets B_PAGING flag to indicate that vnode is not fully associated 1815 * with the buffer. i.e. the bp has not been linked into the vnode or 1816 * ref-counted. 1817 */ 1818void 1819pbgetvp(vp, bp) 1820 register struct vnode *vp; 1821 register struct buf *bp; 1822{ 1823 1824 KASSERT(bp->b_vp == NULL, ("pbgetvp: not free")); 1825 1826 bp->b_vp = vp; 1827 bp->b_flags |= B_PAGING; 1828 bp->b_dev = vn_todev(vp); 1829} 1830 1831/* 1832 * Disassociate a p-buffer from a vnode. 1833 */ 1834void 1835pbrelvp(bp) 1836 register struct buf *bp; 1837{ 1838 1839 KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL")); 1840 1841 /* XXX REMOVE ME */ 1842 VI_LOCK(bp->b_vp); 1843 if (TAILQ_NEXT(bp, b_vnbufs) != NULL) { 1844 panic( 1845 "relpbuf(): b_vp was probably reassignbuf()d %p %x", 1846 bp, 1847 (int)bp->b_flags 1848 ); 1849 } 1850 VI_UNLOCK(bp->b_vp); 1851 bp->b_vp = (struct vnode *) 0; 1852 bp->b_flags &= ~B_PAGING; 1853} 1854 1855/* 1856 * Reassign a buffer from one vnode to another. 1857 * Used to assign file specific control information 1858 * (indirect blocks) to the vnode to which they belong. 1859 */ 1860void 1861reassignbuf(bp, newvp) 1862 register struct buf *bp; 1863 register struct vnode *newvp; 1864{ 1865 int delay; 1866 int s; 1867 1868 if (newvp == NULL) { 1869 printf("reassignbuf: NULL"); 1870 return; 1871 } 1872 ++reassignbufcalls; 1873 1874 /* 1875 * B_PAGING flagged buffers cannot be reassigned because their vp 1876 * is not fully linked in. 1877 */ 1878 if (bp->b_flags & B_PAGING) 1879 panic("cannot reassign paging buffer"); 1880 1881 s = splbio(); 1882 /* 1883 * Delete from old vnode list, if on one. 1884 */ 1885 VI_LOCK(bp->b_vp); 1886 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) { 1887 buf_vlist_remove(bp); 1888 if (bp->b_vp != newvp) { 1889 vdropl(bp->b_vp); 1890 bp->b_vp = NULL; /* for clarification */ 1891 } 1892 } 1893 VI_UNLOCK(bp->b_vp); 1894 /* 1895 * If dirty, put on list of dirty buffers; otherwise insert onto list 1896 * of clean buffers. 1897 */ 1898 VI_LOCK(newvp); 1899 if (bp->b_flags & B_DELWRI) { 1900 if ((newvp->v_iflag & VI_ONWORKLST) == 0) { 1901 switch (newvp->v_type) { 1902 case VDIR: 1903 delay = dirdelay; 1904 break; 1905 case VCHR: 1906 if (newvp->v_rdev->si_mountpoint != NULL) { 1907 delay = metadelay; 1908 break; 1909 } 1910 /* FALLTHROUGH */ 1911 default: 1912 delay = filedelay; 1913 } 1914 vn_syncer_add_to_worklist(newvp, delay); 1915 } 1916 buf_vlist_add(bp, newvp, BX_VNDIRTY); 1917 } else { 1918 buf_vlist_add(bp, newvp, BX_VNCLEAN); 1919 1920 if ((newvp->v_iflag & VI_ONWORKLST) && 1921 TAILQ_EMPTY(&newvp->v_dirtyblkhd)) { 1922 mtx_lock(&sync_mtx); 1923 LIST_REMOVE(newvp, v_synclist); 1924 mtx_unlock(&sync_mtx); 1925 newvp->v_iflag &= ~VI_ONWORKLST; 1926 } 1927 } 1928 if (bp->b_vp != newvp) { 1929 bp->b_vp = newvp; 1930 vholdl(bp->b_vp); 1931 } 1932 VI_UNLOCK(newvp); 1933 splx(s); 1934} 1935 1936/* 1937 * Create a vnode for a device. 1938 * Used for mounting the root filesystem. 1939 */ 1940int 1941bdevvp(dev, vpp) 1942 dev_t dev; 1943 struct vnode **vpp; 1944{ 1945 register struct vnode *vp; 1946 struct vnode *nvp; 1947 int error; 1948 1949 if (dev == NODEV) { 1950 *vpp = NULLVP; 1951 return (ENXIO); 1952 } 1953 if (vfinddev(dev, VCHR, vpp)) 1954 return (0); 1955 error = getnewvnode("none", (struct mount *)0, spec_vnodeop_p, &nvp); 1956 if (error) { 1957 *vpp = NULLVP; 1958 return (error); 1959 } 1960 vp = nvp; 1961 vp->v_type = VCHR; 1962 addalias(vp, dev); 1963 *vpp = vp; 1964 return (0); 1965} 1966 1967static void 1968v_incr_usecount(struct vnode *vp, int delta) 1969{ 1970 vp->v_usecount += delta; 1971 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1972 mtx_lock(&spechash_mtx); 1973 vp->v_rdev->si_usecount += delta; 1974 mtx_unlock(&spechash_mtx); 1975 } 1976} 1977 1978/* 1979 * Add vnode to the alias list hung off the dev_t. 1980 * 1981 * The reason for this gunk is that multiple vnodes can reference 1982 * the same physical device, so checking vp->v_usecount to see 1983 * how many users there are is inadequate; the v_usecount for 1984 * the vnodes need to be accumulated. vcount() does that. 1985 */ 1986struct vnode * 1987addaliasu(nvp, nvp_rdev) 1988 struct vnode *nvp; 1989 udev_t nvp_rdev; 1990{ 1991 struct vnode *ovp; 1992 vop_t **ops; 1993 dev_t dev; 1994 1995 if (nvp->v_type == VBLK) 1996 return (nvp); 1997 if (nvp->v_type != VCHR) 1998 panic("addaliasu on non-special vnode"); 1999 dev = udev2dev(nvp_rdev, 0); 2000 /* 2001 * Check to see if we have a bdevvp vnode with no associated 2002 * filesystem. If so, we want to associate the filesystem of 2003 * the new newly instigated vnode with the bdevvp vnode and 2004 * discard the newly created vnode rather than leaving the 2005 * bdevvp vnode lying around with no associated filesystem. 2006 */ 2007 if (vfinddev(dev, nvp->v_type, &ovp) == 0 || ovp->v_data != NULL) { 2008 addalias(nvp, dev); 2009 return (nvp); 2010 } 2011 /* 2012 * Discard unneeded vnode, but save its node specific data. 2013 * Note that if there is a lock, it is carried over in the 2014 * node specific data to the replacement vnode. 2015 */ 2016 vref(ovp); 2017 ovp->v_data = nvp->v_data; 2018 ovp->v_tag = nvp->v_tag; 2019 nvp->v_data = NULL; 2020 lockdestroy(ovp->v_vnlock); 2021 lockinit(ovp->v_vnlock, PVFS, nvp->v_vnlock->lk_wmesg, 2022 nvp->v_vnlock->lk_timo, nvp->v_vnlock->lk_flags & LK_EXTFLG_MASK); 2023 ops = ovp->v_op; 2024 ovp->v_op = nvp->v_op; 2025 if (VOP_ISLOCKED(nvp, curthread)) { 2026 VOP_UNLOCK(nvp, 0, curthread); 2027 vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread); 2028 } 2029 nvp->v_op = ops; 2030 insmntque(ovp, nvp->v_mount); 2031 vrele(nvp); 2032 vgone(nvp); 2033 return (ovp); 2034} 2035 2036/* This is a local helper function that do the same as addaliasu, but for a 2037 * dev_t instead of an udev_t. */ 2038static void 2039addalias(nvp, dev) 2040 struct vnode *nvp; 2041 dev_t dev; 2042{ 2043 2044 KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode")); 2045 nvp->v_rdev = dev; 2046 VI_LOCK(nvp); 2047 mtx_lock(&spechash_mtx); 2048 SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext); 2049 dev->si_usecount += nvp->v_usecount; 2050 mtx_unlock(&spechash_mtx); 2051 VI_UNLOCK(nvp); 2052} 2053 2054/* 2055 * Grab a particular vnode from the free list, increment its 2056 * reference count and lock it. The vnode lock bit is set if the 2057 * vnode is being eliminated in vgone. The process is awakened 2058 * when the transition is completed, and an error returned to 2059 * indicate that the vnode is no longer usable (possibly having 2060 * been changed to a new filesystem type). 2061 */ 2062int 2063vget(vp, flags, td) 2064 register struct vnode *vp; 2065 int flags; 2066 struct thread *td; 2067{ 2068 int error; 2069 2070 /* 2071 * If the vnode is in the process of being cleaned out for 2072 * another use, we wait for the cleaning to finish and then 2073 * return failure. Cleaning is determined by checking that 2074 * the VI_XLOCK flag is set. 2075 */ 2076 if ((flags & LK_INTERLOCK) == 0) 2077 VI_LOCK(vp); 2078 if (vp->v_iflag & VI_XLOCK && vp->v_vxproc != curthread) { 2079 vp->v_iflag |= VI_XWANT; 2080 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0); 2081 return (ENOENT); 2082 } 2083 2084 v_incr_usecount(vp, 1); 2085 2086 if (VSHOULDBUSY(vp)) 2087 vbusy(vp); 2088 if (flags & LK_TYPE_MASK) { 2089 if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { 2090 /* 2091 * must expand vrele here because we do not want 2092 * to call VOP_INACTIVE if the reference count 2093 * drops back to zero since it was never really 2094 * active. We must remove it from the free list 2095 * before sleeping so that multiple processes do 2096 * not try to recycle it. 2097 */ 2098 VI_LOCK(vp); 2099 v_incr_usecount(vp, -1); 2100 if (VSHOULDFREE(vp)) 2101 vfree(vp); 2102 else 2103 vlruvp(vp); 2104 VI_UNLOCK(vp); 2105 } 2106 return (error); 2107 } 2108 VI_UNLOCK(vp); 2109 return (0); 2110} 2111 2112/* 2113 * Increase the reference count of a vnode. 2114 */ 2115void 2116vref(struct vnode *vp) 2117{ 2118 VI_LOCK(vp); 2119 v_incr_usecount(vp, 1); 2120 VI_UNLOCK(vp); 2121} 2122 2123/* 2124 * Return reference count of a vnode. 2125 * 2126 * The results of this call are only guaranteed when some mechanism other 2127 * than the VI lock is used to stop other processes from gaining references 2128 * to the vnode. This may be the case if the caller holds the only reference. 2129 * This is also useful when stale data is acceptable as race conditions may 2130 * be accounted for by some other means. 2131 */ 2132int 2133vrefcnt(struct vnode *vp) 2134{ 2135 int usecnt; 2136 2137 VI_LOCK(vp); 2138 usecnt = vp->v_usecount; 2139 VI_UNLOCK(vp); 2140 2141 return (usecnt); 2142} 2143 2144 2145/* 2146 * Vnode put/release. 2147 * If count drops to zero, call inactive routine and return to freelist. 2148 */ 2149void 2150vrele(vp) 2151 struct vnode *vp; 2152{ 2153 struct thread *td = curthread; /* XXX */ 2154 2155 KASSERT(vp != NULL, ("vrele: null vp")); 2156 2157 VI_LOCK(vp); 2158 2159 /* Skip this v_writecount check if we're going to panic below. */ 2160 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2161 ("vrele: missed vn_close")); 2162 2163 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2164 vp->v_usecount == 1)) { 2165 v_incr_usecount(vp, -1); 2166 VI_UNLOCK(vp); 2167 2168 return; 2169 } 2170 2171 if (vp->v_usecount == 1) { 2172 v_incr_usecount(vp, -1); 2173 /* 2174 * We must call VOP_INACTIVE with the node locked. Mark 2175 * as VI_DOINGINACT to avoid recursion. 2176 */ 2177 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) { 2178 VI_LOCK(vp); 2179 vp->v_iflag |= VI_DOINGINACT; 2180 VI_UNLOCK(vp); 2181 VOP_INACTIVE(vp, td); 2182 VI_LOCK(vp); 2183 KASSERT(vp->v_iflag & VI_DOINGINACT, 2184 ("vrele: lost VI_DOINGINACT")); 2185 vp->v_iflag &= ~VI_DOINGINACT; 2186 VI_UNLOCK(vp); 2187 } 2188 VI_LOCK(vp); 2189 if (VSHOULDFREE(vp)) 2190 vfree(vp); 2191 else 2192 vlruvp(vp); 2193 VI_UNLOCK(vp); 2194 2195 } else { 2196#ifdef DIAGNOSTIC 2197 vprint("vrele: negative ref count", vp); 2198#endif 2199 VI_UNLOCK(vp); 2200 panic("vrele: negative ref cnt"); 2201 } 2202} 2203 2204/* 2205 * Release an already locked vnode. This give the same effects as 2206 * unlock+vrele(), but takes less time and avoids releasing and 2207 * re-aquiring the lock (as vrele() aquires the lock internally.) 2208 */ 2209void 2210vput(vp) 2211 struct vnode *vp; 2212{ 2213 struct thread *td = curthread; /* XXX */ 2214 2215 GIANT_REQUIRED; 2216 2217 KASSERT(vp != NULL, ("vput: null vp")); 2218 VI_LOCK(vp); 2219 /* Skip this v_writecount check if we're going to panic below. */ 2220 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2221 ("vput: missed vn_close")); 2222 2223 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2224 vp->v_usecount == 1)) { 2225 v_incr_usecount(vp, -1); 2226 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2227 return; 2228 } 2229 2230 if (vp->v_usecount == 1) { 2231 v_incr_usecount(vp, -1); 2232 /* 2233 * We must call VOP_INACTIVE with the node locked, so 2234 * we just need to release the vnode mutex. Mark as 2235 * as VI_DOINGINACT to avoid recursion. 2236 */ 2237 vp->v_iflag |= VI_DOINGINACT; 2238 VI_UNLOCK(vp); 2239 VOP_INACTIVE(vp, td); 2240 VI_LOCK(vp); 2241 KASSERT(vp->v_iflag & VI_DOINGINACT, 2242 ("vput: lost VI_DOINGINACT")); 2243 vp->v_iflag &= ~VI_DOINGINACT; 2244 if (VSHOULDFREE(vp)) 2245 vfree(vp); 2246 else 2247 vlruvp(vp); 2248 VI_UNLOCK(vp); 2249 2250 } else { 2251#ifdef DIAGNOSTIC 2252 vprint("vput: negative ref count", vp); 2253#endif 2254 panic("vput: negative ref cnt"); 2255 } 2256} 2257 2258/* 2259 * Somebody doesn't want the vnode recycled. 2260 */ 2261void 2262vhold(struct vnode *vp) 2263{ 2264 VI_LOCK(vp); 2265 vholdl(vp); 2266 VI_UNLOCK(vp); 2267} 2268 2269void 2270vholdl(vp) 2271 register struct vnode *vp; 2272{ 2273 int s; 2274 2275 s = splbio(); 2276 vp->v_holdcnt++; 2277 if (VSHOULDBUSY(vp)) 2278 vbusy(vp); 2279 splx(s); 2280} 2281 2282/* 2283 * Note that there is one less who cares about this vnode. vdrop() is the 2284 * opposite of vhold(). 2285 */ 2286void 2287vdrop(struct vnode *vp) 2288{ 2289 VI_LOCK(vp); 2290 vdropl(vp); 2291 VI_UNLOCK(vp); 2292} 2293 2294void 2295vdropl(vp) 2296 register struct vnode *vp; 2297{ 2298 int s; 2299 2300 s = splbio(); 2301 if (vp->v_holdcnt <= 0) 2302 panic("vdrop: holdcnt"); 2303 vp->v_holdcnt--; 2304 if (VSHOULDFREE(vp)) 2305 vfree(vp); 2306 else 2307 vlruvp(vp); 2308 splx(s); 2309} 2310 2311/* 2312 * Remove any vnodes in the vnode table belonging to mount point mp. 2313 * 2314 * If FORCECLOSE is not specified, there should not be any active ones, 2315 * return error if any are found (nb: this is a user error, not a 2316 * system error). If FORCECLOSE is specified, detach any active vnodes 2317 * that are found. 2318 * 2319 * If WRITECLOSE is set, only flush out regular file vnodes open for 2320 * writing. 2321 * 2322 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2323 * 2324 * `rootrefs' specifies the base reference count for the root vnode 2325 * of this filesystem. The root vnode is considered busy if its 2326 * v_usecount exceeds this value. On a successful return, vflush() 2327 * will call vrele() on the root vnode exactly rootrefs times. 2328 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2329 * be zero. 2330 */ 2331#ifdef DIAGNOSTIC 2332static int busyprt = 0; /* print out busy vnodes */ 2333SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 2334#endif 2335 2336int 2337vflush(mp, rootrefs, flags) 2338 struct mount *mp; 2339 int rootrefs; 2340 int flags; 2341{ 2342 struct thread *td = curthread; /* XXX */ 2343 struct vnode *vp, *nvp, *rootvp = NULL; 2344 struct vattr vattr; 2345 int busy = 0, error; 2346 2347 if (rootrefs > 0) { 2348 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2349 ("vflush: bad args")); 2350 /* 2351 * Get the filesystem root vnode. We can vput() it 2352 * immediately, since with rootrefs > 0, it won't go away. 2353 */ 2354 if ((error = VFS_ROOT(mp, &rootvp)) != 0) 2355 return (error); 2356 vput(rootvp); 2357 2358 } 2359 mtx_lock(&mntvnode_mtx); 2360loop: 2361 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp; vp = nvp) { 2362 /* 2363 * Make sure this vnode wasn't reclaimed in getnewvnode(). 2364 * Start over if it has (it won't be on the list anymore). 2365 */ 2366 if (vp->v_mount != mp) 2367 goto loop; 2368 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 2369 2370 VI_LOCK(vp); 2371 mtx_unlock(&mntvnode_mtx); 2372 vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_RETRY, td); 2373 /* 2374 * Skip over a vnodes marked VV_SYSTEM. 2375 */ 2376 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2377 VOP_UNLOCK(vp, 0, td); 2378 mtx_lock(&mntvnode_mtx); 2379 continue; 2380 } 2381 /* 2382 * If WRITECLOSE is set, flush out unlinked but still open 2383 * files (even if open only for reading) and regular file 2384 * vnodes open for writing. 2385 */ 2386 if (flags & WRITECLOSE) { 2387 error = VOP_GETATTR(vp, &vattr, td->td_ucred, td); 2388 VI_LOCK(vp); 2389 2390 if ((vp->v_type == VNON || 2391 (error == 0 && vattr.va_nlink > 0)) && 2392 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2393 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2394 mtx_lock(&mntvnode_mtx); 2395 continue; 2396 } 2397 } else 2398 VI_LOCK(vp); 2399 2400 VOP_UNLOCK(vp, 0, td); 2401 2402 /* 2403 * With v_usecount == 0, all we need to do is clear out the 2404 * vnode data structures and we are done. 2405 */ 2406 if (vp->v_usecount == 0) { 2407 vgonel(vp, td); 2408 mtx_lock(&mntvnode_mtx); 2409 continue; 2410 } 2411 2412 /* 2413 * If FORCECLOSE is set, forcibly close the vnode. For block 2414 * or character devices, revert to an anonymous device. For 2415 * all other files, just kill them. 2416 */ 2417 if (flags & FORCECLOSE) { 2418 if (vp->v_type != VCHR) { 2419 vgonel(vp, td); 2420 } else { 2421 vclean(vp, 0, td); 2422 VI_UNLOCK(vp); 2423 vp->v_op = spec_vnodeop_p; 2424 insmntque(vp, (struct mount *) 0); 2425 } 2426 mtx_lock(&mntvnode_mtx); 2427 continue; 2428 } 2429#ifdef DIAGNOSTIC 2430 if (busyprt) 2431 vprint("vflush: busy vnode", vp); 2432#endif 2433 VI_UNLOCK(vp); 2434 mtx_lock(&mntvnode_mtx); 2435 busy++; 2436 } 2437 mtx_unlock(&mntvnode_mtx); 2438 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2439 /* 2440 * If just the root vnode is busy, and if its refcount 2441 * is equal to `rootrefs', then go ahead and kill it. 2442 */ 2443 VI_LOCK(rootvp); 2444 KASSERT(busy > 0, ("vflush: not busy")); 2445 KASSERT(rootvp->v_usecount >= rootrefs, ("vflush: rootrefs")); 2446 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2447 vgonel(rootvp, td); 2448 busy = 0; 2449 } else 2450 VI_UNLOCK(rootvp); 2451 } 2452 if (busy) 2453 return (EBUSY); 2454 for (; rootrefs > 0; rootrefs--) 2455 vrele(rootvp); 2456 return (0); 2457} 2458 2459/* 2460 * This moves a now (likely recyclable) vnode to the end of the 2461 * mountlist. XXX However, it is temporarily disabled until we 2462 * can clean up ffs_sync() and friends, which have loop restart 2463 * conditions which this code causes to operate O(N^2). 2464 */ 2465static void 2466vlruvp(struct vnode *vp) 2467{ 2468#if 0 2469 struct mount *mp; 2470 2471 if ((mp = vp->v_mount) != NULL) { 2472 mtx_lock(&mntvnode_mtx); 2473 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2474 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2475 mtx_unlock(&mntvnode_mtx); 2476 } 2477#endif 2478} 2479 2480/* 2481 * Disassociate the underlying filesystem from a vnode. 2482 */ 2483static void 2484vclean(vp, flags, td) 2485 struct vnode *vp; 2486 int flags; 2487 struct thread *td; 2488{ 2489 int active; 2490 2491 ASSERT_VI_LOCKED(vp, "vclean"); 2492 /* 2493 * Check to see if the vnode is in use. If so we have to reference it 2494 * before we clean it out so that its count cannot fall to zero and 2495 * generate a race against ourselves to recycle it. 2496 */ 2497 if ((active = vp->v_usecount)) 2498 v_incr_usecount(vp, 1); 2499 2500 /* 2501 * Prevent the vnode from being recycled or brought into use while we 2502 * clean it out. 2503 */ 2504 if (vp->v_iflag & VI_XLOCK) 2505 panic("vclean: deadlock"); 2506 vp->v_iflag |= VI_XLOCK; 2507 vp->v_vxproc = curthread; 2508 /* 2509 * Even if the count is zero, the VOP_INACTIVE routine may still 2510 * have the object locked while it cleans it out. The VOP_LOCK 2511 * ensures that the VOP_INACTIVE routine is done with its work. 2512 * For active vnodes, it ensures that no other activity can 2513 * occur while the underlying object is being cleaned out. 2514 */ 2515 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2516 2517 /* 2518 * Clean out any buffers associated with the vnode. 2519 * If the flush fails, just toss the buffers. 2520 */ 2521 if (flags & DOCLOSE) { 2522 struct buf *bp; 2523 VI_LOCK(vp); 2524 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 2525 VI_UNLOCK(vp); 2526 if (bp != NULL) 2527 (void) vn_write_suspend_wait(vp, NULL, V_WAIT); 2528 if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0) 2529 vinvalbuf(vp, 0, NOCRED, td, 0, 0); 2530 } 2531 2532 VOP_DESTROYVOBJECT(vp); 2533 2534 /* 2535 * Any other processes trying to obtain this lock must first 2536 * wait for VXLOCK to clear, then call the new lock operation. 2537 */ 2538 VOP_UNLOCK(vp, 0, td); 2539 2540 /* 2541 * If purging an active vnode, it must be closed and 2542 * deactivated before being reclaimed. Note that the 2543 * VOP_INACTIVE will unlock the vnode. 2544 */ 2545 if (active) { 2546 if (flags & DOCLOSE) 2547 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2548 VI_LOCK(vp); 2549 if ((vp->v_iflag & VI_DOINGINACT) == 0) { 2550 vp->v_iflag |= VI_DOINGINACT; 2551 VI_UNLOCK(vp); 2552 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 2553 panic("vclean: cannot relock."); 2554 VOP_INACTIVE(vp, td); 2555 VI_LOCK(vp); 2556 KASSERT(vp->v_iflag & VI_DOINGINACT, 2557 ("vclean: lost VI_DOINGINACT")); 2558 vp->v_iflag &= ~VI_DOINGINACT; 2559 } 2560 VI_UNLOCK(vp); 2561 } 2562 2563 /* 2564 * Reclaim the vnode. 2565 */ 2566 if (VOP_RECLAIM(vp, td)) 2567 panic("vclean: cannot reclaim"); 2568 2569 if (active) { 2570 /* 2571 * Inline copy of vrele() since VOP_INACTIVE 2572 * has already been called. 2573 */ 2574 VI_LOCK(vp); 2575 v_incr_usecount(vp, -1); 2576 if (vp->v_usecount <= 0) { 2577#ifdef DIAGNOSTIC 2578 if (vp->v_usecount < 0 || vp->v_writecount != 0) { 2579 vprint("vclean: bad ref count", vp); 2580 panic("vclean: ref cnt"); 2581 } 2582#endif 2583 vfree(vp); 2584 } 2585 VI_UNLOCK(vp); 2586 } 2587 2588 cache_purge(vp); 2589 VI_LOCK(vp); 2590 if (VSHOULDFREE(vp)) 2591 vfree(vp); 2592 2593 /* 2594 * Done with purge, reset to the standard lock and 2595 * notify sleepers of the grim news. 2596 */ 2597 vp->v_vnlock = &vp->v_lock; 2598 vp->v_op = dead_vnodeop_p; 2599 if (vp->v_pollinfo != NULL) 2600 vn_pollgone(vp); 2601 vp->v_tag = "none"; 2602 vp->v_iflag &= ~VI_XLOCK; 2603 vp->v_vxproc = NULL; 2604 if (vp->v_iflag & VI_XWANT) { 2605 vp->v_iflag &= ~VI_XWANT; 2606 wakeup(vp); 2607 } 2608} 2609 2610/* 2611 * Eliminate all activity associated with the requested vnode 2612 * and with all vnodes aliased to the requested vnode. 2613 */ 2614int 2615vop_revoke(ap) 2616 struct vop_revoke_args /* { 2617 struct vnode *a_vp; 2618 int a_flags; 2619 } */ *ap; 2620{ 2621 struct vnode *vp, *vq; 2622 dev_t dev; 2623 2624 KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke")); 2625 vp = ap->a_vp; 2626 KASSERT((vp->v_type == VCHR), ("vop_revoke: not VCHR")); 2627 2628 VI_LOCK(vp); 2629 /* 2630 * If a vgone (or vclean) is already in progress, 2631 * wait until it is done and return. 2632 */ 2633 if (vp->v_iflag & VI_XLOCK) { 2634 vp->v_iflag |= VI_XWANT; 2635 msleep(vp, VI_MTX(vp), PINOD | PDROP, 2636 "vop_revokeall", 0); 2637 return (0); 2638 } 2639 VI_UNLOCK(vp); 2640 dev = vp->v_rdev; 2641 for (;;) { 2642 mtx_lock(&spechash_mtx); 2643 vq = SLIST_FIRST(&dev->si_hlist); 2644 mtx_unlock(&spechash_mtx); 2645 if (!vq) 2646 break; 2647 vgone(vq); 2648 } 2649 return (0); 2650} 2651 2652/* 2653 * Recycle an unused vnode to the front of the free list. 2654 * Release the passed interlock if the vnode will be recycled. 2655 */ 2656int 2657vrecycle(vp, inter_lkp, td) 2658 struct vnode *vp; 2659 struct mtx *inter_lkp; 2660 struct thread *td; 2661{ 2662 2663 VI_LOCK(vp); 2664 if (vp->v_usecount == 0) { 2665 if (inter_lkp) { 2666 mtx_unlock(inter_lkp); 2667 } 2668 vgonel(vp, td); 2669 return (1); 2670 } 2671 VI_UNLOCK(vp); 2672 return (0); 2673} 2674 2675/* 2676 * Eliminate all activity associated with a vnode 2677 * in preparation for reuse. 2678 */ 2679void 2680vgone(vp) 2681 register struct vnode *vp; 2682{ 2683 struct thread *td = curthread; /* XXX */ 2684 2685 VI_LOCK(vp); 2686 vgonel(vp, td); 2687} 2688 2689/* 2690 * vgone, with the vp interlock held. 2691 */ 2692void 2693vgonel(vp, td) 2694 struct vnode *vp; 2695 struct thread *td; 2696{ 2697 int s; 2698 2699 /* 2700 * If a vgone (or vclean) is already in progress, 2701 * wait until it is done and return. 2702 */ 2703 ASSERT_VI_LOCKED(vp, "vgonel"); 2704 if (vp->v_iflag & VI_XLOCK) { 2705 vp->v_iflag |= VI_XWANT; 2706 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0); 2707 return; 2708 } 2709 2710 /* 2711 * Clean out the filesystem specific data. 2712 */ 2713 vclean(vp, DOCLOSE, td); 2714 VI_UNLOCK(vp); 2715 2716 /* 2717 * Delete from old mount point vnode list, if on one. 2718 */ 2719 if (vp->v_mount != NULL) 2720 insmntque(vp, (struct mount *)0); 2721 /* 2722 * If special device, remove it from special device alias list 2723 * if it is on one. 2724 */ 2725 if (vp->v_type == VCHR && vp->v_rdev != NULL && vp->v_rdev != NODEV) { 2726 VI_LOCK(vp); 2727 mtx_lock(&spechash_mtx); 2728 SLIST_REMOVE(&vp->v_rdev->si_hlist, vp, vnode, v_specnext); 2729 vp->v_rdev->si_usecount -= vp->v_usecount; 2730 mtx_unlock(&spechash_mtx); 2731 VI_UNLOCK(vp); 2732 vp->v_rdev = NULL; 2733 } 2734 2735 /* 2736 * If it is on the freelist and not already at the head, 2737 * move it to the head of the list. The test of the 2738 * VDOOMED flag and the reference count of zero is because 2739 * it will be removed from the free list by getnewvnode, 2740 * but will not have its reference count incremented until 2741 * after calling vgone. If the reference count were 2742 * incremented first, vgone would (incorrectly) try to 2743 * close the previous instance of the underlying object. 2744 */ 2745 VI_LOCK(vp); 2746 if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) { 2747 s = splbio(); 2748 mtx_lock(&vnode_free_list_mtx); 2749 if (vp->v_iflag & VI_FREE) { 2750 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2751 } else { 2752 vp->v_iflag |= VI_FREE; 2753 freevnodes++; 2754 } 2755 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2756 mtx_unlock(&vnode_free_list_mtx); 2757 splx(s); 2758 } 2759 2760 vp->v_type = VBAD; 2761 VI_UNLOCK(vp); 2762} 2763 2764/* 2765 * Lookup a vnode by device number. 2766 */ 2767int 2768vfinddev(dev, type, vpp) 2769 dev_t dev; 2770 enum vtype type; 2771 struct vnode **vpp; 2772{ 2773 struct vnode *vp; 2774 2775 mtx_lock(&spechash_mtx); 2776 SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) { 2777 if (type == vp->v_type) { 2778 *vpp = vp; 2779 mtx_unlock(&spechash_mtx); 2780 return (1); 2781 } 2782 } 2783 mtx_unlock(&spechash_mtx); 2784 return (0); 2785} 2786 2787/* 2788 * Calculate the total number of references to a special device. 2789 */ 2790int 2791vcount(vp) 2792 struct vnode *vp; 2793{ 2794 int count; 2795 2796 mtx_lock(&spechash_mtx); 2797 count = vp->v_rdev->si_usecount; 2798 mtx_unlock(&spechash_mtx); 2799 return (count); 2800} 2801 2802/* 2803 * Same as above, but using the dev_t as argument 2804 */ 2805int 2806count_dev(dev) 2807 dev_t dev; 2808{ 2809 struct vnode *vp; 2810 2811 vp = SLIST_FIRST(&dev->si_hlist); 2812 if (vp == NULL) 2813 return (0); 2814 return(vcount(vp)); 2815} 2816 2817/* 2818 * Print out a description of a vnode. 2819 */ 2820static char *typename[] = 2821{"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 2822 2823void 2824vprint(label, vp) 2825 char *label; 2826 struct vnode *vp; 2827{ 2828 char buf[96]; 2829 2830 if (label != NULL) 2831 printf("%s: %p: ", label, (void *)vp); 2832 else 2833 printf("%p: ", (void *)vp); 2834 printf("tag %s, type %s, usecount %d, writecount %d, refcount %d,", 2835 vp->v_tag, typename[vp->v_type], vp->v_usecount, 2836 vp->v_writecount, vp->v_holdcnt); 2837 buf[0] = '\0'; 2838 if (vp->v_vflag & VV_ROOT) 2839 strcat(buf, "|VV_ROOT"); 2840 if (vp->v_vflag & VV_TEXT) 2841 strcat(buf, "|VV_TEXT"); 2842 if (vp->v_vflag & VV_SYSTEM) 2843 strcat(buf, "|VV_SYSTEM"); 2844 if (vp->v_iflag & VI_XLOCK) 2845 strcat(buf, "|VI_XLOCK"); 2846 if (vp->v_iflag & VI_XWANT) 2847 strcat(buf, "|VI_XWANT"); 2848 if (vp->v_iflag & VI_BWAIT) 2849 strcat(buf, "|VI_BWAIT"); 2850 if (vp->v_iflag & VI_DOOMED) 2851 strcat(buf, "|VI_DOOMED"); 2852 if (vp->v_iflag & VI_FREE) 2853 strcat(buf, "|VI_FREE"); 2854 if (vp->v_vflag & VV_OBJBUF) 2855 strcat(buf, "|VV_OBJBUF"); 2856 if (buf[0] != '\0') 2857 printf(" flags (%s),", &buf[1]); 2858 lockmgr_printinfo(vp->v_vnlock); 2859 printf("\n"); 2860 if (vp->v_data != NULL) 2861 VOP_PRINT(vp); 2862} 2863 2864#ifdef DDB 2865#include <ddb/ddb.h> 2866/* 2867 * List all of the locked vnodes in the system. 2868 * Called when debugging the kernel. 2869 */ 2870DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 2871{ 2872 struct thread *td = curthread; /* XXX */ 2873 struct mount *mp, *nmp; 2874 struct vnode *vp; 2875 2876 printf("Locked vnodes\n"); 2877 mtx_lock(&mountlist_mtx); 2878 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2879 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 2880 nmp = TAILQ_NEXT(mp, mnt_list); 2881 continue; 2882 } 2883 mtx_lock(&mntvnode_mtx); 2884 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2885 if (VOP_ISLOCKED(vp, NULL)) 2886 vprint(NULL, vp); 2887 } 2888 mtx_unlock(&mntvnode_mtx); 2889 mtx_lock(&mountlist_mtx); 2890 nmp = TAILQ_NEXT(mp, mnt_list); 2891 vfs_unbusy(mp, td); 2892 } 2893 mtx_unlock(&mountlist_mtx); 2894} 2895#endif 2896 2897/* 2898 * Fill in a struct xvfsconf based on a struct vfsconf. 2899 */ 2900static void 2901vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) 2902{ 2903 2904 strcpy(xvfsp->vfc_name, vfsp->vfc_name); 2905 xvfsp->vfc_typenum = vfsp->vfc_typenum; 2906 xvfsp->vfc_refcount = vfsp->vfc_refcount; 2907 xvfsp->vfc_flags = vfsp->vfc_flags; 2908 /* 2909 * These are unused in userland, we keep them 2910 * to not break binary compatibility. 2911 */ 2912 xvfsp->vfc_vfsops = NULL; 2913 xvfsp->vfc_next = NULL; 2914} 2915 2916static int 2917sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 2918{ 2919 struct vfsconf *vfsp; 2920 struct xvfsconf *xvfsp; 2921 int cnt, error, i; 2922 2923 cnt = 0; 2924 for (vfsp = vfsconf; vfsp != NULL; vfsp = vfsp->vfc_next) 2925 cnt++; 2926 xvfsp = malloc(sizeof(struct xvfsconf) * cnt, M_TEMP, M_WAITOK); 2927 /* 2928 * Handle the race that we will have here when struct vfsconf 2929 * will be locked down by using both cnt and checking vfc_next 2930 * against NULL to determine the end of the loop. The race will 2931 * happen because we will have to unlock before calling malloc(). 2932 * We are protected by Giant for now. 2933 */ 2934 i = 0; 2935 for (vfsp = vfsconf; vfsp != NULL && i < cnt; vfsp = vfsp->vfc_next) { 2936 vfsconf2x(vfsp, xvfsp + i); 2937 i++; 2938 } 2939 error = SYSCTL_OUT(req, xvfsp, sizeof(struct xvfsconf) * i); 2940 free(xvfsp, M_TEMP); 2941 return (error); 2942} 2943 2944SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, 2945 "S,xvfsconf", "List of all configured filesystems"); 2946 2947/* 2948 * Top level filesystem related information gathering. 2949 */ 2950static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2951 2952static int 2953vfs_sysctl(SYSCTL_HANDLER_ARGS) 2954{ 2955 int *name = (int *)arg1 - 1; /* XXX */ 2956 u_int namelen = arg2 + 1; /* XXX */ 2957 struct vfsconf *vfsp; 2958 struct xvfsconf xvfsp; 2959 2960 printf("WARNING: userland calling deprecated sysctl, " 2961 "please rebuild world\n"); 2962 2963#if 1 || defined(COMPAT_PRELITE2) 2964 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2965 if (namelen == 1) 2966 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2967#endif 2968 2969 switch (name[1]) { 2970 case VFS_MAXTYPENUM: 2971 if (namelen != 2) 2972 return (ENOTDIR); 2973 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2974 case VFS_CONF: 2975 if (namelen != 3) 2976 return (ENOTDIR); /* overloaded */ 2977 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) 2978 if (vfsp->vfc_typenum == name[2]) 2979 break; 2980 if (vfsp == NULL) 2981 return (EOPNOTSUPP); 2982 vfsconf2x(vfsp, &xvfsp); 2983 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 2984 } 2985 return (EOPNOTSUPP); 2986} 2987 2988SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl, 2989 "Generic filesystem"); 2990 2991#if 1 || defined(COMPAT_PRELITE2) 2992 2993static int 2994sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 2995{ 2996 int error; 2997 struct vfsconf *vfsp; 2998 struct ovfsconf ovfs; 2999 3000 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) { 3001 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 3002 strcpy(ovfs.vfc_name, vfsp->vfc_name); 3003 ovfs.vfc_index = vfsp->vfc_typenum; 3004 ovfs.vfc_refcount = vfsp->vfc_refcount; 3005 ovfs.vfc_flags = vfsp->vfc_flags; 3006 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 3007 if (error) 3008 return error; 3009 } 3010 return 0; 3011} 3012 3013#endif /* 1 || COMPAT_PRELITE2 */ 3014 3015#define KINFO_VNODESLOP 10 3016#ifdef notyet 3017/* 3018 * Dump vnode list (via sysctl). 3019 */ 3020/* ARGSUSED */ 3021static int 3022sysctl_vnode(SYSCTL_HANDLER_ARGS) 3023{ 3024 struct xvnode *xvn; 3025 struct thread *td = req->td; 3026 struct mount *mp; 3027 struct vnode *vp; 3028 int error, len, n; 3029 3030 /* 3031 * Stale numvnodes access is not fatal here. 3032 */ 3033 req->lock = 0; 3034 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 3035 if (!req->oldptr) 3036 /* Make an estimate */ 3037 return (SYSCTL_OUT(req, 0, len)); 3038 3039 sysctl_wire_old_buffer(req, 0); 3040 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 3041 n = 0; 3042 mtx_lock(&mountlist_mtx); 3043 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3044 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 3045 continue; 3046 mtx_lock(&mntvnode_mtx); 3047 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3048 if (n == len) 3049 break; 3050 vref(vp); 3051 xvn[n].xv_size = sizeof *xvn; 3052 xvn[n].xv_vnode = vp; 3053#define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 3054 XV_COPY(usecount); 3055 XV_COPY(writecount); 3056 XV_COPY(holdcnt); 3057 XV_COPY(id); 3058 XV_COPY(mount); 3059 XV_COPY(numoutput); 3060 XV_COPY(type); 3061#undef XV_COPY 3062 xvn[n].xv_flag = vp->v_vflag; 3063 3064 switch (vp->v_type) { 3065 case VREG: 3066 case VDIR: 3067 case VLNK: 3068 xvn[n].xv_dev = vp->v_cachedfs; 3069 xvn[n].xv_ino = vp->v_cachedid; 3070 break; 3071 case VBLK: 3072 case VCHR: 3073 if (vp->v_rdev == NULL) { 3074 vrele(vp); 3075 continue; 3076 } 3077 xvn[n].xv_dev = dev2udev(vp->v_rdev); 3078 break; 3079 case VSOCK: 3080 xvn[n].xv_socket = vp->v_socket; 3081 break; 3082 case VFIFO: 3083 xvn[n].xv_fifo = vp->v_fifoinfo; 3084 break; 3085 case VNON: 3086 case VBAD: 3087 default: 3088 /* shouldn't happen? */ 3089 vrele(vp); 3090 continue; 3091 } 3092 vrele(vp); 3093 ++n; 3094 } 3095 mtx_unlock(&mntvnode_mtx); 3096 mtx_lock(&mountlist_mtx); 3097 vfs_unbusy(mp, td); 3098 if (n == len) 3099 break; 3100 } 3101 mtx_unlock(&mountlist_mtx); 3102 3103 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 3104 free(xvn, M_TEMP); 3105 return (error); 3106} 3107 3108SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 3109 0, 0, sysctl_vnode, "S,xvnode", ""); 3110#endif 3111 3112/* 3113 * Check to see if a filesystem is mounted on a block device. 3114 */ 3115int 3116vfs_mountedon(vp) 3117 struct vnode *vp; 3118{ 3119 3120 if (vp->v_rdev->si_mountpoint != NULL) 3121 return (EBUSY); 3122 return (0); 3123} 3124 3125/* 3126 * Unmount all filesystems. The list is traversed in reverse order 3127 * of mounting to avoid dependencies. 3128 */ 3129void 3130vfs_unmountall() 3131{ 3132 struct mount *mp; 3133 struct thread *td; 3134 int error; 3135 3136 if (curthread != NULL) 3137 td = curthread; 3138 else 3139 td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */ 3140 /* 3141 * Since this only runs when rebooting, it is not interlocked. 3142 */ 3143 while(!TAILQ_EMPTY(&mountlist)) { 3144 mp = TAILQ_LAST(&mountlist, mntlist); 3145 error = dounmount(mp, MNT_FORCE, td); 3146 if (error) { 3147 TAILQ_REMOVE(&mountlist, mp, mnt_list); 3148 printf("unmount of %s failed (", 3149 mp->mnt_stat.f_mntonname); 3150 if (error == EBUSY) 3151 printf("BUSY)\n"); 3152 else 3153 printf("%d)\n", error); 3154 } else { 3155 /* The unmount has removed mp from the mountlist */ 3156 } 3157 } 3158} 3159 3160/* 3161 * perform msync on all vnodes under a mount point 3162 * the mount point must be locked. 3163 */ 3164void 3165vfs_msync(struct mount *mp, int flags) 3166{ 3167 struct vnode *vp, *nvp; 3168 struct vm_object *obj; 3169 int tries; 3170 3171 GIANT_REQUIRED; 3172 3173 tries = 5; 3174 mtx_lock(&mntvnode_mtx); 3175loop: 3176 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp != NULL; vp = nvp) { 3177 if (vp->v_mount != mp) { 3178 if (--tries > 0) 3179 goto loop; 3180 break; 3181 } 3182 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 3183 3184 VI_LOCK(vp); 3185 if (vp->v_iflag & VI_XLOCK) { /* XXX: what if MNT_WAIT? */ 3186 VI_UNLOCK(vp); 3187 continue; 3188 } 3189 3190 if ((vp->v_iflag & VI_OBJDIRTY) && 3191 (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { 3192 mtx_unlock(&mntvnode_mtx); 3193 if (!vget(vp, 3194 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 3195 curthread)) { 3196 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 3197 vput(vp); 3198 mtx_lock(&mntvnode_mtx); 3199 continue; 3200 } 3201 3202 if (VOP_GETVOBJECT(vp, &obj) == 0) { 3203 vm_object_page_clean(obj, 0, 0, 3204 flags == MNT_WAIT ? 3205 OBJPC_SYNC : OBJPC_NOSYNC); 3206 } 3207 vput(vp); 3208 } 3209 mtx_lock(&mntvnode_mtx); 3210 if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) { 3211 if (--tries > 0) 3212 goto loop; 3213 break; 3214 } 3215 } else 3216 VI_UNLOCK(vp); 3217 } 3218 mtx_unlock(&mntvnode_mtx); 3219} 3220 3221/* 3222 * Create the VM object needed for VMIO and mmap support. This 3223 * is done for all VREG files in the system. Some filesystems might 3224 * afford the additional metadata buffering capability of the 3225 * VMIO code by making the device node be VMIO mode also. 3226 * 3227 * vp must be locked when vfs_object_create is called. 3228 */ 3229int 3230vfs_object_create(vp, td, cred) 3231 struct vnode *vp; 3232 struct thread *td; 3233 struct ucred *cred; 3234{ 3235 GIANT_REQUIRED; 3236 return (VOP_CREATEVOBJECT(vp, cred, td)); 3237} 3238 3239/* 3240 * Mark a vnode as free, putting it up for recycling. 3241 */ 3242void 3243vfree(vp) 3244 struct vnode *vp; 3245{ 3246 int s; 3247 3248 ASSERT_VI_LOCKED(vp, "vfree"); 3249 s = splbio(); 3250 mtx_lock(&vnode_free_list_mtx); 3251 KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free")); 3252 if (vp->v_iflag & VI_AGE) { 3253 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 3254 } else { 3255 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 3256 } 3257 freevnodes++; 3258 mtx_unlock(&vnode_free_list_mtx); 3259 vp->v_iflag &= ~VI_AGE; 3260 vp->v_iflag |= VI_FREE; 3261 splx(s); 3262} 3263 3264/* 3265 * Opposite of vfree() - mark a vnode as in use. 3266 */ 3267void 3268vbusy(vp) 3269 struct vnode *vp; 3270{ 3271 int s; 3272 3273 s = splbio(); 3274 ASSERT_VI_LOCKED(vp, "vbusy"); 3275 KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free")); 3276 3277 mtx_lock(&vnode_free_list_mtx); 3278 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 3279 freevnodes--; 3280 mtx_unlock(&vnode_free_list_mtx); 3281 3282 vp->v_iflag &= ~(VI_FREE|VI_AGE); 3283 splx(s); 3284} 3285 3286/* 3287 * Record a process's interest in events which might happen to 3288 * a vnode. Because poll uses the historic select-style interface 3289 * internally, this routine serves as both the ``check for any 3290 * pending events'' and the ``record my interest in future events'' 3291 * functions. (These are done together, while the lock is held, 3292 * to avoid race conditions.) 3293 */ 3294int 3295vn_pollrecord(vp, td, events) 3296 struct vnode *vp; 3297 struct thread *td; 3298 short events; 3299{ 3300 3301 if (vp->v_pollinfo == NULL) 3302 v_addpollinfo(vp); 3303 mtx_lock(&vp->v_pollinfo->vpi_lock); 3304 if (vp->v_pollinfo->vpi_revents & events) { 3305 /* 3306 * This leaves events we are not interested 3307 * in available for the other process which 3308 * which presumably had requested them 3309 * (otherwise they would never have been 3310 * recorded). 3311 */ 3312 events &= vp->v_pollinfo->vpi_revents; 3313 vp->v_pollinfo->vpi_revents &= ~events; 3314 3315 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3316 return events; 3317 } 3318 vp->v_pollinfo->vpi_events |= events; 3319 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 3320 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3321 return 0; 3322} 3323 3324/* 3325 * Note the occurrence of an event. If the VN_POLLEVENT macro is used, 3326 * it is possible for us to miss an event due to race conditions, but 3327 * that condition is expected to be rare, so for the moment it is the 3328 * preferred interface. 3329 */ 3330void 3331vn_pollevent(vp, events) 3332 struct vnode *vp; 3333 short events; 3334{ 3335 3336 if (vp->v_pollinfo == NULL) 3337 v_addpollinfo(vp); 3338 mtx_lock(&vp->v_pollinfo->vpi_lock); 3339 if (vp->v_pollinfo->vpi_events & events) { 3340 /* 3341 * We clear vpi_events so that we don't 3342 * call selwakeup() twice if two events are 3343 * posted before the polling process(es) is 3344 * awakened. This also ensures that we take at 3345 * most one selwakeup() if the polling process 3346 * is no longer interested. However, it does 3347 * mean that only one event can be noticed at 3348 * a time. (Perhaps we should only clear those 3349 * event bits which we note?) XXX 3350 */ 3351 vp->v_pollinfo->vpi_events = 0; /* &= ~events ??? */ 3352 vp->v_pollinfo->vpi_revents |= events; 3353 selwakeup(&vp->v_pollinfo->vpi_selinfo); 3354 } 3355 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3356} 3357 3358/* 3359 * Wake up anyone polling on vp because it is being revoked. 3360 * This depends on dead_poll() returning POLLHUP for correct 3361 * behavior. 3362 */ 3363void 3364vn_pollgone(vp) 3365 struct vnode *vp; 3366{ 3367 3368 mtx_lock(&vp->v_pollinfo->vpi_lock); 3369 VN_KNOTE(vp, NOTE_REVOKE); 3370 if (vp->v_pollinfo->vpi_events) { 3371 vp->v_pollinfo->vpi_events = 0; 3372 selwakeup(&vp->v_pollinfo->vpi_selinfo); 3373 } 3374 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3375} 3376 3377 3378 3379/* 3380 * Routine to create and manage a filesystem syncer vnode. 3381 */ 3382#define sync_close ((int (*)(struct vop_close_args *))nullop) 3383static int sync_fsync(struct vop_fsync_args *); 3384static int sync_inactive(struct vop_inactive_args *); 3385static int sync_reclaim(struct vop_reclaim_args *); 3386 3387static vop_t **sync_vnodeop_p; 3388static struct vnodeopv_entry_desc sync_vnodeop_entries[] = { 3389 { &vop_default_desc, (vop_t *) vop_eopnotsupp }, 3390 { &vop_close_desc, (vop_t *) sync_close }, /* close */ 3391 { &vop_fsync_desc, (vop_t *) sync_fsync }, /* fsync */ 3392 { &vop_inactive_desc, (vop_t *) sync_inactive }, /* inactive */ 3393 { &vop_reclaim_desc, (vop_t *) sync_reclaim }, /* reclaim */ 3394 { &vop_lock_desc, (vop_t *) vop_stdlock }, /* lock */ 3395 { &vop_unlock_desc, (vop_t *) vop_stdunlock }, /* unlock */ 3396 { &vop_islocked_desc, (vop_t *) vop_stdislocked }, /* islocked */ 3397 { NULL, NULL } 3398}; 3399static struct vnodeopv_desc sync_vnodeop_opv_desc = 3400 { &sync_vnodeop_p, sync_vnodeop_entries }; 3401 3402VNODEOP_SET(sync_vnodeop_opv_desc); 3403 3404/* 3405 * Create a new filesystem syncer vnode for the specified mount point. 3406 */ 3407int 3408vfs_allocate_syncvnode(mp) 3409 struct mount *mp; 3410{ 3411 struct vnode *vp; 3412 static long start, incr, next; 3413 int error; 3414 3415 /* Allocate a new vnode */ 3416 if ((error = getnewvnode("syncer", mp, sync_vnodeop_p, &vp)) != 0) { 3417 mp->mnt_syncer = NULL; 3418 return (error); 3419 } 3420 vp->v_type = VNON; 3421 /* 3422 * Place the vnode onto the syncer worklist. We attempt to 3423 * scatter them about on the list so that they will go off 3424 * at evenly distributed times even if all the filesystems 3425 * are mounted at once. 3426 */ 3427 next += incr; 3428 if (next == 0 || next > syncer_maxdelay) { 3429 start /= 2; 3430 incr /= 2; 3431 if (start == 0) { 3432 start = syncer_maxdelay / 2; 3433 incr = syncer_maxdelay; 3434 } 3435 next = start; 3436 } 3437 VI_LOCK(vp); 3438 vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0); 3439 VI_UNLOCK(vp); 3440 mp->mnt_syncer = vp; 3441 return (0); 3442} 3443 3444/* 3445 * Do a lazy sync of the filesystem. 3446 */ 3447static int 3448sync_fsync(ap) 3449 struct vop_fsync_args /* { 3450 struct vnode *a_vp; 3451 struct ucred *a_cred; 3452 int a_waitfor; 3453 struct thread *a_td; 3454 } */ *ap; 3455{ 3456 struct vnode *syncvp = ap->a_vp; 3457 struct mount *mp = syncvp->v_mount; 3458 struct thread *td = ap->a_td; 3459 int error, asyncflag; 3460 3461 /* 3462 * We only need to do something if this is a lazy evaluation. 3463 */ 3464 if (ap->a_waitfor != MNT_LAZY) 3465 return (0); 3466 3467 /* 3468 * Move ourselves to the back of the sync list. 3469 */ 3470 VI_LOCK(syncvp); 3471 vn_syncer_add_to_worklist(syncvp, syncdelay); 3472 VI_UNLOCK(syncvp); 3473 3474 /* 3475 * Walk the list of vnodes pushing all that are dirty and 3476 * not already on the sync list. 3477 */ 3478 mtx_lock(&mountlist_mtx); 3479 if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { 3480 mtx_unlock(&mountlist_mtx); 3481 return (0); 3482 } 3483 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 3484 vfs_unbusy(mp, td); 3485 return (0); 3486 } 3487 asyncflag = mp->mnt_flag & MNT_ASYNC; 3488 mp->mnt_flag &= ~MNT_ASYNC; 3489 vfs_msync(mp, MNT_NOWAIT); 3490 error = VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td); 3491 if (asyncflag) 3492 mp->mnt_flag |= MNT_ASYNC; 3493 vn_finished_write(mp); 3494 vfs_unbusy(mp, td); 3495 return (error); 3496} 3497 3498/* 3499 * The syncer vnode is no referenced. 3500 */ 3501static int 3502sync_inactive(ap) 3503 struct vop_inactive_args /* { 3504 struct vnode *a_vp; 3505 struct thread *a_td; 3506 } */ *ap; 3507{ 3508 3509 VOP_UNLOCK(ap->a_vp, 0, ap->a_td); 3510 vgone(ap->a_vp); 3511 return (0); 3512} 3513 3514/* 3515 * The syncer vnode is no longer needed and is being decommissioned. 3516 * 3517 * Modifications to the worklist must be protected at splbio(). 3518 */ 3519static int 3520sync_reclaim(ap) 3521 struct vop_reclaim_args /* { 3522 struct vnode *a_vp; 3523 } */ *ap; 3524{ 3525 struct vnode *vp = ap->a_vp; 3526 int s; 3527 3528 s = splbio(); 3529 vp->v_mount->mnt_syncer = NULL; 3530 VI_LOCK(vp); 3531 if (vp->v_iflag & VI_ONWORKLST) { 3532 mtx_lock(&sync_mtx); 3533 LIST_REMOVE(vp, v_synclist); 3534 mtx_unlock(&sync_mtx); 3535 vp->v_iflag &= ~VI_ONWORKLST; 3536 } 3537 VI_UNLOCK(vp); 3538 splx(s); 3539 3540 return (0); 3541} 3542 3543/* 3544 * extract the dev_t from a VCHR 3545 */ 3546dev_t 3547vn_todev(vp) 3548 struct vnode *vp; 3549{ 3550 if (vp->v_type != VCHR) 3551 return (NODEV); 3552 return (vp->v_rdev); 3553} 3554 3555/* 3556 * Check if vnode represents a disk device 3557 */ 3558int 3559vn_isdisk(vp, errp) 3560 struct vnode *vp; 3561 int *errp; 3562{ 3563 struct cdevsw *cdevsw; 3564 3565 if (vp->v_type != VCHR) { 3566 if (errp != NULL) 3567 *errp = ENOTBLK; 3568 return (0); 3569 } 3570 if (vp->v_rdev == NULL) { 3571 if (errp != NULL) 3572 *errp = ENXIO; 3573 return (0); 3574 } 3575 cdevsw = devsw(vp->v_rdev); 3576 if (cdevsw == NULL) { 3577 if (errp != NULL) 3578 *errp = ENXIO; 3579 return (0); 3580 } 3581 if (!(cdevsw->d_flags & D_DISK)) { 3582 if (errp != NULL) 3583 *errp = ENOTBLK; 3584 return (0); 3585 } 3586 if (errp != NULL) 3587 *errp = 0; 3588 return (1); 3589} 3590 3591/* 3592 * Free data allocated by namei(); see namei(9) for details. 3593 */ 3594void 3595NDFREE(ndp, flags) 3596 struct nameidata *ndp; 3597 const uint flags; 3598{ 3599 if (!(flags & NDF_NO_FREE_PNBUF) && 3600 (ndp->ni_cnd.cn_flags & HASBUF)) { 3601 uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf); 3602 ndp->ni_cnd.cn_flags &= ~HASBUF; 3603 } 3604 if (!(flags & NDF_NO_DVP_UNLOCK) && 3605 (ndp->ni_cnd.cn_flags & LOCKPARENT) && 3606 ndp->ni_dvp != ndp->ni_vp) 3607 VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread); 3608 if (!(flags & NDF_NO_DVP_RELE) && 3609 (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) { 3610 vrele(ndp->ni_dvp); 3611 ndp->ni_dvp = NULL; 3612 } 3613 if (!(flags & NDF_NO_VP_UNLOCK) && 3614 (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp) 3615 VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread); 3616 if (!(flags & NDF_NO_VP_RELE) && 3617 ndp->ni_vp) { 3618 vrele(ndp->ni_vp); 3619 ndp->ni_vp = NULL; 3620 } 3621 if (!(flags & NDF_NO_STARTDIR_RELE) && 3622 (ndp->ni_cnd.cn_flags & SAVESTART)) { 3623 vrele(ndp->ni_startdir); 3624 ndp->ni_startdir = NULL; 3625 } 3626} 3627 3628/* 3629 * Common filesystem object access control check routine. Accepts a 3630 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3631 * and optional call-by-reference privused argument allowing vaccess() 3632 * to indicate to the caller whether privilege was used to satisfy the 3633 * request (obsoleted). Returns 0 on success, or an errno on failure. 3634 */ 3635int 3636vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused) 3637 enum vtype type; 3638 mode_t file_mode; 3639 uid_t file_uid; 3640 gid_t file_gid; 3641 mode_t acc_mode; 3642 struct ucred *cred; 3643 int *privused; 3644{ 3645 mode_t dac_granted; 3646#ifdef CAPABILITIES 3647 mode_t cap_granted; 3648#endif 3649 3650 /* 3651 * Look for a normal, non-privileged way to access the file/directory 3652 * as requested. If it exists, go with that. 3653 */ 3654 3655 if (privused != NULL) 3656 *privused = 0; 3657 3658 dac_granted = 0; 3659 3660 /* Check the owner. */ 3661 if (cred->cr_uid == file_uid) { 3662 dac_granted |= VADMIN; 3663 if (file_mode & S_IXUSR) 3664 dac_granted |= VEXEC; 3665 if (file_mode & S_IRUSR) 3666 dac_granted |= VREAD; 3667 if (file_mode & S_IWUSR) 3668 dac_granted |= (VWRITE | VAPPEND); 3669 3670 if ((acc_mode & dac_granted) == acc_mode) 3671 return (0); 3672 3673 goto privcheck; 3674 } 3675 3676 /* Otherwise, check the groups (first match) */ 3677 if (groupmember(file_gid, cred)) { 3678 if (file_mode & S_IXGRP) 3679 dac_granted |= VEXEC; 3680 if (file_mode & S_IRGRP) 3681 dac_granted |= VREAD; 3682 if (file_mode & S_IWGRP) 3683 dac_granted |= (VWRITE | VAPPEND); 3684 3685 if ((acc_mode & dac_granted) == acc_mode) 3686 return (0); 3687 3688 goto privcheck; 3689 } 3690 3691 /* Otherwise, check everyone else. */ 3692 if (file_mode & S_IXOTH) 3693 dac_granted |= VEXEC; 3694 if (file_mode & S_IROTH) 3695 dac_granted |= VREAD; 3696 if (file_mode & S_IWOTH) 3697 dac_granted |= (VWRITE | VAPPEND); 3698 if ((acc_mode & dac_granted) == acc_mode) 3699 return (0); 3700 3701privcheck: 3702 if (!suser_cred(cred, PRISON_ROOT)) { 3703 /* XXX audit: privilege used */ 3704 if (privused != NULL) 3705 *privused = 1; 3706 return (0); 3707 } 3708 3709#ifdef CAPABILITIES 3710 /* 3711 * Build a capability mask to determine if the set of capabilities 3712 * satisfies the requirements when combined with the granted mask 3713 * from above. 3714 * For each capability, if the capability is required, bitwise 3715 * or the request type onto the cap_granted mask. 3716 */ 3717 cap_granted = 0; 3718 3719 if (type == VDIR) { 3720 /* 3721 * For directories, use CAP_DAC_READ_SEARCH to satisfy 3722 * VEXEC requests, instead of CAP_DAC_EXECUTE. 3723 */ 3724 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3725 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) 3726 cap_granted |= VEXEC; 3727 } else { 3728 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3729 !cap_check(cred, NULL, CAP_DAC_EXECUTE, PRISON_ROOT)) 3730 cap_granted |= VEXEC; 3731 } 3732 3733 if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && 3734 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) 3735 cap_granted |= VREAD; 3736 3737 if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3738 !cap_check(cred, NULL, CAP_DAC_WRITE, PRISON_ROOT)) 3739 cap_granted |= (VWRITE | VAPPEND); 3740 3741 if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3742 !cap_check(cred, NULL, CAP_FOWNER, PRISON_ROOT)) 3743 cap_granted |= VADMIN; 3744 3745 if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { 3746 /* XXX audit: privilege used */ 3747 if (privused != NULL) 3748 *privused = 1; 3749 return (0); 3750 } 3751#endif 3752 3753 return ((acc_mode & VADMIN) ? EPERM : EACCES); 3754} 3755 3756/* 3757 * Credential check based on process requesting service, and per-attribute 3758 * permissions. 3759 */ 3760int 3761extattr_check_cred(struct vnode *vp, int attrnamespace, 3762 struct ucred *cred, struct thread *td, int access) 3763{ 3764 3765 /* 3766 * Kernel-invoked always succeeds. 3767 */ 3768 if (cred == NOCRED) 3769 return (0); 3770 3771 /* 3772 * Do not allow privileged processes in jail to directly 3773 * manipulate system attributes. 3774 * 3775 * XXX What capability should apply here? 3776 * Probably CAP_SYS_SETFFLAG. 3777 */ 3778 switch (attrnamespace) { 3779 case EXTATTR_NAMESPACE_SYSTEM: 3780 /* Potentially should be: return (EPERM); */ 3781 return (suser_cred(cred, 0)); 3782 case EXTATTR_NAMESPACE_USER: 3783 return (VOP_ACCESS(vp, access, cred, td)); 3784 default: 3785 return (EPERM); 3786 } 3787} 3788