vfs_subr.c revision 136992
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 * 4. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 35 */ 36 37/* 38 * External virtual filesystem routines 39 */ 40 41#include <sys/cdefs.h> 42__FBSDID("$FreeBSD: head/sys/kern/vfs_subr.c 136992 2004-10-27 08:05:02Z phk $"); 43 44#include "opt_ddb.h" 45#include "opt_mac.h" 46 47#include <sys/param.h> 48#include <sys/systm.h> 49#include <sys/bio.h> 50#include <sys/buf.h> 51#include <sys/conf.h> 52#include <sys/event.h> 53#include <sys/eventhandler.h> 54#include <sys/extattr.h> 55#include <sys/fcntl.h> 56#include <sys/kdb.h> 57#include <sys/kernel.h> 58#include <sys/kthread.h> 59#include <sys/mac.h> 60#include <sys/malloc.h> 61#include <sys/mount.h> 62#include <sys/namei.h> 63#include <sys/reboot.h> 64#include <sys/sleepqueue.h> 65#include <sys/stat.h> 66#include <sys/sysctl.h> 67#include <sys/syslog.h> 68#include <sys/vmmeter.h> 69#include <sys/vnode.h> 70 71#include <vm/vm.h> 72#include <vm/vm_object.h> 73#include <vm/vm_extern.h> 74#include <vm/pmap.h> 75#include <vm/vm_map.h> 76#include <vm/vm_page.h> 77#include <vm/vm_kern.h> 78#include <vm/uma.h> 79 80static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure"); 81 82static void addalias(struct vnode *vp, struct cdev *nvp_rdev); 83static void delmntque(struct vnode *vp); 84static void insmntque(struct vnode *vp, struct mount *mp); 85static void vclean(struct vnode *vp, int flags, struct thread *td); 86static void vlruvp(struct vnode *vp); 87static int flushbuflist(struct buf *blist, int flags, struct vnode *vp, 88 int slpflag, int slptimeo, int *errorp); 89static void syncer_shutdown(void *arg, int howto); 90static int vtryrecycle(struct vnode *vp); 91static void vx_lock(struct vnode *vp); 92static void vx_unlock(struct vnode *vp); 93static void vgonechrl(struct vnode *vp, struct thread *td); 94 95 96/* 97 * Number of vnodes in existence. Increased whenever getnewvnode() 98 * allocates a new vnode, never decreased. 99 */ 100static unsigned long numvnodes; 101 102SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); 103 104/* 105 * Conversion tables for conversion from vnode types to inode formats 106 * and back. 107 */ 108enum vtype iftovt_tab[16] = { 109 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 110 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 111}; 112int vttoif_tab[9] = { 113 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 114 S_IFSOCK, S_IFIFO, S_IFMT, 115}; 116 117/* 118 * List of vnodes that are ready for recycling. 119 */ 120static TAILQ_HEAD(freelst, vnode) vnode_free_list; 121 122/* 123 * Minimum number of free vnodes. If there are fewer than this free vnodes, 124 * getnewvnode() will return a newly allocated vnode. 125 */ 126static u_long wantfreevnodes = 25; 127SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 128/* Number of vnodes in the free list. */ 129static u_long freevnodes; 130SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); 131 132/* 133 * Various variables used for debugging the new implementation of 134 * reassignbuf(). 135 * XXX these are probably of (very) limited utility now. 136 */ 137static int reassignbufcalls; 138SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); 139static int nameileafonly; 140SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, ""); 141 142/* 143 * Cache for the mount type id assigned to NFS. This is used for 144 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 145 */ 146int nfs_mount_type = -1; 147 148/* To keep more than one thread at a time from running vfs_getnewfsid */ 149static struct mtx mntid_mtx; 150 151/* 152 * Lock for any access to the following: 153 * vnode_free_list 154 * numvnodes 155 * freevnodes 156 */ 157static struct mtx vnode_free_list_mtx; 158 159/* Publicly exported FS */ 160struct nfs_public nfs_pub; 161 162/* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 163static uma_zone_t vnode_zone; 164static uma_zone_t vnodepoll_zone; 165 166/* Set to 1 to print out reclaim of active vnodes */ 167int prtactive; 168 169/* 170 * The workitem queue. 171 * 172 * It is useful to delay writes of file data and filesystem metadata 173 * for tens of seconds so that quickly created and deleted files need 174 * not waste disk bandwidth being created and removed. To realize this, 175 * we append vnodes to a "workitem" queue. When running with a soft 176 * updates implementation, most pending metadata dependencies should 177 * not wait for more than a few seconds. Thus, mounted on block devices 178 * are delayed only about a half the time that file data is delayed. 179 * Similarly, directory updates are more critical, so are only delayed 180 * about a third the time that file data is delayed. Thus, there are 181 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 182 * one each second (driven off the filesystem syncer process). The 183 * syncer_delayno variable indicates the next queue that is to be processed. 184 * Items that need to be processed soon are placed in this queue: 185 * 186 * syncer_workitem_pending[syncer_delayno] 187 * 188 * A delay of fifteen seconds is done by placing the request fifteen 189 * entries later in the queue: 190 * 191 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 192 * 193 */ 194static int syncer_delayno; 195static long syncer_mask; 196LIST_HEAD(synclist, bufobj); 197static struct synclist *syncer_workitem_pending; 198/* 199 * The sync_mtx protects: 200 * bo->bo_synclist 201 * sync_vnode_count 202 * syncer_delayno 203 * syncer_state 204 * syncer_workitem_pending 205 * syncer_worklist_len 206 * rushjob 207 */ 208static struct mtx sync_mtx; 209 210#define SYNCER_MAXDELAY 32 211static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 212static int syncdelay = 30; /* max time to delay syncing data */ 213static int filedelay = 30; /* time to delay syncing files */ 214SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); 215static int dirdelay = 29; /* time to delay syncing directories */ 216SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); 217static int metadelay = 28; /* time to delay syncing metadata */ 218SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); 219static int rushjob; /* number of slots to run ASAP */ 220static int stat_rush_requests; /* number of times I/O speeded up */ 221SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); 222 223/* 224 * When shutting down the syncer, run it at four times normal speed. 225 */ 226#define SYNCER_SHUTDOWN_SPEEDUP 4 227static int sync_vnode_count; 228static int syncer_worklist_len; 229static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 230 syncer_state; 231 232/* 233 * Number of vnodes we want to exist at any one time. This is mostly used 234 * to size hash tables in vnode-related code. It is normally not used in 235 * getnewvnode(), as wantfreevnodes is normally nonzero.) 236 * 237 * XXX desiredvnodes is historical cruft and should not exist. 238 */ 239int desiredvnodes; 240SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 241 &desiredvnodes, 0, "Maximum number of vnodes"); 242static int minvnodes; 243SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 244 &minvnodes, 0, "Minimum number of vnodes"); 245static int vnlru_nowhere; 246SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 247 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 248 249/* Hook for calling soft updates. */ 250int (*softdep_process_worklist_hook)(struct mount *); 251 252/* 253 * Initialize the vnode management data structures. 254 */ 255#ifndef MAXVNODES_MAX 256#define MAXVNODES_MAX 100000 257#endif 258static void 259vntblinit(void *dummy __unused) 260{ 261 262 /* 263 * Desiredvnodes is a function of the physical memory size and 264 * the kernel's heap size. Specifically, desiredvnodes scales 265 * in proportion to the physical memory size until two fifths 266 * of the kernel's heap size is consumed by vnodes and vm 267 * objects. 268 */ 269 desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size / 270 (5 * (sizeof(struct vm_object) + sizeof(struct vnode)))); 271 if (desiredvnodes > MAXVNODES_MAX) { 272 if (bootverbose) 273 printf("Reducing kern.maxvnodes %d -> %d\n", 274 desiredvnodes, MAXVNODES_MAX); 275 desiredvnodes = MAXVNODES_MAX; 276 } 277 minvnodes = desiredvnodes / 4; 278 mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF); 279 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 280 TAILQ_INIT(&vnode_free_list); 281 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 282 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 283 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 284 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 285 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 286 /* 287 * Initialize the filesystem syncer. 288 */ 289 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 290 &syncer_mask); 291 syncer_maxdelay = syncer_mask + 1; 292 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 293} 294SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) 295 296 297/* 298 * Mark a mount point as busy. Used to synchronize access and to delay 299 * unmounting. Interlock is not released on failure. 300 */ 301int 302vfs_busy(mp, flags, interlkp, td) 303 struct mount *mp; 304 int flags; 305 struct mtx *interlkp; 306 struct thread *td; 307{ 308 int lkflags; 309 310 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 311 if (flags & LK_NOWAIT) 312 return (ENOENT); 313 mp->mnt_kern_flag |= MNTK_MWAIT; 314 /* 315 * Since all busy locks are shared except the exclusive 316 * lock granted when unmounting, the only place that a 317 * wakeup needs to be done is at the release of the 318 * exclusive lock at the end of dounmount. 319 */ 320 msleep(mp, interlkp, PVFS, "vfs_busy", 0); 321 return (ENOENT); 322 } 323 lkflags = LK_SHARED | LK_NOPAUSE; 324 if (interlkp) 325 lkflags |= LK_INTERLOCK; 326 if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td)) 327 panic("vfs_busy: unexpected lock failure"); 328 return (0); 329} 330 331/* 332 * Free a busy filesystem. 333 */ 334void 335vfs_unbusy(mp, td) 336 struct mount *mp; 337 struct thread *td; 338{ 339 340 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); 341} 342 343/* 344 * Lookup a mount point by filesystem identifier. 345 */ 346struct mount * 347vfs_getvfs(fsid) 348 fsid_t *fsid; 349{ 350 register struct mount *mp; 351 352 mtx_lock(&mountlist_mtx); 353 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 354 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 355 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 356 mtx_unlock(&mountlist_mtx); 357 return (mp); 358 } 359 } 360 mtx_unlock(&mountlist_mtx); 361 return ((struct mount *) 0); 362} 363 364/* 365 * Check if a user can access priveledged mount options. 366 */ 367int 368vfs_suser(struct mount *mp, struct thread *td) 369{ 370 int error; 371 372 if ((mp->mnt_flag & MNT_USER) == 0 || 373 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 374 if ((error = suser(td)) != 0) 375 return (error); 376 } 377 return (0); 378} 379 380/* 381 * Get a new unique fsid. Try to make its val[0] unique, since this value 382 * will be used to create fake device numbers for stat(). Also try (but 383 * not so hard) make its val[0] unique mod 2^16, since some emulators only 384 * support 16-bit device numbers. We end up with unique val[0]'s for the 385 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 386 * 387 * Keep in mind that several mounts may be running in parallel. Starting 388 * the search one past where the previous search terminated is both a 389 * micro-optimization and a defense against returning the same fsid to 390 * different mounts. 391 */ 392void 393vfs_getnewfsid(mp) 394 struct mount *mp; 395{ 396 static u_int16_t mntid_base; 397 fsid_t tfsid; 398 int mtype; 399 400 mtx_lock(&mntid_mtx); 401 mtype = mp->mnt_vfc->vfc_typenum; 402 tfsid.val[1] = mtype; 403 mtype = (mtype & 0xFF) << 24; 404 for (;;) { 405 tfsid.val[0] = makedev(255, 406 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 407 mntid_base++; 408 if (vfs_getvfs(&tfsid) == NULL) 409 break; 410 } 411 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 412 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 413 mtx_unlock(&mntid_mtx); 414} 415 416/* 417 * Knob to control the precision of file timestamps: 418 * 419 * 0 = seconds only; nanoseconds zeroed. 420 * 1 = seconds and nanoseconds, accurate within 1/HZ. 421 * 2 = seconds and nanoseconds, truncated to microseconds. 422 * >=3 = seconds and nanoseconds, maximum precision. 423 */ 424enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 425 426static int timestamp_precision = TSP_SEC; 427SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 428 ×tamp_precision, 0, ""); 429 430/* 431 * Get a current timestamp. 432 */ 433void 434vfs_timestamp(tsp) 435 struct timespec *tsp; 436{ 437 struct timeval tv; 438 439 switch (timestamp_precision) { 440 case TSP_SEC: 441 tsp->tv_sec = time_second; 442 tsp->tv_nsec = 0; 443 break; 444 case TSP_HZ: 445 getnanotime(tsp); 446 break; 447 case TSP_USEC: 448 microtime(&tv); 449 TIMEVAL_TO_TIMESPEC(&tv, tsp); 450 break; 451 case TSP_NSEC: 452 default: 453 nanotime(tsp); 454 break; 455 } 456} 457 458/* 459 * Set vnode attributes to VNOVAL 460 */ 461void 462vattr_null(vap) 463 register struct vattr *vap; 464{ 465 466 vap->va_type = VNON; 467 vap->va_size = VNOVAL; 468 vap->va_bytes = VNOVAL; 469 vap->va_mode = VNOVAL; 470 vap->va_nlink = VNOVAL; 471 vap->va_uid = VNOVAL; 472 vap->va_gid = VNOVAL; 473 vap->va_fsid = VNOVAL; 474 vap->va_fileid = VNOVAL; 475 vap->va_blocksize = VNOVAL; 476 vap->va_rdev = VNOVAL; 477 vap->va_atime.tv_sec = VNOVAL; 478 vap->va_atime.tv_nsec = VNOVAL; 479 vap->va_mtime.tv_sec = VNOVAL; 480 vap->va_mtime.tv_nsec = VNOVAL; 481 vap->va_ctime.tv_sec = VNOVAL; 482 vap->va_ctime.tv_nsec = VNOVAL; 483 vap->va_birthtime.tv_sec = VNOVAL; 484 vap->va_birthtime.tv_nsec = VNOVAL; 485 vap->va_flags = VNOVAL; 486 vap->va_gen = VNOVAL; 487 vap->va_vaflags = 0; 488} 489 490/* 491 * This routine is called when we have too many vnodes. It attempts 492 * to free <count> vnodes and will potentially free vnodes that still 493 * have VM backing store (VM backing store is typically the cause 494 * of a vnode blowout so we want to do this). Therefore, this operation 495 * is not considered cheap. 496 * 497 * A number of conditions may prevent a vnode from being reclaimed. 498 * the buffer cache may have references on the vnode, a directory 499 * vnode may still have references due to the namei cache representing 500 * underlying files, or the vnode may be in active use. It is not 501 * desireable to reuse such vnodes. These conditions may cause the 502 * number of vnodes to reach some minimum value regardless of what 503 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 504 */ 505static int 506vlrureclaim(struct mount *mp) 507{ 508 struct vnode *vp; 509 int done; 510 int trigger; 511 int usevnodes; 512 int count; 513 514 /* 515 * Calculate the trigger point, don't allow user 516 * screwups to blow us up. This prevents us from 517 * recycling vnodes with lots of resident pages. We 518 * aren't trying to free memory, we are trying to 519 * free vnodes. 520 */ 521 usevnodes = desiredvnodes; 522 if (usevnodes <= 0) 523 usevnodes = 1; 524 trigger = cnt.v_page_count * 2 / usevnodes; 525 526 done = 0; 527 MNT_ILOCK(mp); 528 count = mp->mnt_nvnodelistsize / 10 + 1; 529 while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) { 530 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 531 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 532 533 if (vp->v_type != VNON && 534 vp->v_type != VBAD && 535 VI_TRYLOCK(vp)) { 536 if (VMIGHTFREE(vp) && /* critical path opt */ 537 (vp->v_object == NULL || 538 vp->v_object->resident_page_count < trigger)) { 539 MNT_IUNLOCK(mp); 540 vgonel(vp, curthread); 541 done++; 542 MNT_ILOCK(mp); 543 } else 544 VI_UNLOCK(vp); 545 } 546 --count; 547 } 548 MNT_IUNLOCK(mp); 549 return done; 550} 551 552/* 553 * Attempt to recycle vnodes in a context that is always safe to block. 554 * Calling vlrurecycle() from the bowels of filesystem code has some 555 * interesting deadlock problems. 556 */ 557static struct proc *vnlruproc; 558static int vnlruproc_sig; 559 560static void 561vnlru_proc(void) 562{ 563 struct mount *mp, *nmp; 564 int done; 565 struct proc *p = vnlruproc; 566 struct thread *td = FIRST_THREAD_IN_PROC(p); 567 568 mtx_lock(&Giant); 569 570 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 571 SHUTDOWN_PRI_FIRST); 572 573 for (;;) { 574 kthread_suspend_check(p); 575 mtx_lock(&vnode_free_list_mtx); 576 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { 577 mtx_unlock(&vnode_free_list_mtx); 578 vnlruproc_sig = 0; 579 wakeup(&vnlruproc_sig); 580 tsleep(vnlruproc, PVFS, "vlruwt", hz); 581 continue; 582 } 583 mtx_unlock(&vnode_free_list_mtx); 584 done = 0; 585 mtx_lock(&mountlist_mtx); 586 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 587 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 588 nmp = TAILQ_NEXT(mp, mnt_list); 589 continue; 590 } 591 done += vlrureclaim(mp); 592 mtx_lock(&mountlist_mtx); 593 nmp = TAILQ_NEXT(mp, mnt_list); 594 vfs_unbusy(mp, td); 595 } 596 mtx_unlock(&mountlist_mtx); 597 if (done == 0) { 598#if 0 599 /* These messages are temporary debugging aids */ 600 if (vnlru_nowhere < 5) 601 printf("vnlru process getting nowhere..\n"); 602 else if (vnlru_nowhere == 5) 603 printf("vnlru process messages stopped.\n"); 604#endif 605 vnlru_nowhere++; 606 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 607 } 608 } 609} 610 611static struct kproc_desc vnlru_kp = { 612 "vnlru", 613 vnlru_proc, 614 &vnlruproc 615}; 616SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 617 618 619/* 620 * Routines having to do with the management of the vnode table. 621 */ 622 623/* 624 * Check to see if a free vnode can be recycled. If it can, 625 * recycle it and return it with the vnode interlock held. 626 */ 627static int 628vtryrecycle(struct vnode *vp) 629{ 630 struct thread *td = curthread; 631 vm_object_t object; 632 struct mount *vnmp; 633 int error; 634 635 /* Don't recycle if we can't get the interlock */ 636 if (!VI_TRYLOCK(vp)) 637 return (EWOULDBLOCK); 638 /* 639 * This vnode may found and locked via some other list, if so we 640 * can't recycle it yet. 641 */ 642 if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 643 return (EWOULDBLOCK); 644 /* 645 * Don't recycle if its filesystem is being suspended. 646 */ 647 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 648 VOP_UNLOCK(vp, 0, td); 649 return (EBUSY); 650 } 651 652 /* 653 * Don't recycle if we still have cached pages. 654 */ 655 if (VOP_GETVOBJECT(vp, &object) == 0) { 656 VM_OBJECT_LOCK(object); 657 if (object->resident_page_count || 658 object->ref_count) { 659 VM_OBJECT_UNLOCK(object); 660 error = EBUSY; 661 goto done; 662 } 663 VM_OBJECT_UNLOCK(object); 664 } 665 if (LIST_FIRST(&vp->v_cache_src)) { 666 /* 667 * note: nameileafonly sysctl is temporary, 668 * for debugging only, and will eventually be 669 * removed. 670 */ 671 if (nameileafonly > 0) { 672 /* 673 * Do not reuse namei-cached directory 674 * vnodes that have cached 675 * subdirectories. 676 */ 677 if (cache_leaf_test(vp) < 0) { 678 error = EISDIR; 679 goto done; 680 } 681 } else if (nameileafonly < 0 || 682 vmiodirenable == 0) { 683 /* 684 * Do not reuse namei-cached directory 685 * vnodes if nameileafonly is -1 or 686 * if VMIO backing for directories is 687 * turned off (otherwise we reuse them 688 * too quickly). 689 */ 690 error = EBUSY; 691 goto done; 692 } 693 } 694 /* 695 * If we got this far, we need to acquire the interlock and see if 696 * anyone picked up this vnode from another list. If not, we will 697 * mark it with XLOCK via vgonel() so that anyone who does find it 698 * will skip over it. 699 */ 700 VI_LOCK(vp); 701 if (VSHOULDBUSY(vp) && (vp->v_iflag & VI_XLOCK) == 0) { 702 VI_UNLOCK(vp); 703 error = EBUSY; 704 goto done; 705 } 706 mtx_lock(&vnode_free_list_mtx); 707 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 708 vp->v_iflag &= ~VI_FREE; 709 mtx_unlock(&vnode_free_list_mtx); 710 vp->v_iflag |= VI_DOOMED; 711 if ((vp->v_type != VBAD) || (vp->v_data != NULL)) { 712 VOP_UNLOCK(vp, 0, td); 713 vgonel(vp, td); 714 VI_LOCK(vp); 715 } else 716 VOP_UNLOCK(vp, 0, td); 717 vn_finished_write(vnmp); 718 return (0); 719done: 720 VOP_UNLOCK(vp, 0, td); 721 vn_finished_write(vnmp); 722 return (error); 723} 724 725/* 726 * Return the next vnode from the free list. 727 */ 728int 729getnewvnode(tag, mp, vops, vpp) 730 const char *tag; 731 struct mount *mp; 732 vop_t **vops; 733 struct vnode **vpp; 734{ 735 struct vnode *vp = NULL; 736 struct vpollinfo *pollinfo = NULL; 737 struct bufobj *bo; 738 739 mtx_lock(&vnode_free_list_mtx); 740 741 /* 742 * Try to reuse vnodes if we hit the max. This situation only 743 * occurs in certain large-memory (2G+) situations. We cannot 744 * attempt to directly reclaim vnodes due to nasty recursion 745 * problems. 746 */ 747 while (numvnodes - freevnodes > desiredvnodes) { 748 if (vnlruproc_sig == 0) { 749 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 750 wakeup(vnlruproc); 751 } 752 mtx_unlock(&vnode_free_list_mtx); 753 tsleep(&vnlruproc_sig, PVFS, "vlruwk", hz); 754 mtx_lock(&vnode_free_list_mtx); 755 } 756 757 /* 758 * Attempt to reuse a vnode already on the free list, allocating 759 * a new vnode if we can't find one or if we have not reached a 760 * good minimum for good LRU performance. 761 */ 762 763 if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) { 764 int error; 765 int count; 766 767 for (count = 0; count < freevnodes; count++) { 768 vp = TAILQ_FIRST(&vnode_free_list); 769 770 KASSERT(vp->v_usecount == 0 && 771 (vp->v_iflag & VI_DOINGINACT) == 0, 772 ("getnewvnode: free vnode isn't")); 773 774 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 775 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 776 mtx_unlock(&vnode_free_list_mtx); 777 error = vtryrecycle(vp); 778 mtx_lock(&vnode_free_list_mtx); 779 if (error == 0) 780 break; 781 vp = NULL; 782 } 783 } 784 if (vp) { 785 freevnodes--; 786 bo = &vp->v_bufobj; 787 mtx_unlock(&vnode_free_list_mtx); 788 789#ifdef INVARIANTS 790 { 791 if (vp->v_data) 792 printf("cleaned vnode isn't, " 793 "address %p, inode %p\n", 794 vp, vp->v_data); 795 if (bo->bo_numoutput) 796 panic("Clean vnode has pending I/O's"); 797 if (vp->v_writecount != 0) 798 panic("Non-zero write count"); 799 } 800#endif 801 if ((pollinfo = vp->v_pollinfo) != NULL) { 802 /* 803 * To avoid lock order reversals, the call to 804 * uma_zfree() must be delayed until the vnode 805 * interlock is released. 806 */ 807 vp->v_pollinfo = NULL; 808 } 809#ifdef MAC 810 mac_destroy_vnode(vp); 811#endif 812 vp->v_iflag = 0; 813 vp->v_vflag = 0; 814 vp->v_lastw = 0; 815 vp->v_lasta = 0; 816 vp->v_cstart = 0; 817 vp->v_clen = 0; 818 vp->v_socket = 0; 819 lockdestroy(vp->v_vnlock); 820 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 821 KASSERT(bo->bo_clean.bv_cnt == 0, ("cleanbufcnt not 0")); 822 KASSERT(bo->bo_clean.bv_root == NULL, ("cleanblkroot not NULL")); 823 KASSERT(bo->bo_dirty.bv_cnt == 0, ("dirtybufcnt not 0")); 824 KASSERT(bo->bo_dirty.bv_root == NULL, ("dirtyblkroot not NULL")); 825 } else { 826 numvnodes++; 827 mtx_unlock(&vnode_free_list_mtx); 828 829 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); 830 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 831 VI_LOCK(vp); 832 vp->v_dd = vp; 833 bo = &vp->v_bufobj; 834 bo->bo_private = vp; 835 bo->bo_mtx = &vp->v_interlock; 836 vp->v_vnlock = &vp->v_lock; 837 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 838 cache_purge(vp); /* Sets up v_id. */ 839 LIST_INIT(&vp->v_cache_src); 840 TAILQ_INIT(&vp->v_cache_dst); 841 } 842 843 TAILQ_INIT(&bo->bo_clean.bv_hd); 844 TAILQ_INIT(&bo->bo_dirty.bv_hd); 845 bo->bo_ops = &buf_ops_bio; 846 vp->v_type = VNON; 847 vp->v_tag = tag; 848 vp->v_op = vops; 849 *vpp = vp; 850 vp->v_usecount = 1; 851 vp->v_data = 0; 852 vp->v_cachedid = -1; 853 VI_UNLOCK(vp); 854 if (pollinfo != NULL) { 855 knlist_destroy(&pollinfo->vpi_selinfo.si_note); 856 mtx_destroy(&pollinfo->vpi_lock); 857 uma_zfree(vnodepoll_zone, pollinfo); 858 } 859#ifdef MAC 860 mac_init_vnode(vp); 861 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 862 mac_associate_vnode_singlelabel(mp, vp); 863 else if (mp == NULL) 864 printf("NULL mp in getnewvnode()\n"); 865#endif 866 delmntque(vp); 867 if (mp != NULL) { 868 insmntque(vp, mp); 869 bo->bo_bsize = mp->mnt_stat.f_iosize; 870 } 871 872 return (0); 873} 874 875/* 876 * Delete from old mount point vnode list, if on one. 877 */ 878static void 879delmntque(struct vnode *vp) 880{ 881 struct mount *mp; 882 883 if (vp->v_mount == NULL) 884 return; 885 mp = vp->v_mount; 886 MNT_ILOCK(mp); 887 vp->v_mount = NULL; 888 KASSERT(mp->mnt_nvnodelistsize > 0, 889 ("bad mount point vnode list size")); 890 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 891 mp->mnt_nvnodelistsize--; 892 MNT_IUNLOCK(mp); 893} 894 895/* 896 * Insert into list of vnodes for the new mount point, if available. 897 */ 898static void 899insmntque(struct vnode *vp, struct mount *mp) 900{ 901 902 vp->v_mount = mp; 903 KASSERT(mp != NULL, ("Don't call insmntque(foo, NULL)")); 904 MNT_ILOCK(vp->v_mount); 905 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 906 mp->mnt_nvnodelistsize++; 907 MNT_IUNLOCK(vp->v_mount); 908} 909 910/* 911 * Flush out and invalidate all buffers associated with a vnode. 912 * Called with the underlying object locked. 913 */ 914int 915vinvalbuf(vp, flags, cred, td, slpflag, slptimeo) 916 struct vnode *vp; 917 int flags; 918 struct ucred *cred; 919 struct thread *td; 920 int slpflag, slptimeo; 921{ 922 struct buf *blist; 923 int error; 924 vm_object_t object; 925 struct bufobj *bo; 926 927 GIANT_REQUIRED; 928 929 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 930 931 bo = &vp->v_bufobj; 932 BO_LOCK(bo); 933 if (flags & V_SAVE) { 934 error = bufobj_wwait(bo, slpflag, slptimeo); 935 if (error) { 936 VI_UNLOCK(vp); 937 return (error); 938 } 939 if (bo->bo_dirty.bv_cnt > 0) { 940 VI_UNLOCK(vp); 941 if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0) 942 return (error); 943 /* 944 * XXX We could save a lock/unlock if this was only 945 * enabled under INVARIANTS 946 */ 947 VI_LOCK(vp); 948 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 949 panic("vinvalbuf: dirty bufs"); 950 } 951 } 952 /* 953 * If you alter this loop please notice that interlock is dropped and 954 * reacquired in flushbuflist. Special care is needed to ensure that 955 * no race conditions occur from this. 956 */ 957 for (error = 0;;) { 958 blist = TAILQ_FIRST(&vp->v_bufobj.bo_clean.bv_hd); 959 if (blist != NULL && 960 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 961 if (error) 962 break; 963 continue; 964 } 965 blist = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd); 966 if (blist != NULL && 967 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 968 if (error) 969 break; 970 continue; 971 } 972 break; 973 } 974 if (error) { 975 VI_UNLOCK(vp); 976 return (error); 977 } 978 979 /* 980 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 981 * have write I/O in-progress but if there is a VM object then the 982 * VM object can also have read-I/O in-progress. 983 */ 984 do { 985 bufobj_wwait(bo, 0, 0); 986 VI_UNLOCK(vp); 987 if (VOP_GETVOBJECT(vp, &object) == 0) { 988 VM_OBJECT_LOCK(object); 989 vm_object_pip_wait(object, "vnvlbx"); 990 VM_OBJECT_UNLOCK(object); 991 } 992 VI_LOCK(vp); 993 } while (bo->bo_numoutput > 0); 994 VI_UNLOCK(vp); 995 996 /* 997 * Destroy the copy in the VM cache, too. 998 */ 999 if (VOP_GETVOBJECT(vp, &object) == 0) { 1000 VM_OBJECT_LOCK(object); 1001 vm_object_page_remove(object, 0, 0, 1002 (flags & V_SAVE) ? TRUE : FALSE); 1003 VM_OBJECT_UNLOCK(object); 1004 } 1005 1006#ifdef INVARIANTS 1007 VI_LOCK(vp); 1008 if ((flags & (V_ALT | V_NORMAL)) == 0 && 1009 (vp->v_bufobj.bo_dirty.bv_cnt > 0 || 1010 vp->v_bufobj.bo_clean.bv_cnt > 0)) 1011 panic("vinvalbuf: flush failed"); 1012 VI_UNLOCK(vp); 1013#endif 1014 return (0); 1015} 1016 1017/* 1018 * Flush out buffers on the specified list. 1019 * 1020 */ 1021static int 1022flushbuflist(blist, flags, vp, slpflag, slptimeo, errorp) 1023 struct buf *blist; 1024 int flags; 1025 struct vnode *vp; 1026 int slpflag, slptimeo; 1027 int *errorp; 1028{ 1029 struct buf *bp, *nbp; 1030 int found, error; 1031 1032 ASSERT_VI_LOCKED(vp, "flushbuflist"); 1033 1034 for (found = 0, bp = blist; bp; bp = nbp) { 1035 nbp = TAILQ_NEXT(bp, b_bobufs); 1036 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1037 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1038 continue; 1039 } 1040 found += 1; 1041 error = BUF_TIMELOCK(bp, 1042 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp), 1043 "flushbuf", slpflag, slptimeo); 1044 if (error) { 1045 if (error != ENOLCK) 1046 *errorp = error; 1047 goto done; 1048 } 1049 /* 1050 * XXX Since there are no node locks for NFS, I 1051 * believe there is a slight chance that a delayed 1052 * write will occur while sleeping just above, so 1053 * check for it. Note that vfs_bio_awrite expects 1054 * buffers to reside on a queue, while bwrite and 1055 * brelse do not. 1056 */ 1057 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1058 (flags & V_SAVE)) { 1059 1060 if (bp->b_vp == vp) { 1061 if (bp->b_flags & B_CLUSTEROK) { 1062 vfs_bio_awrite(bp); 1063 } else { 1064 bremfree(bp); 1065 bp->b_flags |= B_ASYNC; 1066 bwrite(bp); 1067 } 1068 } else { 1069 bremfree(bp); 1070 (void) bwrite(bp); 1071 } 1072 goto done; 1073 } 1074 bremfree(bp); 1075 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 1076 bp->b_flags &= ~B_ASYNC; 1077 brelse(bp); 1078 VI_LOCK(vp); 1079 } 1080 return (found); 1081done: 1082 VI_LOCK(vp); 1083 return (found); 1084} 1085 1086/* 1087 * Truncate a file's buffer and pages to a specified length. This 1088 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1089 * sync activity. 1090 */ 1091int 1092vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, off_t length, int blksize) 1093{ 1094 struct buf *bp, *nbp; 1095 int anyfreed; 1096 int trunclbn; 1097 struct bufobj *bo; 1098 1099 /* 1100 * Round up to the *next* lbn. 1101 */ 1102 trunclbn = (length + blksize - 1) / blksize; 1103 1104 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1105restart: 1106 VI_LOCK(vp); 1107 bo = &vp->v_bufobj; 1108 anyfreed = 1; 1109 for (;anyfreed;) { 1110 anyfreed = 0; 1111 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 1112 if (bp->b_lblkno < trunclbn) 1113 continue; 1114 if (BUF_LOCK(bp, 1115 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1116 VI_MTX(vp)) == ENOLCK) 1117 goto restart; 1118 1119 bremfree(bp); 1120 bp->b_flags |= (B_INVAL | B_RELBUF); 1121 bp->b_flags &= ~B_ASYNC; 1122 brelse(bp); 1123 anyfreed = 1; 1124 1125 if (nbp != NULL && 1126 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1127 (nbp->b_vp != vp) || 1128 (nbp->b_flags & B_DELWRI))) { 1129 goto restart; 1130 } 1131 VI_LOCK(vp); 1132 } 1133 1134 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1135 if (bp->b_lblkno < trunclbn) 1136 continue; 1137 if (BUF_LOCK(bp, 1138 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1139 VI_MTX(vp)) == ENOLCK) 1140 goto restart; 1141 bremfree(bp); 1142 bp->b_flags |= (B_INVAL | B_RELBUF); 1143 bp->b_flags &= ~B_ASYNC; 1144 brelse(bp); 1145 anyfreed = 1; 1146 if (nbp != NULL && 1147 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1148 (nbp->b_vp != vp) || 1149 (nbp->b_flags & B_DELWRI) == 0)) { 1150 goto restart; 1151 } 1152 VI_LOCK(vp); 1153 } 1154 } 1155 1156 if (length > 0) { 1157restartsync: 1158 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1159 if (bp->b_lblkno > 0) 1160 continue; 1161 /* 1162 * Since we hold the vnode lock this should only 1163 * fail if we're racing with the buf daemon. 1164 */ 1165 if (BUF_LOCK(bp, 1166 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1167 VI_MTX(vp)) == ENOLCK) { 1168 goto restart; 1169 } 1170 KASSERT((bp->b_flags & B_DELWRI), 1171 ("buf(%p) on dirty queue without DELWRI", bp)); 1172 1173 bremfree(bp); 1174 bawrite(bp); 1175 VI_LOCK(vp); 1176 goto restartsync; 1177 } 1178 } 1179 1180 bufobj_wwait(bo, 0, 0); 1181 VI_UNLOCK(vp); 1182 vnode_pager_setsize(vp, length); 1183 1184 return (0); 1185} 1186 1187/* 1188 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1189 * a vnode. 1190 * 1191 * NOTE: We have to deal with the special case of a background bitmap 1192 * buffer, a situation where two buffers will have the same logical 1193 * block offset. We want (1) only the foreground buffer to be accessed 1194 * in a lookup and (2) must differentiate between the foreground and 1195 * background buffer in the splay tree algorithm because the splay 1196 * tree cannot normally handle multiple entities with the same 'index'. 1197 * We accomplish this by adding differentiating flags to the splay tree's 1198 * numerical domain. 1199 */ 1200static 1201struct buf * 1202buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1203{ 1204 struct buf dummy; 1205 struct buf *lefttreemax, *righttreemin, *y; 1206 1207 if (root == NULL) 1208 return (NULL); 1209 lefttreemax = righttreemin = &dummy; 1210 for (;;) { 1211 if (lblkno < root->b_lblkno || 1212 (lblkno == root->b_lblkno && 1213 (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1214 if ((y = root->b_left) == NULL) 1215 break; 1216 if (lblkno < y->b_lblkno) { 1217 /* Rotate right. */ 1218 root->b_left = y->b_right; 1219 y->b_right = root; 1220 root = y; 1221 if ((y = root->b_left) == NULL) 1222 break; 1223 } 1224 /* Link into the new root's right tree. */ 1225 righttreemin->b_left = root; 1226 righttreemin = root; 1227 } else if (lblkno > root->b_lblkno || 1228 (lblkno == root->b_lblkno && 1229 (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { 1230 if ((y = root->b_right) == NULL) 1231 break; 1232 if (lblkno > y->b_lblkno) { 1233 /* Rotate left. */ 1234 root->b_right = y->b_left; 1235 y->b_left = root; 1236 root = y; 1237 if ((y = root->b_right) == NULL) 1238 break; 1239 } 1240 /* Link into the new root's left tree. */ 1241 lefttreemax->b_right = root; 1242 lefttreemax = root; 1243 } else { 1244 break; 1245 } 1246 root = y; 1247 } 1248 /* Assemble the new root. */ 1249 lefttreemax->b_right = root->b_left; 1250 righttreemin->b_left = root->b_right; 1251 root->b_left = dummy.b_right; 1252 root->b_right = dummy.b_left; 1253 return (root); 1254} 1255 1256static void 1257buf_vlist_remove(struct buf *bp) 1258{ 1259 struct buf *root; 1260 struct bufv *bv; 1261 1262 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1263 ASSERT_BO_LOCKED(bp->b_bufobj); 1264 if (bp->b_xflags & BX_VNDIRTY) 1265 bv = &bp->b_bufobj->bo_dirty; 1266 else 1267 bv = &bp->b_bufobj->bo_clean; 1268 if (bp != bv->bv_root) { 1269 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1270 KASSERT(root == bp, ("splay lookup failed in remove")); 1271 } 1272 if (bp->b_left == NULL) { 1273 root = bp->b_right; 1274 } else { 1275 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1276 root->b_right = bp->b_right; 1277 } 1278 bv->bv_root = root; 1279 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 1280 bv->bv_cnt--; 1281 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1282} 1283 1284/* 1285 * Add the buffer to the sorted clean or dirty block list using a 1286 * splay tree algorithm. 1287 * 1288 * NOTE: xflags is passed as a constant, optimizing this inline function! 1289 */ 1290static void 1291buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 1292{ 1293 struct buf *root; 1294 struct bufv *bv; 1295 1296 ASSERT_BO_LOCKED(bo); 1297 bp->b_xflags |= xflags; 1298 if (xflags & BX_VNDIRTY) 1299 bv = &bo->bo_dirty; 1300 else 1301 bv = &bo->bo_clean; 1302 1303 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1304 if (root == NULL) { 1305 bp->b_left = NULL; 1306 bp->b_right = NULL; 1307 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 1308 } else if (bp->b_lblkno < root->b_lblkno || 1309 (bp->b_lblkno == root->b_lblkno && 1310 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1311 bp->b_left = root->b_left; 1312 bp->b_right = root; 1313 root->b_left = NULL; 1314 TAILQ_INSERT_BEFORE(root, bp, b_bobufs); 1315 } else { 1316 bp->b_right = root->b_right; 1317 bp->b_left = root; 1318 root->b_right = NULL; 1319 TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs); 1320 } 1321 bv->bv_cnt++; 1322 bv->bv_root = bp; 1323} 1324 1325/* 1326 * Lookup a buffer using the splay tree. Note that we specifically avoid 1327 * shadow buffers used in background bitmap writes. 1328 * 1329 * This code isn't quite efficient as it could be because we are maintaining 1330 * two sorted lists and do not know which list the block resides in. 1331 * 1332 * During a "make buildworld" the desired buffer is found at one of 1333 * the roots more than 60% of the time. Thus, checking both roots 1334 * before performing either splay eliminates unnecessary splays on the 1335 * first tree splayed. 1336 */ 1337struct buf * 1338gbincore(struct bufobj *bo, daddr_t lblkno) 1339{ 1340 struct buf *bp; 1341 1342 GIANT_REQUIRED; 1343 1344 ASSERT_BO_LOCKED(bo); 1345 if ((bp = bo->bo_clean.bv_root) != NULL && 1346 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1347 return (bp); 1348 if ((bp = bo->bo_dirty.bv_root) != NULL && 1349 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1350 return (bp); 1351 if ((bp = bo->bo_clean.bv_root) != NULL) { 1352 bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp); 1353 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1354 return (bp); 1355 } 1356 if ((bp = bo->bo_dirty.bv_root) != NULL) { 1357 bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp); 1358 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1359 return (bp); 1360 } 1361 return (NULL); 1362} 1363 1364/* 1365 * Associate a buffer with a vnode. 1366 */ 1367void 1368bgetvp(struct vnode *vp, struct buf *bp) 1369{ 1370 1371 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 1372 1373 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1374 ("bgetvp: bp already attached! %p", bp)); 1375 1376 ASSERT_VI_LOCKED(vp, "bgetvp"); 1377 vholdl(vp); 1378 bp->b_vp = vp; 1379 bp->b_bufobj = &vp->v_bufobj; 1380 bp->b_dev = vn_todev(vp); 1381 /* 1382 * Insert onto list for new vnode. 1383 */ 1384 buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN); 1385} 1386 1387/* 1388 * Disassociate a buffer from a vnode. 1389 */ 1390void 1391brelvp(struct buf *bp) 1392{ 1393 struct bufobj *bo; 1394 struct vnode *vp; 1395 1396 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1397 1398 /* 1399 * Delete from old vnode list, if on one. 1400 */ 1401 vp = bp->b_vp; /* XXX */ 1402 bo = bp->b_bufobj; 1403 BO_LOCK(bo); 1404 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1405 buf_vlist_remove(bp); 1406 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1407 bo->bo_flag &= ~BO_ONWORKLST; 1408 mtx_lock(&sync_mtx); 1409 LIST_REMOVE(bo, bo_synclist); 1410 syncer_worklist_len--; 1411 mtx_unlock(&sync_mtx); 1412 } 1413 vdropl(vp); 1414 bp->b_vp = NULL; 1415 bp->b_bufobj = NULL; 1416 BO_UNLOCK(bo); 1417} 1418 1419/* 1420 * Add an item to the syncer work queue. 1421 */ 1422static void 1423vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 1424{ 1425 int slot; 1426 1427 ASSERT_BO_LOCKED(bo); 1428 1429 mtx_lock(&sync_mtx); 1430 if (bo->bo_flag & BO_ONWORKLST) 1431 LIST_REMOVE(bo, bo_synclist); 1432 else { 1433 bo->bo_flag |= BO_ONWORKLST; 1434 syncer_worklist_len++; 1435 } 1436 1437 if (delay > syncer_maxdelay - 2) 1438 delay = syncer_maxdelay - 2; 1439 slot = (syncer_delayno + delay) & syncer_mask; 1440 1441 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 1442 mtx_unlock(&sync_mtx); 1443} 1444 1445static int 1446sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 1447{ 1448 int error, len; 1449 1450 mtx_lock(&sync_mtx); 1451 len = syncer_worklist_len - sync_vnode_count; 1452 mtx_unlock(&sync_mtx); 1453 error = SYSCTL_OUT(req, &len, sizeof(len)); 1454 return (error); 1455} 1456 1457SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0, 1458 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 1459 1460struct proc *updateproc; 1461static void sched_sync(void); 1462static struct kproc_desc up_kp = { 1463 "syncer", 1464 sched_sync, 1465 &updateproc 1466}; 1467SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 1468 1469/* 1470 * System filesystem synchronizer daemon. 1471 */ 1472static void 1473sched_sync(void) 1474{ 1475 struct synclist *next; 1476 struct synclist *slp; 1477 struct vnode *vp; 1478 struct bufobj *bo; 1479 struct mount *mp; 1480 long starttime; 1481 struct thread *td = FIRST_THREAD_IN_PROC(updateproc); 1482 static int dummychan; 1483 int last_work_seen; 1484 int net_worklist_len; 1485 int syncer_final_iter; 1486 int first_printf; 1487 1488 mtx_lock(&Giant); 1489 last_work_seen = 0; 1490 syncer_final_iter = 0; 1491 first_printf = 1; 1492 syncer_state = SYNCER_RUNNING; 1493 starttime = time_second; 1494 1495 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 1496 SHUTDOWN_PRI_LAST); 1497 1498 for (;;) { 1499 mtx_lock(&sync_mtx); 1500 if (syncer_state == SYNCER_FINAL_DELAY && 1501 syncer_final_iter == 0) { 1502 mtx_unlock(&sync_mtx); 1503 kthread_suspend_check(td->td_proc); 1504 mtx_lock(&sync_mtx); 1505 } 1506 net_worklist_len = syncer_worklist_len - sync_vnode_count; 1507 if (syncer_state != SYNCER_RUNNING && 1508 starttime != time_second) { 1509 if (first_printf) { 1510 printf("\nSyncing disks, vnodes remaining..."); 1511 first_printf = 0; 1512 } 1513 printf("%d ", net_worklist_len); 1514 } 1515 starttime = time_second; 1516 1517 /* 1518 * Push files whose dirty time has expired. Be careful 1519 * of interrupt race on slp queue. 1520 * 1521 * Skip over empty worklist slots when shutting down. 1522 */ 1523 do { 1524 slp = &syncer_workitem_pending[syncer_delayno]; 1525 syncer_delayno += 1; 1526 if (syncer_delayno == syncer_maxdelay) 1527 syncer_delayno = 0; 1528 next = &syncer_workitem_pending[syncer_delayno]; 1529 /* 1530 * If the worklist has wrapped since the 1531 * it was emptied of all but syncer vnodes, 1532 * switch to the FINAL_DELAY state and run 1533 * for one more second. 1534 */ 1535 if (syncer_state == SYNCER_SHUTTING_DOWN && 1536 net_worklist_len == 0 && 1537 last_work_seen == syncer_delayno) { 1538 syncer_state = SYNCER_FINAL_DELAY; 1539 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 1540 } 1541 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 1542 syncer_worklist_len > 0); 1543 1544 /* 1545 * Keep track of the last time there was anything 1546 * on the worklist other than syncer vnodes. 1547 * Return to the SHUTTING_DOWN state if any 1548 * new work appears. 1549 */ 1550 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 1551 last_work_seen = syncer_delayno; 1552 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 1553 syncer_state = SYNCER_SHUTTING_DOWN; 1554 while ((bo = LIST_FIRST(slp)) != NULL) { 1555 vp = bo->bo_private; /* XXX */ 1556 if (VOP_ISLOCKED(vp, NULL) != 0 || 1557 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1558 LIST_REMOVE(bo, bo_synclist); 1559 LIST_INSERT_HEAD(next, bo, bo_synclist); 1560 continue; 1561 } 1562 if (VI_TRYLOCK(vp) == 0) { 1563 LIST_REMOVE(bo, bo_synclist); 1564 LIST_INSERT_HEAD(next, bo, bo_synclist); 1565 vn_finished_write(mp); 1566 continue; 1567 } 1568 /* 1569 * We use vhold in case the vnode does not 1570 * successfully sync. vhold prevents the vnode from 1571 * going away when we unlock the sync_mtx so that 1572 * we can acquire the vnode interlock. 1573 */ 1574 vholdl(vp); 1575 mtx_unlock(&sync_mtx); 1576 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, td); 1577 (void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td); 1578 VOP_UNLOCK(vp, 0, td); 1579 vn_finished_write(mp); 1580 VI_LOCK(vp); 1581 if ((bo->bo_flag & BO_ONWORKLST) != 0) { 1582 /* 1583 * Put us back on the worklist. The worklist 1584 * routine will remove us from our current 1585 * position and then add us back in at a later 1586 * position. 1587 */ 1588 vn_syncer_add_to_worklist(bo, syncdelay); 1589 } 1590 vdropl(vp); 1591 VI_UNLOCK(vp); 1592 mtx_lock(&sync_mtx); 1593 } 1594 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 1595 syncer_final_iter--; 1596 mtx_unlock(&sync_mtx); 1597 1598 /* 1599 * Do soft update processing. 1600 */ 1601 if (softdep_process_worklist_hook != NULL) 1602 (*softdep_process_worklist_hook)(NULL); 1603 1604 /* 1605 * The variable rushjob allows the kernel to speed up the 1606 * processing of the filesystem syncer process. A rushjob 1607 * value of N tells the filesystem syncer to process the next 1608 * N seconds worth of work on its queue ASAP. Currently rushjob 1609 * is used by the soft update code to speed up the filesystem 1610 * syncer process when the incore state is getting so far 1611 * ahead of the disk that the kernel memory pool is being 1612 * threatened with exhaustion. 1613 */ 1614 mtx_lock(&sync_mtx); 1615 if (rushjob > 0) { 1616 rushjob -= 1; 1617 mtx_unlock(&sync_mtx); 1618 continue; 1619 } 1620 mtx_unlock(&sync_mtx); 1621 /* 1622 * Just sleep for a short period if time between 1623 * iterations when shutting down to allow some I/O 1624 * to happen. 1625 * 1626 * If it has taken us less than a second to process the 1627 * current work, then wait. Otherwise start right over 1628 * again. We can still lose time if any single round 1629 * takes more than two seconds, but it does not really 1630 * matter as we are just trying to generally pace the 1631 * filesystem activity. 1632 */ 1633 if (syncer_state != SYNCER_RUNNING) 1634 tsleep(&dummychan, PPAUSE, "syncfnl", 1635 hz / SYNCER_SHUTDOWN_SPEEDUP); 1636 else if (time_second == starttime) 1637 tsleep(&lbolt, PPAUSE, "syncer", 0); 1638 } 1639} 1640 1641/* 1642 * Request the syncer daemon to speed up its work. 1643 * We never push it to speed up more than half of its 1644 * normal turn time, otherwise it could take over the cpu. 1645 */ 1646int 1647speedup_syncer() 1648{ 1649 struct thread *td; 1650 int ret = 0; 1651 1652 td = FIRST_THREAD_IN_PROC(updateproc); 1653 sleepq_remove(td, &lbolt); 1654 mtx_lock(&sync_mtx); 1655 if (rushjob < syncdelay / 2) { 1656 rushjob += 1; 1657 stat_rush_requests += 1; 1658 ret = 1; 1659 } 1660 mtx_unlock(&sync_mtx); 1661 return (ret); 1662} 1663 1664/* 1665 * Tell the syncer to speed up its work and run though its work 1666 * list several times, then tell it to shut down. 1667 */ 1668static void 1669syncer_shutdown(void *arg, int howto) 1670{ 1671 struct thread *td; 1672 1673 if (howto & RB_NOSYNC) 1674 return; 1675 td = FIRST_THREAD_IN_PROC(updateproc); 1676 sleepq_remove(td, &lbolt); 1677 mtx_lock(&sync_mtx); 1678 syncer_state = SYNCER_SHUTTING_DOWN; 1679 rushjob = 0; 1680 mtx_unlock(&sync_mtx); 1681 kproc_shutdown(arg, howto); 1682} 1683 1684/* 1685 * Associate a p-buffer with a vnode. 1686 * 1687 * Also sets B_PAGING flag to indicate that vnode is not fully associated 1688 * with the buffer. i.e. the bp has not been linked into the vnode or 1689 * ref-counted. 1690 */ 1691void 1692pbgetvp(vp, bp) 1693 register struct vnode *vp; 1694 register struct buf *bp; 1695{ 1696 1697 KASSERT(bp->b_vp == NULL, ("pbgetvp: not free")); 1698 1699 bp->b_vp = vp; 1700 bp->b_flags |= B_PAGING; 1701 bp->b_dev = vn_todev(vp); 1702 bp->b_bufobj = &vp->v_bufobj; 1703} 1704 1705/* 1706 * Disassociate a p-buffer from a vnode. 1707 */ 1708void 1709pbrelvp(bp) 1710 register struct buf *bp; 1711{ 1712 1713 KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL")); 1714 KASSERT(bp->b_bufobj != NULL, ("pbrelvp: NULL bufobj")); 1715 1716 /* XXX REMOVE ME */ 1717 BO_LOCK(bp->b_bufobj); 1718 if (TAILQ_NEXT(bp, b_bobufs) != NULL) { 1719 panic( 1720 "relpbuf(): b_vp was probably reassignbuf()d %p %x", 1721 bp, 1722 (int)bp->b_flags 1723 ); 1724 } 1725 BO_UNLOCK(bp->b_bufobj); 1726 bp->b_vp = NULL; 1727 bp->b_bufobj = NULL; 1728 bp->b_flags &= ~B_PAGING; 1729} 1730 1731/* 1732 * Reassign a buffer from one vnode to another. 1733 * Used to assign file specific control information 1734 * (indirect blocks) to the vnode to which they belong. 1735 */ 1736void 1737reassignbuf(struct buf *bp) 1738{ 1739 struct vnode *vp; 1740 struct bufobj *bo; 1741 int delay; 1742 1743 vp = bp->b_vp; 1744 bo = bp->b_bufobj; 1745 ++reassignbufcalls; 1746 1747 /* 1748 * B_PAGING flagged buffers cannot be reassigned because their vp 1749 * is not fully linked in. 1750 */ 1751 if (bp->b_flags & B_PAGING) 1752 panic("cannot reassign paging buffer"); 1753 1754 /* 1755 * Delete from old vnode list, if on one. 1756 */ 1757 VI_LOCK(vp); 1758 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1759 buf_vlist_remove(bp); 1760 /* 1761 * If dirty, put on list of dirty buffers; otherwise insert onto list 1762 * of clean buffers. 1763 */ 1764 if (bp->b_flags & B_DELWRI) { 1765 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 1766 switch (vp->v_type) { 1767 case VDIR: 1768 delay = dirdelay; 1769 break; 1770 case VCHR: 1771 delay = metadelay; 1772 break; 1773 default: 1774 delay = filedelay; 1775 } 1776 vn_syncer_add_to_worklist(bo, delay); 1777 } 1778 buf_vlist_add(bp, bo, BX_VNDIRTY); 1779 } else { 1780 buf_vlist_add(bp, bo, BX_VNCLEAN); 1781 1782 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1783 mtx_lock(&sync_mtx); 1784 LIST_REMOVE(bo, bo_synclist); 1785 syncer_worklist_len--; 1786 mtx_unlock(&sync_mtx); 1787 bo->bo_flag &= ~BO_ONWORKLST; 1788 } 1789 } 1790 VI_UNLOCK(vp); 1791} 1792 1793/* 1794 * Create a vnode for a device. 1795 * Used for mounting the root filesystem. 1796 */ 1797int 1798bdevvp(dev, vpp) 1799 struct cdev *dev; 1800 struct vnode **vpp; 1801{ 1802 register struct vnode *vp; 1803 struct vnode *nvp; 1804 int error; 1805 1806 if (dev == NULL) { 1807 *vpp = NULLVP; 1808 return (ENXIO); 1809 } 1810 if (vfinddev(dev, vpp)) 1811 return (0); 1812 1813 error = getnewvnode("none", (struct mount *)0, devfs_specop_p, &nvp); 1814 if (error) { 1815 *vpp = NULLVP; 1816 return (error); 1817 } 1818 vp = nvp; 1819 vp->v_type = VCHR; 1820 vp->v_bufobj.bo_bsize = DEV_BSIZE; 1821 addalias(vp, dev); 1822 *vpp = vp; 1823 return (0); 1824} 1825 1826static void 1827v_incr_usecount(struct vnode *vp, int delta) 1828{ 1829 1830 vp->v_usecount += delta; 1831 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1832 dev_lock(); 1833 vp->v_rdev->si_usecount += delta; 1834 dev_unlock(); 1835 } 1836} 1837 1838/* 1839 * Add vnode to the alias list hung off the struct cdev *. 1840 * 1841 * The reason for this gunk is that multiple vnodes can reference 1842 * the same physical device, so checking vp->v_usecount to see 1843 * how many users there are is inadequate; the v_usecount for 1844 * the vnodes need to be accumulated. vcount() does that. 1845 */ 1846struct vnode * 1847addaliasu(nvp, nvp_rdev) 1848 struct vnode *nvp; 1849 dev_t nvp_rdev; 1850{ 1851 struct vnode *ovp; 1852 vop_t **ops; 1853 struct cdev *dev; 1854 1855 if (nvp->v_type == VBLK) 1856 return (nvp); 1857 if (nvp->v_type != VCHR) 1858 panic("addaliasu on non-special vnode"); 1859 dev = findcdev(nvp_rdev); 1860 if (dev == NULL) 1861 return (nvp); 1862 /* 1863 * Check to see if we have a bdevvp vnode with no associated 1864 * filesystem. If so, we want to associate the filesystem of 1865 * the new newly instigated vnode with the bdevvp vnode and 1866 * discard the newly created vnode rather than leaving the 1867 * bdevvp vnode lying around with no associated filesystem. 1868 */ 1869 if (vfinddev(dev, &ovp) == 0 || ovp->v_data != NULL) { 1870 addalias(nvp, dev); 1871 return (nvp); 1872 } 1873 /* 1874 * Discard unneeded vnode, but save its node specific data. 1875 * Note that if there is a lock, it is carried over in the 1876 * node specific data to the replacement vnode. 1877 */ 1878 vref(ovp); 1879 ovp->v_data = nvp->v_data; 1880 ovp->v_tag = nvp->v_tag; 1881 nvp->v_data = NULL; 1882 lockdestroy(ovp->v_vnlock); 1883 lockinit(ovp->v_vnlock, PVFS, nvp->v_vnlock->lk_wmesg, 1884 nvp->v_vnlock->lk_timo, nvp->v_vnlock->lk_flags & LK_EXTFLG_MASK); 1885 ops = ovp->v_op; 1886 ovp->v_op = nvp->v_op; 1887 if (VOP_ISLOCKED(nvp, curthread)) { 1888 VOP_UNLOCK(nvp, 0, curthread); 1889 vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread); 1890 } 1891 nvp->v_op = ops; 1892 delmntque(ovp); 1893 insmntque(ovp, nvp->v_mount); 1894 vrele(nvp); 1895 vgone(nvp); 1896 return (ovp); 1897} 1898 1899/* This is a local helper function that do the same as addaliasu, but for a 1900 * struct cdev *instead of an dev_t. */ 1901static void 1902addalias(nvp, dev) 1903 struct vnode *nvp; 1904 struct cdev *dev; 1905{ 1906 1907 KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode")); 1908 VI_LOCK(nvp); 1909 dev_lock(); 1910 dev->si_refcount++; 1911 nvp->v_rdev = dev; 1912 SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext); 1913 dev->si_usecount += nvp->v_usecount; 1914 dev_unlock(); 1915 VI_UNLOCK(nvp); 1916} 1917 1918/* 1919 * Grab a particular vnode from the free list, increment its 1920 * reference count and lock it. The vnode lock bit is set if the 1921 * vnode is being eliminated in vgone. The process is awakened 1922 * when the transition is completed, and an error returned to 1923 * indicate that the vnode is no longer usable (possibly having 1924 * been changed to a new filesystem type). 1925 */ 1926int 1927vget(vp, flags, td) 1928 register struct vnode *vp; 1929 int flags; 1930 struct thread *td; 1931{ 1932 int error; 1933 1934 /* 1935 * If the vnode is in the process of being cleaned out for 1936 * another use, we wait for the cleaning to finish and then 1937 * return failure. Cleaning is determined by checking that 1938 * the VI_XLOCK flag is set. 1939 */ 1940 if ((flags & LK_INTERLOCK) == 0) 1941 VI_LOCK(vp); 1942 if (vp->v_iflag & VI_XLOCK && vp->v_vxthread != curthread) { 1943 if ((flags & LK_NOWAIT) == 0) { 1944 vp->v_iflag |= VI_XWANT; 1945 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0); 1946 return (ENOENT); 1947 } 1948 VI_UNLOCK(vp); 1949 return (EBUSY); 1950 } 1951 1952 v_incr_usecount(vp, 1); 1953 1954 if (VSHOULDBUSY(vp)) 1955 vbusy(vp); 1956 if (flags & LK_TYPE_MASK) { 1957 if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { 1958 /* 1959 * must expand vrele here because we do not want 1960 * to call VOP_INACTIVE if the reference count 1961 * drops back to zero since it was never really 1962 * active. We must remove it from the free list 1963 * before sleeping so that multiple processes do 1964 * not try to recycle it. 1965 */ 1966 VI_LOCK(vp); 1967 v_incr_usecount(vp, -1); 1968 if (VSHOULDFREE(vp)) 1969 vfree(vp); 1970 else 1971 vlruvp(vp); 1972 VI_UNLOCK(vp); 1973 } 1974 return (error); 1975 } 1976 VI_UNLOCK(vp); 1977 return (0); 1978} 1979 1980/* 1981 * Increase the reference count of a vnode. 1982 */ 1983void 1984vref(struct vnode *vp) 1985{ 1986 1987 VI_LOCK(vp); 1988 v_incr_usecount(vp, 1); 1989 VI_UNLOCK(vp); 1990} 1991 1992/* 1993 * Return reference count of a vnode. 1994 * 1995 * The results of this call are only guaranteed when some mechanism other 1996 * than the VI lock is used to stop other processes from gaining references 1997 * to the vnode. This may be the case if the caller holds the only reference. 1998 * This is also useful when stale data is acceptable as race conditions may 1999 * be accounted for by some other means. 2000 */ 2001int 2002vrefcnt(struct vnode *vp) 2003{ 2004 int usecnt; 2005 2006 VI_LOCK(vp); 2007 usecnt = vp->v_usecount; 2008 VI_UNLOCK(vp); 2009 2010 return (usecnt); 2011} 2012 2013 2014/* 2015 * Vnode put/release. 2016 * If count drops to zero, call inactive routine and return to freelist. 2017 */ 2018void 2019vrele(vp) 2020 struct vnode *vp; 2021{ 2022 struct thread *td = curthread; /* XXX */ 2023 2024 GIANT_REQUIRED; 2025 2026 KASSERT(vp != NULL, ("vrele: null vp")); 2027 2028 VI_LOCK(vp); 2029 2030 /* Skip this v_writecount check if we're going to panic below. */ 2031 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2032 ("vrele: missed vn_close")); 2033 2034 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2035 vp->v_usecount == 1)) { 2036 v_incr_usecount(vp, -1); 2037 VI_UNLOCK(vp); 2038 2039 return; 2040 } 2041 2042 if (vp->v_usecount == 1) { 2043 v_incr_usecount(vp, -1); 2044 /* 2045 * We must call VOP_INACTIVE with the node locked. Mark 2046 * as VI_DOINGINACT to avoid recursion. 2047 */ 2048 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) { 2049 VI_LOCK(vp); 2050 vp->v_iflag |= VI_DOINGINACT; 2051 VI_UNLOCK(vp); 2052 VOP_INACTIVE(vp, td); 2053 VI_LOCK(vp); 2054 KASSERT(vp->v_iflag & VI_DOINGINACT, 2055 ("vrele: lost VI_DOINGINACT")); 2056 vp->v_iflag &= ~VI_DOINGINACT; 2057 } else 2058 VI_LOCK(vp); 2059 if (VSHOULDFREE(vp)) 2060 vfree(vp); 2061 else 2062 vlruvp(vp); 2063 VI_UNLOCK(vp); 2064 2065 } else { 2066#ifdef DIAGNOSTIC 2067 vprint("vrele: negative ref count", vp); 2068#endif 2069 VI_UNLOCK(vp); 2070 panic("vrele: negative ref cnt"); 2071 } 2072} 2073 2074/* 2075 * Release an already locked vnode. This give the same effects as 2076 * unlock+vrele(), but takes less time and avoids releasing and 2077 * re-aquiring the lock (as vrele() aquires the lock internally.) 2078 */ 2079void 2080vput(vp) 2081 struct vnode *vp; 2082{ 2083 struct thread *td = curthread; /* XXX */ 2084 2085 GIANT_REQUIRED; 2086 2087 KASSERT(vp != NULL, ("vput: null vp")); 2088 VI_LOCK(vp); 2089 /* Skip this v_writecount check if we're going to panic below. */ 2090 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2091 ("vput: missed vn_close")); 2092 2093 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2094 vp->v_usecount == 1)) { 2095 v_incr_usecount(vp, -1); 2096 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2097 return; 2098 } 2099 2100 if (vp->v_usecount == 1) { 2101 v_incr_usecount(vp, -1); 2102 /* 2103 * We must call VOP_INACTIVE with the node locked, so 2104 * we just need to release the vnode mutex. Mark as 2105 * as VI_DOINGINACT to avoid recursion. 2106 */ 2107 vp->v_iflag |= VI_DOINGINACT; 2108 VI_UNLOCK(vp); 2109 VOP_INACTIVE(vp, td); 2110 VI_LOCK(vp); 2111 KASSERT(vp->v_iflag & VI_DOINGINACT, 2112 ("vput: lost VI_DOINGINACT")); 2113 vp->v_iflag &= ~VI_DOINGINACT; 2114 if (VSHOULDFREE(vp)) 2115 vfree(vp); 2116 else 2117 vlruvp(vp); 2118 VI_UNLOCK(vp); 2119 2120 } else { 2121#ifdef DIAGNOSTIC 2122 vprint("vput: negative ref count", vp); 2123#endif 2124 panic("vput: negative ref cnt"); 2125 } 2126} 2127 2128/* 2129 * Somebody doesn't want the vnode recycled. 2130 */ 2131void 2132vhold(struct vnode *vp) 2133{ 2134 2135 VI_LOCK(vp); 2136 vholdl(vp); 2137 VI_UNLOCK(vp); 2138} 2139 2140void 2141vholdl(struct vnode *vp) 2142{ 2143 2144 vp->v_holdcnt++; 2145 if (VSHOULDBUSY(vp)) 2146 vbusy(vp); 2147} 2148 2149/* 2150 * Note that there is one less who cares about this vnode. vdrop() is the 2151 * opposite of vhold(). 2152 */ 2153void 2154vdrop(struct vnode *vp) 2155{ 2156 2157 VI_LOCK(vp); 2158 vdropl(vp); 2159 VI_UNLOCK(vp); 2160} 2161 2162void 2163vdropl(vp) 2164 register struct vnode *vp; 2165{ 2166 2167 if (vp->v_holdcnt <= 0) 2168 panic("vdrop: holdcnt"); 2169 vp->v_holdcnt--; 2170 if (VSHOULDFREE(vp)) 2171 vfree(vp); 2172 else 2173 vlruvp(vp); 2174} 2175 2176/* 2177 * Remove any vnodes in the vnode table belonging to mount point mp. 2178 * 2179 * If FORCECLOSE is not specified, there should not be any active ones, 2180 * return error if any are found (nb: this is a user error, not a 2181 * system error). If FORCECLOSE is specified, detach any active vnodes 2182 * that are found. 2183 * 2184 * If WRITECLOSE is set, only flush out regular file vnodes open for 2185 * writing. 2186 * 2187 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2188 * 2189 * `rootrefs' specifies the base reference count for the root vnode 2190 * of this filesystem. The root vnode is considered busy if its 2191 * v_usecount exceeds this value. On a successful return, vflush(, td) 2192 * will call vrele() on the root vnode exactly rootrefs times. 2193 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2194 * be zero. 2195 */ 2196#ifdef DIAGNOSTIC 2197static int busyprt = 0; /* print out busy vnodes */ 2198SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 2199#endif 2200 2201int 2202vflush(mp, rootrefs, flags, td) 2203 struct mount *mp; 2204 int rootrefs; 2205 int flags; 2206 struct thread *td; 2207{ 2208 struct vnode *vp, *nvp, *rootvp = NULL; 2209 struct vattr vattr; 2210 int busy = 0, error; 2211 2212 if (rootrefs > 0) { 2213 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2214 ("vflush: bad args")); 2215 /* 2216 * Get the filesystem root vnode. We can vput() it 2217 * immediately, since with rootrefs > 0, it won't go away. 2218 */ 2219 if ((error = VFS_ROOT(mp, &rootvp, td)) != 0) 2220 return (error); 2221 vput(rootvp); 2222 2223 } 2224 MNT_ILOCK(mp); 2225loop: 2226 MNT_VNODE_FOREACH(vp, mp, nvp) { 2227 2228 VI_LOCK(vp); 2229 MNT_IUNLOCK(mp); 2230 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td); 2231 if (error) { 2232 MNT_ILOCK(mp); 2233 goto loop; 2234 } 2235 /* 2236 * Skip over a vnodes marked VV_SYSTEM. 2237 */ 2238 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2239 VOP_UNLOCK(vp, 0, td); 2240 MNT_ILOCK(mp); 2241 continue; 2242 } 2243 /* 2244 * If WRITECLOSE is set, flush out unlinked but still open 2245 * files (even if open only for reading) and regular file 2246 * vnodes open for writing. 2247 */ 2248 if (flags & WRITECLOSE) { 2249 error = VOP_GETATTR(vp, &vattr, td->td_ucred, td); 2250 VI_LOCK(vp); 2251 2252 if ((vp->v_type == VNON || 2253 (error == 0 && vattr.va_nlink > 0)) && 2254 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2255 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2256 MNT_ILOCK(mp); 2257 continue; 2258 } 2259 } else 2260 VI_LOCK(vp); 2261 2262 VOP_UNLOCK(vp, 0, td); 2263 2264 /* 2265 * With v_usecount == 0, all we need to do is clear out the 2266 * vnode data structures and we are done. 2267 */ 2268 if (vp->v_usecount == 0) { 2269 vgonel(vp, td); 2270 MNT_ILOCK(mp); 2271 continue; 2272 } 2273 2274 /* 2275 * If FORCECLOSE is set, forcibly close the vnode. For block 2276 * or character devices, revert to an anonymous device. For 2277 * all other files, just kill them. 2278 */ 2279 if (flags & FORCECLOSE) { 2280 if (vp->v_type != VCHR) 2281 vgonel(vp, td); 2282 else 2283 vgonechrl(vp, td); 2284 MNT_ILOCK(mp); 2285 continue; 2286 } 2287#ifdef DIAGNOSTIC 2288 if (busyprt) 2289 vprint("vflush: busy vnode", vp); 2290#endif 2291 VI_UNLOCK(vp); 2292 MNT_ILOCK(mp); 2293 busy++; 2294 } 2295 MNT_IUNLOCK(mp); 2296 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2297 /* 2298 * If just the root vnode is busy, and if its refcount 2299 * is equal to `rootrefs', then go ahead and kill it. 2300 */ 2301 VI_LOCK(rootvp); 2302 KASSERT(busy > 0, ("vflush: not busy")); 2303 KASSERT(rootvp->v_usecount >= rootrefs, 2304 ("vflush: usecount %d < rootrefs %d", 2305 rootvp->v_usecount, rootrefs)); 2306 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2307 vgonel(rootvp, td); 2308 busy = 0; 2309 } else 2310 VI_UNLOCK(rootvp); 2311 } 2312 if (busy) 2313 return (EBUSY); 2314 for (; rootrefs > 0; rootrefs--) 2315 vrele(rootvp); 2316 return (0); 2317} 2318 2319/* 2320 * This moves a now (likely recyclable) vnode to the end of the 2321 * mountlist. XXX However, it is temporarily disabled until we 2322 * can clean up ffs_sync() and friends, which have loop restart 2323 * conditions which this code causes to operate O(N^2). 2324 */ 2325static void 2326vlruvp(struct vnode *vp) 2327{ 2328#if 0 2329 struct mount *mp; 2330 2331 if ((mp = vp->v_mount) != NULL) { 2332 MNT_ILOCK(mp); 2333 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2334 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2335 MNT_IUNLOCK(mp); 2336 } 2337#endif 2338} 2339 2340static void 2341vx_lock(struct vnode *vp) 2342{ 2343 2344 ASSERT_VI_LOCKED(vp, "vx_lock"); 2345 2346 /* 2347 * Prevent the vnode from being recycled or brought into use while we 2348 * clean it out. 2349 */ 2350 if (vp->v_iflag & VI_XLOCK) 2351 panic("vclean: deadlock"); 2352 vp->v_iflag |= VI_XLOCK; 2353 vp->v_vxthread = curthread; 2354} 2355 2356static void 2357vx_unlock(struct vnode *vp) 2358{ 2359 ASSERT_VI_LOCKED(vp, "vx_unlock"); 2360 vp->v_iflag &= ~VI_XLOCK; 2361 vp->v_vxthread = NULL; 2362 if (vp->v_iflag & VI_XWANT) { 2363 vp->v_iflag &= ~VI_XWANT; 2364 wakeup(vp); 2365 } 2366} 2367 2368/* 2369 * Disassociate the underlying filesystem from a vnode. 2370 */ 2371static void 2372vclean(vp, flags, td) 2373 struct vnode *vp; 2374 int flags; 2375 struct thread *td; 2376{ 2377 int active; 2378 2379 ASSERT_VI_LOCKED(vp, "vclean"); 2380 /* 2381 * Check to see if the vnode is in use. If so we have to reference it 2382 * before we clean it out so that its count cannot fall to zero and 2383 * generate a race against ourselves to recycle it. 2384 */ 2385 if ((active = vp->v_usecount)) 2386 v_incr_usecount(vp, 1); 2387 2388 /* 2389 * Even if the count is zero, the VOP_INACTIVE routine may still 2390 * have the object locked while it cleans it out. The VOP_LOCK 2391 * ensures that the VOP_INACTIVE routine is done with its work. 2392 * For active vnodes, it ensures that no other activity can 2393 * occur while the underlying object is being cleaned out. 2394 */ 2395 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2396 2397 /* 2398 * Clean out any buffers associated with the vnode. 2399 * If the flush fails, just toss the buffers. 2400 */ 2401 if (flags & DOCLOSE) { 2402 struct buf *bp; 2403 bp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd); 2404 if (bp != NULL) 2405 (void) vn_write_suspend_wait(vp, NULL, V_WAIT); 2406 if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0) 2407 vinvalbuf(vp, 0, NOCRED, td, 0, 0); 2408 } 2409 2410 VOP_DESTROYVOBJECT(vp); 2411 2412 /* 2413 * Any other processes trying to obtain this lock must first 2414 * wait for VXLOCK to clear, then call the new lock operation. 2415 */ 2416 VOP_UNLOCK(vp, 0, td); 2417 2418 /* 2419 * If purging an active vnode, it must be closed and 2420 * deactivated before being reclaimed. Note that the 2421 * VOP_INACTIVE will unlock the vnode. 2422 */ 2423 if (active) { 2424 if (flags & DOCLOSE) 2425 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2426 VI_LOCK(vp); 2427 if ((vp->v_iflag & VI_DOINGINACT) == 0) { 2428 vp->v_iflag |= VI_DOINGINACT; 2429 VI_UNLOCK(vp); 2430 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 2431 panic("vclean: cannot relock."); 2432 VOP_INACTIVE(vp, td); 2433 VI_LOCK(vp); 2434 KASSERT(vp->v_iflag & VI_DOINGINACT, 2435 ("vclean: lost VI_DOINGINACT")); 2436 vp->v_iflag &= ~VI_DOINGINACT; 2437 } 2438 VI_UNLOCK(vp); 2439 } 2440 /* 2441 * Reclaim the vnode. 2442 */ 2443 if (VOP_RECLAIM(vp, td)) 2444 panic("vclean: cannot reclaim"); 2445 2446 if (active) { 2447 /* 2448 * Inline copy of vrele() since VOP_INACTIVE 2449 * has already been called. 2450 */ 2451 VI_LOCK(vp); 2452 v_incr_usecount(vp, -1); 2453 if (vp->v_usecount <= 0) { 2454#ifdef INVARIANTS 2455 if (vp->v_usecount < 0 || vp->v_writecount != 0) { 2456 vprint("vclean: bad ref count", vp); 2457 panic("vclean: ref cnt"); 2458 } 2459#endif 2460 if (VSHOULDFREE(vp)) 2461 vfree(vp); 2462 } 2463 VI_UNLOCK(vp); 2464 } 2465 /* 2466 * Delete from old mount point vnode list. 2467 */ 2468 delmntque(vp); 2469 cache_purge(vp); 2470 VI_LOCK(vp); 2471 if (VSHOULDFREE(vp)) 2472 vfree(vp); 2473 2474 /* 2475 * Done with purge, reset to the standard lock and 2476 * notify sleepers of the grim news. 2477 */ 2478 vp->v_vnlock = &vp->v_lock; 2479 vp->v_op = dead_vnodeop_p; 2480 if (vp->v_pollinfo != NULL) 2481 vn_pollgone(vp); 2482 vp->v_tag = "none"; 2483} 2484 2485/* 2486 * Eliminate all activity associated with the requested vnode 2487 * and with all vnodes aliased to the requested vnode. 2488 */ 2489int 2490vop_revoke(ap) 2491 struct vop_revoke_args /* { 2492 struct vnode *a_vp; 2493 int a_flags; 2494 } */ *ap; 2495{ 2496 struct vnode *vp, *vq; 2497 struct cdev *dev; 2498 2499 KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke")); 2500 vp = ap->a_vp; 2501 KASSERT((vp->v_type == VCHR), ("vop_revoke: not VCHR")); 2502 2503 VI_LOCK(vp); 2504 /* 2505 * If a vgone (or vclean) is already in progress, 2506 * wait until it is done and return. 2507 */ 2508 if (vp->v_iflag & VI_XLOCK) { 2509 vp->v_iflag |= VI_XWANT; 2510 msleep(vp, VI_MTX(vp), PINOD | PDROP, 2511 "vop_revokeall", 0); 2512 return (0); 2513 } 2514 VI_UNLOCK(vp); 2515 dev = vp->v_rdev; 2516 for (;;) { 2517 dev_lock(); 2518 vq = SLIST_FIRST(&dev->si_hlist); 2519 dev_unlock(); 2520 if (vq == NULL) 2521 break; 2522 vgone(vq); 2523 } 2524 return (0); 2525} 2526 2527/* 2528 * Recycle an unused vnode to the front of the free list. 2529 * Release the passed interlock if the vnode will be recycled. 2530 */ 2531int 2532vrecycle(vp, inter_lkp, td) 2533 struct vnode *vp; 2534 struct mtx *inter_lkp; 2535 struct thread *td; 2536{ 2537 2538 VI_LOCK(vp); 2539 if (vp->v_usecount == 0) { 2540 if (inter_lkp) { 2541 mtx_unlock(inter_lkp); 2542 } 2543 vgonel(vp, td); 2544 return (1); 2545 } 2546 VI_UNLOCK(vp); 2547 return (0); 2548} 2549 2550/* 2551 * Eliminate all activity associated with a vnode 2552 * in preparation for reuse. 2553 */ 2554void 2555vgone(vp) 2556 register struct vnode *vp; 2557{ 2558 struct thread *td = curthread; /* XXX */ 2559 2560 VI_LOCK(vp); 2561 vgonel(vp, td); 2562} 2563 2564/* 2565 * Disassociate a character device from the its underlying filesystem and 2566 * attach it to spec. This is for use when the chr device is still active 2567 * and the filesystem is going away. 2568 */ 2569static void 2570vgonechrl(struct vnode *vp, struct thread *td) 2571{ 2572 ASSERT_VI_LOCKED(vp, "vgonechrl"); 2573 vx_lock(vp); 2574 /* 2575 * This is a custom version of vclean() which does not tearm down 2576 * the bufs or vm objects held by this vnode. This allows filesystems 2577 * to continue using devices which were discovered via another 2578 * filesystem that has been unmounted. 2579 */ 2580 if (vp->v_usecount != 0) { 2581 v_incr_usecount(vp, 1); 2582 /* 2583 * Ensure that no other activity can occur while the 2584 * underlying object is being cleaned out. 2585 */ 2586 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2587 /* 2588 * Any other processes trying to obtain this lock must first 2589 * wait for VXLOCK to clear, then call the new lock operation. 2590 */ 2591 VOP_UNLOCK(vp, 0, td); 2592 vp->v_vnlock = &vp->v_lock; 2593 vp->v_tag = "orphanchr"; 2594 vp->v_op = devfs_specop_p; 2595 delmntque(vp); 2596 cache_purge(vp); 2597 vrele(vp); 2598 VI_LOCK(vp); 2599 } else 2600 vclean(vp, 0, td); 2601 vp->v_op = devfs_specop_p; 2602 vx_unlock(vp); 2603 VI_UNLOCK(vp); 2604} 2605 2606/* 2607 * vgone, with the vp interlock held. 2608 */ 2609void 2610vgonel(vp, td) 2611 struct vnode *vp; 2612 struct thread *td; 2613{ 2614 /* 2615 * If a vgone (or vclean) is already in progress, 2616 * wait until it is done and return. 2617 */ 2618 ASSERT_VI_LOCKED(vp, "vgonel"); 2619 if (vp->v_iflag & VI_XLOCK) { 2620 vp->v_iflag |= VI_XWANT; 2621 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0); 2622 return; 2623 } 2624 vx_lock(vp); 2625 2626 /* 2627 * Clean out the filesystem specific data. 2628 */ 2629 vclean(vp, DOCLOSE, td); 2630 VI_UNLOCK(vp); 2631 2632 /* 2633 * If special device, remove it from special device alias list 2634 * if it is on one. 2635 */ 2636 VI_LOCK(vp); 2637 if (vp->v_type == VCHR && vp->v_rdev != NULL) 2638 dev_rel(vp); 2639 2640 /* 2641 * If it is on the freelist and not already at the head, 2642 * move it to the head of the list. The test of the 2643 * VDOOMED flag and the reference count of zero is because 2644 * it will be removed from the free list by getnewvnode, 2645 * but will not have its reference count incremented until 2646 * after calling vgone. If the reference count were 2647 * incremented first, vgone would (incorrectly) try to 2648 * close the previous instance of the underlying object. 2649 */ 2650 if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) { 2651 mtx_lock(&vnode_free_list_mtx); 2652 if (vp->v_iflag & VI_FREE) { 2653 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2654 } else { 2655 vp->v_iflag |= VI_FREE; 2656 freevnodes++; 2657 } 2658 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2659 mtx_unlock(&vnode_free_list_mtx); 2660 } 2661 2662 vp->v_type = VBAD; 2663 vx_unlock(vp); 2664 VI_UNLOCK(vp); 2665} 2666 2667/* 2668 * Lookup a vnode by device number. 2669 */ 2670int 2671vfinddev(dev, vpp) 2672 struct cdev *dev; 2673 struct vnode **vpp; 2674{ 2675 struct vnode *vp; 2676 2677 dev_lock(); 2678 SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) { 2679 *vpp = vp; 2680 dev_unlock(); 2681 return (1); 2682 } 2683 dev_unlock(); 2684 return (0); 2685} 2686 2687/* 2688 * Calculate the total number of references to a special device. 2689 */ 2690int 2691vcount(vp) 2692 struct vnode *vp; 2693{ 2694 int count; 2695 2696 dev_lock(); 2697 count = vp->v_rdev->si_usecount; 2698 dev_unlock(); 2699 return (count); 2700} 2701 2702/* 2703 * Same as above, but using the struct cdev *as argument 2704 */ 2705int 2706count_dev(dev) 2707 struct cdev *dev; 2708{ 2709 int count; 2710 2711 dev_lock(); 2712 count = dev->si_usecount; 2713 dev_unlock(); 2714 return(count); 2715} 2716 2717/* 2718 * Print out a description of a vnode. 2719 */ 2720static char *typename[] = 2721{"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 2722 2723void 2724vprint(label, vp) 2725 char *label; 2726 struct vnode *vp; 2727{ 2728 char buf[96]; 2729 2730 if (label != NULL) 2731 printf("%s: %p: ", label, (void *)vp); 2732 else 2733 printf("%p: ", (void *)vp); 2734 printf("tag %s, type %s, usecount %d, writecount %d, refcount %d,", 2735 vp->v_tag, typename[vp->v_type], vp->v_usecount, 2736 vp->v_writecount, vp->v_holdcnt); 2737 buf[0] = '\0'; 2738 if (vp->v_vflag & VV_ROOT) 2739 strcat(buf, "|VV_ROOT"); 2740 if (vp->v_vflag & VV_TEXT) 2741 strcat(buf, "|VV_TEXT"); 2742 if (vp->v_vflag & VV_SYSTEM) 2743 strcat(buf, "|VV_SYSTEM"); 2744 if (vp->v_iflag & VI_XLOCK) 2745 strcat(buf, "|VI_XLOCK"); 2746 if (vp->v_iflag & VI_XWANT) 2747 strcat(buf, "|VI_XWANT"); 2748 if (vp->v_iflag & VI_DOOMED) 2749 strcat(buf, "|VI_DOOMED"); 2750 if (vp->v_iflag & VI_FREE) 2751 strcat(buf, "|VI_FREE"); 2752 if (vp->v_vflag & VV_OBJBUF) 2753 strcat(buf, "|VV_OBJBUF"); 2754 if (buf[0] != '\0') 2755 printf(" flags (%s),", &buf[1]); 2756 lockmgr_printinfo(vp->v_vnlock); 2757 printf("\n"); 2758 if (vp->v_data != NULL) 2759 VOP_PRINT(vp); 2760} 2761 2762#ifdef DDB 2763#include <ddb/ddb.h> 2764/* 2765 * List all of the locked vnodes in the system. 2766 * Called when debugging the kernel. 2767 */ 2768DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 2769{ 2770 struct mount *mp, *nmp; 2771 struct vnode *vp; 2772 2773 /* 2774 * Note: because this is DDB, we can't obey the locking semantics 2775 * for these structures, which means we could catch an inconsistent 2776 * state and dereference a nasty pointer. Not much to be done 2777 * about that. 2778 */ 2779 printf("Locked vnodes\n"); 2780 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2781 nmp = TAILQ_NEXT(mp, mnt_list); 2782 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2783 if (VOP_ISLOCKED(vp, NULL)) 2784 vprint(NULL, vp); 2785 } 2786 nmp = TAILQ_NEXT(mp, mnt_list); 2787 } 2788} 2789#endif 2790 2791/* 2792 * Fill in a struct xvfsconf based on a struct vfsconf. 2793 */ 2794static void 2795vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) 2796{ 2797 2798 strcpy(xvfsp->vfc_name, vfsp->vfc_name); 2799 xvfsp->vfc_typenum = vfsp->vfc_typenum; 2800 xvfsp->vfc_refcount = vfsp->vfc_refcount; 2801 xvfsp->vfc_flags = vfsp->vfc_flags; 2802 /* 2803 * These are unused in userland, we keep them 2804 * to not break binary compatibility. 2805 */ 2806 xvfsp->vfc_vfsops = NULL; 2807 xvfsp->vfc_next = NULL; 2808} 2809 2810/* 2811 * Top level filesystem related information gathering. 2812 */ 2813static int 2814sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 2815{ 2816 struct vfsconf *vfsp; 2817 struct xvfsconf xvfsp; 2818 int error; 2819 2820 error = 0; 2821 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2822 vfsconf2x(vfsp, &xvfsp); 2823 error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp); 2824 if (error) 2825 break; 2826 } 2827 return (error); 2828} 2829 2830SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, 2831 "S,xvfsconf", "List of all configured filesystems"); 2832 2833#ifndef BURN_BRIDGES 2834static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2835 2836static int 2837vfs_sysctl(SYSCTL_HANDLER_ARGS) 2838{ 2839 int *name = (int *)arg1 - 1; /* XXX */ 2840 u_int namelen = arg2 + 1; /* XXX */ 2841 struct vfsconf *vfsp; 2842 struct xvfsconf xvfsp; 2843 2844 printf("WARNING: userland calling deprecated sysctl, " 2845 "please rebuild world\n"); 2846 2847#if 1 || defined(COMPAT_PRELITE2) 2848 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2849 if (namelen == 1) 2850 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2851#endif 2852 2853 switch (name[1]) { 2854 case VFS_MAXTYPENUM: 2855 if (namelen != 2) 2856 return (ENOTDIR); 2857 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2858 case VFS_CONF: 2859 if (namelen != 3) 2860 return (ENOTDIR); /* overloaded */ 2861 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) 2862 if (vfsp->vfc_typenum == name[2]) 2863 break; 2864 if (vfsp == NULL) 2865 return (EOPNOTSUPP); 2866 vfsconf2x(vfsp, &xvfsp); 2867 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 2868 } 2869 return (EOPNOTSUPP); 2870} 2871 2872SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl, 2873 "Generic filesystem"); 2874 2875#if 1 || defined(COMPAT_PRELITE2) 2876 2877static int 2878sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 2879{ 2880 int error; 2881 struct vfsconf *vfsp; 2882 struct ovfsconf ovfs; 2883 2884 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2885 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 2886 strcpy(ovfs.vfc_name, vfsp->vfc_name); 2887 ovfs.vfc_index = vfsp->vfc_typenum; 2888 ovfs.vfc_refcount = vfsp->vfc_refcount; 2889 ovfs.vfc_flags = vfsp->vfc_flags; 2890 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 2891 if (error) 2892 return error; 2893 } 2894 return 0; 2895} 2896 2897#endif /* 1 || COMPAT_PRELITE2 */ 2898#endif /* !BURN_BRIDGES */ 2899 2900#define KINFO_VNODESLOP 10 2901#ifdef notyet 2902/* 2903 * Dump vnode list (via sysctl). 2904 */ 2905/* ARGSUSED */ 2906static int 2907sysctl_vnode(SYSCTL_HANDLER_ARGS) 2908{ 2909 struct xvnode *xvn; 2910 struct thread *td = req->td; 2911 struct mount *mp; 2912 struct vnode *vp; 2913 int error, len, n; 2914 2915 /* 2916 * Stale numvnodes access is not fatal here. 2917 */ 2918 req->lock = 0; 2919 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 2920 if (!req->oldptr) 2921 /* Make an estimate */ 2922 return (SYSCTL_OUT(req, 0, len)); 2923 2924 error = sysctl_wire_old_buffer(req, 0); 2925 if (error != 0) 2926 return (error); 2927 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 2928 n = 0; 2929 mtx_lock(&mountlist_mtx); 2930 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 2931 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 2932 continue; 2933 MNT_ILOCK(mp); 2934 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2935 if (n == len) 2936 break; 2937 vref(vp); 2938 xvn[n].xv_size = sizeof *xvn; 2939 xvn[n].xv_vnode = vp; 2940#define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 2941 XV_COPY(usecount); 2942 XV_COPY(writecount); 2943 XV_COPY(holdcnt); 2944 XV_COPY(id); 2945 XV_COPY(mount); 2946 XV_COPY(numoutput); 2947 XV_COPY(type); 2948#undef XV_COPY 2949 xvn[n].xv_flag = vp->v_vflag; 2950 2951 switch (vp->v_type) { 2952 case VREG: 2953 case VDIR: 2954 case VLNK: 2955 xvn[n].xv_dev = vp->v_cachedfs; 2956 xvn[n].xv_ino = vp->v_cachedid; 2957 break; 2958 case VBLK: 2959 case VCHR: 2960 if (vp->v_rdev == NULL) { 2961 vrele(vp); 2962 continue; 2963 } 2964 xvn[n].xv_dev = dev2udev(vp->v_rdev); 2965 break; 2966 case VSOCK: 2967 xvn[n].xv_socket = vp->v_socket; 2968 break; 2969 case VFIFO: 2970 xvn[n].xv_fifo = vp->v_fifoinfo; 2971 break; 2972 case VNON: 2973 case VBAD: 2974 default: 2975 /* shouldn't happen? */ 2976 vrele(vp); 2977 continue; 2978 } 2979 vrele(vp); 2980 ++n; 2981 } 2982 MNT_IUNLOCK(mp); 2983 mtx_lock(&mountlist_mtx); 2984 vfs_unbusy(mp, td); 2985 if (n == len) 2986 break; 2987 } 2988 mtx_unlock(&mountlist_mtx); 2989 2990 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 2991 free(xvn, M_TEMP); 2992 return (error); 2993} 2994 2995SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 2996 0, 0, sysctl_vnode, "S,xvnode", ""); 2997#endif 2998 2999/* 3000 * Check to see if a filesystem is mounted on a block device. 3001 */ 3002int 3003vfs_mountedon(vp) 3004 struct vnode *vp; 3005{ 3006 3007 if (vp->v_rdev->si_mountpoint != NULL) 3008 return (EBUSY); 3009 return (0); 3010} 3011 3012/* 3013 * Unmount all filesystems. The list is traversed in reverse order 3014 * of mounting to avoid dependencies. 3015 */ 3016void 3017vfs_unmountall() 3018{ 3019 struct mount *mp; 3020 struct thread *td; 3021 int error; 3022 3023 if (curthread != NULL) 3024 td = curthread; 3025 else 3026 td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */ 3027 /* 3028 * Since this only runs when rebooting, it is not interlocked. 3029 */ 3030 while(!TAILQ_EMPTY(&mountlist)) { 3031 mp = TAILQ_LAST(&mountlist, mntlist); 3032 error = dounmount(mp, MNT_FORCE, td); 3033 if (error) { 3034 TAILQ_REMOVE(&mountlist, mp, mnt_list); 3035 printf("unmount of %s failed (", 3036 mp->mnt_stat.f_mntonname); 3037 if (error == EBUSY) 3038 printf("BUSY)\n"); 3039 else 3040 printf("%d)\n", error); 3041 } else { 3042 /* The unmount has removed mp from the mountlist */ 3043 } 3044 } 3045} 3046 3047/* 3048 * perform msync on all vnodes under a mount point 3049 * the mount point must be locked. 3050 */ 3051void 3052vfs_msync(struct mount *mp, int flags) 3053{ 3054 struct vnode *vp, *nvp; 3055 struct vm_object *obj; 3056 int tries; 3057 3058 GIANT_REQUIRED; 3059 3060 tries = 5; 3061 MNT_ILOCK(mp); 3062loop: 3063 TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) { 3064 if (vp->v_mount != mp) { 3065 if (--tries > 0) 3066 goto loop; 3067 break; 3068 } 3069 3070 VI_LOCK(vp); 3071 if (vp->v_iflag & VI_XLOCK) { 3072 VI_UNLOCK(vp); 3073 continue; 3074 } 3075 3076 if ((vp->v_iflag & VI_OBJDIRTY) && 3077 (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { 3078 MNT_IUNLOCK(mp); 3079 if (!vget(vp, 3080 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 3081 curthread)) { 3082 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 3083 vput(vp); 3084 MNT_ILOCK(mp); 3085 continue; 3086 } 3087 3088 if (VOP_GETVOBJECT(vp, &obj) == 0) { 3089 VM_OBJECT_LOCK(obj); 3090 vm_object_page_clean(obj, 0, 0, 3091 flags == MNT_WAIT ? 3092 OBJPC_SYNC : OBJPC_NOSYNC); 3093 VM_OBJECT_UNLOCK(obj); 3094 } 3095 vput(vp); 3096 } 3097 MNT_ILOCK(mp); 3098 if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) { 3099 if (--tries > 0) 3100 goto loop; 3101 break; 3102 } 3103 } else 3104 VI_UNLOCK(vp); 3105 } 3106 MNT_IUNLOCK(mp); 3107} 3108 3109/* 3110 * Create the VM object needed for VMIO and mmap support. This 3111 * is done for all VREG files in the system. Some filesystems might 3112 * afford the additional metadata buffering capability of the 3113 * VMIO code by making the device node be VMIO mode also. 3114 * 3115 * vp must be locked when vfs_object_create is called. 3116 */ 3117int 3118vfs_object_create(struct vnode *vp, struct thread *td, struct ucred *cred) 3119{ 3120 3121 GIANT_REQUIRED; 3122 return (VOP_CREATEVOBJECT(vp, cred, td)); 3123} 3124 3125/* 3126 * Mark a vnode as free, putting it up for recycling. 3127 */ 3128void 3129vfree(struct vnode *vp) 3130{ 3131 3132 ASSERT_VI_LOCKED(vp, "vfree"); 3133 mtx_lock(&vnode_free_list_mtx); 3134 KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free")); 3135 if (vp->v_iflag & VI_AGE) { 3136 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 3137 } else { 3138 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 3139 } 3140 freevnodes++; 3141 mtx_unlock(&vnode_free_list_mtx); 3142 vp->v_iflag &= ~VI_AGE; 3143 vp->v_iflag |= VI_FREE; 3144} 3145 3146/* 3147 * Opposite of vfree() - mark a vnode as in use. 3148 */ 3149void 3150vbusy(struct vnode *vp) 3151{ 3152 3153 ASSERT_VI_LOCKED(vp, "vbusy"); 3154 KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free")); 3155 3156 mtx_lock(&vnode_free_list_mtx); 3157 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 3158 freevnodes--; 3159 mtx_unlock(&vnode_free_list_mtx); 3160 3161 vp->v_iflag &= ~(VI_FREE|VI_AGE); 3162} 3163 3164/* 3165 * Initalize per-vnode helper structure to hold poll-related state. 3166 */ 3167void 3168v_addpollinfo(struct vnode *vp) 3169{ 3170 struct vpollinfo *vi; 3171 3172 vi = uma_zalloc(vnodepoll_zone, M_WAITOK); 3173 if (vp->v_pollinfo != NULL) { 3174 uma_zfree(vnodepoll_zone, vi); 3175 return; 3176 } 3177 vp->v_pollinfo = vi; 3178 mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 3179 knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, 3180 &vp->v_pollinfo->vpi_lock); 3181} 3182 3183/* 3184 * Record a process's interest in events which might happen to 3185 * a vnode. Because poll uses the historic select-style interface 3186 * internally, this routine serves as both the ``check for any 3187 * pending events'' and the ``record my interest in future events'' 3188 * functions. (These are done together, while the lock is held, 3189 * to avoid race conditions.) 3190 */ 3191int 3192vn_pollrecord(vp, td, events) 3193 struct vnode *vp; 3194 struct thread *td; 3195 short events; 3196{ 3197 3198 if (vp->v_pollinfo == NULL) 3199 v_addpollinfo(vp); 3200 mtx_lock(&vp->v_pollinfo->vpi_lock); 3201 if (vp->v_pollinfo->vpi_revents & events) { 3202 /* 3203 * This leaves events we are not interested 3204 * in available for the other process which 3205 * which presumably had requested them 3206 * (otherwise they would never have been 3207 * recorded). 3208 */ 3209 events &= vp->v_pollinfo->vpi_revents; 3210 vp->v_pollinfo->vpi_revents &= ~events; 3211 3212 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3213 return events; 3214 } 3215 vp->v_pollinfo->vpi_events |= events; 3216 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 3217 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3218 return 0; 3219} 3220 3221/* 3222 * Note the occurrence of an event. If the VN_POLLEVENT macro is used, 3223 * it is possible for us to miss an event due to race conditions, but 3224 * that condition is expected to be rare, so for the moment it is the 3225 * preferred interface. 3226 */ 3227void 3228vn_pollevent(vp, events) 3229 struct vnode *vp; 3230 short events; 3231{ 3232 3233 if (vp->v_pollinfo == NULL) 3234 v_addpollinfo(vp); 3235 mtx_lock(&vp->v_pollinfo->vpi_lock); 3236 if (vp->v_pollinfo->vpi_events & events) { 3237 /* 3238 * We clear vpi_events so that we don't 3239 * call selwakeup() twice if two events are 3240 * posted before the polling process(es) is 3241 * awakened. This also ensures that we take at 3242 * most one selwakeup() if the polling process 3243 * is no longer interested. However, it does 3244 * mean that only one event can be noticed at 3245 * a time. (Perhaps we should only clear those 3246 * event bits which we note?) XXX 3247 */ 3248 vp->v_pollinfo->vpi_events = 0; /* &= ~events ??? */ 3249 vp->v_pollinfo->vpi_revents |= events; 3250 selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO); 3251 } 3252 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3253} 3254 3255/* 3256 * Wake up anyone polling on vp because it is being revoked. 3257 * This depends on dead_poll() returning POLLHUP for correct 3258 * behavior. 3259 */ 3260void 3261vn_pollgone(vp) 3262 struct vnode *vp; 3263{ 3264 3265 mtx_lock(&vp->v_pollinfo->vpi_lock); 3266 VN_KNOTE_LOCKED(vp, NOTE_REVOKE); 3267 if (vp->v_pollinfo->vpi_events) { 3268 vp->v_pollinfo->vpi_events = 0; 3269 selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO); 3270 } 3271 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3272} 3273 3274 3275 3276/* 3277 * Routine to create and manage a filesystem syncer vnode. 3278 */ 3279#define sync_close ((int (*)(struct vop_close_args *))nullop) 3280static int sync_fsync(struct vop_fsync_args *); 3281static int sync_inactive(struct vop_inactive_args *); 3282static int sync_reclaim(struct vop_reclaim_args *); 3283 3284static vop_t **sync_vnodeop_p; 3285static struct vnodeopv_entry_desc sync_vnodeop_entries[] = { 3286 { &vop_default_desc, (vop_t *) vop_eopnotsupp }, 3287 { &vop_close_desc, (vop_t *) sync_close }, /* close */ 3288 { &vop_fsync_desc, (vop_t *) sync_fsync }, /* fsync */ 3289 { &vop_inactive_desc, (vop_t *) sync_inactive }, /* inactive */ 3290 { &vop_reclaim_desc, (vop_t *) sync_reclaim }, /* reclaim */ 3291 { &vop_lock_desc, (vop_t *) vop_stdlock }, /* lock */ 3292 { &vop_unlock_desc, (vop_t *) vop_stdunlock }, /* unlock */ 3293 { &vop_islocked_desc, (vop_t *) vop_stdislocked }, /* islocked */ 3294 { NULL, NULL } 3295}; 3296static struct vnodeopv_desc sync_vnodeop_opv_desc = 3297 { &sync_vnodeop_p, sync_vnodeop_entries }; 3298 3299VNODEOP_SET(sync_vnodeop_opv_desc); 3300 3301/* 3302 * Create a new filesystem syncer vnode for the specified mount point. 3303 */ 3304int 3305vfs_allocate_syncvnode(mp) 3306 struct mount *mp; 3307{ 3308 struct vnode *vp; 3309 static long start, incr, next; 3310 int error; 3311 3312 /* Allocate a new vnode */ 3313 if ((error = getnewvnode("syncer", mp, sync_vnodeop_p, &vp)) != 0) { 3314 mp->mnt_syncer = NULL; 3315 return (error); 3316 } 3317 vp->v_type = VNON; 3318 /* 3319 * Place the vnode onto the syncer worklist. We attempt to 3320 * scatter them about on the list so that they will go off 3321 * at evenly distributed times even if all the filesystems 3322 * are mounted at once. 3323 */ 3324 next += incr; 3325 if (next == 0 || next > syncer_maxdelay) { 3326 start /= 2; 3327 incr /= 2; 3328 if (start == 0) { 3329 start = syncer_maxdelay / 2; 3330 incr = syncer_maxdelay; 3331 } 3332 next = start; 3333 } 3334 VI_LOCK(vp); 3335 vn_syncer_add_to_worklist(&vp->v_bufobj, 3336 syncdelay > 0 ? next % syncdelay : 0); 3337 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 3338 mtx_lock(&sync_mtx); 3339 sync_vnode_count++; 3340 mtx_unlock(&sync_mtx); 3341 VI_UNLOCK(vp); 3342 mp->mnt_syncer = vp; 3343 return (0); 3344} 3345 3346/* 3347 * Do a lazy sync of the filesystem. 3348 */ 3349static int 3350sync_fsync(ap) 3351 struct vop_fsync_args /* { 3352 struct vnode *a_vp; 3353 struct ucred *a_cred; 3354 int a_waitfor; 3355 struct thread *a_td; 3356 } */ *ap; 3357{ 3358 struct vnode *syncvp = ap->a_vp; 3359 struct mount *mp = syncvp->v_mount; 3360 struct thread *td = ap->a_td; 3361 int error, asyncflag; 3362 struct bufobj *bo; 3363 3364 /* 3365 * We only need to do something if this is a lazy evaluation. 3366 */ 3367 if (ap->a_waitfor != MNT_LAZY) 3368 return (0); 3369 3370 /* 3371 * Move ourselves to the back of the sync list. 3372 */ 3373 bo = &syncvp->v_bufobj; 3374 BO_LOCK(bo); 3375 vn_syncer_add_to_worklist(bo, syncdelay); 3376 BO_UNLOCK(bo); 3377 3378 /* 3379 * Walk the list of vnodes pushing all that are dirty and 3380 * not already on the sync list. 3381 */ 3382 mtx_lock(&mountlist_mtx); 3383 if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { 3384 mtx_unlock(&mountlist_mtx); 3385 return (0); 3386 } 3387 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 3388 vfs_unbusy(mp, td); 3389 return (0); 3390 } 3391 asyncflag = mp->mnt_flag & MNT_ASYNC; 3392 mp->mnt_flag &= ~MNT_ASYNC; 3393 vfs_msync(mp, MNT_NOWAIT); 3394 error = VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td); 3395 if (asyncflag) 3396 mp->mnt_flag |= MNT_ASYNC; 3397 vn_finished_write(mp); 3398 vfs_unbusy(mp, td); 3399 return (error); 3400} 3401 3402/* 3403 * The syncer vnode is no referenced. 3404 */ 3405static int 3406sync_inactive(ap) 3407 struct vop_inactive_args /* { 3408 struct vnode *a_vp; 3409 struct thread *a_td; 3410 } */ *ap; 3411{ 3412 3413 VOP_UNLOCK(ap->a_vp, 0, ap->a_td); 3414 vgone(ap->a_vp); 3415 return (0); 3416} 3417 3418/* 3419 * The syncer vnode is no longer needed and is being decommissioned. 3420 * 3421 * Modifications to the worklist must be protected by sync_mtx. 3422 */ 3423static int 3424sync_reclaim(ap) 3425 struct vop_reclaim_args /* { 3426 struct vnode *a_vp; 3427 } */ *ap; 3428{ 3429 struct vnode *vp = ap->a_vp; 3430 struct bufobj *bo; 3431 3432 VI_LOCK(vp); 3433 bo = &vp->v_bufobj; 3434 vp->v_mount->mnt_syncer = NULL; 3435 if (bo->bo_flag & BO_ONWORKLST) { 3436 mtx_lock(&sync_mtx); 3437 LIST_REMOVE(bo, bo_synclist); 3438 syncer_worklist_len--; 3439 sync_vnode_count--; 3440 mtx_unlock(&sync_mtx); 3441 bo->bo_flag &= ~BO_ONWORKLST; 3442 } 3443 VI_UNLOCK(vp); 3444 3445 return (0); 3446} 3447 3448/* 3449 * extract the struct cdev *from a VCHR 3450 */ 3451struct cdev * 3452vn_todev(vp) 3453 struct vnode *vp; 3454{ 3455 3456 if (vp->v_type != VCHR) 3457 return (NULL); 3458 return (vp->v_rdev); 3459} 3460 3461/* 3462 * Check if vnode represents a disk device 3463 */ 3464int 3465vn_isdisk(vp, errp) 3466 struct vnode *vp; 3467 int *errp; 3468{ 3469 int error; 3470 3471 error = 0; 3472 dev_lock(); 3473 if (vp->v_type != VCHR) 3474 error = ENOTBLK; 3475 else if (vp->v_rdev == NULL) 3476 error = ENXIO; 3477 else if (vp->v_rdev->si_devsw == NULL) 3478 error = ENXIO; 3479 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 3480 error = ENOTBLK; 3481 dev_unlock(); 3482 if (errp != NULL) 3483 *errp = error; 3484 return (error == 0); 3485} 3486 3487/* 3488 * Free data allocated by namei(); see namei(9) for details. 3489 */ 3490void 3491NDFREE(ndp, flags) 3492 struct nameidata *ndp; 3493 const u_int flags; 3494{ 3495 3496 if (!(flags & NDF_NO_FREE_PNBUF) && 3497 (ndp->ni_cnd.cn_flags & HASBUF)) { 3498 uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf); 3499 ndp->ni_cnd.cn_flags &= ~HASBUF; 3500 } 3501 if (!(flags & NDF_NO_DVP_UNLOCK) && 3502 (ndp->ni_cnd.cn_flags & LOCKPARENT) && 3503 ndp->ni_dvp != ndp->ni_vp) 3504 VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread); 3505 if (!(flags & NDF_NO_DVP_RELE) && 3506 (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) { 3507 vrele(ndp->ni_dvp); 3508 ndp->ni_dvp = NULL; 3509 } 3510 if (!(flags & NDF_NO_VP_UNLOCK) && 3511 (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp) 3512 VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread); 3513 if (!(flags & NDF_NO_VP_RELE) && 3514 ndp->ni_vp) { 3515 vrele(ndp->ni_vp); 3516 ndp->ni_vp = NULL; 3517 } 3518 if (!(flags & NDF_NO_STARTDIR_RELE) && 3519 (ndp->ni_cnd.cn_flags & SAVESTART)) { 3520 vrele(ndp->ni_startdir); 3521 ndp->ni_startdir = NULL; 3522 } 3523} 3524 3525/* 3526 * Common filesystem object access control check routine. Accepts a 3527 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3528 * and optional call-by-reference privused argument allowing vaccess() 3529 * to indicate to the caller whether privilege was used to satisfy the 3530 * request (obsoleted). Returns 0 on success, or an errno on failure. 3531 */ 3532int 3533vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused) 3534 enum vtype type; 3535 mode_t file_mode; 3536 uid_t file_uid; 3537 gid_t file_gid; 3538 mode_t acc_mode; 3539 struct ucred *cred; 3540 int *privused; 3541{ 3542 mode_t dac_granted; 3543#ifdef CAPABILITIES 3544 mode_t cap_granted; 3545#endif 3546 3547 /* 3548 * Look for a normal, non-privileged way to access the file/directory 3549 * as requested. If it exists, go with that. 3550 */ 3551 3552 if (privused != NULL) 3553 *privused = 0; 3554 3555 dac_granted = 0; 3556 3557 /* Check the owner. */ 3558 if (cred->cr_uid == file_uid) { 3559 dac_granted |= VADMIN; 3560 if (file_mode & S_IXUSR) 3561 dac_granted |= VEXEC; 3562 if (file_mode & S_IRUSR) 3563 dac_granted |= VREAD; 3564 if (file_mode & S_IWUSR) 3565 dac_granted |= (VWRITE | VAPPEND); 3566 3567 if ((acc_mode & dac_granted) == acc_mode) 3568 return (0); 3569 3570 goto privcheck; 3571 } 3572 3573 /* Otherwise, check the groups (first match) */ 3574 if (groupmember(file_gid, cred)) { 3575 if (file_mode & S_IXGRP) 3576 dac_granted |= VEXEC; 3577 if (file_mode & S_IRGRP) 3578 dac_granted |= VREAD; 3579 if (file_mode & S_IWGRP) 3580 dac_granted |= (VWRITE | VAPPEND); 3581 3582 if ((acc_mode & dac_granted) == acc_mode) 3583 return (0); 3584 3585 goto privcheck; 3586 } 3587 3588 /* Otherwise, check everyone else. */ 3589 if (file_mode & S_IXOTH) 3590 dac_granted |= VEXEC; 3591 if (file_mode & S_IROTH) 3592 dac_granted |= VREAD; 3593 if (file_mode & S_IWOTH) 3594 dac_granted |= (VWRITE | VAPPEND); 3595 if ((acc_mode & dac_granted) == acc_mode) 3596 return (0); 3597 3598privcheck: 3599 if (!suser_cred(cred, SUSER_ALLOWJAIL)) { 3600 /* XXX audit: privilege used */ 3601 if (privused != NULL) 3602 *privused = 1; 3603 return (0); 3604 } 3605 3606#ifdef CAPABILITIES 3607 /* 3608 * Build a capability mask to determine if the set of capabilities 3609 * satisfies the requirements when combined with the granted mask 3610 * from above. 3611 * For each capability, if the capability is required, bitwise 3612 * or the request type onto the cap_granted mask. 3613 */ 3614 cap_granted = 0; 3615 3616 if (type == VDIR) { 3617 /* 3618 * For directories, use CAP_DAC_READ_SEARCH to satisfy 3619 * VEXEC requests, instead of CAP_DAC_EXECUTE. 3620 */ 3621 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3622 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL)) 3623 cap_granted |= VEXEC; 3624 } else { 3625 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3626 !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL)) 3627 cap_granted |= VEXEC; 3628 } 3629 3630 if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && 3631 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL)) 3632 cap_granted |= VREAD; 3633 3634 if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3635 !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL)) 3636 cap_granted |= (VWRITE | VAPPEND); 3637 3638 if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3639 !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL)) 3640 cap_granted |= VADMIN; 3641 3642 if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { 3643 /* XXX audit: privilege used */ 3644 if (privused != NULL) 3645 *privused = 1; 3646 return (0); 3647 } 3648#endif 3649 3650 return ((acc_mode & VADMIN) ? EPERM : EACCES); 3651} 3652 3653/* 3654 * Credential check based on process requesting service, and per-attribute 3655 * permissions. 3656 */ 3657int 3658extattr_check_cred(struct vnode *vp, int attrnamespace, 3659 struct ucred *cred, struct thread *td, int access) 3660{ 3661 3662 /* 3663 * Kernel-invoked always succeeds. 3664 */ 3665 if (cred == NOCRED) 3666 return (0); 3667 3668 /* 3669 * Do not allow privileged processes in jail to directly 3670 * manipulate system attributes. 3671 * 3672 * XXX What capability should apply here? 3673 * Probably CAP_SYS_SETFFLAG. 3674 */ 3675 switch (attrnamespace) { 3676 case EXTATTR_NAMESPACE_SYSTEM: 3677 /* Potentially should be: return (EPERM); */ 3678 return (suser_cred(cred, 0)); 3679 case EXTATTR_NAMESPACE_USER: 3680 return (VOP_ACCESS(vp, access, cred, td)); 3681 default: 3682 return (EPERM); 3683 } 3684} 3685 3686#ifdef DEBUG_VFS_LOCKS 3687/* 3688 * This only exists to supress warnings from unlocked specfs accesses. It is 3689 * no longer ok to have an unlocked VFS. 3690 */ 3691#define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD) 3692 3693int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 3694SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, ""); 3695 3696int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 3697SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, ""); 3698 3699int vfs_badlock_print = 1; /* Print lock violations. */ 3700SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, ""); 3701 3702#ifdef KDB 3703int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 3704SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, ""); 3705#endif 3706 3707static void 3708vfs_badlock(const char *msg, const char *str, struct vnode *vp) 3709{ 3710 3711#ifdef KDB 3712 if (vfs_badlock_backtrace) 3713 kdb_backtrace(); 3714#endif 3715 if (vfs_badlock_print) 3716 printf("%s: %p %s\n", str, (void *)vp, msg); 3717 if (vfs_badlock_ddb) 3718 kdb_enter("lock violation"); 3719} 3720 3721void 3722assert_vi_locked(struct vnode *vp, const char *str) 3723{ 3724 3725 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 3726 vfs_badlock("interlock is not locked but should be", str, vp); 3727} 3728 3729void 3730assert_vi_unlocked(struct vnode *vp, const char *str) 3731{ 3732 3733 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 3734 vfs_badlock("interlock is locked but should not be", str, vp); 3735} 3736 3737void 3738assert_vop_locked(struct vnode *vp, const char *str) 3739{ 3740 3741 if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0) 3742 vfs_badlock("is not locked but should be", str, vp); 3743} 3744 3745void 3746assert_vop_unlocked(struct vnode *vp, const char *str) 3747{ 3748 3749 if (vp && !IGNORE_LOCK(vp) && 3750 VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE) 3751 vfs_badlock("is locked but should not be", str, vp); 3752} 3753 3754#if 0 3755void 3756assert_vop_elocked(struct vnode *vp, const char *str) 3757{ 3758 3759 if (vp && !IGNORE_LOCK(vp) && 3760 VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE) 3761 vfs_badlock("is not exclusive locked but should be", str, vp); 3762} 3763 3764void 3765assert_vop_elocked_other(struct vnode *vp, const char *str) 3766{ 3767 3768 if (vp && !IGNORE_LOCK(vp) && 3769 VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER) 3770 vfs_badlock("is not exclusive locked by another thread", 3771 str, vp); 3772} 3773 3774void 3775assert_vop_slocked(struct vnode *vp, const char *str) 3776{ 3777 3778 if (vp && !IGNORE_LOCK(vp) && 3779 VOP_ISLOCKED(vp, curthread) != LK_SHARED) 3780 vfs_badlock("is not locked shared but should be", str, vp); 3781} 3782#endif /* 0 */ 3783 3784void 3785vop_rename_pre(void *ap) 3786{ 3787 struct vop_rename_args *a = ap; 3788 3789 if (a->a_tvp) 3790 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 3791 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 3792 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 3793 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 3794 3795 /* Check the source (from). */ 3796 if (a->a_tdvp != a->a_fdvp) 3797 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 3798 if (a->a_tvp != a->a_fvp) 3799 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked"); 3800 3801 /* Check the target. */ 3802 if (a->a_tvp) 3803 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 3804 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 3805} 3806 3807void 3808vop_strategy_pre(void *ap) 3809{ 3810 struct vop_strategy_args *a; 3811 struct buf *bp; 3812 3813 a = ap; 3814 bp = a->a_bp; 3815 3816 /* 3817 * Cluster ops lock their component buffers but not the IO container. 3818 */ 3819 if ((bp->b_flags & B_CLUSTER) != 0) 3820 return; 3821 3822 if (BUF_REFCNT(bp) < 1) { 3823 if (vfs_badlock_print) 3824 printf( 3825 "VOP_STRATEGY: bp is not locked but should be\n"); 3826 if (vfs_badlock_ddb) 3827 kdb_enter("lock violation"); 3828 } 3829} 3830 3831void 3832vop_lookup_pre(void *ap) 3833{ 3834 struct vop_lookup_args *a; 3835 struct vnode *dvp; 3836 3837 a = ap; 3838 dvp = a->a_dvp; 3839 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3840 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 3841} 3842 3843void 3844vop_lookup_post(void *ap, int rc) 3845{ 3846 struct vop_lookup_args *a; 3847 struct componentname *cnp; 3848 struct vnode *dvp; 3849 struct vnode *vp; 3850 int flags; 3851 3852 a = ap; 3853 dvp = a->a_dvp; 3854 cnp = a->a_cnp; 3855 vp = *(a->a_vpp); 3856 flags = cnp->cn_flags; 3857 3858 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3859 3860 /* 3861 * If this is the last path component for this lookup and LOCKPARENT 3862 * is set, OR if there is an error the directory has to be locked. 3863 */ 3864 if ((flags & LOCKPARENT) && (flags & ISLASTCN)) 3865 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)"); 3866 else if (rc != 0) 3867 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)"); 3868 else if (dvp != vp) 3869 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)"); 3870 if (flags & PDIRUNLOCK) 3871 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)"); 3872} 3873 3874void 3875vop_lock_pre(void *ap) 3876{ 3877 struct vop_lock_args *a = ap; 3878 3879 if ((a->a_flags & LK_INTERLOCK) == 0) 3880 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3881 else 3882 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 3883} 3884 3885void 3886vop_lock_post(void *ap, int rc) 3887{ 3888 struct vop_lock_args *a = ap; 3889 3890 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3891 if (rc == 0) 3892 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 3893} 3894 3895void 3896vop_unlock_pre(void *ap) 3897{ 3898 struct vop_unlock_args *a = ap; 3899 3900 if (a->a_flags & LK_INTERLOCK) 3901 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 3902 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 3903} 3904 3905void 3906vop_unlock_post(void *ap, int rc) 3907{ 3908 struct vop_unlock_args *a = ap; 3909 3910 if (a->a_flags & LK_INTERLOCK) 3911 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 3912} 3913#endif /* DEBUG_VFS_LOCKS */ 3914 3915static struct knlist fs_knlist; 3916 3917static void 3918vfs_event_init(void *arg) 3919{ 3920 knlist_init(&fs_knlist, NULL); 3921} 3922/* XXX - correct order? */ 3923SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 3924 3925void 3926vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused) 3927{ 3928 3929 KNOTE_UNLOCKED(&fs_knlist, event); 3930} 3931 3932static int filt_fsattach(struct knote *kn); 3933static void filt_fsdetach(struct knote *kn); 3934static int filt_fsevent(struct knote *kn, long hint); 3935 3936struct filterops fs_filtops = 3937 { 0, filt_fsattach, filt_fsdetach, filt_fsevent }; 3938 3939static int 3940filt_fsattach(struct knote *kn) 3941{ 3942 3943 kn->kn_flags |= EV_CLEAR; 3944 knlist_add(&fs_knlist, kn, 0); 3945 return (0); 3946} 3947 3948static void 3949filt_fsdetach(struct knote *kn) 3950{ 3951 3952 knlist_remove(&fs_knlist, kn, 0); 3953} 3954 3955static int 3956filt_fsevent(struct knote *kn, long hint) 3957{ 3958 3959 kn->kn_fflags |= hint; 3960 return (kn->kn_fflags != 0); 3961} 3962 3963static int 3964sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 3965{ 3966 struct vfsidctl vc; 3967 int error; 3968 struct mount *mp; 3969 3970 error = SYSCTL_IN(req, &vc, sizeof(vc)); 3971 if (error) 3972 return (error); 3973 if (vc.vc_vers != VFS_CTL_VERS1) 3974 return (EINVAL); 3975 mp = vfs_getvfs(&vc.vc_fsid); 3976 if (mp == NULL) 3977 return (ENOENT); 3978 /* ensure that a specific sysctl goes to the right filesystem. */ 3979 if (strcmp(vc.vc_fstypename, "*") != 0 && 3980 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 3981 return (EINVAL); 3982 } 3983 VCTLTOREQ(&vc, req); 3984 return (VFS_SYSCTL(mp, vc.vc_op, req)); 3985} 3986 3987SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR, 3988 NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid"); 3989 3990/* 3991 * Function to initialize a va_filerev field sensibly. 3992 * XXX: Wouldn't a random number make a lot more sense ?? 3993 */ 3994u_quad_t 3995init_va_filerev(void) 3996{ 3997 struct bintime bt; 3998 3999 getbinuptime(&bt); 4000 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 4001} 4002