vfs_subr.c revision 137050
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 137050 2004-10-29 11:15:08Z 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_vnode = vp; 836 bo->bo_mtx = &vp->v_interlock; 837 vp->v_vnlock = &vp->v_lock; 838 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 839 cache_purge(vp); /* Sets up v_id. */ 840 LIST_INIT(&vp->v_cache_src); 841 TAILQ_INIT(&vp->v_cache_dst); 842 } 843 844 TAILQ_INIT(&bo->bo_clean.bv_hd); 845 TAILQ_INIT(&bo->bo_dirty.bv_hd); 846 bo->bo_ops = &buf_ops_bio; 847 vp->v_type = VNON; 848 vp->v_tag = tag; 849 vp->v_op = vops; 850 *vpp = vp; 851 vp->v_usecount = 1; 852 vp->v_data = 0; 853 vp->v_cachedid = -1; 854 VI_UNLOCK(vp); 855 if (pollinfo != NULL) { 856 knlist_destroy(&pollinfo->vpi_selinfo.si_note); 857 mtx_destroy(&pollinfo->vpi_lock); 858 uma_zfree(vnodepoll_zone, pollinfo); 859 } 860#ifdef MAC 861 mac_init_vnode(vp); 862 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 863 mac_associate_vnode_singlelabel(mp, vp); 864 else if (mp == NULL) 865 printf("NULL mp in getnewvnode()\n"); 866#endif 867 delmntque(vp); 868 if (mp != NULL) { 869 insmntque(vp, mp); 870 bo->bo_bsize = mp->mnt_stat.f_iosize; 871 } 872 873 return (0); 874} 875 876/* 877 * Delete from old mount point vnode list, if on one. 878 */ 879static void 880delmntque(struct vnode *vp) 881{ 882 struct mount *mp; 883 884 if (vp->v_mount == NULL) 885 return; 886 mp = vp->v_mount; 887 MNT_ILOCK(mp); 888 vp->v_mount = NULL; 889 KASSERT(mp->mnt_nvnodelistsize > 0, 890 ("bad mount point vnode list size")); 891 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 892 mp->mnt_nvnodelistsize--; 893 MNT_IUNLOCK(mp); 894} 895 896/* 897 * Insert into list of vnodes for the new mount point, if available. 898 */ 899static void 900insmntque(struct vnode *vp, struct mount *mp) 901{ 902 903 vp->v_mount = mp; 904 KASSERT(mp != NULL, ("Don't call insmntque(foo, NULL)")); 905 MNT_ILOCK(vp->v_mount); 906 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 907 mp->mnt_nvnodelistsize++; 908 MNT_IUNLOCK(vp->v_mount); 909} 910 911/* 912 * Flush out and invalidate all buffers associated with a vnode. 913 * Called with the underlying object locked. 914 */ 915int 916vinvalbuf(vp, flags, cred, td, slpflag, slptimeo) 917 struct vnode *vp; 918 int flags; 919 struct ucred *cred; 920 struct thread *td; 921 int slpflag, slptimeo; 922{ 923 struct buf *blist; 924 int error; 925 vm_object_t object; 926 struct bufobj *bo; 927 928 GIANT_REQUIRED; 929 930 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 931 932 bo = &vp->v_bufobj; 933 BO_LOCK(bo); 934 if (flags & V_SAVE) { 935 error = bufobj_wwait(bo, slpflag, slptimeo); 936 if (error) { 937 VI_UNLOCK(vp); 938 return (error); 939 } 940 if (bo->bo_dirty.bv_cnt > 0) { 941 VI_UNLOCK(vp); 942 if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0) 943 return (error); 944 /* 945 * XXX We could save a lock/unlock if this was only 946 * enabled under INVARIANTS 947 */ 948 VI_LOCK(vp); 949 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 950 panic("vinvalbuf: dirty bufs"); 951 } 952 } 953 /* 954 * If you alter this loop please notice that interlock is dropped and 955 * reacquired in flushbuflist. Special care is needed to ensure that 956 * no race conditions occur from this. 957 */ 958 for (error = 0;;) { 959 blist = TAILQ_FIRST(&vp->v_bufobj.bo_clean.bv_hd); 960 if (blist != NULL && 961 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 962 if (error) 963 break; 964 continue; 965 } 966 blist = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd); 967 if (blist != NULL && 968 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 969 if (error) 970 break; 971 continue; 972 } 973 break; 974 } 975 if (error) { 976 VI_UNLOCK(vp); 977 return (error); 978 } 979 980 /* 981 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 982 * have write I/O in-progress but if there is a VM object then the 983 * VM object can also have read-I/O in-progress. 984 */ 985 do { 986 bufobj_wwait(bo, 0, 0); 987 VI_UNLOCK(vp); 988 if (VOP_GETVOBJECT(vp, &object) == 0) { 989 VM_OBJECT_LOCK(object); 990 vm_object_pip_wait(object, "vnvlbx"); 991 VM_OBJECT_UNLOCK(object); 992 } 993 VI_LOCK(vp); 994 } while (bo->bo_numoutput > 0); 995 VI_UNLOCK(vp); 996 997 /* 998 * Destroy the copy in the VM cache, too. 999 */ 1000 if (VOP_GETVOBJECT(vp, &object) == 0) { 1001 VM_OBJECT_LOCK(object); 1002 vm_object_page_remove(object, 0, 0, 1003 (flags & V_SAVE) ? TRUE : FALSE); 1004 VM_OBJECT_UNLOCK(object); 1005 } 1006 1007#ifdef INVARIANTS 1008 VI_LOCK(vp); 1009 if ((flags & (V_ALT | V_NORMAL)) == 0 && 1010 (vp->v_bufobj.bo_dirty.bv_cnt > 0 || 1011 vp->v_bufobj.bo_clean.bv_cnt > 0)) 1012 panic("vinvalbuf: flush failed"); 1013 VI_UNLOCK(vp); 1014#endif 1015 return (0); 1016} 1017 1018/* 1019 * Flush out buffers on the specified list. 1020 * 1021 */ 1022static int 1023flushbuflist(blist, flags, vp, slpflag, slptimeo, errorp) 1024 struct buf *blist; 1025 int flags; 1026 struct vnode *vp; 1027 int slpflag, slptimeo; 1028 int *errorp; 1029{ 1030 struct buf *bp, *nbp; 1031 int found, error; 1032 1033 ASSERT_VI_LOCKED(vp, "flushbuflist"); 1034 1035 for (found = 0, bp = blist; bp; bp = nbp) { 1036 nbp = TAILQ_NEXT(bp, b_bobufs); 1037 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1038 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1039 continue; 1040 } 1041 found += 1; 1042 error = BUF_TIMELOCK(bp, 1043 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp), 1044 "flushbuf", slpflag, slptimeo); 1045 if (error) { 1046 if (error != ENOLCK) 1047 *errorp = error; 1048 goto done; 1049 } 1050 /* 1051 * XXX Since there are no node locks for NFS, I 1052 * believe there is a slight chance that a delayed 1053 * write will occur while sleeping just above, so 1054 * check for it. Note that vfs_bio_awrite expects 1055 * buffers to reside on a queue, while bwrite and 1056 * brelse do not. 1057 */ 1058 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1059 (flags & V_SAVE)) { 1060 1061 if (bp->b_vp == vp) { 1062 if (bp->b_flags & B_CLUSTEROK) { 1063 vfs_bio_awrite(bp); 1064 } else { 1065 bremfree(bp); 1066 bp->b_flags |= B_ASYNC; 1067 bwrite(bp); 1068 } 1069 } else { 1070 bremfree(bp); 1071 (void) bwrite(bp); 1072 } 1073 goto done; 1074 } 1075 bremfree(bp); 1076 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 1077 bp->b_flags &= ~B_ASYNC; 1078 brelse(bp); 1079 VI_LOCK(vp); 1080 } 1081 return (found); 1082done: 1083 VI_LOCK(vp); 1084 return (found); 1085} 1086 1087/* 1088 * Truncate a file's buffer and pages to a specified length. This 1089 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1090 * sync activity. 1091 */ 1092int 1093vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, off_t length, int blksize) 1094{ 1095 struct buf *bp, *nbp; 1096 int anyfreed; 1097 int trunclbn; 1098 struct bufobj *bo; 1099 1100 /* 1101 * Round up to the *next* lbn. 1102 */ 1103 trunclbn = (length + blksize - 1) / blksize; 1104 1105 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1106restart: 1107 VI_LOCK(vp); 1108 bo = &vp->v_bufobj; 1109 anyfreed = 1; 1110 for (;anyfreed;) { 1111 anyfreed = 0; 1112 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 1113 if (bp->b_lblkno < trunclbn) 1114 continue; 1115 if (BUF_LOCK(bp, 1116 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1117 VI_MTX(vp)) == ENOLCK) 1118 goto restart; 1119 1120 bremfree(bp); 1121 bp->b_flags |= (B_INVAL | B_RELBUF); 1122 bp->b_flags &= ~B_ASYNC; 1123 brelse(bp); 1124 anyfreed = 1; 1125 1126 if (nbp != NULL && 1127 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1128 (nbp->b_vp != vp) || 1129 (nbp->b_flags & B_DELWRI))) { 1130 goto restart; 1131 } 1132 VI_LOCK(vp); 1133 } 1134 1135 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1136 if (bp->b_lblkno < trunclbn) 1137 continue; 1138 if (BUF_LOCK(bp, 1139 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1140 VI_MTX(vp)) == ENOLCK) 1141 goto restart; 1142 bremfree(bp); 1143 bp->b_flags |= (B_INVAL | B_RELBUF); 1144 bp->b_flags &= ~B_ASYNC; 1145 brelse(bp); 1146 anyfreed = 1; 1147 if (nbp != NULL && 1148 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1149 (nbp->b_vp != vp) || 1150 (nbp->b_flags & B_DELWRI) == 0)) { 1151 goto restart; 1152 } 1153 VI_LOCK(vp); 1154 } 1155 } 1156 1157 if (length > 0) { 1158restartsync: 1159 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1160 if (bp->b_lblkno > 0) 1161 continue; 1162 /* 1163 * Since we hold the vnode lock this should only 1164 * fail if we're racing with the buf daemon. 1165 */ 1166 if (BUF_LOCK(bp, 1167 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1168 VI_MTX(vp)) == ENOLCK) { 1169 goto restart; 1170 } 1171 KASSERT((bp->b_flags & B_DELWRI), 1172 ("buf(%p) on dirty queue without DELWRI", bp)); 1173 1174 bremfree(bp); 1175 bawrite(bp); 1176 VI_LOCK(vp); 1177 goto restartsync; 1178 } 1179 } 1180 1181 bufobj_wwait(bo, 0, 0); 1182 VI_UNLOCK(vp); 1183 vnode_pager_setsize(vp, length); 1184 1185 return (0); 1186} 1187 1188/* 1189 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1190 * a vnode. 1191 * 1192 * NOTE: We have to deal with the special case of a background bitmap 1193 * buffer, a situation where two buffers will have the same logical 1194 * block offset. We want (1) only the foreground buffer to be accessed 1195 * in a lookup and (2) must differentiate between the foreground and 1196 * background buffer in the splay tree algorithm because the splay 1197 * tree cannot normally handle multiple entities with the same 'index'. 1198 * We accomplish this by adding differentiating flags to the splay tree's 1199 * numerical domain. 1200 */ 1201static 1202struct buf * 1203buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1204{ 1205 struct buf dummy; 1206 struct buf *lefttreemax, *righttreemin, *y; 1207 1208 if (root == NULL) 1209 return (NULL); 1210 lefttreemax = righttreemin = &dummy; 1211 for (;;) { 1212 if (lblkno < root->b_lblkno || 1213 (lblkno == root->b_lblkno && 1214 (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1215 if ((y = root->b_left) == NULL) 1216 break; 1217 if (lblkno < y->b_lblkno) { 1218 /* Rotate right. */ 1219 root->b_left = y->b_right; 1220 y->b_right = root; 1221 root = y; 1222 if ((y = root->b_left) == NULL) 1223 break; 1224 } 1225 /* Link into the new root's right tree. */ 1226 righttreemin->b_left = root; 1227 righttreemin = root; 1228 } else if (lblkno > root->b_lblkno || 1229 (lblkno == root->b_lblkno && 1230 (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { 1231 if ((y = root->b_right) == NULL) 1232 break; 1233 if (lblkno > y->b_lblkno) { 1234 /* Rotate left. */ 1235 root->b_right = y->b_left; 1236 y->b_left = root; 1237 root = y; 1238 if ((y = root->b_right) == NULL) 1239 break; 1240 } 1241 /* Link into the new root's left tree. */ 1242 lefttreemax->b_right = root; 1243 lefttreemax = root; 1244 } else { 1245 break; 1246 } 1247 root = y; 1248 } 1249 /* Assemble the new root. */ 1250 lefttreemax->b_right = root->b_left; 1251 righttreemin->b_left = root->b_right; 1252 root->b_left = dummy.b_right; 1253 root->b_right = dummy.b_left; 1254 return (root); 1255} 1256 1257static void 1258buf_vlist_remove(struct buf *bp) 1259{ 1260 struct buf *root; 1261 struct bufv *bv; 1262 1263 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1264 ASSERT_BO_LOCKED(bp->b_bufobj); 1265 if (bp->b_xflags & BX_VNDIRTY) 1266 bv = &bp->b_bufobj->bo_dirty; 1267 else 1268 bv = &bp->b_bufobj->bo_clean; 1269 if (bp != bv->bv_root) { 1270 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1271 KASSERT(root == bp, ("splay lookup failed in remove")); 1272 } 1273 if (bp->b_left == NULL) { 1274 root = bp->b_right; 1275 } else { 1276 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1277 root->b_right = bp->b_right; 1278 } 1279 bv->bv_root = root; 1280 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 1281 bv->bv_cnt--; 1282 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1283} 1284 1285/* 1286 * Add the buffer to the sorted clean or dirty block list using a 1287 * splay tree algorithm. 1288 * 1289 * NOTE: xflags is passed as a constant, optimizing this inline function! 1290 */ 1291static void 1292buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 1293{ 1294 struct buf *root; 1295 struct bufv *bv; 1296 1297 ASSERT_BO_LOCKED(bo); 1298 bp->b_xflags |= xflags; 1299 if (xflags & BX_VNDIRTY) 1300 bv = &bo->bo_dirty; 1301 else 1302 bv = &bo->bo_clean; 1303 1304 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1305 if (root == NULL) { 1306 bp->b_left = NULL; 1307 bp->b_right = NULL; 1308 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 1309 } else if (bp->b_lblkno < root->b_lblkno || 1310 (bp->b_lblkno == root->b_lblkno && 1311 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1312 bp->b_left = root->b_left; 1313 bp->b_right = root; 1314 root->b_left = NULL; 1315 TAILQ_INSERT_BEFORE(root, bp, b_bobufs); 1316 } else { 1317 bp->b_right = root->b_right; 1318 bp->b_left = root; 1319 root->b_right = NULL; 1320 TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs); 1321 } 1322 bv->bv_cnt++; 1323 bv->bv_root = bp; 1324} 1325 1326/* 1327 * Lookup a buffer using the splay tree. Note that we specifically avoid 1328 * shadow buffers used in background bitmap writes. 1329 * 1330 * This code isn't quite efficient as it could be because we are maintaining 1331 * two sorted lists and do not know which list the block resides in. 1332 * 1333 * During a "make buildworld" the desired buffer is found at one of 1334 * the roots more than 60% of the time. Thus, checking both roots 1335 * before performing either splay eliminates unnecessary splays on the 1336 * first tree splayed. 1337 */ 1338struct buf * 1339gbincore(struct bufobj *bo, daddr_t lblkno) 1340{ 1341 struct buf *bp; 1342 1343 GIANT_REQUIRED; 1344 1345 ASSERT_BO_LOCKED(bo); 1346 if ((bp = bo->bo_clean.bv_root) != NULL && 1347 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1348 return (bp); 1349 if ((bp = bo->bo_dirty.bv_root) != NULL && 1350 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1351 return (bp); 1352 if ((bp = bo->bo_clean.bv_root) != NULL) { 1353 bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp); 1354 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1355 return (bp); 1356 } 1357 if ((bp = bo->bo_dirty.bv_root) != NULL) { 1358 bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp); 1359 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1360 return (bp); 1361 } 1362 return (NULL); 1363} 1364 1365/* 1366 * Associate a buffer with a vnode. 1367 */ 1368void 1369bgetvp(struct vnode *vp, struct buf *bp) 1370{ 1371 1372 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 1373 1374 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1375 ("bgetvp: bp already attached! %p", bp)); 1376 1377 ASSERT_VI_LOCKED(vp, "bgetvp"); 1378 vholdl(vp); 1379 bp->b_vp = vp; 1380 bp->b_bufobj = &vp->v_bufobj; 1381 bp->b_dev = vn_todev(vp); 1382 /* 1383 * Insert onto list for new vnode. 1384 */ 1385 buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN); 1386} 1387 1388/* 1389 * Disassociate a buffer from a vnode. 1390 */ 1391void 1392brelvp(struct buf *bp) 1393{ 1394 struct bufobj *bo; 1395 struct vnode *vp; 1396 1397 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1398 1399 /* 1400 * Delete from old vnode list, if on one. 1401 */ 1402 vp = bp->b_vp; /* XXX */ 1403 bo = bp->b_bufobj; 1404 BO_LOCK(bo); 1405 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1406 buf_vlist_remove(bp); 1407 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1408 bo->bo_flag &= ~BO_ONWORKLST; 1409 mtx_lock(&sync_mtx); 1410 LIST_REMOVE(bo, bo_synclist); 1411 syncer_worklist_len--; 1412 mtx_unlock(&sync_mtx); 1413 } 1414 vdropl(vp); 1415 bp->b_vp = NULL; 1416 bp->b_bufobj = NULL; 1417 BO_UNLOCK(bo); 1418} 1419 1420/* 1421 * Add an item to the syncer work queue. 1422 */ 1423static void 1424vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 1425{ 1426 int slot; 1427 1428 ASSERT_BO_LOCKED(bo); 1429 1430 mtx_lock(&sync_mtx); 1431 if (bo->bo_flag & BO_ONWORKLST) 1432 LIST_REMOVE(bo, bo_synclist); 1433 else { 1434 bo->bo_flag |= BO_ONWORKLST; 1435 syncer_worklist_len++; 1436 } 1437 1438 if (delay > syncer_maxdelay - 2) 1439 delay = syncer_maxdelay - 2; 1440 slot = (syncer_delayno + delay) & syncer_mask; 1441 1442 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 1443 mtx_unlock(&sync_mtx); 1444} 1445 1446static int 1447sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 1448{ 1449 int error, len; 1450 1451 mtx_lock(&sync_mtx); 1452 len = syncer_worklist_len - sync_vnode_count; 1453 mtx_unlock(&sync_mtx); 1454 error = SYSCTL_OUT(req, &len, sizeof(len)); 1455 return (error); 1456} 1457 1458SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0, 1459 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 1460 1461struct proc *updateproc; 1462static void sched_sync(void); 1463static struct kproc_desc up_kp = { 1464 "syncer", 1465 sched_sync, 1466 &updateproc 1467}; 1468SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 1469 1470/* 1471 * System filesystem synchronizer daemon. 1472 */ 1473static void 1474sched_sync(void) 1475{ 1476 struct synclist *next; 1477 struct synclist *slp; 1478 struct vnode *vp; 1479 struct bufobj *bo; 1480 struct mount *mp; 1481 long starttime; 1482 struct thread *td = FIRST_THREAD_IN_PROC(updateproc); 1483 static int dummychan; 1484 int last_work_seen; 1485 int net_worklist_len; 1486 int syncer_final_iter; 1487 int first_printf; 1488 1489 mtx_lock(&Giant); 1490 last_work_seen = 0; 1491 syncer_final_iter = 0; 1492 first_printf = 1; 1493 syncer_state = SYNCER_RUNNING; 1494 starttime = time_second; 1495 1496 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 1497 SHUTDOWN_PRI_LAST); 1498 1499 for (;;) { 1500 mtx_lock(&sync_mtx); 1501 if (syncer_state == SYNCER_FINAL_DELAY && 1502 syncer_final_iter == 0) { 1503 mtx_unlock(&sync_mtx); 1504 kthread_suspend_check(td->td_proc); 1505 mtx_lock(&sync_mtx); 1506 } 1507 net_worklist_len = syncer_worklist_len - sync_vnode_count; 1508 if (syncer_state != SYNCER_RUNNING && 1509 starttime != time_second) { 1510 if (first_printf) { 1511 printf("\nSyncing disks, vnodes remaining..."); 1512 first_printf = 0; 1513 } 1514 printf("%d ", net_worklist_len); 1515 } 1516 starttime = time_second; 1517 1518 /* 1519 * Push files whose dirty time has expired. Be careful 1520 * of interrupt race on slp queue. 1521 * 1522 * Skip over empty worklist slots when shutting down. 1523 */ 1524 do { 1525 slp = &syncer_workitem_pending[syncer_delayno]; 1526 syncer_delayno += 1; 1527 if (syncer_delayno == syncer_maxdelay) 1528 syncer_delayno = 0; 1529 next = &syncer_workitem_pending[syncer_delayno]; 1530 /* 1531 * If the worklist has wrapped since the 1532 * it was emptied of all but syncer vnodes, 1533 * switch to the FINAL_DELAY state and run 1534 * for one more second. 1535 */ 1536 if (syncer_state == SYNCER_SHUTTING_DOWN && 1537 net_worklist_len == 0 && 1538 last_work_seen == syncer_delayno) { 1539 syncer_state = SYNCER_FINAL_DELAY; 1540 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 1541 } 1542 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 1543 syncer_worklist_len > 0); 1544 1545 /* 1546 * Keep track of the last time there was anything 1547 * on the worklist other than syncer vnodes. 1548 * Return to the SHUTTING_DOWN state if any 1549 * new work appears. 1550 */ 1551 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 1552 last_work_seen = syncer_delayno; 1553 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 1554 syncer_state = SYNCER_SHUTTING_DOWN; 1555 while ((bo = LIST_FIRST(slp)) != NULL) { 1556 vp = bo->__bo_vnode; /* XXX */ 1557 if (VOP_ISLOCKED(vp, NULL) != 0 || 1558 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1559 LIST_REMOVE(bo, bo_synclist); 1560 LIST_INSERT_HEAD(next, bo, bo_synclist); 1561 continue; 1562 } 1563 if (VI_TRYLOCK(vp) == 0) { 1564 LIST_REMOVE(bo, bo_synclist); 1565 LIST_INSERT_HEAD(next, bo, bo_synclist); 1566 vn_finished_write(mp); 1567 continue; 1568 } 1569 /* 1570 * We use vhold in case the vnode does not 1571 * successfully sync. vhold prevents the vnode from 1572 * going away when we unlock the sync_mtx so that 1573 * we can acquire the vnode interlock. 1574 */ 1575 vholdl(vp); 1576 mtx_unlock(&sync_mtx); 1577 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, td); 1578 (void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td); 1579 VOP_UNLOCK(vp, 0, td); 1580 vn_finished_write(mp); 1581 VI_LOCK(vp); 1582 if ((bo->bo_flag & BO_ONWORKLST) != 0) { 1583 /* 1584 * Put us back on the worklist. The worklist 1585 * routine will remove us from our current 1586 * position and then add us back in at a later 1587 * position. 1588 */ 1589 vn_syncer_add_to_worklist(bo, syncdelay); 1590 } 1591 vdropl(vp); 1592 VI_UNLOCK(vp); 1593 mtx_lock(&sync_mtx); 1594 } 1595 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 1596 syncer_final_iter--; 1597 mtx_unlock(&sync_mtx); 1598 1599 /* 1600 * Do soft update processing. 1601 */ 1602 if (softdep_process_worklist_hook != NULL) 1603 (*softdep_process_worklist_hook)(NULL); 1604 1605 /* 1606 * The variable rushjob allows the kernel to speed up the 1607 * processing of the filesystem syncer process. A rushjob 1608 * value of N tells the filesystem syncer to process the next 1609 * N seconds worth of work on its queue ASAP. Currently rushjob 1610 * is used by the soft update code to speed up the filesystem 1611 * syncer process when the incore state is getting so far 1612 * ahead of the disk that the kernel memory pool is being 1613 * threatened with exhaustion. 1614 */ 1615 mtx_lock(&sync_mtx); 1616 if (rushjob > 0) { 1617 rushjob -= 1; 1618 mtx_unlock(&sync_mtx); 1619 continue; 1620 } 1621 mtx_unlock(&sync_mtx); 1622 /* 1623 * Just sleep for a short period if time between 1624 * iterations when shutting down to allow some I/O 1625 * to happen. 1626 * 1627 * If it has taken us less than a second to process the 1628 * current work, then wait. Otherwise start right over 1629 * again. We can still lose time if any single round 1630 * takes more than two seconds, but it does not really 1631 * matter as we are just trying to generally pace the 1632 * filesystem activity. 1633 */ 1634 if (syncer_state != SYNCER_RUNNING) 1635 tsleep(&dummychan, PPAUSE, "syncfnl", 1636 hz / SYNCER_SHUTDOWN_SPEEDUP); 1637 else if (time_second == starttime) 1638 tsleep(&lbolt, PPAUSE, "syncer", 0); 1639 } 1640} 1641 1642/* 1643 * Request the syncer daemon to speed up its work. 1644 * We never push it to speed up more than half of its 1645 * normal turn time, otherwise it could take over the cpu. 1646 */ 1647int 1648speedup_syncer() 1649{ 1650 struct thread *td; 1651 int ret = 0; 1652 1653 td = FIRST_THREAD_IN_PROC(updateproc); 1654 sleepq_remove(td, &lbolt); 1655 mtx_lock(&sync_mtx); 1656 if (rushjob < syncdelay / 2) { 1657 rushjob += 1; 1658 stat_rush_requests += 1; 1659 ret = 1; 1660 } 1661 mtx_unlock(&sync_mtx); 1662 return (ret); 1663} 1664 1665/* 1666 * Tell the syncer to speed up its work and run though its work 1667 * list several times, then tell it to shut down. 1668 */ 1669static void 1670syncer_shutdown(void *arg, int howto) 1671{ 1672 struct thread *td; 1673 1674 if (howto & RB_NOSYNC) 1675 return; 1676 td = FIRST_THREAD_IN_PROC(updateproc); 1677 sleepq_remove(td, &lbolt); 1678 mtx_lock(&sync_mtx); 1679 syncer_state = SYNCER_SHUTTING_DOWN; 1680 rushjob = 0; 1681 mtx_unlock(&sync_mtx); 1682 kproc_shutdown(arg, howto); 1683} 1684 1685/* 1686 * Associate a p-buffer with a vnode. 1687 * 1688 * Also sets B_PAGING flag to indicate that vnode is not fully associated 1689 * with the buffer. i.e. the bp has not been linked into the vnode or 1690 * ref-counted. 1691 */ 1692void 1693pbgetvp(vp, bp) 1694 register struct vnode *vp; 1695 register struct buf *bp; 1696{ 1697 1698 KASSERT(bp->b_vp == NULL, ("pbgetvp: not free")); 1699 1700 bp->b_vp = vp; 1701 bp->b_flags |= B_PAGING; 1702 bp->b_dev = vn_todev(vp); 1703 bp->b_bufobj = &vp->v_bufobj; 1704} 1705 1706/* 1707 * Disassociate a p-buffer from a vnode. 1708 */ 1709void 1710pbrelvp(bp) 1711 register struct buf *bp; 1712{ 1713 1714 KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL")); 1715 KASSERT(bp->b_bufobj != NULL, ("pbrelvp: NULL bufobj")); 1716 1717 /* XXX REMOVE ME */ 1718 BO_LOCK(bp->b_bufobj); 1719 if (TAILQ_NEXT(bp, b_bobufs) != NULL) { 1720 panic( 1721 "relpbuf(): b_vp was probably reassignbuf()d %p %x", 1722 bp, 1723 (int)bp->b_flags 1724 ); 1725 } 1726 BO_UNLOCK(bp->b_bufobj); 1727 bp->b_vp = NULL; 1728 bp->b_bufobj = NULL; 1729 bp->b_flags &= ~B_PAGING; 1730} 1731 1732/* 1733 * Reassign a buffer from one vnode to another. 1734 * Used to assign file specific control information 1735 * (indirect blocks) to the vnode to which they belong. 1736 */ 1737void 1738reassignbuf(struct buf *bp) 1739{ 1740 struct vnode *vp; 1741 struct bufobj *bo; 1742 int delay; 1743 1744 vp = bp->b_vp; 1745 bo = bp->b_bufobj; 1746 ++reassignbufcalls; 1747 1748 /* 1749 * B_PAGING flagged buffers cannot be reassigned because their vp 1750 * is not fully linked in. 1751 */ 1752 if (bp->b_flags & B_PAGING) 1753 panic("cannot reassign paging buffer"); 1754 1755 /* 1756 * Delete from old vnode list, if on one. 1757 */ 1758 VI_LOCK(vp); 1759 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1760 buf_vlist_remove(bp); 1761 /* 1762 * If dirty, put on list of dirty buffers; otherwise insert onto list 1763 * of clean buffers. 1764 */ 1765 if (bp->b_flags & B_DELWRI) { 1766 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 1767 switch (vp->v_type) { 1768 case VDIR: 1769 delay = dirdelay; 1770 break; 1771 case VCHR: 1772 delay = metadelay; 1773 break; 1774 default: 1775 delay = filedelay; 1776 } 1777 vn_syncer_add_to_worklist(bo, delay); 1778 } 1779 buf_vlist_add(bp, bo, BX_VNDIRTY); 1780 } else { 1781 buf_vlist_add(bp, bo, BX_VNCLEAN); 1782 1783 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1784 mtx_lock(&sync_mtx); 1785 LIST_REMOVE(bo, bo_synclist); 1786 syncer_worklist_len--; 1787 mtx_unlock(&sync_mtx); 1788 bo->bo_flag &= ~BO_ONWORKLST; 1789 } 1790 } 1791 VI_UNLOCK(vp); 1792} 1793 1794/* 1795 * Create a vnode for a device. 1796 * Used for mounting the root filesystem. 1797 */ 1798int 1799bdevvp(dev, vpp) 1800 struct cdev *dev; 1801 struct vnode **vpp; 1802{ 1803 register struct vnode *vp; 1804 struct vnode *nvp; 1805 int error; 1806 1807 if (dev == NULL) { 1808 *vpp = NULLVP; 1809 return (ENXIO); 1810 } 1811 if (vfinddev(dev, vpp)) 1812 return (0); 1813 1814 error = getnewvnode("none", (struct mount *)0, devfs_specop_p, &nvp); 1815 if (error) { 1816 *vpp = NULLVP; 1817 return (error); 1818 } 1819 vp = nvp; 1820 vp->v_type = VCHR; 1821 vp->v_bufobj.bo_bsize = DEV_BSIZE; 1822 addalias(vp, dev); 1823 *vpp = vp; 1824 return (0); 1825} 1826 1827static void 1828v_incr_usecount(struct vnode *vp, int delta) 1829{ 1830 1831 vp->v_usecount += delta; 1832 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1833 dev_lock(); 1834 vp->v_rdev->si_usecount += delta; 1835 dev_unlock(); 1836 } 1837} 1838 1839/* 1840 * Add vnode to the alias list hung off the struct cdev *. 1841 * 1842 * The reason for this gunk is that multiple vnodes can reference 1843 * the same physical device, so checking vp->v_usecount to see 1844 * how many users there are is inadequate; the v_usecount for 1845 * the vnodes need to be accumulated. vcount() does that. 1846 */ 1847struct vnode * 1848addaliasu(nvp, nvp_rdev) 1849 struct vnode *nvp; 1850 dev_t nvp_rdev; 1851{ 1852 struct vnode *ovp; 1853 vop_t **ops; 1854 struct cdev *dev; 1855 1856 if (nvp->v_type == VBLK) 1857 return (nvp); 1858 if (nvp->v_type != VCHR) 1859 panic("addaliasu on non-special vnode"); 1860 dev = findcdev(nvp_rdev); 1861 if (dev == NULL) 1862 return (nvp); 1863 /* 1864 * Check to see if we have a bdevvp vnode with no associated 1865 * filesystem. If so, we want to associate the filesystem of 1866 * the new newly instigated vnode with the bdevvp vnode and 1867 * discard the newly created vnode rather than leaving the 1868 * bdevvp vnode lying around with no associated filesystem. 1869 */ 1870 if (vfinddev(dev, &ovp) == 0 || ovp->v_data != NULL) { 1871 addalias(nvp, dev); 1872 return (nvp); 1873 } 1874 /* 1875 * Discard unneeded vnode, but save its node specific data. 1876 * Note that if there is a lock, it is carried over in the 1877 * node specific data to the replacement vnode. 1878 */ 1879 vref(ovp); 1880 ovp->v_data = nvp->v_data; 1881 ovp->v_tag = nvp->v_tag; 1882 nvp->v_data = NULL; 1883 lockdestroy(ovp->v_vnlock); 1884 lockinit(ovp->v_vnlock, PVFS, nvp->v_vnlock->lk_wmesg, 1885 nvp->v_vnlock->lk_timo, nvp->v_vnlock->lk_flags & LK_EXTFLG_MASK); 1886 ops = ovp->v_op; 1887 ovp->v_op = nvp->v_op; 1888 if (VOP_ISLOCKED(nvp, curthread)) { 1889 VOP_UNLOCK(nvp, 0, curthread); 1890 vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread); 1891 } 1892 nvp->v_op = ops; 1893 delmntque(ovp); 1894 insmntque(ovp, nvp->v_mount); 1895 vrele(nvp); 1896 vgone(nvp); 1897 return (ovp); 1898} 1899 1900/* This is a local helper function that do the same as addaliasu, but for a 1901 * struct cdev *instead of an dev_t. */ 1902static void 1903addalias(nvp, dev) 1904 struct vnode *nvp; 1905 struct cdev *dev; 1906{ 1907 1908 KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode")); 1909 VI_LOCK(nvp); 1910 dev_lock(); 1911 dev->si_refcount++; 1912 nvp->v_rdev = dev; 1913 SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext); 1914 dev->si_usecount += nvp->v_usecount; 1915 dev_unlock(); 1916 VI_UNLOCK(nvp); 1917} 1918 1919/* 1920 * Grab a particular vnode from the free list, increment its 1921 * reference count and lock it. The vnode lock bit is set if the 1922 * vnode is being eliminated in vgone. The process is awakened 1923 * when the transition is completed, and an error returned to 1924 * indicate that the vnode is no longer usable (possibly having 1925 * been changed to a new filesystem type). 1926 */ 1927int 1928vget(vp, flags, td) 1929 register struct vnode *vp; 1930 int flags; 1931 struct thread *td; 1932{ 1933 int error; 1934 1935 /* 1936 * If the vnode is in the process of being cleaned out for 1937 * another use, we wait for the cleaning to finish and then 1938 * return failure. Cleaning is determined by checking that 1939 * the VI_XLOCK flag is set. 1940 */ 1941 if ((flags & LK_INTERLOCK) == 0) 1942 VI_LOCK(vp); 1943 if (vp->v_iflag & VI_XLOCK && vp->v_vxthread != curthread) { 1944 if ((flags & LK_NOWAIT) == 0) { 1945 vp->v_iflag |= VI_XWANT; 1946 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0); 1947 return (ENOENT); 1948 } 1949 VI_UNLOCK(vp); 1950 return (EBUSY); 1951 } 1952 1953 v_incr_usecount(vp, 1); 1954 1955 if (VSHOULDBUSY(vp)) 1956 vbusy(vp); 1957 if (flags & LK_TYPE_MASK) { 1958 if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { 1959 /* 1960 * must expand vrele here because we do not want 1961 * to call VOP_INACTIVE if the reference count 1962 * drops back to zero since it was never really 1963 * active. We must remove it from the free list 1964 * before sleeping so that multiple processes do 1965 * not try to recycle it. 1966 */ 1967 VI_LOCK(vp); 1968 v_incr_usecount(vp, -1); 1969 if (VSHOULDFREE(vp)) 1970 vfree(vp); 1971 else 1972 vlruvp(vp); 1973 VI_UNLOCK(vp); 1974 } 1975 return (error); 1976 } 1977 VI_UNLOCK(vp); 1978 return (0); 1979} 1980 1981/* 1982 * Increase the reference count of a vnode. 1983 */ 1984void 1985vref(struct vnode *vp) 1986{ 1987 1988 VI_LOCK(vp); 1989 v_incr_usecount(vp, 1); 1990 VI_UNLOCK(vp); 1991} 1992 1993/* 1994 * Return reference count of a vnode. 1995 * 1996 * The results of this call are only guaranteed when some mechanism other 1997 * than the VI lock is used to stop other processes from gaining references 1998 * to the vnode. This may be the case if the caller holds the only reference. 1999 * This is also useful when stale data is acceptable as race conditions may 2000 * be accounted for by some other means. 2001 */ 2002int 2003vrefcnt(struct vnode *vp) 2004{ 2005 int usecnt; 2006 2007 VI_LOCK(vp); 2008 usecnt = vp->v_usecount; 2009 VI_UNLOCK(vp); 2010 2011 return (usecnt); 2012} 2013 2014 2015/* 2016 * Vnode put/release. 2017 * If count drops to zero, call inactive routine and return to freelist. 2018 */ 2019void 2020vrele(vp) 2021 struct vnode *vp; 2022{ 2023 struct thread *td = curthread; /* XXX */ 2024 2025 GIANT_REQUIRED; 2026 2027 KASSERT(vp != NULL, ("vrele: null vp")); 2028 2029 VI_LOCK(vp); 2030 2031 /* Skip this v_writecount check if we're going to panic below. */ 2032 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2033 ("vrele: missed vn_close")); 2034 2035 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2036 vp->v_usecount == 1)) { 2037 v_incr_usecount(vp, -1); 2038 VI_UNLOCK(vp); 2039 2040 return; 2041 } 2042 2043 if (vp->v_usecount == 1) { 2044 v_incr_usecount(vp, -1); 2045 /* 2046 * We must call VOP_INACTIVE with the node locked. Mark 2047 * as VI_DOINGINACT to avoid recursion. 2048 */ 2049 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) { 2050 VI_LOCK(vp); 2051 vp->v_iflag |= VI_DOINGINACT; 2052 VI_UNLOCK(vp); 2053 VOP_INACTIVE(vp, td); 2054 VI_LOCK(vp); 2055 KASSERT(vp->v_iflag & VI_DOINGINACT, 2056 ("vrele: lost VI_DOINGINACT")); 2057 vp->v_iflag &= ~VI_DOINGINACT; 2058 } else 2059 VI_LOCK(vp); 2060 if (VSHOULDFREE(vp)) 2061 vfree(vp); 2062 else 2063 vlruvp(vp); 2064 VI_UNLOCK(vp); 2065 2066 } else { 2067#ifdef DIAGNOSTIC 2068 vprint("vrele: negative ref count", vp); 2069#endif 2070 VI_UNLOCK(vp); 2071 panic("vrele: negative ref cnt"); 2072 } 2073} 2074 2075/* 2076 * Release an already locked vnode. This give the same effects as 2077 * unlock+vrele(), but takes less time and avoids releasing and 2078 * re-aquiring the lock (as vrele() aquires the lock internally.) 2079 */ 2080void 2081vput(vp) 2082 struct vnode *vp; 2083{ 2084 struct thread *td = curthread; /* XXX */ 2085 2086 GIANT_REQUIRED; 2087 2088 KASSERT(vp != NULL, ("vput: null vp")); 2089 VI_LOCK(vp); 2090 /* Skip this v_writecount check if we're going to panic below. */ 2091 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2092 ("vput: missed vn_close")); 2093 2094 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2095 vp->v_usecount == 1)) { 2096 v_incr_usecount(vp, -1); 2097 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2098 return; 2099 } 2100 2101 if (vp->v_usecount == 1) { 2102 v_incr_usecount(vp, -1); 2103 /* 2104 * We must call VOP_INACTIVE with the node locked, so 2105 * we just need to release the vnode mutex. Mark as 2106 * as VI_DOINGINACT to avoid recursion. 2107 */ 2108 vp->v_iflag |= VI_DOINGINACT; 2109 VI_UNLOCK(vp); 2110 VOP_INACTIVE(vp, td); 2111 VI_LOCK(vp); 2112 KASSERT(vp->v_iflag & VI_DOINGINACT, 2113 ("vput: lost VI_DOINGINACT")); 2114 vp->v_iflag &= ~VI_DOINGINACT; 2115 if (VSHOULDFREE(vp)) 2116 vfree(vp); 2117 else 2118 vlruvp(vp); 2119 VI_UNLOCK(vp); 2120 2121 } else { 2122#ifdef DIAGNOSTIC 2123 vprint("vput: negative ref count", vp); 2124#endif 2125 panic("vput: negative ref cnt"); 2126 } 2127} 2128 2129/* 2130 * Somebody doesn't want the vnode recycled. 2131 */ 2132void 2133vhold(struct vnode *vp) 2134{ 2135 2136 VI_LOCK(vp); 2137 vholdl(vp); 2138 VI_UNLOCK(vp); 2139} 2140 2141void 2142vholdl(struct vnode *vp) 2143{ 2144 2145 vp->v_holdcnt++; 2146 if (VSHOULDBUSY(vp)) 2147 vbusy(vp); 2148} 2149 2150/* 2151 * Note that there is one less who cares about this vnode. vdrop() is the 2152 * opposite of vhold(). 2153 */ 2154void 2155vdrop(struct vnode *vp) 2156{ 2157 2158 VI_LOCK(vp); 2159 vdropl(vp); 2160 VI_UNLOCK(vp); 2161} 2162 2163void 2164vdropl(vp) 2165 register struct vnode *vp; 2166{ 2167 2168 if (vp->v_holdcnt <= 0) 2169 panic("vdrop: holdcnt"); 2170 vp->v_holdcnt--; 2171 if (VSHOULDFREE(vp)) 2172 vfree(vp); 2173 else 2174 vlruvp(vp); 2175} 2176 2177/* 2178 * Remove any vnodes in the vnode table belonging to mount point mp. 2179 * 2180 * If FORCECLOSE is not specified, there should not be any active ones, 2181 * return error if any are found (nb: this is a user error, not a 2182 * system error). If FORCECLOSE is specified, detach any active vnodes 2183 * that are found. 2184 * 2185 * If WRITECLOSE is set, only flush out regular file vnodes open for 2186 * writing. 2187 * 2188 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2189 * 2190 * `rootrefs' specifies the base reference count for the root vnode 2191 * of this filesystem. The root vnode is considered busy if its 2192 * v_usecount exceeds this value. On a successful return, vflush(, td) 2193 * will call vrele() on the root vnode exactly rootrefs times. 2194 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2195 * be zero. 2196 */ 2197#ifdef DIAGNOSTIC 2198static int busyprt = 0; /* print out busy vnodes */ 2199SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 2200#endif 2201 2202int 2203vflush(mp, rootrefs, flags, td) 2204 struct mount *mp; 2205 int rootrefs; 2206 int flags; 2207 struct thread *td; 2208{ 2209 struct vnode *vp, *nvp, *rootvp = NULL; 2210 struct vattr vattr; 2211 int busy = 0, error; 2212 2213 if (rootrefs > 0) { 2214 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2215 ("vflush: bad args")); 2216 /* 2217 * Get the filesystem root vnode. We can vput() it 2218 * immediately, since with rootrefs > 0, it won't go away. 2219 */ 2220 if ((error = VFS_ROOT(mp, &rootvp, td)) != 0) 2221 return (error); 2222 vput(rootvp); 2223 2224 } 2225 MNT_ILOCK(mp); 2226loop: 2227 MNT_VNODE_FOREACH(vp, mp, nvp) { 2228 2229 VI_LOCK(vp); 2230 MNT_IUNLOCK(mp); 2231 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td); 2232 if (error) { 2233 MNT_ILOCK(mp); 2234 goto loop; 2235 } 2236 /* 2237 * Skip over a vnodes marked VV_SYSTEM. 2238 */ 2239 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2240 VOP_UNLOCK(vp, 0, td); 2241 MNT_ILOCK(mp); 2242 continue; 2243 } 2244 /* 2245 * If WRITECLOSE is set, flush out unlinked but still open 2246 * files (even if open only for reading) and regular file 2247 * vnodes open for writing. 2248 */ 2249 if (flags & WRITECLOSE) { 2250 error = VOP_GETATTR(vp, &vattr, td->td_ucred, td); 2251 VI_LOCK(vp); 2252 2253 if ((vp->v_type == VNON || 2254 (error == 0 && vattr.va_nlink > 0)) && 2255 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2256 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2257 MNT_ILOCK(mp); 2258 continue; 2259 } 2260 } else 2261 VI_LOCK(vp); 2262 2263 VOP_UNLOCK(vp, 0, td); 2264 2265 /* 2266 * With v_usecount == 0, all we need to do is clear out the 2267 * vnode data structures and we are done. 2268 */ 2269 if (vp->v_usecount == 0) { 2270 vgonel(vp, td); 2271 MNT_ILOCK(mp); 2272 continue; 2273 } 2274 2275 /* 2276 * If FORCECLOSE is set, forcibly close the vnode. For block 2277 * or character devices, revert to an anonymous device. For 2278 * all other files, just kill them. 2279 */ 2280 if (flags & FORCECLOSE) { 2281 if (vp->v_type != VCHR) 2282 vgonel(vp, td); 2283 else 2284 vgonechrl(vp, td); 2285 MNT_ILOCK(mp); 2286 continue; 2287 } 2288#ifdef DIAGNOSTIC 2289 if (busyprt) 2290 vprint("vflush: busy vnode", vp); 2291#endif 2292 VI_UNLOCK(vp); 2293 MNT_ILOCK(mp); 2294 busy++; 2295 } 2296 MNT_IUNLOCK(mp); 2297 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2298 /* 2299 * If just the root vnode is busy, and if its refcount 2300 * is equal to `rootrefs', then go ahead and kill it. 2301 */ 2302 VI_LOCK(rootvp); 2303 KASSERT(busy > 0, ("vflush: not busy")); 2304 KASSERT(rootvp->v_usecount >= rootrefs, 2305 ("vflush: usecount %d < rootrefs %d", 2306 rootvp->v_usecount, rootrefs)); 2307 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2308 vgonel(rootvp, td); 2309 busy = 0; 2310 } else 2311 VI_UNLOCK(rootvp); 2312 } 2313 if (busy) 2314 return (EBUSY); 2315 for (; rootrefs > 0; rootrefs--) 2316 vrele(rootvp); 2317 return (0); 2318} 2319 2320/* 2321 * This moves a now (likely recyclable) vnode to the end of the 2322 * mountlist. XXX However, it is temporarily disabled until we 2323 * can clean up ffs_sync() and friends, which have loop restart 2324 * conditions which this code causes to operate O(N^2). 2325 */ 2326static void 2327vlruvp(struct vnode *vp) 2328{ 2329#if 0 2330 struct mount *mp; 2331 2332 if ((mp = vp->v_mount) != NULL) { 2333 MNT_ILOCK(mp); 2334 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2335 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2336 MNT_IUNLOCK(mp); 2337 } 2338#endif 2339} 2340 2341static void 2342vx_lock(struct vnode *vp) 2343{ 2344 2345 ASSERT_VI_LOCKED(vp, "vx_lock"); 2346 2347 /* 2348 * Prevent the vnode from being recycled or brought into use while we 2349 * clean it out. 2350 */ 2351 if (vp->v_iflag & VI_XLOCK) 2352 panic("vclean: deadlock"); 2353 vp->v_iflag |= VI_XLOCK; 2354 vp->v_vxthread = curthread; 2355} 2356 2357static void 2358vx_unlock(struct vnode *vp) 2359{ 2360 ASSERT_VI_LOCKED(vp, "vx_unlock"); 2361 vp->v_iflag &= ~VI_XLOCK; 2362 vp->v_vxthread = NULL; 2363 if (vp->v_iflag & VI_XWANT) { 2364 vp->v_iflag &= ~VI_XWANT; 2365 wakeup(vp); 2366 } 2367} 2368 2369/* 2370 * Disassociate the underlying filesystem from a vnode. 2371 */ 2372static void 2373vclean(vp, flags, td) 2374 struct vnode *vp; 2375 int flags; 2376 struct thread *td; 2377{ 2378 int active; 2379 2380 ASSERT_VI_LOCKED(vp, "vclean"); 2381 /* 2382 * Check to see if the vnode is in use. If so we have to reference it 2383 * before we clean it out so that its count cannot fall to zero and 2384 * generate a race against ourselves to recycle it. 2385 */ 2386 if ((active = vp->v_usecount)) 2387 v_incr_usecount(vp, 1); 2388 2389 /* 2390 * Even if the count is zero, the VOP_INACTIVE routine may still 2391 * have the object locked while it cleans it out. The VOP_LOCK 2392 * ensures that the VOP_INACTIVE routine is done with its work. 2393 * For active vnodes, it ensures that no other activity can 2394 * occur while the underlying object is being cleaned out. 2395 */ 2396 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2397 2398 /* 2399 * Clean out any buffers associated with the vnode. 2400 * If the flush fails, just toss the buffers. 2401 */ 2402 if (flags & DOCLOSE) { 2403 struct buf *bp; 2404 bp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd); 2405 if (bp != NULL) 2406 (void) vn_write_suspend_wait(vp, NULL, V_WAIT); 2407 if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0) 2408 vinvalbuf(vp, 0, NOCRED, td, 0, 0); 2409 } 2410 2411 VOP_DESTROYVOBJECT(vp); 2412 2413 /* 2414 * Any other processes trying to obtain this lock must first 2415 * wait for VXLOCK to clear, then call the new lock operation. 2416 */ 2417 VOP_UNLOCK(vp, 0, td); 2418 2419 /* 2420 * If purging an active vnode, it must be closed and 2421 * deactivated before being reclaimed. Note that the 2422 * VOP_INACTIVE will unlock the vnode. 2423 */ 2424 if (active) { 2425 if (flags & DOCLOSE) 2426 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2427 VI_LOCK(vp); 2428 if ((vp->v_iflag & VI_DOINGINACT) == 0) { 2429 vp->v_iflag |= VI_DOINGINACT; 2430 VI_UNLOCK(vp); 2431 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 2432 panic("vclean: cannot relock."); 2433 VOP_INACTIVE(vp, td); 2434 VI_LOCK(vp); 2435 KASSERT(vp->v_iflag & VI_DOINGINACT, 2436 ("vclean: lost VI_DOINGINACT")); 2437 vp->v_iflag &= ~VI_DOINGINACT; 2438 } 2439 VI_UNLOCK(vp); 2440 } 2441 /* 2442 * Reclaim the vnode. 2443 */ 2444 if (VOP_RECLAIM(vp, td)) 2445 panic("vclean: cannot reclaim"); 2446 2447 if (active) { 2448 /* 2449 * Inline copy of vrele() since VOP_INACTIVE 2450 * has already been called. 2451 */ 2452 VI_LOCK(vp); 2453 v_incr_usecount(vp, -1); 2454 if (vp->v_usecount <= 0) { 2455#ifdef INVARIANTS 2456 if (vp->v_usecount < 0 || vp->v_writecount != 0) { 2457 vprint("vclean: bad ref count", vp); 2458 panic("vclean: ref cnt"); 2459 } 2460#endif 2461 if (VSHOULDFREE(vp)) 2462 vfree(vp); 2463 } 2464 VI_UNLOCK(vp); 2465 } 2466 /* 2467 * Delete from old mount point vnode list. 2468 */ 2469 delmntque(vp); 2470 cache_purge(vp); 2471 VI_LOCK(vp); 2472 if (VSHOULDFREE(vp)) 2473 vfree(vp); 2474 2475 /* 2476 * Done with purge, reset to the standard lock and 2477 * notify sleepers of the grim news. 2478 */ 2479 vp->v_vnlock = &vp->v_lock; 2480 vp->v_op = dead_vnodeop_p; 2481 if (vp->v_pollinfo != NULL) 2482 vn_pollgone(vp); 2483 vp->v_tag = "none"; 2484} 2485 2486/* 2487 * Eliminate all activity associated with the requested vnode 2488 * and with all vnodes aliased to the requested vnode. 2489 */ 2490int 2491vop_revoke(ap) 2492 struct vop_revoke_args /* { 2493 struct vnode *a_vp; 2494 int a_flags; 2495 } */ *ap; 2496{ 2497 struct vnode *vp, *vq; 2498 struct cdev *dev; 2499 2500 KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke")); 2501 vp = ap->a_vp; 2502 KASSERT((vp->v_type == VCHR), ("vop_revoke: not VCHR")); 2503 2504 VI_LOCK(vp); 2505 /* 2506 * If a vgone (or vclean) is already in progress, 2507 * wait until it is done and return. 2508 */ 2509 if (vp->v_iflag & VI_XLOCK) { 2510 vp->v_iflag |= VI_XWANT; 2511 msleep(vp, VI_MTX(vp), PINOD | PDROP, 2512 "vop_revokeall", 0); 2513 return (0); 2514 } 2515 VI_UNLOCK(vp); 2516 dev = vp->v_rdev; 2517 for (;;) { 2518 dev_lock(); 2519 vq = SLIST_FIRST(&dev->si_hlist); 2520 dev_unlock(); 2521 if (vq == NULL) 2522 break; 2523 vgone(vq); 2524 } 2525 return (0); 2526} 2527 2528/* 2529 * Recycle an unused vnode to the front of the free list. 2530 * Release the passed interlock if the vnode will be recycled. 2531 */ 2532int 2533vrecycle(vp, inter_lkp, td) 2534 struct vnode *vp; 2535 struct mtx *inter_lkp; 2536 struct thread *td; 2537{ 2538 2539 VI_LOCK(vp); 2540 if (vp->v_usecount == 0) { 2541 if (inter_lkp) { 2542 mtx_unlock(inter_lkp); 2543 } 2544 vgonel(vp, td); 2545 return (1); 2546 } 2547 VI_UNLOCK(vp); 2548 return (0); 2549} 2550 2551/* 2552 * Eliminate all activity associated with a vnode 2553 * in preparation for reuse. 2554 */ 2555void 2556vgone(vp) 2557 register struct vnode *vp; 2558{ 2559 struct thread *td = curthread; /* XXX */ 2560 2561 VI_LOCK(vp); 2562 vgonel(vp, td); 2563} 2564 2565/* 2566 * Disassociate a character device from the its underlying filesystem and 2567 * attach it to spec. This is for use when the chr device is still active 2568 * and the filesystem is going away. 2569 */ 2570static void 2571vgonechrl(struct vnode *vp, struct thread *td) 2572{ 2573 ASSERT_VI_LOCKED(vp, "vgonechrl"); 2574 vx_lock(vp); 2575 /* 2576 * This is a custom version of vclean() which does not tearm down 2577 * the bufs or vm objects held by this vnode. This allows filesystems 2578 * to continue using devices which were discovered via another 2579 * filesystem that has been unmounted. 2580 */ 2581 if (vp->v_usecount != 0) { 2582 v_incr_usecount(vp, 1); 2583 /* 2584 * Ensure that no other activity can occur while the 2585 * underlying object is being cleaned out. 2586 */ 2587 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2588 /* 2589 * Any other processes trying to obtain this lock must first 2590 * wait for VXLOCK to clear, then call the new lock operation. 2591 */ 2592 VOP_UNLOCK(vp, 0, td); 2593 vp->v_vnlock = &vp->v_lock; 2594 vp->v_tag = "orphanchr"; 2595 vp->v_op = devfs_specop_p; 2596 delmntque(vp); 2597 cache_purge(vp); 2598 vrele(vp); 2599 VI_LOCK(vp); 2600 } else 2601 vclean(vp, 0, td); 2602 vp->v_op = devfs_specop_p; 2603 vx_unlock(vp); 2604 VI_UNLOCK(vp); 2605} 2606 2607/* 2608 * vgone, with the vp interlock held. 2609 */ 2610void 2611vgonel(vp, td) 2612 struct vnode *vp; 2613 struct thread *td; 2614{ 2615 /* 2616 * If a vgone (or vclean) is already in progress, 2617 * wait until it is done and return. 2618 */ 2619 ASSERT_VI_LOCKED(vp, "vgonel"); 2620 if (vp->v_iflag & VI_XLOCK) { 2621 vp->v_iflag |= VI_XWANT; 2622 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0); 2623 return; 2624 } 2625 vx_lock(vp); 2626 2627 /* 2628 * Clean out the filesystem specific data. 2629 */ 2630 vclean(vp, DOCLOSE, td); 2631 VI_UNLOCK(vp); 2632 2633 /* 2634 * If special device, remove it from special device alias list 2635 * if it is on one. 2636 */ 2637 VI_LOCK(vp); 2638 if (vp->v_type == VCHR && vp->v_rdev != NULL) 2639 dev_rel(vp); 2640 2641 /* 2642 * If it is on the freelist and not already at the head, 2643 * move it to the head of the list. The test of the 2644 * VDOOMED flag and the reference count of zero is because 2645 * it will be removed from the free list by getnewvnode, 2646 * but will not have its reference count incremented until 2647 * after calling vgone. If the reference count were 2648 * incremented first, vgone would (incorrectly) try to 2649 * close the previous instance of the underlying object. 2650 */ 2651 if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) { 2652 mtx_lock(&vnode_free_list_mtx); 2653 if (vp->v_iflag & VI_FREE) { 2654 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2655 } else { 2656 vp->v_iflag |= VI_FREE; 2657 freevnodes++; 2658 } 2659 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2660 mtx_unlock(&vnode_free_list_mtx); 2661 } 2662 2663 vp->v_type = VBAD; 2664 vx_unlock(vp); 2665 VI_UNLOCK(vp); 2666} 2667 2668/* 2669 * Lookup a vnode by device number. 2670 */ 2671int 2672vfinddev(dev, vpp) 2673 struct cdev *dev; 2674 struct vnode **vpp; 2675{ 2676 struct vnode *vp; 2677 2678 dev_lock(); 2679 SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) { 2680 *vpp = vp; 2681 dev_unlock(); 2682 return (1); 2683 } 2684 dev_unlock(); 2685 return (0); 2686} 2687 2688/* 2689 * Calculate the total number of references to a special device. 2690 */ 2691int 2692vcount(vp) 2693 struct vnode *vp; 2694{ 2695 int count; 2696 2697 dev_lock(); 2698 count = vp->v_rdev->si_usecount; 2699 dev_unlock(); 2700 return (count); 2701} 2702 2703/* 2704 * Same as above, but using the struct cdev *as argument 2705 */ 2706int 2707count_dev(dev) 2708 struct cdev *dev; 2709{ 2710 int count; 2711 2712 dev_lock(); 2713 count = dev->si_usecount; 2714 dev_unlock(); 2715 return(count); 2716} 2717 2718/* 2719 * Print out a description of a vnode. 2720 */ 2721static char *typename[] = 2722{"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 2723 2724void 2725vprint(label, vp) 2726 char *label; 2727 struct vnode *vp; 2728{ 2729 char buf[96]; 2730 2731 if (label != NULL) 2732 printf("%s: %p: ", label, (void *)vp); 2733 else 2734 printf("%p: ", (void *)vp); 2735 printf("tag %s, type %s, usecount %d, writecount %d, refcount %d,", 2736 vp->v_tag, typename[vp->v_type], vp->v_usecount, 2737 vp->v_writecount, vp->v_holdcnt); 2738 buf[0] = '\0'; 2739 if (vp->v_vflag & VV_ROOT) 2740 strcat(buf, "|VV_ROOT"); 2741 if (vp->v_vflag & VV_TEXT) 2742 strcat(buf, "|VV_TEXT"); 2743 if (vp->v_vflag & VV_SYSTEM) 2744 strcat(buf, "|VV_SYSTEM"); 2745 if (vp->v_iflag & VI_XLOCK) 2746 strcat(buf, "|VI_XLOCK"); 2747 if (vp->v_iflag & VI_XWANT) 2748 strcat(buf, "|VI_XWANT"); 2749 if (vp->v_iflag & VI_DOOMED) 2750 strcat(buf, "|VI_DOOMED"); 2751 if (vp->v_iflag & VI_FREE) 2752 strcat(buf, "|VI_FREE"); 2753 if (vp->v_vflag & VV_OBJBUF) 2754 strcat(buf, "|VV_OBJBUF"); 2755 if (buf[0] != '\0') 2756 printf(" flags (%s),", &buf[1]); 2757 lockmgr_printinfo(vp->v_vnlock); 2758 printf("\n"); 2759 if (vp->v_data != NULL) 2760 VOP_PRINT(vp); 2761} 2762 2763#ifdef DDB 2764#include <ddb/ddb.h> 2765/* 2766 * List all of the locked vnodes in the system. 2767 * Called when debugging the kernel. 2768 */ 2769DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 2770{ 2771 struct mount *mp, *nmp; 2772 struct vnode *vp; 2773 2774 /* 2775 * Note: because this is DDB, we can't obey the locking semantics 2776 * for these structures, which means we could catch an inconsistent 2777 * state and dereference a nasty pointer. Not much to be done 2778 * about that. 2779 */ 2780 printf("Locked vnodes\n"); 2781 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2782 nmp = TAILQ_NEXT(mp, mnt_list); 2783 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2784 if (VOP_ISLOCKED(vp, NULL)) 2785 vprint(NULL, vp); 2786 } 2787 nmp = TAILQ_NEXT(mp, mnt_list); 2788 } 2789} 2790#endif 2791 2792/* 2793 * Fill in a struct xvfsconf based on a struct vfsconf. 2794 */ 2795static void 2796vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) 2797{ 2798 2799 strcpy(xvfsp->vfc_name, vfsp->vfc_name); 2800 xvfsp->vfc_typenum = vfsp->vfc_typenum; 2801 xvfsp->vfc_refcount = vfsp->vfc_refcount; 2802 xvfsp->vfc_flags = vfsp->vfc_flags; 2803 /* 2804 * These are unused in userland, we keep them 2805 * to not break binary compatibility. 2806 */ 2807 xvfsp->vfc_vfsops = NULL; 2808 xvfsp->vfc_next = NULL; 2809} 2810 2811/* 2812 * Top level filesystem related information gathering. 2813 */ 2814static int 2815sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 2816{ 2817 struct vfsconf *vfsp; 2818 struct xvfsconf xvfsp; 2819 int error; 2820 2821 error = 0; 2822 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2823 vfsconf2x(vfsp, &xvfsp); 2824 error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp); 2825 if (error) 2826 break; 2827 } 2828 return (error); 2829} 2830 2831SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, 2832 "S,xvfsconf", "List of all configured filesystems"); 2833 2834#ifndef BURN_BRIDGES 2835static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2836 2837static int 2838vfs_sysctl(SYSCTL_HANDLER_ARGS) 2839{ 2840 int *name = (int *)arg1 - 1; /* XXX */ 2841 u_int namelen = arg2 + 1; /* XXX */ 2842 struct vfsconf *vfsp; 2843 struct xvfsconf xvfsp; 2844 2845 printf("WARNING: userland calling deprecated sysctl, " 2846 "please rebuild world\n"); 2847 2848#if 1 || defined(COMPAT_PRELITE2) 2849 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2850 if (namelen == 1) 2851 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2852#endif 2853 2854 switch (name[1]) { 2855 case VFS_MAXTYPENUM: 2856 if (namelen != 2) 2857 return (ENOTDIR); 2858 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2859 case VFS_CONF: 2860 if (namelen != 3) 2861 return (ENOTDIR); /* overloaded */ 2862 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) 2863 if (vfsp->vfc_typenum == name[2]) 2864 break; 2865 if (vfsp == NULL) 2866 return (EOPNOTSUPP); 2867 vfsconf2x(vfsp, &xvfsp); 2868 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 2869 } 2870 return (EOPNOTSUPP); 2871} 2872 2873SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl, 2874 "Generic filesystem"); 2875 2876#if 1 || defined(COMPAT_PRELITE2) 2877 2878static int 2879sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 2880{ 2881 int error; 2882 struct vfsconf *vfsp; 2883 struct ovfsconf ovfs; 2884 2885 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2886 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 2887 strcpy(ovfs.vfc_name, vfsp->vfc_name); 2888 ovfs.vfc_index = vfsp->vfc_typenum; 2889 ovfs.vfc_refcount = vfsp->vfc_refcount; 2890 ovfs.vfc_flags = vfsp->vfc_flags; 2891 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 2892 if (error) 2893 return error; 2894 } 2895 return 0; 2896} 2897 2898#endif /* 1 || COMPAT_PRELITE2 */ 2899#endif /* !BURN_BRIDGES */ 2900 2901#define KINFO_VNODESLOP 10 2902#ifdef notyet 2903/* 2904 * Dump vnode list (via sysctl). 2905 */ 2906/* ARGSUSED */ 2907static int 2908sysctl_vnode(SYSCTL_HANDLER_ARGS) 2909{ 2910 struct xvnode *xvn; 2911 struct thread *td = req->td; 2912 struct mount *mp; 2913 struct vnode *vp; 2914 int error, len, n; 2915 2916 /* 2917 * Stale numvnodes access is not fatal here. 2918 */ 2919 req->lock = 0; 2920 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 2921 if (!req->oldptr) 2922 /* Make an estimate */ 2923 return (SYSCTL_OUT(req, 0, len)); 2924 2925 error = sysctl_wire_old_buffer(req, 0); 2926 if (error != 0) 2927 return (error); 2928 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 2929 n = 0; 2930 mtx_lock(&mountlist_mtx); 2931 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 2932 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 2933 continue; 2934 MNT_ILOCK(mp); 2935 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2936 if (n == len) 2937 break; 2938 vref(vp); 2939 xvn[n].xv_size = sizeof *xvn; 2940 xvn[n].xv_vnode = vp; 2941#define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 2942 XV_COPY(usecount); 2943 XV_COPY(writecount); 2944 XV_COPY(holdcnt); 2945 XV_COPY(id); 2946 XV_COPY(mount); 2947 XV_COPY(numoutput); 2948 XV_COPY(type); 2949#undef XV_COPY 2950 xvn[n].xv_flag = vp->v_vflag; 2951 2952 switch (vp->v_type) { 2953 case VREG: 2954 case VDIR: 2955 case VLNK: 2956 xvn[n].xv_dev = vp->v_cachedfs; 2957 xvn[n].xv_ino = vp->v_cachedid; 2958 break; 2959 case VBLK: 2960 case VCHR: 2961 if (vp->v_rdev == NULL) { 2962 vrele(vp); 2963 continue; 2964 } 2965 xvn[n].xv_dev = dev2udev(vp->v_rdev); 2966 break; 2967 case VSOCK: 2968 xvn[n].xv_socket = vp->v_socket; 2969 break; 2970 case VFIFO: 2971 xvn[n].xv_fifo = vp->v_fifoinfo; 2972 break; 2973 case VNON: 2974 case VBAD: 2975 default: 2976 /* shouldn't happen? */ 2977 vrele(vp); 2978 continue; 2979 } 2980 vrele(vp); 2981 ++n; 2982 } 2983 MNT_IUNLOCK(mp); 2984 mtx_lock(&mountlist_mtx); 2985 vfs_unbusy(mp, td); 2986 if (n == len) 2987 break; 2988 } 2989 mtx_unlock(&mountlist_mtx); 2990 2991 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 2992 free(xvn, M_TEMP); 2993 return (error); 2994} 2995 2996SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 2997 0, 0, sysctl_vnode, "S,xvnode", ""); 2998#endif 2999 3000/* 3001 * Unmount all filesystems. The list is traversed in reverse order 3002 * of mounting to avoid dependencies. 3003 */ 3004void 3005vfs_unmountall() 3006{ 3007 struct mount *mp; 3008 struct thread *td; 3009 int error; 3010 3011 if (curthread != NULL) 3012 td = curthread; 3013 else 3014 td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */ 3015 /* 3016 * Since this only runs when rebooting, it is not interlocked. 3017 */ 3018 while(!TAILQ_EMPTY(&mountlist)) { 3019 mp = TAILQ_LAST(&mountlist, mntlist); 3020 error = dounmount(mp, MNT_FORCE, td); 3021 if (error) { 3022 TAILQ_REMOVE(&mountlist, mp, mnt_list); 3023 printf("unmount of %s failed (", 3024 mp->mnt_stat.f_mntonname); 3025 if (error == EBUSY) 3026 printf("BUSY)\n"); 3027 else 3028 printf("%d)\n", error); 3029 } else { 3030 /* The unmount has removed mp from the mountlist */ 3031 } 3032 } 3033} 3034 3035/* 3036 * perform msync on all vnodes under a mount point 3037 * the mount point must be locked. 3038 */ 3039void 3040vfs_msync(struct mount *mp, int flags) 3041{ 3042 struct vnode *vp, *nvp; 3043 struct vm_object *obj; 3044 int tries; 3045 3046 GIANT_REQUIRED; 3047 3048 tries = 5; 3049 MNT_ILOCK(mp); 3050loop: 3051 TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) { 3052 if (vp->v_mount != mp) { 3053 if (--tries > 0) 3054 goto loop; 3055 break; 3056 } 3057 3058 VI_LOCK(vp); 3059 if (vp->v_iflag & VI_XLOCK) { 3060 VI_UNLOCK(vp); 3061 continue; 3062 } 3063 3064 if ((vp->v_iflag & VI_OBJDIRTY) && 3065 (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { 3066 MNT_IUNLOCK(mp); 3067 if (!vget(vp, 3068 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 3069 curthread)) { 3070 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 3071 vput(vp); 3072 MNT_ILOCK(mp); 3073 continue; 3074 } 3075 3076 if (VOP_GETVOBJECT(vp, &obj) == 0) { 3077 VM_OBJECT_LOCK(obj); 3078 vm_object_page_clean(obj, 0, 0, 3079 flags == MNT_WAIT ? 3080 OBJPC_SYNC : OBJPC_NOSYNC); 3081 VM_OBJECT_UNLOCK(obj); 3082 } 3083 vput(vp); 3084 } 3085 MNT_ILOCK(mp); 3086 if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) { 3087 if (--tries > 0) 3088 goto loop; 3089 break; 3090 } 3091 } else 3092 VI_UNLOCK(vp); 3093 } 3094 MNT_IUNLOCK(mp); 3095} 3096 3097/* 3098 * Create the VM object needed for VMIO and mmap support. This 3099 * is done for all VREG files in the system. Some filesystems might 3100 * afford the additional metadata buffering capability of the 3101 * VMIO code by making the device node be VMIO mode also. 3102 * 3103 * vp must be locked when vfs_object_create is called. 3104 */ 3105int 3106vfs_object_create(struct vnode *vp, struct thread *td, struct ucred *cred) 3107{ 3108 3109 GIANT_REQUIRED; 3110 return (VOP_CREATEVOBJECT(vp, cred, td)); 3111} 3112 3113/* 3114 * Mark a vnode as free, putting it up for recycling. 3115 */ 3116void 3117vfree(struct vnode *vp) 3118{ 3119 3120 ASSERT_VI_LOCKED(vp, "vfree"); 3121 mtx_lock(&vnode_free_list_mtx); 3122 KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free")); 3123 if (vp->v_iflag & VI_AGE) { 3124 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 3125 } else { 3126 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 3127 } 3128 freevnodes++; 3129 mtx_unlock(&vnode_free_list_mtx); 3130 vp->v_iflag &= ~VI_AGE; 3131 vp->v_iflag |= VI_FREE; 3132} 3133 3134/* 3135 * Opposite of vfree() - mark a vnode as in use. 3136 */ 3137void 3138vbusy(struct vnode *vp) 3139{ 3140 3141 ASSERT_VI_LOCKED(vp, "vbusy"); 3142 KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free")); 3143 3144 mtx_lock(&vnode_free_list_mtx); 3145 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 3146 freevnodes--; 3147 mtx_unlock(&vnode_free_list_mtx); 3148 3149 vp->v_iflag &= ~(VI_FREE|VI_AGE); 3150} 3151 3152/* 3153 * Initalize per-vnode helper structure to hold poll-related state. 3154 */ 3155void 3156v_addpollinfo(struct vnode *vp) 3157{ 3158 struct vpollinfo *vi; 3159 3160 vi = uma_zalloc(vnodepoll_zone, M_WAITOK); 3161 if (vp->v_pollinfo != NULL) { 3162 uma_zfree(vnodepoll_zone, vi); 3163 return; 3164 } 3165 vp->v_pollinfo = vi; 3166 mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 3167 knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, 3168 &vp->v_pollinfo->vpi_lock); 3169} 3170 3171/* 3172 * Record a process's interest in events which might happen to 3173 * a vnode. Because poll uses the historic select-style interface 3174 * internally, this routine serves as both the ``check for any 3175 * pending events'' and the ``record my interest in future events'' 3176 * functions. (These are done together, while the lock is held, 3177 * to avoid race conditions.) 3178 */ 3179int 3180vn_pollrecord(vp, td, events) 3181 struct vnode *vp; 3182 struct thread *td; 3183 short events; 3184{ 3185 3186 if (vp->v_pollinfo == NULL) 3187 v_addpollinfo(vp); 3188 mtx_lock(&vp->v_pollinfo->vpi_lock); 3189 if (vp->v_pollinfo->vpi_revents & events) { 3190 /* 3191 * This leaves events we are not interested 3192 * in available for the other process which 3193 * which presumably had requested them 3194 * (otherwise they would never have been 3195 * recorded). 3196 */ 3197 events &= vp->v_pollinfo->vpi_revents; 3198 vp->v_pollinfo->vpi_revents &= ~events; 3199 3200 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3201 return events; 3202 } 3203 vp->v_pollinfo->vpi_events |= events; 3204 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 3205 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3206 return 0; 3207} 3208 3209/* 3210 * Note the occurrence of an event. If the VN_POLLEVENT macro is used, 3211 * it is possible for us to miss an event due to race conditions, but 3212 * that condition is expected to be rare, so for the moment it is the 3213 * preferred interface. 3214 */ 3215void 3216vn_pollevent(vp, events) 3217 struct vnode *vp; 3218 short events; 3219{ 3220 3221 if (vp->v_pollinfo == NULL) 3222 v_addpollinfo(vp); 3223 mtx_lock(&vp->v_pollinfo->vpi_lock); 3224 if (vp->v_pollinfo->vpi_events & events) { 3225 /* 3226 * We clear vpi_events so that we don't 3227 * call selwakeup() twice if two events are 3228 * posted before the polling process(es) is 3229 * awakened. This also ensures that we take at 3230 * most one selwakeup() if the polling process 3231 * is no longer interested. However, it does 3232 * mean that only one event can be noticed at 3233 * a time. (Perhaps we should only clear those 3234 * event bits which we note?) XXX 3235 */ 3236 vp->v_pollinfo->vpi_events = 0; /* &= ~events ??? */ 3237 vp->v_pollinfo->vpi_revents |= events; 3238 selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO); 3239 } 3240 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3241} 3242 3243/* 3244 * Wake up anyone polling on vp because it is being revoked. 3245 * This depends on dead_poll() returning POLLHUP for correct 3246 * behavior. 3247 */ 3248void 3249vn_pollgone(vp) 3250 struct vnode *vp; 3251{ 3252 3253 mtx_lock(&vp->v_pollinfo->vpi_lock); 3254 VN_KNOTE_LOCKED(vp, NOTE_REVOKE); 3255 if (vp->v_pollinfo->vpi_events) { 3256 vp->v_pollinfo->vpi_events = 0; 3257 selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO); 3258 } 3259 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3260} 3261 3262 3263 3264/* 3265 * Routine to create and manage a filesystem syncer vnode. 3266 */ 3267#define sync_close ((int (*)(struct vop_close_args *))nullop) 3268static int sync_fsync(struct vop_fsync_args *); 3269static int sync_inactive(struct vop_inactive_args *); 3270static int sync_reclaim(struct vop_reclaim_args *); 3271 3272static vop_t **sync_vnodeop_p; 3273static struct vnodeopv_entry_desc sync_vnodeop_entries[] = { 3274 { &vop_default_desc, (vop_t *) vop_eopnotsupp }, 3275 { &vop_close_desc, (vop_t *) sync_close }, /* close */ 3276 { &vop_fsync_desc, (vop_t *) sync_fsync }, /* fsync */ 3277 { &vop_inactive_desc, (vop_t *) sync_inactive }, /* inactive */ 3278 { &vop_reclaim_desc, (vop_t *) sync_reclaim }, /* reclaim */ 3279 { &vop_lock_desc, (vop_t *) vop_stdlock }, /* lock */ 3280 { &vop_unlock_desc, (vop_t *) vop_stdunlock }, /* unlock */ 3281 { &vop_islocked_desc, (vop_t *) vop_stdislocked }, /* islocked */ 3282 { NULL, NULL } 3283}; 3284static struct vnodeopv_desc sync_vnodeop_opv_desc = 3285 { &sync_vnodeop_p, sync_vnodeop_entries }; 3286 3287VNODEOP_SET(sync_vnodeop_opv_desc); 3288 3289/* 3290 * Create a new filesystem syncer vnode for the specified mount point. 3291 */ 3292int 3293vfs_allocate_syncvnode(mp) 3294 struct mount *mp; 3295{ 3296 struct vnode *vp; 3297 static long start, incr, next; 3298 int error; 3299 3300 /* Allocate a new vnode */ 3301 if ((error = getnewvnode("syncer", mp, sync_vnodeop_p, &vp)) != 0) { 3302 mp->mnt_syncer = NULL; 3303 return (error); 3304 } 3305 vp->v_type = VNON; 3306 /* 3307 * Place the vnode onto the syncer worklist. We attempt to 3308 * scatter them about on the list so that they will go off 3309 * at evenly distributed times even if all the filesystems 3310 * are mounted at once. 3311 */ 3312 next += incr; 3313 if (next == 0 || next > syncer_maxdelay) { 3314 start /= 2; 3315 incr /= 2; 3316 if (start == 0) { 3317 start = syncer_maxdelay / 2; 3318 incr = syncer_maxdelay; 3319 } 3320 next = start; 3321 } 3322 VI_LOCK(vp); 3323 vn_syncer_add_to_worklist(&vp->v_bufobj, 3324 syncdelay > 0 ? next % syncdelay : 0); 3325 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 3326 mtx_lock(&sync_mtx); 3327 sync_vnode_count++; 3328 mtx_unlock(&sync_mtx); 3329 VI_UNLOCK(vp); 3330 mp->mnt_syncer = vp; 3331 return (0); 3332} 3333 3334/* 3335 * Do a lazy sync of the filesystem. 3336 */ 3337static int 3338sync_fsync(ap) 3339 struct vop_fsync_args /* { 3340 struct vnode *a_vp; 3341 struct ucred *a_cred; 3342 int a_waitfor; 3343 struct thread *a_td; 3344 } */ *ap; 3345{ 3346 struct vnode *syncvp = ap->a_vp; 3347 struct mount *mp = syncvp->v_mount; 3348 struct thread *td = ap->a_td; 3349 int error, asyncflag; 3350 struct bufobj *bo; 3351 3352 /* 3353 * We only need to do something if this is a lazy evaluation. 3354 */ 3355 if (ap->a_waitfor != MNT_LAZY) 3356 return (0); 3357 3358 /* 3359 * Move ourselves to the back of the sync list. 3360 */ 3361 bo = &syncvp->v_bufobj; 3362 BO_LOCK(bo); 3363 vn_syncer_add_to_worklist(bo, syncdelay); 3364 BO_UNLOCK(bo); 3365 3366 /* 3367 * Walk the list of vnodes pushing all that are dirty and 3368 * not already on the sync list. 3369 */ 3370 mtx_lock(&mountlist_mtx); 3371 if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { 3372 mtx_unlock(&mountlist_mtx); 3373 return (0); 3374 } 3375 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 3376 vfs_unbusy(mp, td); 3377 return (0); 3378 } 3379 asyncflag = mp->mnt_flag & MNT_ASYNC; 3380 mp->mnt_flag &= ~MNT_ASYNC; 3381 vfs_msync(mp, MNT_NOWAIT); 3382 error = VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td); 3383 if (asyncflag) 3384 mp->mnt_flag |= MNT_ASYNC; 3385 vn_finished_write(mp); 3386 vfs_unbusy(mp, td); 3387 return (error); 3388} 3389 3390/* 3391 * The syncer vnode is no referenced. 3392 */ 3393static int 3394sync_inactive(ap) 3395 struct vop_inactive_args /* { 3396 struct vnode *a_vp; 3397 struct thread *a_td; 3398 } */ *ap; 3399{ 3400 3401 VOP_UNLOCK(ap->a_vp, 0, ap->a_td); 3402 vgone(ap->a_vp); 3403 return (0); 3404} 3405 3406/* 3407 * The syncer vnode is no longer needed and is being decommissioned. 3408 * 3409 * Modifications to the worklist must be protected by sync_mtx. 3410 */ 3411static int 3412sync_reclaim(ap) 3413 struct vop_reclaim_args /* { 3414 struct vnode *a_vp; 3415 } */ *ap; 3416{ 3417 struct vnode *vp = ap->a_vp; 3418 struct bufobj *bo; 3419 3420 VI_LOCK(vp); 3421 bo = &vp->v_bufobj; 3422 vp->v_mount->mnt_syncer = NULL; 3423 if (bo->bo_flag & BO_ONWORKLST) { 3424 mtx_lock(&sync_mtx); 3425 LIST_REMOVE(bo, bo_synclist); 3426 syncer_worklist_len--; 3427 sync_vnode_count--; 3428 mtx_unlock(&sync_mtx); 3429 bo->bo_flag &= ~BO_ONWORKLST; 3430 } 3431 VI_UNLOCK(vp); 3432 3433 return (0); 3434} 3435 3436/* 3437 * extract the struct cdev *from a VCHR 3438 */ 3439struct cdev * 3440vn_todev(vp) 3441 struct vnode *vp; 3442{ 3443 3444 if (vp->v_type != VCHR) 3445 return (NULL); 3446 return (vp->v_rdev); 3447} 3448 3449/* 3450 * Check if vnode represents a disk device 3451 */ 3452int 3453vn_isdisk(vp, errp) 3454 struct vnode *vp; 3455 int *errp; 3456{ 3457 int error; 3458 3459 error = 0; 3460 dev_lock(); 3461 if (vp->v_type != VCHR) 3462 error = ENOTBLK; 3463 else if (vp->v_rdev == NULL) 3464 error = ENXIO; 3465 else if (vp->v_rdev->si_devsw == NULL) 3466 error = ENXIO; 3467 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 3468 error = ENOTBLK; 3469 dev_unlock(); 3470 if (errp != NULL) 3471 *errp = error; 3472 return (error == 0); 3473} 3474 3475/* 3476 * Free data allocated by namei(); see namei(9) for details. 3477 */ 3478void 3479NDFREE(ndp, flags) 3480 struct nameidata *ndp; 3481 const u_int flags; 3482{ 3483 3484 if (!(flags & NDF_NO_FREE_PNBUF) && 3485 (ndp->ni_cnd.cn_flags & HASBUF)) { 3486 uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf); 3487 ndp->ni_cnd.cn_flags &= ~HASBUF; 3488 } 3489 if (!(flags & NDF_NO_DVP_UNLOCK) && 3490 (ndp->ni_cnd.cn_flags & LOCKPARENT) && 3491 ndp->ni_dvp != ndp->ni_vp) 3492 VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread); 3493 if (!(flags & NDF_NO_DVP_RELE) && 3494 (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) { 3495 vrele(ndp->ni_dvp); 3496 ndp->ni_dvp = NULL; 3497 } 3498 if (!(flags & NDF_NO_VP_UNLOCK) && 3499 (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp) 3500 VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread); 3501 if (!(flags & NDF_NO_VP_RELE) && 3502 ndp->ni_vp) { 3503 vrele(ndp->ni_vp); 3504 ndp->ni_vp = NULL; 3505 } 3506 if (!(flags & NDF_NO_STARTDIR_RELE) && 3507 (ndp->ni_cnd.cn_flags & SAVESTART)) { 3508 vrele(ndp->ni_startdir); 3509 ndp->ni_startdir = NULL; 3510 } 3511} 3512 3513/* 3514 * Common filesystem object access control check routine. Accepts a 3515 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3516 * and optional call-by-reference privused argument allowing vaccess() 3517 * to indicate to the caller whether privilege was used to satisfy the 3518 * request (obsoleted). Returns 0 on success, or an errno on failure. 3519 */ 3520int 3521vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused) 3522 enum vtype type; 3523 mode_t file_mode; 3524 uid_t file_uid; 3525 gid_t file_gid; 3526 mode_t acc_mode; 3527 struct ucred *cred; 3528 int *privused; 3529{ 3530 mode_t dac_granted; 3531#ifdef CAPABILITIES 3532 mode_t cap_granted; 3533#endif 3534 3535 /* 3536 * Look for a normal, non-privileged way to access the file/directory 3537 * as requested. If it exists, go with that. 3538 */ 3539 3540 if (privused != NULL) 3541 *privused = 0; 3542 3543 dac_granted = 0; 3544 3545 /* Check the owner. */ 3546 if (cred->cr_uid == file_uid) { 3547 dac_granted |= VADMIN; 3548 if (file_mode & S_IXUSR) 3549 dac_granted |= VEXEC; 3550 if (file_mode & S_IRUSR) 3551 dac_granted |= VREAD; 3552 if (file_mode & S_IWUSR) 3553 dac_granted |= (VWRITE | VAPPEND); 3554 3555 if ((acc_mode & dac_granted) == acc_mode) 3556 return (0); 3557 3558 goto privcheck; 3559 } 3560 3561 /* Otherwise, check the groups (first match) */ 3562 if (groupmember(file_gid, cred)) { 3563 if (file_mode & S_IXGRP) 3564 dac_granted |= VEXEC; 3565 if (file_mode & S_IRGRP) 3566 dac_granted |= VREAD; 3567 if (file_mode & S_IWGRP) 3568 dac_granted |= (VWRITE | VAPPEND); 3569 3570 if ((acc_mode & dac_granted) == acc_mode) 3571 return (0); 3572 3573 goto privcheck; 3574 } 3575 3576 /* Otherwise, check everyone else. */ 3577 if (file_mode & S_IXOTH) 3578 dac_granted |= VEXEC; 3579 if (file_mode & S_IROTH) 3580 dac_granted |= VREAD; 3581 if (file_mode & S_IWOTH) 3582 dac_granted |= (VWRITE | VAPPEND); 3583 if ((acc_mode & dac_granted) == acc_mode) 3584 return (0); 3585 3586privcheck: 3587 if (!suser_cred(cred, SUSER_ALLOWJAIL)) { 3588 /* XXX audit: privilege used */ 3589 if (privused != NULL) 3590 *privused = 1; 3591 return (0); 3592 } 3593 3594#ifdef CAPABILITIES 3595 /* 3596 * Build a capability mask to determine if the set of capabilities 3597 * satisfies the requirements when combined with the granted mask 3598 * from above. 3599 * For each capability, if the capability is required, bitwise 3600 * or the request type onto the cap_granted mask. 3601 */ 3602 cap_granted = 0; 3603 3604 if (type == VDIR) { 3605 /* 3606 * For directories, use CAP_DAC_READ_SEARCH to satisfy 3607 * VEXEC requests, instead of CAP_DAC_EXECUTE. 3608 */ 3609 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3610 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL)) 3611 cap_granted |= VEXEC; 3612 } else { 3613 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3614 !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL)) 3615 cap_granted |= VEXEC; 3616 } 3617 3618 if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && 3619 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL)) 3620 cap_granted |= VREAD; 3621 3622 if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3623 !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL)) 3624 cap_granted |= (VWRITE | VAPPEND); 3625 3626 if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3627 !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL)) 3628 cap_granted |= VADMIN; 3629 3630 if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { 3631 /* XXX audit: privilege used */ 3632 if (privused != NULL) 3633 *privused = 1; 3634 return (0); 3635 } 3636#endif 3637 3638 return ((acc_mode & VADMIN) ? EPERM : EACCES); 3639} 3640 3641/* 3642 * Credential check based on process requesting service, and per-attribute 3643 * permissions. 3644 */ 3645int 3646extattr_check_cred(struct vnode *vp, int attrnamespace, 3647 struct ucred *cred, struct thread *td, int access) 3648{ 3649 3650 /* 3651 * Kernel-invoked always succeeds. 3652 */ 3653 if (cred == NOCRED) 3654 return (0); 3655 3656 /* 3657 * Do not allow privileged processes in jail to directly 3658 * manipulate system attributes. 3659 * 3660 * XXX What capability should apply here? 3661 * Probably CAP_SYS_SETFFLAG. 3662 */ 3663 switch (attrnamespace) { 3664 case EXTATTR_NAMESPACE_SYSTEM: 3665 /* Potentially should be: return (EPERM); */ 3666 return (suser_cred(cred, 0)); 3667 case EXTATTR_NAMESPACE_USER: 3668 return (VOP_ACCESS(vp, access, cred, td)); 3669 default: 3670 return (EPERM); 3671 } 3672} 3673 3674#ifdef DEBUG_VFS_LOCKS 3675/* 3676 * This only exists to supress warnings from unlocked specfs accesses. It is 3677 * no longer ok to have an unlocked VFS. 3678 */ 3679#define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD) 3680 3681int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 3682SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, ""); 3683 3684int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 3685SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, ""); 3686 3687int vfs_badlock_print = 1; /* Print lock violations. */ 3688SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, ""); 3689 3690#ifdef KDB 3691int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 3692SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, ""); 3693#endif 3694 3695static void 3696vfs_badlock(const char *msg, const char *str, struct vnode *vp) 3697{ 3698 3699#ifdef KDB 3700 if (vfs_badlock_backtrace) 3701 kdb_backtrace(); 3702#endif 3703 if (vfs_badlock_print) 3704 printf("%s: %p %s\n", str, (void *)vp, msg); 3705 if (vfs_badlock_ddb) 3706 kdb_enter("lock violation"); 3707} 3708 3709void 3710assert_vi_locked(struct vnode *vp, const char *str) 3711{ 3712 3713 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 3714 vfs_badlock("interlock is not locked but should be", str, vp); 3715} 3716 3717void 3718assert_vi_unlocked(struct vnode *vp, const char *str) 3719{ 3720 3721 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 3722 vfs_badlock("interlock is locked but should not be", str, vp); 3723} 3724 3725void 3726assert_vop_locked(struct vnode *vp, const char *str) 3727{ 3728 3729 if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0) 3730 vfs_badlock("is not locked but should be", str, vp); 3731} 3732 3733void 3734assert_vop_unlocked(struct vnode *vp, const char *str) 3735{ 3736 3737 if (vp && !IGNORE_LOCK(vp) && 3738 VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE) 3739 vfs_badlock("is locked but should not be", str, vp); 3740} 3741 3742#if 0 3743void 3744assert_vop_elocked(struct vnode *vp, const char *str) 3745{ 3746 3747 if (vp && !IGNORE_LOCK(vp) && 3748 VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE) 3749 vfs_badlock("is not exclusive locked but should be", str, vp); 3750} 3751 3752void 3753assert_vop_elocked_other(struct vnode *vp, const char *str) 3754{ 3755 3756 if (vp && !IGNORE_LOCK(vp) && 3757 VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER) 3758 vfs_badlock("is not exclusive locked by another thread", 3759 str, vp); 3760} 3761 3762void 3763assert_vop_slocked(struct vnode *vp, const char *str) 3764{ 3765 3766 if (vp && !IGNORE_LOCK(vp) && 3767 VOP_ISLOCKED(vp, curthread) != LK_SHARED) 3768 vfs_badlock("is not locked shared but should be", str, vp); 3769} 3770#endif /* 0 */ 3771 3772void 3773vop_rename_pre(void *ap) 3774{ 3775 struct vop_rename_args *a = ap; 3776 3777 if (a->a_tvp) 3778 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 3779 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 3780 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 3781 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 3782 3783 /* Check the source (from). */ 3784 if (a->a_tdvp != a->a_fdvp) 3785 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 3786 if (a->a_tvp != a->a_fvp) 3787 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked"); 3788 3789 /* Check the target. */ 3790 if (a->a_tvp) 3791 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 3792 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 3793} 3794 3795void 3796vop_strategy_pre(void *ap) 3797{ 3798 struct vop_strategy_args *a; 3799 struct buf *bp; 3800 3801 a = ap; 3802 bp = a->a_bp; 3803 3804 /* 3805 * Cluster ops lock their component buffers but not the IO container. 3806 */ 3807 if ((bp->b_flags & B_CLUSTER) != 0) 3808 return; 3809 3810 if (BUF_REFCNT(bp) < 1) { 3811 if (vfs_badlock_print) 3812 printf( 3813 "VOP_STRATEGY: bp is not locked but should be\n"); 3814 if (vfs_badlock_ddb) 3815 kdb_enter("lock violation"); 3816 } 3817} 3818 3819void 3820vop_lookup_pre(void *ap) 3821{ 3822 struct vop_lookup_args *a; 3823 struct vnode *dvp; 3824 3825 a = ap; 3826 dvp = a->a_dvp; 3827 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3828 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 3829} 3830 3831void 3832vop_lookup_post(void *ap, int rc) 3833{ 3834 struct vop_lookup_args *a; 3835 struct componentname *cnp; 3836 struct vnode *dvp; 3837 struct vnode *vp; 3838 int flags; 3839 3840 a = ap; 3841 dvp = a->a_dvp; 3842 cnp = a->a_cnp; 3843 vp = *(a->a_vpp); 3844 flags = cnp->cn_flags; 3845 3846 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3847 3848 /* 3849 * If this is the last path component for this lookup and LOCKPARENT 3850 * is set, OR if there is an error the directory has to be locked. 3851 */ 3852 if ((flags & LOCKPARENT) && (flags & ISLASTCN)) 3853 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)"); 3854 else if (rc != 0) 3855 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)"); 3856 else if (dvp != vp) 3857 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)"); 3858 if (flags & PDIRUNLOCK) 3859 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)"); 3860} 3861 3862void 3863vop_lock_pre(void *ap) 3864{ 3865 struct vop_lock_args *a = ap; 3866 3867 if ((a->a_flags & LK_INTERLOCK) == 0) 3868 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3869 else 3870 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 3871} 3872 3873void 3874vop_lock_post(void *ap, int rc) 3875{ 3876 struct vop_lock_args *a = ap; 3877 3878 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3879 if (rc == 0) 3880 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 3881} 3882 3883void 3884vop_unlock_pre(void *ap) 3885{ 3886 struct vop_unlock_args *a = ap; 3887 3888 if (a->a_flags & LK_INTERLOCK) 3889 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 3890 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 3891} 3892 3893void 3894vop_unlock_post(void *ap, int rc) 3895{ 3896 struct vop_unlock_args *a = ap; 3897 3898 if (a->a_flags & LK_INTERLOCK) 3899 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 3900} 3901#endif /* DEBUG_VFS_LOCKS */ 3902 3903static struct knlist fs_knlist; 3904 3905static void 3906vfs_event_init(void *arg) 3907{ 3908 knlist_init(&fs_knlist, NULL); 3909} 3910/* XXX - correct order? */ 3911SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 3912 3913void 3914vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused) 3915{ 3916 3917 KNOTE_UNLOCKED(&fs_knlist, event); 3918} 3919 3920static int filt_fsattach(struct knote *kn); 3921static void filt_fsdetach(struct knote *kn); 3922static int filt_fsevent(struct knote *kn, long hint); 3923 3924struct filterops fs_filtops = 3925 { 0, filt_fsattach, filt_fsdetach, filt_fsevent }; 3926 3927static int 3928filt_fsattach(struct knote *kn) 3929{ 3930 3931 kn->kn_flags |= EV_CLEAR; 3932 knlist_add(&fs_knlist, kn, 0); 3933 return (0); 3934} 3935 3936static void 3937filt_fsdetach(struct knote *kn) 3938{ 3939 3940 knlist_remove(&fs_knlist, kn, 0); 3941} 3942 3943static int 3944filt_fsevent(struct knote *kn, long hint) 3945{ 3946 3947 kn->kn_fflags |= hint; 3948 return (kn->kn_fflags != 0); 3949} 3950 3951static int 3952sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 3953{ 3954 struct vfsidctl vc; 3955 int error; 3956 struct mount *mp; 3957 3958 error = SYSCTL_IN(req, &vc, sizeof(vc)); 3959 if (error) 3960 return (error); 3961 if (vc.vc_vers != VFS_CTL_VERS1) 3962 return (EINVAL); 3963 mp = vfs_getvfs(&vc.vc_fsid); 3964 if (mp == NULL) 3965 return (ENOENT); 3966 /* ensure that a specific sysctl goes to the right filesystem. */ 3967 if (strcmp(vc.vc_fstypename, "*") != 0 && 3968 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 3969 return (EINVAL); 3970 } 3971 VCTLTOREQ(&vc, req); 3972 return (VFS_SYSCTL(mp, vc.vc_op, req)); 3973} 3974 3975SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR, 3976 NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid"); 3977 3978/* 3979 * Function to initialize a va_filerev field sensibly. 3980 * XXX: Wouldn't a random number make a lot more sense ?? 3981 */ 3982u_quad_t 3983init_va_filerev(void) 3984{ 3985 struct bintime bt; 3986 3987 getbinuptime(&bt); 3988 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 3989} 3990