vfs_subr.c revision 114074
1/* 2 * Copyright (c) 1989, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 39 * $FreeBSD: head/sys/kern/vfs_subr.c 114074 2003-04-26 18:33:18Z alc $ 40 */ 41 42/* 43 * External virtual filesystem routines 44 */ 45#include "opt_ddb.h" 46#include "opt_mac.h" 47 48#include <sys/param.h> 49#include <sys/systm.h> 50#include <sys/bio.h> 51#include <sys/buf.h> 52#include <sys/conf.h> 53#include <sys/eventhandler.h> 54#include <sys/extattr.h> 55#include <sys/fcntl.h> 56#include <sys/kernel.h> 57#include <sys/kthread.h> 58#include <sys/mac.h> 59#include <sys/malloc.h> 60#include <sys/mount.h> 61#include <sys/namei.h> 62#include <sys/stat.h> 63#include <sys/sysctl.h> 64#include <sys/syslog.h> 65#include <sys/vmmeter.h> 66#include <sys/vnode.h> 67 68#include <vm/vm.h> 69#include <vm/vm_object.h> 70#include <vm/vm_extern.h> 71#include <vm/pmap.h> 72#include <vm/vm_map.h> 73#include <vm/vm_page.h> 74#include <vm/uma.h> 75 76static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure"); 77 78static void addalias(struct vnode *vp, dev_t nvp_rdev); 79static void insmntque(struct vnode *vp, struct mount *mp); 80static void vclean(struct vnode *vp, int flags, struct thread *td); 81static void vlruvp(struct vnode *vp); 82static int flushbuflist(struct buf *blist, int flags, struct vnode *vp, 83 int slpflag, int slptimeo, int *errorp); 84static int vcanrecycle(struct vnode *vp, struct mount **vnmpp); 85 86 87/* 88 * Number of vnodes in existence. Increased whenever getnewvnode() 89 * allocates a new vnode, never decreased. 90 */ 91static unsigned long numvnodes; 92 93SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); 94 95/* 96 * Conversion tables for conversion from vnode types to inode formats 97 * and back. 98 */ 99enum vtype iftovt_tab[16] = { 100 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 101 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 102}; 103int vttoif_tab[9] = { 104 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 105 S_IFSOCK, S_IFIFO, S_IFMT, 106}; 107 108/* 109 * List of vnodes that are ready for recycling. 110 */ 111static TAILQ_HEAD(freelst, vnode) vnode_free_list; 112 113/* 114 * Minimum number of free vnodes. If there are fewer than this free vnodes, 115 * getnewvnode() will return a newly allocated vnode. 116 */ 117static u_long wantfreevnodes = 25; 118SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 119/* Number of vnodes in the free list. */ 120static u_long freevnodes; 121SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); 122 123/* 124 * Various variables used for debugging the new implementation of 125 * reassignbuf(). 126 * XXX these are probably of (very) limited utility now. 127 */ 128static int reassignbufcalls; 129SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); 130static int nameileafonly; 131SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, ""); 132 133/* 134 * Cache for the mount type id assigned to NFS. This is used for 135 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 136 */ 137int nfs_mount_type = -1; 138 139/* To keep more than one thread at a time from running vfs_getnewfsid */ 140static struct mtx mntid_mtx; 141 142/* 143 * Lock for any access to the following: 144 * vnode_free_list 145 * numvnodes 146 * freevnodes 147 */ 148static struct mtx vnode_free_list_mtx; 149 150/* 151 * For any iteration/modification of dev->si_hlist (linked through 152 * v_specnext) 153 */ 154static struct mtx spechash_mtx; 155 156/* Publicly exported FS */ 157struct nfs_public nfs_pub; 158 159/* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 160static uma_zone_t vnode_zone; 161static uma_zone_t vnodepoll_zone; 162 163/* Set to 1 to print out reclaim of active vnodes */ 164int prtactive; 165 166/* 167 * The workitem queue. 168 * 169 * It is useful to delay writes of file data and filesystem metadata 170 * for tens of seconds so that quickly created and deleted files need 171 * not waste disk bandwidth being created and removed. To realize this, 172 * we append vnodes to a "workitem" queue. When running with a soft 173 * updates implementation, most pending metadata dependencies should 174 * not wait for more than a few seconds. Thus, mounted on block devices 175 * are delayed only about a half the time that file data is delayed. 176 * Similarly, directory updates are more critical, so are only delayed 177 * about a third the time that file data is delayed. Thus, there are 178 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 179 * one each second (driven off the filesystem syncer process). The 180 * syncer_delayno variable indicates the next queue that is to be processed. 181 * Items that need to be processed soon are placed in this queue: 182 * 183 * syncer_workitem_pending[syncer_delayno] 184 * 185 * A delay of fifteen seconds is done by placing the request fifteen 186 * entries later in the queue: 187 * 188 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 189 * 190 */ 191static int syncer_delayno; 192static long syncer_mask; 193LIST_HEAD(synclist, vnode); 194static struct synclist *syncer_workitem_pending; 195/* 196 * The sync_mtx protects: 197 * vp->v_synclist 198 * syncer_delayno 199 * syncer_workitem_pending 200 * rushjob 201 */ 202static struct mtx sync_mtx; 203 204#define SYNCER_MAXDELAY 32 205static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 206static int syncdelay = 30; /* max time to delay syncing data */ 207static int filedelay = 30; /* time to delay syncing files */ 208SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); 209static int dirdelay = 29; /* time to delay syncing directories */ 210SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); 211static int metadelay = 28; /* time to delay syncing metadata */ 212SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); 213static int rushjob; /* number of slots to run ASAP */ 214static int stat_rush_requests; /* number of times I/O speeded up */ 215SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); 216 217/* 218 * Number of vnodes we want to exist at any one time. This is mostly used 219 * to size hash tables in vnode-related code. It is normally not used in 220 * getnewvnode(), as wantfreevnodes is normally nonzero.) 221 * 222 * XXX desiredvnodes is historical cruft and should not exist. 223 */ 224int desiredvnodes; 225SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 226 &desiredvnodes, 0, "Maximum number of vnodes"); 227static int minvnodes; 228SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 229 &minvnodes, 0, "Minimum number of vnodes"); 230static int vnlru_nowhere; 231SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, &vnlru_nowhere, 0, 232 "Number of times the vnlru process ran without success"); 233 234/* Hook for calling soft updates */ 235int (*softdep_process_worklist_hook)(struct mount *); 236 237/* 238 * This only exists to supress warnings from unlocked specfs accesses. It is 239 * no longer ok to have an unlocked VFS. 240 */ 241#define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD) 242 243/* Print lock violations */ 244int vfs_badlock_print = 1; 245 246/* Panic on violation */ 247int vfs_badlock_panic = 1; 248 249/* Check for interlock across VOPs */ 250int vfs_badlock_mutex = 1; 251 252static void 253vfs_badlock(char *msg, char *str, struct vnode *vp) 254{ 255 if (vfs_badlock_print) 256 printf("%s: %p %s\n", str, vp, msg); 257 if (vfs_badlock_panic) 258 Debugger("Lock violation.\n"); 259} 260 261void 262assert_vi_unlocked(struct vnode *vp, char *str) 263{ 264 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 265 vfs_badlock("interlock is locked but should not be", str, vp); 266} 267 268void 269assert_vi_locked(struct vnode *vp, char *str) 270{ 271 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 272 vfs_badlock("interlock is not locked but should be", str, vp); 273} 274 275void 276assert_vop_locked(struct vnode *vp, char *str) 277{ 278 if (vp && !IGNORE_LOCK(vp) && !VOP_ISLOCKED(vp, NULL)) 279 vfs_badlock("is not locked but should be", str, vp); 280} 281 282void 283assert_vop_unlocked(struct vnode *vp, char *str) 284{ 285 if (vp && !IGNORE_LOCK(vp) && 286 VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE) 287 vfs_badlock("is locked but should not be", str, vp); 288} 289 290void 291assert_vop_elocked(struct vnode *vp, char *str) 292{ 293 if (vp && !IGNORE_LOCK(vp) && 294 VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE) 295 vfs_badlock("is not exclusive locked but should be", str, vp); 296} 297 298void 299assert_vop_elocked_other(struct vnode *vp, char *str) 300{ 301 if (vp && !IGNORE_LOCK(vp) && 302 VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER) 303 vfs_badlock("is not exclusive locked by another thread", 304 str, vp); 305} 306 307void 308assert_vop_slocked(struct vnode *vp, char *str) 309{ 310 if (vp && !IGNORE_LOCK(vp) && 311 VOP_ISLOCKED(vp, curthread) != LK_SHARED) 312 vfs_badlock("is not locked shared but should be", str, vp); 313} 314 315void 316vop_rename_pre(void *ap) 317{ 318 struct vop_rename_args *a = ap; 319 320 if (a->a_tvp) 321 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 322 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 323 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 324 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 325 326 /* Check the source (from) */ 327 if (a->a_tdvp != a->a_fdvp) 328 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked.\n"); 329 if (a->a_tvp != a->a_fvp) 330 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked.\n"); 331 332 /* Check the target */ 333 if (a->a_tvp) 334 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked.\n"); 335 336 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked.\n"); 337} 338 339void 340vop_strategy_pre(void *ap) 341{ 342 struct vop_strategy_args *a = ap; 343 struct buf *bp; 344 345 bp = a->a_bp; 346 347 /* 348 * Cluster ops lock their component buffers but not the IO container. 349 */ 350 if ((bp->b_flags & B_CLUSTER) != 0) 351 return; 352 353 if (BUF_REFCNT(bp) < 1) { 354 if (vfs_badlock_print) 355 printf("VOP_STRATEGY: bp is not locked but should be.\n"); 356 if (vfs_badlock_panic) 357 Debugger("Lock violation.\n"); 358 } 359} 360 361void 362vop_lookup_pre(void *ap) 363{ 364 struct vop_lookup_args *a = ap; 365 struct vnode *dvp; 366 367 dvp = a->a_dvp; 368 369 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 370 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 371} 372 373void 374vop_lookup_post(void *ap, int rc) 375{ 376 struct vop_lookup_args *a = ap; 377 struct componentname *cnp; 378 struct vnode *dvp; 379 struct vnode *vp; 380 int flags; 381 382 dvp = a->a_dvp; 383 cnp = a->a_cnp; 384 vp = *(a->a_vpp); 385 flags = cnp->cn_flags; 386 387 388 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 389 /* 390 * If this is the last path component for this lookup and LOCPARENT 391 * is set, OR if there is an error the directory has to be locked. 392 */ 393 if ((flags & LOCKPARENT) && (flags & ISLASTCN)) 394 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)"); 395 else if (rc != 0) 396 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)"); 397 else if (dvp != vp) 398 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)"); 399 400 if (flags & PDIRUNLOCK) 401 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)"); 402} 403 404void 405vop_unlock_pre(void *ap) 406{ 407 struct vop_unlock_args *a = ap; 408 409 if (a->a_flags & LK_INTERLOCK) 410 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 411 412 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 413} 414 415void 416vop_unlock_post(void *ap, int rc) 417{ 418 struct vop_unlock_args *a = ap; 419 420 if (a->a_flags & LK_INTERLOCK) 421 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 422} 423 424void 425vop_lock_pre(void *ap) 426{ 427 struct vop_lock_args *a = ap; 428 429 if ((a->a_flags & LK_INTERLOCK) == 0) 430 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 431 else 432 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 433} 434 435void 436vop_lock_post(void *ap, int rc) 437{ 438 struct vop_lock_args *a; 439 440 a = ap; 441 442 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 443 if (rc == 0) 444 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 445} 446 447void 448v_addpollinfo(struct vnode *vp) 449{ 450 vp->v_pollinfo = uma_zalloc(vnodepoll_zone, M_WAITOK); 451 mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 452} 453 454/* 455 * Initialize the vnode management data structures. 456 */ 457static void 458vntblinit(void *dummy __unused) 459{ 460 461 desiredvnodes = maxproc + cnt.v_page_count / 4; 462 minvnodes = desiredvnodes / 4; 463 mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF); 464 mtx_init(&mntvnode_mtx, "mntvnode", NULL, MTX_DEF); 465 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 466 mtx_init(&spechash_mtx, "spechash", NULL, MTX_DEF); 467 TAILQ_INIT(&vnode_free_list); 468 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 469 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 470 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 471 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 472 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 473 /* 474 * Initialize the filesystem syncer. 475 */ 476 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 477 &syncer_mask); 478 syncer_maxdelay = syncer_mask + 1; 479 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 480} 481SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) 482 483 484/* 485 * Mark a mount point as busy. Used to synchronize access and to delay 486 * unmounting. Interlock is not released on failure. 487 */ 488int 489vfs_busy(mp, flags, interlkp, td) 490 struct mount *mp; 491 int flags; 492 struct mtx *interlkp; 493 struct thread *td; 494{ 495 int lkflags; 496 497 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 498 if (flags & LK_NOWAIT) 499 return (ENOENT); 500 mp->mnt_kern_flag |= MNTK_MWAIT; 501 /* 502 * Since all busy locks are shared except the exclusive 503 * lock granted when unmounting, the only place that a 504 * wakeup needs to be done is at the release of the 505 * exclusive lock at the end of dounmount. 506 */ 507 msleep(mp, interlkp, PVFS, "vfs_busy", 0); 508 return (ENOENT); 509 } 510 lkflags = LK_SHARED | LK_NOPAUSE; 511 if (interlkp) 512 lkflags |= LK_INTERLOCK; 513 if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td)) 514 panic("vfs_busy: unexpected lock failure"); 515 return (0); 516} 517 518/* 519 * Free a busy filesystem. 520 */ 521void 522vfs_unbusy(mp, td) 523 struct mount *mp; 524 struct thread *td; 525{ 526 527 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); 528} 529 530/* 531 * Lookup a mount point by filesystem identifier. 532 */ 533struct mount * 534vfs_getvfs(fsid) 535 fsid_t *fsid; 536{ 537 register struct mount *mp; 538 539 mtx_lock(&mountlist_mtx); 540 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 541 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 542 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 543 mtx_unlock(&mountlist_mtx); 544 return (mp); 545 } 546 } 547 mtx_unlock(&mountlist_mtx); 548 return ((struct mount *) 0); 549} 550 551/* 552 * Get a new unique fsid. Try to make its val[0] unique, since this value 553 * will be used to create fake device numbers for stat(). Also try (but 554 * not so hard) make its val[0] unique mod 2^16, since some emulators only 555 * support 16-bit device numbers. We end up with unique val[0]'s for the 556 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 557 * 558 * Keep in mind that several mounts may be running in parallel. Starting 559 * the search one past where the previous search terminated is both a 560 * micro-optimization and a defense against returning the same fsid to 561 * different mounts. 562 */ 563void 564vfs_getnewfsid(mp) 565 struct mount *mp; 566{ 567 static u_int16_t mntid_base; 568 fsid_t tfsid; 569 int mtype; 570 571 mtx_lock(&mntid_mtx); 572 mtype = mp->mnt_vfc->vfc_typenum; 573 tfsid.val[1] = mtype; 574 mtype = (mtype & 0xFF) << 24; 575 for (;;) { 576 tfsid.val[0] = makeudev(255, 577 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 578 mntid_base++; 579 if (vfs_getvfs(&tfsid) == NULL) 580 break; 581 } 582 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 583 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 584 mtx_unlock(&mntid_mtx); 585} 586 587/* 588 * Knob to control the precision of file timestamps: 589 * 590 * 0 = seconds only; nanoseconds zeroed. 591 * 1 = seconds and nanoseconds, accurate within 1/HZ. 592 * 2 = seconds and nanoseconds, truncated to microseconds. 593 * >=3 = seconds and nanoseconds, maximum precision. 594 */ 595enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 596 597static int timestamp_precision = TSP_SEC; 598SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 599 ×tamp_precision, 0, ""); 600 601/* 602 * Get a current timestamp. 603 */ 604void 605vfs_timestamp(tsp) 606 struct timespec *tsp; 607{ 608 struct timeval tv; 609 610 switch (timestamp_precision) { 611 case TSP_SEC: 612 tsp->tv_sec = time_second; 613 tsp->tv_nsec = 0; 614 break; 615 case TSP_HZ: 616 getnanotime(tsp); 617 break; 618 case TSP_USEC: 619 microtime(&tv); 620 TIMEVAL_TO_TIMESPEC(&tv, tsp); 621 break; 622 case TSP_NSEC: 623 default: 624 nanotime(tsp); 625 break; 626 } 627} 628 629/* 630 * Set vnode attributes to VNOVAL 631 */ 632void 633vattr_null(vap) 634 register struct vattr *vap; 635{ 636 637 vap->va_type = VNON; 638 vap->va_size = VNOVAL; 639 vap->va_bytes = VNOVAL; 640 vap->va_mode = VNOVAL; 641 vap->va_nlink = VNOVAL; 642 vap->va_uid = VNOVAL; 643 vap->va_gid = VNOVAL; 644 vap->va_fsid = VNOVAL; 645 vap->va_fileid = VNOVAL; 646 vap->va_blocksize = VNOVAL; 647 vap->va_rdev = VNOVAL; 648 vap->va_atime.tv_sec = VNOVAL; 649 vap->va_atime.tv_nsec = VNOVAL; 650 vap->va_mtime.tv_sec = VNOVAL; 651 vap->va_mtime.tv_nsec = VNOVAL; 652 vap->va_ctime.tv_sec = VNOVAL; 653 vap->va_ctime.tv_nsec = VNOVAL; 654 vap->va_birthtime.tv_sec = VNOVAL; 655 vap->va_birthtime.tv_nsec = VNOVAL; 656 vap->va_flags = VNOVAL; 657 vap->va_gen = VNOVAL; 658 vap->va_vaflags = 0; 659} 660 661/* 662 * This routine is called when we have too many vnodes. It attempts 663 * to free <count> vnodes and will potentially free vnodes that still 664 * have VM backing store (VM backing store is typically the cause 665 * of a vnode blowout so we want to do this). Therefore, this operation 666 * is not considered cheap. 667 * 668 * A number of conditions may prevent a vnode from being reclaimed. 669 * the buffer cache may have references on the vnode, a directory 670 * vnode may still have references due to the namei cache representing 671 * underlying files, or the vnode may be in active use. It is not 672 * desireable to reuse such vnodes. These conditions may cause the 673 * number of vnodes to reach some minimum value regardless of what 674 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 675 */ 676static int 677vlrureclaim(struct mount *mp) 678{ 679 struct vnode *vp; 680 int done; 681 int trigger; 682 int usevnodes; 683 int count; 684 685 /* 686 * Calculate the trigger point, don't allow user 687 * screwups to blow us up. This prevents us from 688 * recycling vnodes with lots of resident pages. We 689 * aren't trying to free memory, we are trying to 690 * free vnodes. 691 */ 692 usevnodes = desiredvnodes; 693 if (usevnodes <= 0) 694 usevnodes = 1; 695 trigger = cnt.v_page_count * 2 / usevnodes; 696 697 done = 0; 698 mtx_lock(&mntvnode_mtx); 699 count = mp->mnt_nvnodelistsize / 10 + 1; 700 while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) { 701 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 702 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 703 704 if (vp->v_type != VNON && 705 vp->v_type != VBAD && 706 VI_TRYLOCK(vp)) { 707 if (VMIGHTFREE(vp) && /* critical path opt */ 708 (vp->v_object == NULL || 709 vp->v_object->resident_page_count < trigger)) { 710 mtx_unlock(&mntvnode_mtx); 711 vgonel(vp, curthread); 712 done++; 713 mtx_lock(&mntvnode_mtx); 714 } else 715 VI_UNLOCK(vp); 716 } 717 --count; 718 } 719 mtx_unlock(&mntvnode_mtx); 720 return done; 721} 722 723/* 724 * Attempt to recycle vnodes in a context that is always safe to block. 725 * Calling vlrurecycle() from the bowels of filesystem code has some 726 * interesting deadlock problems. 727 */ 728static struct proc *vnlruproc; 729static int vnlruproc_sig; 730 731static void 732vnlru_proc(void) 733{ 734 struct mount *mp, *nmp; 735 int s; 736 int done; 737 struct proc *p = vnlruproc; 738 struct thread *td = FIRST_THREAD_IN_PROC(p); /* XXXKSE */ 739 740 mtx_lock(&Giant); 741 742 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 743 SHUTDOWN_PRI_FIRST); 744 745 s = splbio(); 746 for (;;) { 747 kthread_suspend_check(p); 748 mtx_lock(&vnode_free_list_mtx); 749 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { 750 mtx_unlock(&vnode_free_list_mtx); 751 vnlruproc_sig = 0; 752 wakeup(&vnlruproc_sig); 753 tsleep(vnlruproc, PVFS, "vlruwt", hz); 754 continue; 755 } 756 mtx_unlock(&vnode_free_list_mtx); 757 done = 0; 758 mtx_lock(&mountlist_mtx); 759 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 760 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 761 nmp = TAILQ_NEXT(mp, mnt_list); 762 continue; 763 } 764 done += vlrureclaim(mp); 765 mtx_lock(&mountlist_mtx); 766 nmp = TAILQ_NEXT(mp, mnt_list); 767 vfs_unbusy(mp, td); 768 } 769 mtx_unlock(&mountlist_mtx); 770 if (done == 0) { 771#if 0 772 /* These messages are temporary debugging aids */ 773 if (vnlru_nowhere < 5) 774 printf("vnlru process getting nowhere..\n"); 775 else if (vnlru_nowhere == 5) 776 printf("vnlru process messages stopped.\n"); 777#endif 778 vnlru_nowhere++; 779 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 780 } 781 } 782 splx(s); 783} 784 785static struct kproc_desc vnlru_kp = { 786 "vnlru", 787 vnlru_proc, 788 &vnlruproc 789}; 790SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 791 792 793/* 794 * Routines having to do with the management of the vnode table. 795 */ 796 797/* 798 * Check to see if a free vnode can be recycled. If it can, 799 * return it locked with the vn lock, but not interlock. Also 800 * get the vn_start_write lock. Otherwise indicate the error. 801 */ 802static int 803vcanrecycle(struct vnode *vp, struct mount **vnmpp) 804{ 805 struct thread *td = curthread; 806 vm_object_t object; 807 int error; 808 809 /* Don't recycle if we can't get the interlock */ 810 if (!VI_TRYLOCK(vp)) 811 return (EWOULDBLOCK); 812 813 /* We should be able to immediately acquire this */ 814 /* XXX This looks like it should panic if it fails */ 815 if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td) != 0) { 816 if (VOP_ISLOCKED(vp, td)) 817 panic("vcanrecycle: locked vnode"); 818 return (EWOULDBLOCK); 819 } 820 821 /* 822 * Don't recycle if its filesystem is being suspended. 823 */ 824 if (vn_start_write(vp, vnmpp, V_NOWAIT) != 0) { 825 error = EBUSY; 826 goto done; 827 } 828 829 /* 830 * Don't recycle if we still have cached pages. 831 */ 832 if (VOP_GETVOBJECT(vp, &object) == 0 && 833 (object->resident_page_count || 834 object->ref_count)) { 835 error = EBUSY; 836 goto done; 837 } 838 if (LIST_FIRST(&vp->v_cache_src)) { 839 /* 840 * note: nameileafonly sysctl is temporary, 841 * for debugging only, and will eventually be 842 * removed. 843 */ 844 if (nameileafonly > 0) { 845 /* 846 * Do not reuse namei-cached directory 847 * vnodes that have cached 848 * subdirectories. 849 */ 850 if (cache_leaf_test(vp) < 0) { 851 error = EISDIR; 852 goto done; 853 } 854 } else if (nameileafonly < 0 || 855 vmiodirenable == 0) { 856 /* 857 * Do not reuse namei-cached directory 858 * vnodes if nameileafonly is -1 or 859 * if VMIO backing for directories is 860 * turned off (otherwise we reuse them 861 * too quickly). 862 */ 863 error = EBUSY; 864 goto done; 865 } 866 } 867 return (0); 868done: 869 VOP_UNLOCK(vp, 0, td); 870 return (error); 871} 872 873/* 874 * Return the next vnode from the free list. 875 */ 876int 877getnewvnode(tag, mp, vops, vpp) 878 const char *tag; 879 struct mount *mp; 880 vop_t **vops; 881 struct vnode **vpp; 882{ 883 int s; 884 struct thread *td = curthread; /* XXX */ 885 struct vnode *vp = NULL; 886 struct vpollinfo *pollinfo = NULL; 887 struct mount *vnmp; 888 889 s = splbio(); 890 mtx_lock(&vnode_free_list_mtx); 891 892 /* 893 * Try to reuse vnodes if we hit the max. This situation only 894 * occurs in certain large-memory (2G+) situations. We cannot 895 * attempt to directly reclaim vnodes due to nasty recursion 896 * problems. 897 */ 898 while (numvnodes - freevnodes > desiredvnodes) { 899 if (vnlruproc_sig == 0) { 900 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 901 wakeup(vnlruproc); 902 } 903 mtx_unlock(&vnode_free_list_mtx); 904 tsleep(&vnlruproc_sig, PVFS, "vlruwk", hz); 905 mtx_lock(&vnode_free_list_mtx); 906 } 907 908 /* 909 * Attempt to reuse a vnode already on the free list, allocating 910 * a new vnode if we can't find one or if we have not reached a 911 * good minimum for good LRU performance. 912 */ 913 914 if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) { 915 int error; 916 int count; 917 918 for (count = 0; count < freevnodes; count++) { 919 vp = TAILQ_FIRST(&vnode_free_list); 920 921 KASSERT(vp->v_usecount == 0 && 922 (vp->v_iflag & VI_DOINGINACT) == 0, 923 ("getnewvnode: free vnode isn't")); 924 925 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 926 /* 927 * We have to drop the free list mtx to avoid lock 928 * order reversals with interlock. 929 */ 930 mtx_unlock(&vnode_free_list_mtx); 931 error = vcanrecycle(vp, &vnmp); 932 mtx_lock(&vnode_free_list_mtx); 933 if (error == 0) 934 break; 935 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 936 vp = NULL; 937 } 938 } 939 if (vp) { 940 freevnodes--; 941 mtx_unlock(&vnode_free_list_mtx); 942 943 cache_purge(vp); 944 VI_LOCK(vp); 945 vp->v_iflag |= VI_DOOMED; 946 vp->v_iflag &= ~VI_FREE; 947 if (vp->v_type != VBAD) { 948 VOP_UNLOCK(vp, 0, td); 949 vgonel(vp, td); 950 VI_LOCK(vp); 951 } else { 952 VOP_UNLOCK(vp, 0, td); 953 } 954 vn_finished_write(vnmp); 955 956#ifdef INVARIANTS 957 { 958 if (vp->v_data) 959 panic("cleaned vnode isn't"); 960 if (vp->v_numoutput) 961 panic("Clean vnode has pending I/O's"); 962 if (vp->v_writecount != 0) 963 panic("Non-zero write count"); 964 } 965#endif 966 if ((pollinfo = vp->v_pollinfo) != NULL) { 967 /* 968 * To avoid lock order reversals, the call to 969 * uma_zfree() must be delayed until the vnode 970 * interlock is released. 971 */ 972 vp->v_pollinfo = NULL; 973 } 974#ifdef MAC 975 mac_destroy_vnode(vp); 976#endif 977 vp->v_iflag = 0; 978 vp->v_vflag = 0; 979 vp->v_lastw = 0; 980 vp->v_lasta = 0; 981 vp->v_cstart = 0; 982 vp->v_clen = 0; 983 vp->v_socket = 0; 984 lockdestroy(vp->v_vnlock); 985 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 986 KASSERT(vp->v_cleanbufcnt == 0, ("cleanbufcnt not 0")); 987 KASSERT(vp->v_cleanblkroot == NULL, ("cleanblkroot not NULL")); 988 KASSERT(vp->v_dirtybufcnt == 0, ("dirtybufcnt not 0")); 989 KASSERT(vp->v_dirtyblkroot == NULL, ("dirtyblkroot not NULL")); 990 } else { 991 numvnodes++; 992 mtx_unlock(&vnode_free_list_mtx); 993 994 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); 995 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 996 VI_LOCK(vp); 997 vp->v_dd = vp; 998 vp->v_vnlock = &vp->v_lock; 999 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 1000 cache_purge(vp); 1001 LIST_INIT(&vp->v_cache_src); 1002 TAILQ_INIT(&vp->v_cache_dst); 1003 } 1004 1005 TAILQ_INIT(&vp->v_cleanblkhd); 1006 TAILQ_INIT(&vp->v_dirtyblkhd); 1007 vp->v_type = VNON; 1008 vp->v_tag = tag; 1009 vp->v_op = vops; 1010 *vpp = vp; 1011 vp->v_usecount = 1; 1012 vp->v_data = 0; 1013 vp->v_cachedid = -1; 1014 VI_UNLOCK(vp); 1015 if (pollinfo != NULL) { 1016 mtx_destroy(&pollinfo->vpi_lock); 1017 uma_zfree(vnodepoll_zone, pollinfo); 1018 } 1019#ifdef MAC 1020 mac_init_vnode(vp); 1021 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 1022 mac_associate_vnode_singlelabel(mp, vp); 1023#endif 1024 insmntque(vp, mp); 1025 1026 return (0); 1027} 1028 1029/* 1030 * Move a vnode from one mount queue to another. 1031 */ 1032static void 1033insmntque(vp, mp) 1034 register struct vnode *vp; 1035 register struct mount *mp; 1036{ 1037 1038 mtx_lock(&mntvnode_mtx); 1039 /* 1040 * Delete from old mount point vnode list, if on one. 1041 */ 1042 if (vp->v_mount != NULL) { 1043 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0, 1044 ("bad mount point vnode list size")); 1045 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes); 1046 vp->v_mount->mnt_nvnodelistsize--; 1047 } 1048 /* 1049 * Insert into list of vnodes for the new mount point, if available. 1050 */ 1051 if ((vp->v_mount = mp) == NULL) { 1052 mtx_unlock(&mntvnode_mtx); 1053 return; 1054 } 1055 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1056 mp->mnt_nvnodelistsize++; 1057 mtx_unlock(&mntvnode_mtx); 1058} 1059 1060/* 1061 * Update outstanding I/O count and do wakeup if requested. 1062 */ 1063void 1064vwakeup(bp) 1065 register struct buf *bp; 1066{ 1067 register struct vnode *vp; 1068 1069 bp->b_flags &= ~B_WRITEINPROG; 1070 if ((vp = bp->b_vp)) { 1071 VI_LOCK(vp); 1072 vp->v_numoutput--; 1073 if (vp->v_numoutput < 0) 1074 panic("vwakeup: neg numoutput"); 1075 if ((vp->v_numoutput == 0) && (vp->v_iflag & VI_BWAIT)) { 1076 vp->v_iflag &= ~VI_BWAIT; 1077 wakeup(&vp->v_numoutput); 1078 } 1079 VI_UNLOCK(vp); 1080 } 1081} 1082 1083/* 1084 * Flush out and invalidate all buffers associated with a vnode. 1085 * Called with the underlying object locked. 1086 */ 1087int 1088vinvalbuf(vp, flags, cred, td, slpflag, slptimeo) 1089 struct vnode *vp; 1090 int flags; 1091 struct ucred *cred; 1092 struct thread *td; 1093 int slpflag, slptimeo; 1094{ 1095 struct buf *blist; 1096 int s, error; 1097 vm_object_t object; 1098 1099 GIANT_REQUIRED; 1100 1101 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1102 1103 VI_LOCK(vp); 1104 if (flags & V_SAVE) { 1105 s = splbio(); 1106 while (vp->v_numoutput) { 1107 vp->v_iflag |= VI_BWAIT; 1108 error = msleep(&vp->v_numoutput, VI_MTX(vp), 1109 slpflag | (PRIBIO + 1), "vinvlbuf", slptimeo); 1110 if (error) { 1111 VI_UNLOCK(vp); 1112 splx(s); 1113 return (error); 1114 } 1115 } 1116 if (!TAILQ_EMPTY(&vp->v_dirtyblkhd)) { 1117 splx(s); 1118 VI_UNLOCK(vp); 1119 if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0) 1120 return (error); 1121 /* 1122 * XXX We could save a lock/unlock if this was only 1123 * enabled under INVARIANTS 1124 */ 1125 VI_LOCK(vp); 1126 s = splbio(); 1127 if (vp->v_numoutput > 0 || 1128 !TAILQ_EMPTY(&vp->v_dirtyblkhd)) 1129 panic("vinvalbuf: dirty bufs"); 1130 } 1131 splx(s); 1132 } 1133 s = splbio(); 1134 /* 1135 * If you alter this loop please notice that interlock is dropped and 1136 * reacquired in flushbuflist. Special care is needed to ensure that 1137 * no race conditions occur from this. 1138 */ 1139 for (error = 0;;) { 1140 if ((blist = TAILQ_FIRST(&vp->v_cleanblkhd)) != 0 && 1141 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 1142 if (error) 1143 break; 1144 continue; 1145 } 1146 if ((blist = TAILQ_FIRST(&vp->v_dirtyblkhd)) != 0 && 1147 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 1148 if (error) 1149 break; 1150 continue; 1151 } 1152 break; 1153 } 1154 if (error) { 1155 splx(s); 1156 VI_UNLOCK(vp); 1157 return (error); 1158 } 1159 1160 /* 1161 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1162 * have write I/O in-progress but if there is a VM object then the 1163 * VM object can also have read-I/O in-progress. 1164 */ 1165 do { 1166 while (vp->v_numoutput > 0) { 1167 vp->v_iflag |= VI_BWAIT; 1168 msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vnvlbv", 0); 1169 } 1170 VI_UNLOCK(vp); 1171 if (VOP_GETVOBJECT(vp, &object) == 0) { 1172 VM_OBJECT_LOCK(object); 1173 vm_object_pip_wait(object, "vnvlbx"); 1174 VM_OBJECT_UNLOCK(object); 1175 } 1176 VI_LOCK(vp); 1177 } while (vp->v_numoutput > 0); 1178 VI_UNLOCK(vp); 1179 1180 splx(s); 1181 1182 /* 1183 * Destroy the copy in the VM cache, too. 1184 */ 1185 if (VOP_GETVOBJECT(vp, &object) == 0) { 1186 VM_OBJECT_LOCK(object); 1187 vm_object_page_remove(object, 0, 0, 1188 (flags & V_SAVE) ? TRUE : FALSE); 1189 VM_OBJECT_UNLOCK(object); 1190 } 1191 1192#ifdef INVARIANTS 1193 VI_LOCK(vp); 1194 if ((flags & (V_ALT | V_NORMAL)) == 0 && 1195 (!TAILQ_EMPTY(&vp->v_dirtyblkhd) || 1196 !TAILQ_EMPTY(&vp->v_cleanblkhd))) 1197 panic("vinvalbuf: flush failed"); 1198 VI_UNLOCK(vp); 1199#endif 1200 return (0); 1201} 1202 1203/* 1204 * Flush out buffers on the specified list. 1205 * 1206 */ 1207static int 1208flushbuflist(blist, flags, vp, slpflag, slptimeo, errorp) 1209 struct buf *blist; 1210 int flags; 1211 struct vnode *vp; 1212 int slpflag, slptimeo; 1213 int *errorp; 1214{ 1215 struct buf *bp, *nbp; 1216 int found, error; 1217 1218 ASSERT_VI_LOCKED(vp, "flushbuflist"); 1219 1220 for (found = 0, bp = blist; bp; bp = nbp) { 1221 nbp = TAILQ_NEXT(bp, b_vnbufs); 1222 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1223 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1224 continue; 1225 } 1226 found += 1; 1227 error = BUF_TIMELOCK(bp, 1228 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp), 1229 "flushbuf", slpflag, slptimeo); 1230 if (error) { 1231 if (error != ENOLCK) 1232 *errorp = error; 1233 goto done; 1234 } 1235 /* 1236 * XXX Since there are no node locks for NFS, I 1237 * believe there is a slight chance that a delayed 1238 * write will occur while sleeping just above, so 1239 * check for it. Note that vfs_bio_awrite expects 1240 * buffers to reside on a queue, while BUF_WRITE and 1241 * brelse do not. 1242 */ 1243 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1244 (flags & V_SAVE)) { 1245 1246 if (bp->b_vp == vp) { 1247 if (bp->b_flags & B_CLUSTEROK) { 1248 vfs_bio_awrite(bp); 1249 } else { 1250 bremfree(bp); 1251 bp->b_flags |= B_ASYNC; 1252 BUF_WRITE(bp); 1253 } 1254 } else { 1255 bremfree(bp); 1256 (void) BUF_WRITE(bp); 1257 } 1258 goto done; 1259 } 1260 bremfree(bp); 1261 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 1262 bp->b_flags &= ~B_ASYNC; 1263 brelse(bp); 1264 VI_LOCK(vp); 1265 } 1266 return (found); 1267done: 1268 VI_LOCK(vp); 1269 return (found); 1270} 1271 1272/* 1273 * Truncate a file's buffer and pages to a specified length. This 1274 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1275 * sync activity. 1276 */ 1277int 1278vtruncbuf(vp, cred, td, length, blksize) 1279 register struct vnode *vp; 1280 struct ucred *cred; 1281 struct thread *td; 1282 off_t length; 1283 int blksize; 1284{ 1285 register struct buf *bp; 1286 struct buf *nbp; 1287 int s, anyfreed; 1288 int trunclbn; 1289 1290 /* 1291 * Round up to the *next* lbn. 1292 */ 1293 trunclbn = (length + blksize - 1) / blksize; 1294 1295 s = splbio(); 1296 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1297restart: 1298 VI_LOCK(vp); 1299 anyfreed = 1; 1300 for (;anyfreed;) { 1301 anyfreed = 0; 1302 for (bp = TAILQ_FIRST(&vp->v_cleanblkhd); bp; bp = nbp) { 1303 nbp = TAILQ_NEXT(bp, b_vnbufs); 1304 if (bp->b_lblkno >= trunclbn) { 1305 if (BUF_LOCK(bp, 1306 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1307 VI_MTX(vp)) == ENOLCK) 1308 goto restart; 1309 1310 bremfree(bp); 1311 bp->b_flags |= (B_INVAL | B_RELBUF); 1312 bp->b_flags &= ~B_ASYNC; 1313 brelse(bp); 1314 anyfreed = 1; 1315 1316 if (nbp && 1317 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1318 (nbp->b_vp != vp) || 1319 (nbp->b_flags & B_DELWRI))) { 1320 goto restart; 1321 } 1322 VI_LOCK(vp); 1323 } 1324 } 1325 1326 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1327 nbp = TAILQ_NEXT(bp, b_vnbufs); 1328 if (bp->b_lblkno >= trunclbn) { 1329 if (BUF_LOCK(bp, 1330 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1331 VI_MTX(vp)) == ENOLCK) 1332 goto restart; 1333 bremfree(bp); 1334 bp->b_flags |= (B_INVAL | B_RELBUF); 1335 bp->b_flags &= ~B_ASYNC; 1336 brelse(bp); 1337 anyfreed = 1; 1338 if (nbp && 1339 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1340 (nbp->b_vp != vp) || 1341 (nbp->b_flags & B_DELWRI) == 0)) { 1342 goto restart; 1343 } 1344 VI_LOCK(vp); 1345 } 1346 } 1347 } 1348 1349 if (length > 0) { 1350restartsync: 1351 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1352 nbp = TAILQ_NEXT(bp, b_vnbufs); 1353 if (bp->b_lblkno > 0) 1354 continue; 1355 /* 1356 * Since we hold the vnode lock this should only 1357 * fail if we're racing with the buf daemon. 1358 */ 1359 if (BUF_LOCK(bp, 1360 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1361 VI_MTX(vp)) == ENOLCK) { 1362 goto restart; 1363 } 1364 KASSERT((bp->b_flags & B_DELWRI), 1365 ("buf(%p) on dirty queue without DELWRI.", bp)); 1366 1367 bremfree(bp); 1368 bawrite(bp); 1369 VI_LOCK(vp); 1370 goto restartsync; 1371 } 1372 } 1373 1374 while (vp->v_numoutput > 0) { 1375 vp->v_iflag |= VI_BWAIT; 1376 msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vbtrunc", 0); 1377 } 1378 VI_UNLOCK(vp); 1379 splx(s); 1380 1381 vnode_pager_setsize(vp, length); 1382 1383 return (0); 1384} 1385 1386/* 1387 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1388 * a vnode. 1389 * 1390 * NOTE: We have to deal with the special case of a background bitmap 1391 * buffer, a situation where two buffers will have the same logical 1392 * block offset. We want (1) only the foreground buffer to be accessed 1393 * in a lookup and (2) must differentiate between the foreground and 1394 * background buffer in the splay tree algorithm because the splay 1395 * tree cannot normally handle multiple entities with the same 'index'. 1396 * We accomplish this by adding differentiating flags to the splay tree's 1397 * numerical domain. 1398 */ 1399static 1400struct buf * 1401buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1402{ 1403 struct buf dummy; 1404 struct buf *lefttreemax, *righttreemin, *y; 1405 1406 if (root == NULL) 1407 return (NULL); 1408 lefttreemax = righttreemin = &dummy; 1409 for (;;) { 1410 if (lblkno < root->b_lblkno || 1411 (lblkno == root->b_lblkno && 1412 (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1413 if ((y = root->b_left) == NULL) 1414 break; 1415 if (lblkno < y->b_lblkno) { 1416 /* Rotate right. */ 1417 root->b_left = y->b_right; 1418 y->b_right = root; 1419 root = y; 1420 if ((y = root->b_left) == NULL) 1421 break; 1422 } 1423 /* Link into the new root's right tree. */ 1424 righttreemin->b_left = root; 1425 righttreemin = root; 1426 } else if (lblkno > root->b_lblkno || 1427 (lblkno == root->b_lblkno && 1428 (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { 1429 if ((y = root->b_right) == NULL) 1430 break; 1431 if (lblkno > y->b_lblkno) { 1432 /* Rotate left. */ 1433 root->b_right = y->b_left; 1434 y->b_left = root; 1435 root = y; 1436 if ((y = root->b_right) == NULL) 1437 break; 1438 } 1439 /* Link into the new root's left tree. */ 1440 lefttreemax->b_right = root; 1441 lefttreemax = root; 1442 } else { 1443 break; 1444 } 1445 root = y; 1446 } 1447 /* Assemble the new root. */ 1448 lefttreemax->b_right = root->b_left; 1449 righttreemin->b_left = root->b_right; 1450 root->b_left = dummy.b_right; 1451 root->b_right = dummy.b_left; 1452 return (root); 1453} 1454 1455static 1456void 1457buf_vlist_remove(struct buf *bp) 1458{ 1459 struct vnode *vp = bp->b_vp; 1460 struct buf *root; 1461 1462 ASSERT_VI_LOCKED(vp, "buf_vlist_remove"); 1463 if (bp->b_xflags & BX_VNDIRTY) { 1464 if (bp != vp->v_dirtyblkroot) { 1465 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot); 1466 KASSERT(root == bp, ("splay lookup failed during dirty remove")); 1467 } 1468 if (bp->b_left == NULL) { 1469 root = bp->b_right; 1470 } else { 1471 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1472 root->b_right = bp->b_right; 1473 } 1474 vp->v_dirtyblkroot = root; 1475 TAILQ_REMOVE(&vp->v_dirtyblkhd, bp, b_vnbufs); 1476 vp->v_dirtybufcnt--; 1477 } else { 1478 /* KASSERT(bp->b_xflags & BX_VNCLEAN, ("bp wasn't clean")); */ 1479 if (bp != vp->v_cleanblkroot) { 1480 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot); 1481 KASSERT(root == bp, ("splay lookup failed during clean remove")); 1482 } 1483 if (bp->b_left == NULL) { 1484 root = bp->b_right; 1485 } else { 1486 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1487 root->b_right = bp->b_right; 1488 } 1489 vp->v_cleanblkroot = root; 1490 TAILQ_REMOVE(&vp->v_cleanblkhd, bp, b_vnbufs); 1491 vp->v_cleanbufcnt--; 1492 } 1493 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1494} 1495 1496/* 1497 * Add the buffer to the sorted clean or dirty block list using a 1498 * splay tree algorithm. 1499 * 1500 * NOTE: xflags is passed as a constant, optimizing this inline function! 1501 */ 1502static 1503void 1504buf_vlist_add(struct buf *bp, struct vnode *vp, b_xflags_t xflags) 1505{ 1506 struct buf *root; 1507 1508 ASSERT_VI_LOCKED(vp, "buf_vlist_add"); 1509 bp->b_xflags |= xflags; 1510 if (xflags & BX_VNDIRTY) { 1511 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot); 1512 if (root == NULL) { 1513 bp->b_left = NULL; 1514 bp->b_right = NULL; 1515 TAILQ_INSERT_TAIL(&vp->v_dirtyblkhd, bp, b_vnbufs); 1516 } else if (bp->b_lblkno < root->b_lblkno || 1517 (bp->b_lblkno == root->b_lblkno && 1518 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1519 bp->b_left = root->b_left; 1520 bp->b_right = root; 1521 root->b_left = NULL; 1522 TAILQ_INSERT_BEFORE(root, bp, b_vnbufs); 1523 } else { 1524 bp->b_right = root->b_right; 1525 bp->b_left = root; 1526 root->b_right = NULL; 1527 TAILQ_INSERT_AFTER(&vp->v_dirtyblkhd, 1528 root, bp, b_vnbufs); 1529 } 1530 vp->v_dirtybufcnt++; 1531 vp->v_dirtyblkroot = bp; 1532 } else { 1533 /* KASSERT(xflags & BX_VNCLEAN, ("xflags not clean")); */ 1534 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot); 1535 if (root == NULL) { 1536 bp->b_left = NULL; 1537 bp->b_right = NULL; 1538 TAILQ_INSERT_TAIL(&vp->v_cleanblkhd, bp, b_vnbufs); 1539 } else if (bp->b_lblkno < root->b_lblkno || 1540 (bp->b_lblkno == root->b_lblkno && 1541 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1542 bp->b_left = root->b_left; 1543 bp->b_right = root; 1544 root->b_left = NULL; 1545 TAILQ_INSERT_BEFORE(root, bp, b_vnbufs); 1546 } else { 1547 bp->b_right = root->b_right; 1548 bp->b_left = root; 1549 root->b_right = NULL; 1550 TAILQ_INSERT_AFTER(&vp->v_cleanblkhd, 1551 root, bp, b_vnbufs); 1552 } 1553 vp->v_cleanbufcnt++; 1554 vp->v_cleanblkroot = bp; 1555 } 1556} 1557 1558/* 1559 * Lookup a buffer using the splay tree. Note that we specifically avoid 1560 * shadow buffers used in background bitmap writes. 1561 * 1562 * This code isn't quite efficient as it could be because we are maintaining 1563 * two sorted lists and do not know which list the block resides in. 1564 */ 1565struct buf * 1566gbincore(struct vnode *vp, daddr_t lblkno) 1567{ 1568 struct buf *bp; 1569 1570 GIANT_REQUIRED; 1571 1572 ASSERT_VI_LOCKED(vp, "gbincore"); 1573 bp = vp->v_cleanblkroot = buf_splay(lblkno, 0, vp->v_cleanblkroot); 1574 if (bp && bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1575 return(bp); 1576 bp = vp->v_dirtyblkroot = buf_splay(lblkno, 0, vp->v_dirtyblkroot); 1577 if (bp && bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1578 return(bp); 1579 return(NULL); 1580} 1581 1582/* 1583 * Associate a buffer with a vnode. 1584 */ 1585void 1586bgetvp(vp, bp) 1587 register struct vnode *vp; 1588 register struct buf *bp; 1589{ 1590 int s; 1591 1592 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 1593 1594 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1595 ("bgetvp: bp already attached! %p", bp)); 1596 1597 ASSERT_VI_LOCKED(vp, "bgetvp"); 1598 vholdl(vp); 1599 bp->b_vp = vp; 1600 bp->b_dev = vn_todev(vp); 1601 /* 1602 * Insert onto list for new vnode. 1603 */ 1604 s = splbio(); 1605 buf_vlist_add(bp, vp, BX_VNCLEAN); 1606 splx(s); 1607} 1608 1609/* 1610 * Disassociate a buffer from a vnode. 1611 */ 1612void 1613brelvp(bp) 1614 register struct buf *bp; 1615{ 1616 struct vnode *vp; 1617 int s; 1618 1619 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1620 1621 /* 1622 * Delete from old vnode list, if on one. 1623 */ 1624 vp = bp->b_vp; 1625 s = splbio(); 1626 VI_LOCK(vp); 1627 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1628 buf_vlist_remove(bp); 1629 if ((vp->v_iflag & VI_ONWORKLST) && TAILQ_EMPTY(&vp->v_dirtyblkhd)) { 1630 vp->v_iflag &= ~VI_ONWORKLST; 1631 mtx_lock(&sync_mtx); 1632 LIST_REMOVE(vp, v_synclist); 1633 mtx_unlock(&sync_mtx); 1634 } 1635 vdropl(vp); 1636 VI_UNLOCK(vp); 1637 bp->b_vp = (struct vnode *) 0; 1638 if (bp->b_object) 1639 bp->b_object = NULL; 1640 splx(s); 1641} 1642 1643/* 1644 * Add an item to the syncer work queue. 1645 */ 1646static void 1647vn_syncer_add_to_worklist(struct vnode *vp, int delay) 1648{ 1649 int s, slot; 1650 1651 s = splbio(); 1652 ASSERT_VI_LOCKED(vp, "vn_syncer_add_to_worklist"); 1653 1654 mtx_lock(&sync_mtx); 1655 if (vp->v_iflag & VI_ONWORKLST) 1656 LIST_REMOVE(vp, v_synclist); 1657 else 1658 vp->v_iflag |= VI_ONWORKLST; 1659 1660 if (delay > syncer_maxdelay - 2) 1661 delay = syncer_maxdelay - 2; 1662 slot = (syncer_delayno + delay) & syncer_mask; 1663 1664 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist); 1665 mtx_unlock(&sync_mtx); 1666 1667 splx(s); 1668} 1669 1670struct proc *updateproc; 1671static void sched_sync(void); 1672static struct kproc_desc up_kp = { 1673 "syncer", 1674 sched_sync, 1675 &updateproc 1676}; 1677SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 1678 1679/* 1680 * System filesystem synchronizer daemon. 1681 */ 1682static void 1683sched_sync(void) 1684{ 1685 struct synclist *slp; 1686 struct vnode *vp; 1687 struct mount *mp; 1688 long starttime; 1689 int s; 1690 struct thread *td = FIRST_THREAD_IN_PROC(updateproc); /* XXXKSE */ 1691 1692 mtx_lock(&Giant); 1693 1694 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, td->td_proc, 1695 SHUTDOWN_PRI_LAST); 1696 1697 for (;;) { 1698 kthread_suspend_check(td->td_proc); 1699 1700 starttime = time_second; 1701 1702 /* 1703 * Push files whose dirty time has expired. Be careful 1704 * of interrupt race on slp queue. 1705 */ 1706 s = splbio(); 1707 mtx_lock(&sync_mtx); 1708 slp = &syncer_workitem_pending[syncer_delayno]; 1709 syncer_delayno += 1; 1710 if (syncer_delayno == syncer_maxdelay) 1711 syncer_delayno = 0; 1712 splx(s); 1713 1714 while ((vp = LIST_FIRST(slp)) != NULL) { 1715 mtx_unlock(&sync_mtx); 1716 if (VOP_ISLOCKED(vp, NULL) == 0 && 1717 vn_start_write(vp, &mp, V_NOWAIT) == 0) { 1718 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 1719 (void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td); 1720 VOP_UNLOCK(vp, 0, td); 1721 vn_finished_write(mp); 1722 } 1723 s = splbio(); 1724 mtx_lock(&sync_mtx); 1725 if (LIST_FIRST(slp) == vp) { 1726 mtx_unlock(&sync_mtx); 1727 /* 1728 * Note: VFS vnodes can remain on the 1729 * worklist too with no dirty blocks, but 1730 * since sync_fsync() moves it to a different 1731 * slot we are safe. 1732 */ 1733 VI_LOCK(vp); 1734 if (TAILQ_EMPTY(&vp->v_dirtyblkhd) && 1735 !vn_isdisk(vp, NULL)) { 1736 panic("sched_sync: fsync failed " 1737 "vp %p tag %s", vp, vp->v_tag); 1738 } 1739 /* 1740 * Put us back on the worklist. The worklist 1741 * routine will remove us from our current 1742 * position and then add us back in at a later 1743 * position. 1744 */ 1745 vn_syncer_add_to_worklist(vp, syncdelay); 1746 VI_UNLOCK(vp); 1747 mtx_lock(&sync_mtx); 1748 } 1749 splx(s); 1750 } 1751 mtx_unlock(&sync_mtx); 1752 1753 /* 1754 * Do soft update processing. 1755 */ 1756 if (softdep_process_worklist_hook != NULL) 1757 (*softdep_process_worklist_hook)(NULL); 1758 1759 /* 1760 * The variable rushjob allows the kernel to speed up the 1761 * processing of the filesystem syncer process. A rushjob 1762 * value of N tells the filesystem syncer to process the next 1763 * N seconds worth of work on its queue ASAP. Currently rushjob 1764 * is used by the soft update code to speed up the filesystem 1765 * syncer process when the incore state is getting so far 1766 * ahead of the disk that the kernel memory pool is being 1767 * threatened with exhaustion. 1768 */ 1769 mtx_lock(&sync_mtx); 1770 if (rushjob > 0) { 1771 rushjob -= 1; 1772 mtx_unlock(&sync_mtx); 1773 continue; 1774 } 1775 mtx_unlock(&sync_mtx); 1776 /* 1777 * If it has taken us less than a second to process the 1778 * current work, then wait. Otherwise start right over 1779 * again. We can still lose time if any single round 1780 * takes more than two seconds, but it does not really 1781 * matter as we are just trying to generally pace the 1782 * filesystem activity. 1783 */ 1784 if (time_second == starttime) 1785 tsleep(&lbolt, PPAUSE, "syncer", 0); 1786 } 1787} 1788 1789/* 1790 * Request the syncer daemon to speed up its work. 1791 * We never push it to speed up more than half of its 1792 * normal turn time, otherwise it could take over the cpu. 1793 * XXXKSE only one update? 1794 */ 1795int 1796speedup_syncer() 1797{ 1798 struct thread *td; 1799 int ret = 0; 1800 1801 td = FIRST_THREAD_IN_PROC(updateproc); 1802 mtx_lock_spin(&sched_lock); 1803 if (td->td_wchan == &lbolt) { 1804 unsleep(td); 1805 TD_CLR_SLEEPING(td); 1806 setrunnable(td); 1807 } 1808 mtx_unlock_spin(&sched_lock); 1809 mtx_lock(&sync_mtx); 1810 if (rushjob < syncdelay / 2) { 1811 rushjob += 1; 1812 stat_rush_requests += 1; 1813 ret = 1; 1814 } 1815 mtx_unlock(&sync_mtx); 1816 return (ret); 1817} 1818 1819/* 1820 * Associate a p-buffer with a vnode. 1821 * 1822 * Also sets B_PAGING flag to indicate that vnode is not fully associated 1823 * with the buffer. i.e. the bp has not been linked into the vnode or 1824 * ref-counted. 1825 */ 1826void 1827pbgetvp(vp, bp) 1828 register struct vnode *vp; 1829 register struct buf *bp; 1830{ 1831 1832 KASSERT(bp->b_vp == NULL, ("pbgetvp: not free")); 1833 1834 bp->b_vp = vp; 1835 bp->b_flags |= B_PAGING; 1836 bp->b_dev = vn_todev(vp); 1837} 1838 1839/* 1840 * Disassociate a p-buffer from a vnode. 1841 */ 1842void 1843pbrelvp(bp) 1844 register struct buf *bp; 1845{ 1846 1847 KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL")); 1848 1849 /* XXX REMOVE ME */ 1850 VI_LOCK(bp->b_vp); 1851 if (TAILQ_NEXT(bp, b_vnbufs) != NULL) { 1852 panic( 1853 "relpbuf(): b_vp was probably reassignbuf()d %p %x", 1854 bp, 1855 (int)bp->b_flags 1856 ); 1857 } 1858 VI_UNLOCK(bp->b_vp); 1859 bp->b_vp = (struct vnode *) 0; 1860 bp->b_flags &= ~B_PAGING; 1861} 1862 1863/* 1864 * Reassign a buffer from one vnode to another. 1865 * Used to assign file specific control information 1866 * (indirect blocks) to the vnode to which they belong. 1867 */ 1868void 1869reassignbuf(bp, newvp) 1870 register struct buf *bp; 1871 register struct vnode *newvp; 1872{ 1873 int delay; 1874 int s; 1875 1876 if (newvp == NULL) { 1877 printf("reassignbuf: NULL"); 1878 return; 1879 } 1880 ++reassignbufcalls; 1881 1882 /* 1883 * B_PAGING flagged buffers cannot be reassigned because their vp 1884 * is not fully linked in. 1885 */ 1886 if (bp->b_flags & B_PAGING) 1887 panic("cannot reassign paging buffer"); 1888 1889 s = splbio(); 1890 /* 1891 * Delete from old vnode list, if on one. 1892 */ 1893 VI_LOCK(bp->b_vp); 1894 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) { 1895 buf_vlist_remove(bp); 1896 if (bp->b_vp != newvp) { 1897 vdropl(bp->b_vp); 1898 bp->b_vp = NULL; /* for clarification */ 1899 } 1900 } 1901 VI_UNLOCK(bp->b_vp); 1902 /* 1903 * If dirty, put on list of dirty buffers; otherwise insert onto list 1904 * of clean buffers. 1905 */ 1906 VI_LOCK(newvp); 1907 if (bp->b_flags & B_DELWRI) { 1908 if ((newvp->v_iflag & VI_ONWORKLST) == 0) { 1909 switch (newvp->v_type) { 1910 case VDIR: 1911 delay = dirdelay; 1912 break; 1913 case VCHR: 1914 if (newvp->v_rdev->si_mountpoint != NULL) { 1915 delay = metadelay; 1916 break; 1917 } 1918 /* FALLTHROUGH */ 1919 default: 1920 delay = filedelay; 1921 } 1922 vn_syncer_add_to_worklist(newvp, delay); 1923 } 1924 buf_vlist_add(bp, newvp, BX_VNDIRTY); 1925 } else { 1926 buf_vlist_add(bp, newvp, BX_VNCLEAN); 1927 1928 if ((newvp->v_iflag & VI_ONWORKLST) && 1929 TAILQ_EMPTY(&newvp->v_dirtyblkhd)) { 1930 mtx_lock(&sync_mtx); 1931 LIST_REMOVE(newvp, v_synclist); 1932 mtx_unlock(&sync_mtx); 1933 newvp->v_iflag &= ~VI_ONWORKLST; 1934 } 1935 } 1936 if (bp->b_vp != newvp) { 1937 bp->b_vp = newvp; 1938 vholdl(bp->b_vp); 1939 } 1940 VI_UNLOCK(newvp); 1941 splx(s); 1942} 1943 1944/* 1945 * Create a vnode for a device. 1946 * Used for mounting the root filesystem. 1947 */ 1948int 1949bdevvp(dev, vpp) 1950 dev_t dev; 1951 struct vnode **vpp; 1952{ 1953 register struct vnode *vp; 1954 struct vnode *nvp; 1955 int error; 1956 1957 if (dev == NODEV) { 1958 *vpp = NULLVP; 1959 return (ENXIO); 1960 } 1961 if (vfinddev(dev, VCHR, vpp)) 1962 return (0); 1963 error = getnewvnode("none", (struct mount *)0, spec_vnodeop_p, &nvp); 1964 if (error) { 1965 *vpp = NULLVP; 1966 return (error); 1967 } 1968 vp = nvp; 1969 vp->v_type = VCHR; 1970 addalias(vp, dev); 1971 *vpp = vp; 1972 return (0); 1973} 1974 1975static void 1976v_incr_usecount(struct vnode *vp, int delta) 1977{ 1978 vp->v_usecount += delta; 1979 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1980 mtx_lock(&spechash_mtx); 1981 vp->v_rdev->si_usecount += delta; 1982 mtx_unlock(&spechash_mtx); 1983 } 1984} 1985 1986/* 1987 * Add vnode to the alias list hung off the dev_t. 1988 * 1989 * The reason for this gunk is that multiple vnodes can reference 1990 * the same physical device, so checking vp->v_usecount to see 1991 * how many users there are is inadequate; the v_usecount for 1992 * the vnodes need to be accumulated. vcount() does that. 1993 */ 1994struct vnode * 1995addaliasu(nvp, nvp_rdev) 1996 struct vnode *nvp; 1997 udev_t nvp_rdev; 1998{ 1999 struct vnode *ovp; 2000 vop_t **ops; 2001 dev_t dev; 2002 2003 if (nvp->v_type == VBLK) 2004 return (nvp); 2005 if (nvp->v_type != VCHR) 2006 panic("addaliasu on non-special vnode"); 2007 dev = udev2dev(nvp_rdev, 0); 2008 /* 2009 * Check to see if we have a bdevvp vnode with no associated 2010 * filesystem. If so, we want to associate the filesystem of 2011 * the new newly instigated vnode with the bdevvp vnode and 2012 * discard the newly created vnode rather than leaving the 2013 * bdevvp vnode lying around with no associated filesystem. 2014 */ 2015 if (vfinddev(dev, nvp->v_type, &ovp) == 0 || ovp->v_data != NULL) { 2016 addalias(nvp, dev); 2017 return (nvp); 2018 } 2019 /* 2020 * Discard unneeded vnode, but save its node specific data. 2021 * Note that if there is a lock, it is carried over in the 2022 * node specific data to the replacement vnode. 2023 */ 2024 vref(ovp); 2025 ovp->v_data = nvp->v_data; 2026 ovp->v_tag = nvp->v_tag; 2027 nvp->v_data = NULL; 2028 lockdestroy(ovp->v_vnlock); 2029 lockinit(ovp->v_vnlock, PVFS, nvp->v_vnlock->lk_wmesg, 2030 nvp->v_vnlock->lk_timo, nvp->v_vnlock->lk_flags & LK_EXTFLG_MASK); 2031 ops = ovp->v_op; 2032 ovp->v_op = nvp->v_op; 2033 if (VOP_ISLOCKED(nvp, curthread)) { 2034 VOP_UNLOCK(nvp, 0, curthread); 2035 vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread); 2036 } 2037 nvp->v_op = ops; 2038 insmntque(ovp, nvp->v_mount); 2039 vrele(nvp); 2040 vgone(nvp); 2041 return (ovp); 2042} 2043 2044/* This is a local helper function that do the same as addaliasu, but for a 2045 * dev_t instead of an udev_t. */ 2046static void 2047addalias(nvp, dev) 2048 struct vnode *nvp; 2049 dev_t dev; 2050{ 2051 2052 KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode")); 2053 nvp->v_rdev = dev; 2054 VI_LOCK(nvp); 2055 mtx_lock(&spechash_mtx); 2056 SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext); 2057 dev->si_usecount += nvp->v_usecount; 2058 mtx_unlock(&spechash_mtx); 2059 VI_UNLOCK(nvp); 2060} 2061 2062/* 2063 * Grab a particular vnode from the free list, increment its 2064 * reference count and lock it. The vnode lock bit is set if the 2065 * vnode is being eliminated in vgone. The process is awakened 2066 * when the transition is completed, and an error returned to 2067 * indicate that the vnode is no longer usable (possibly having 2068 * been changed to a new filesystem type). 2069 */ 2070int 2071vget(vp, flags, td) 2072 register struct vnode *vp; 2073 int flags; 2074 struct thread *td; 2075{ 2076 int error; 2077 2078 /* 2079 * If the vnode is in the process of being cleaned out for 2080 * another use, we wait for the cleaning to finish and then 2081 * return failure. Cleaning is determined by checking that 2082 * the VI_XLOCK flag is set. 2083 */ 2084 if ((flags & LK_INTERLOCK) == 0) 2085 VI_LOCK(vp); 2086 if (vp->v_iflag & VI_XLOCK && vp->v_vxproc != curthread) { 2087 vp->v_iflag |= VI_XWANT; 2088 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0); 2089 return (ENOENT); 2090 } 2091 2092 v_incr_usecount(vp, 1); 2093 2094 if (VSHOULDBUSY(vp)) 2095 vbusy(vp); 2096 if (flags & LK_TYPE_MASK) { 2097 if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { 2098 /* 2099 * must expand vrele here because we do not want 2100 * to call VOP_INACTIVE if the reference count 2101 * drops back to zero since it was never really 2102 * active. We must remove it from the free list 2103 * before sleeping so that multiple processes do 2104 * not try to recycle it. 2105 */ 2106 VI_LOCK(vp); 2107 v_incr_usecount(vp, -1); 2108 if (VSHOULDFREE(vp)) 2109 vfree(vp); 2110 else 2111 vlruvp(vp); 2112 VI_UNLOCK(vp); 2113 } 2114 return (error); 2115 } 2116 VI_UNLOCK(vp); 2117 return (0); 2118} 2119 2120/* 2121 * Increase the reference count of a vnode. 2122 */ 2123void 2124vref(struct vnode *vp) 2125{ 2126 VI_LOCK(vp); 2127 v_incr_usecount(vp, 1); 2128 VI_UNLOCK(vp); 2129} 2130 2131/* 2132 * Return reference count of a vnode. 2133 * 2134 * The results of this call are only guaranteed when some mechanism other 2135 * than the VI lock is used to stop other processes from gaining references 2136 * to the vnode. This may be the case if the caller holds the only reference. 2137 * This is also useful when stale data is acceptable as race conditions may 2138 * be accounted for by some other means. 2139 */ 2140int 2141vrefcnt(struct vnode *vp) 2142{ 2143 int usecnt; 2144 2145 VI_LOCK(vp); 2146 usecnt = vp->v_usecount; 2147 VI_UNLOCK(vp); 2148 2149 return (usecnt); 2150} 2151 2152 2153/* 2154 * Vnode put/release. 2155 * If count drops to zero, call inactive routine and return to freelist. 2156 */ 2157void 2158vrele(vp) 2159 struct vnode *vp; 2160{ 2161 struct thread *td = curthread; /* XXX */ 2162 2163 KASSERT(vp != NULL, ("vrele: null vp")); 2164 2165 VI_LOCK(vp); 2166 2167 /* Skip this v_writecount check if we're going to panic below. */ 2168 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2169 ("vrele: missed vn_close")); 2170 2171 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2172 vp->v_usecount == 1)) { 2173 v_incr_usecount(vp, -1); 2174 VI_UNLOCK(vp); 2175 2176 return; 2177 } 2178 2179 if (vp->v_usecount == 1) { 2180 v_incr_usecount(vp, -1); 2181 /* 2182 * We must call VOP_INACTIVE with the node locked. Mark 2183 * as VI_DOINGINACT to avoid recursion. 2184 */ 2185 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) { 2186 VI_LOCK(vp); 2187 vp->v_iflag |= VI_DOINGINACT; 2188 VI_UNLOCK(vp); 2189 VOP_INACTIVE(vp, td); 2190 VI_LOCK(vp); 2191 KASSERT(vp->v_iflag & VI_DOINGINACT, 2192 ("vrele: lost VI_DOINGINACT")); 2193 vp->v_iflag &= ~VI_DOINGINACT; 2194 VI_UNLOCK(vp); 2195 } 2196 VI_LOCK(vp); 2197 if (VSHOULDFREE(vp)) 2198 vfree(vp); 2199 else 2200 vlruvp(vp); 2201 VI_UNLOCK(vp); 2202 2203 } else { 2204#ifdef DIAGNOSTIC 2205 vprint("vrele: negative ref count", vp); 2206#endif 2207 VI_UNLOCK(vp); 2208 panic("vrele: negative ref cnt"); 2209 } 2210} 2211 2212/* 2213 * Release an already locked vnode. This give the same effects as 2214 * unlock+vrele(), but takes less time and avoids releasing and 2215 * re-aquiring the lock (as vrele() aquires the lock internally.) 2216 */ 2217void 2218vput(vp) 2219 struct vnode *vp; 2220{ 2221 struct thread *td = curthread; /* XXX */ 2222 2223 GIANT_REQUIRED; 2224 2225 KASSERT(vp != NULL, ("vput: null vp")); 2226 VI_LOCK(vp); 2227 /* Skip this v_writecount check if we're going to panic below. */ 2228 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2229 ("vput: missed vn_close")); 2230 2231 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2232 vp->v_usecount == 1)) { 2233 v_incr_usecount(vp, -1); 2234 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2235 return; 2236 } 2237 2238 if (vp->v_usecount == 1) { 2239 v_incr_usecount(vp, -1); 2240 /* 2241 * We must call VOP_INACTIVE with the node locked, so 2242 * we just need to release the vnode mutex. Mark as 2243 * as VI_DOINGINACT to avoid recursion. 2244 */ 2245 vp->v_iflag |= VI_DOINGINACT; 2246 VI_UNLOCK(vp); 2247 VOP_INACTIVE(vp, td); 2248 VI_LOCK(vp); 2249 KASSERT(vp->v_iflag & VI_DOINGINACT, 2250 ("vput: lost VI_DOINGINACT")); 2251 vp->v_iflag &= ~VI_DOINGINACT; 2252 if (VSHOULDFREE(vp)) 2253 vfree(vp); 2254 else 2255 vlruvp(vp); 2256 VI_UNLOCK(vp); 2257 2258 } else { 2259#ifdef DIAGNOSTIC 2260 vprint("vput: negative ref count", vp); 2261#endif 2262 panic("vput: negative ref cnt"); 2263 } 2264} 2265 2266/* 2267 * Somebody doesn't want the vnode recycled. 2268 */ 2269void 2270vhold(struct vnode *vp) 2271{ 2272 VI_LOCK(vp); 2273 vholdl(vp); 2274 VI_UNLOCK(vp); 2275} 2276 2277void 2278vholdl(vp) 2279 register struct vnode *vp; 2280{ 2281 int s; 2282 2283 s = splbio(); 2284 vp->v_holdcnt++; 2285 if (VSHOULDBUSY(vp)) 2286 vbusy(vp); 2287 splx(s); 2288} 2289 2290/* 2291 * Note that there is one less who cares about this vnode. vdrop() is the 2292 * opposite of vhold(). 2293 */ 2294void 2295vdrop(struct vnode *vp) 2296{ 2297 VI_LOCK(vp); 2298 vdropl(vp); 2299 VI_UNLOCK(vp); 2300} 2301 2302void 2303vdropl(vp) 2304 register struct vnode *vp; 2305{ 2306 int s; 2307 2308 s = splbio(); 2309 if (vp->v_holdcnt <= 0) 2310 panic("vdrop: holdcnt"); 2311 vp->v_holdcnt--; 2312 if (VSHOULDFREE(vp)) 2313 vfree(vp); 2314 else 2315 vlruvp(vp); 2316 splx(s); 2317} 2318 2319/* 2320 * Remove any vnodes in the vnode table belonging to mount point mp. 2321 * 2322 * If FORCECLOSE is not specified, there should not be any active ones, 2323 * return error if any are found (nb: this is a user error, not a 2324 * system error). If FORCECLOSE is specified, detach any active vnodes 2325 * that are found. 2326 * 2327 * If WRITECLOSE is set, only flush out regular file vnodes open for 2328 * writing. 2329 * 2330 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2331 * 2332 * `rootrefs' specifies the base reference count for the root vnode 2333 * of this filesystem. The root vnode is considered busy if its 2334 * v_usecount exceeds this value. On a successful return, vflush() 2335 * will call vrele() on the root vnode exactly rootrefs times. 2336 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2337 * be zero. 2338 */ 2339#ifdef DIAGNOSTIC 2340static int busyprt = 0; /* print out busy vnodes */ 2341SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 2342#endif 2343 2344int 2345vflush(mp, rootrefs, flags) 2346 struct mount *mp; 2347 int rootrefs; 2348 int flags; 2349{ 2350 struct thread *td = curthread; /* XXX */ 2351 struct vnode *vp, *nvp, *rootvp = NULL; 2352 struct vattr vattr; 2353 int busy = 0, error; 2354 2355 if (rootrefs > 0) { 2356 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2357 ("vflush: bad args")); 2358 /* 2359 * Get the filesystem root vnode. We can vput() it 2360 * immediately, since with rootrefs > 0, it won't go away. 2361 */ 2362 if ((error = VFS_ROOT(mp, &rootvp)) != 0) 2363 return (error); 2364 vput(rootvp); 2365 2366 } 2367 mtx_lock(&mntvnode_mtx); 2368loop: 2369 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp; vp = nvp) { 2370 /* 2371 * Make sure this vnode wasn't reclaimed in getnewvnode(). 2372 * Start over if it has (it won't be on the list anymore). 2373 */ 2374 if (vp->v_mount != mp) 2375 goto loop; 2376 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 2377 2378 VI_LOCK(vp); 2379 mtx_unlock(&mntvnode_mtx); 2380 vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_RETRY, td); 2381 /* 2382 * Skip over a vnodes marked VV_SYSTEM. 2383 */ 2384 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2385 VOP_UNLOCK(vp, 0, td); 2386 mtx_lock(&mntvnode_mtx); 2387 continue; 2388 } 2389 /* 2390 * If WRITECLOSE is set, flush out unlinked but still open 2391 * files (even if open only for reading) and regular file 2392 * vnodes open for writing. 2393 */ 2394 if (flags & WRITECLOSE) { 2395 error = VOP_GETATTR(vp, &vattr, td->td_ucred, td); 2396 VI_LOCK(vp); 2397 2398 if ((vp->v_type == VNON || 2399 (error == 0 && vattr.va_nlink > 0)) && 2400 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2401 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2402 mtx_lock(&mntvnode_mtx); 2403 continue; 2404 } 2405 } else 2406 VI_LOCK(vp); 2407 2408 VOP_UNLOCK(vp, 0, td); 2409 2410 /* 2411 * With v_usecount == 0, all we need to do is clear out the 2412 * vnode data structures and we are done. 2413 */ 2414 if (vp->v_usecount == 0) { 2415 vgonel(vp, td); 2416 mtx_lock(&mntvnode_mtx); 2417 continue; 2418 } 2419 2420 /* 2421 * If FORCECLOSE is set, forcibly close the vnode. For block 2422 * or character devices, revert to an anonymous device. For 2423 * all other files, just kill them. 2424 */ 2425 if (flags & FORCECLOSE) { 2426 if (vp->v_type != VCHR) { 2427 vgonel(vp, td); 2428 } else { 2429 vclean(vp, 0, td); 2430 VI_UNLOCK(vp); 2431 vp->v_op = spec_vnodeop_p; 2432 insmntque(vp, (struct mount *) 0); 2433 } 2434 mtx_lock(&mntvnode_mtx); 2435 continue; 2436 } 2437#ifdef DIAGNOSTIC 2438 if (busyprt) 2439 vprint("vflush: busy vnode", vp); 2440#endif 2441 VI_UNLOCK(vp); 2442 mtx_lock(&mntvnode_mtx); 2443 busy++; 2444 } 2445 mtx_unlock(&mntvnode_mtx); 2446 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2447 /* 2448 * If just the root vnode is busy, and if its refcount 2449 * is equal to `rootrefs', then go ahead and kill it. 2450 */ 2451 VI_LOCK(rootvp); 2452 KASSERT(busy > 0, ("vflush: not busy")); 2453 KASSERT(rootvp->v_usecount >= rootrefs, ("vflush: rootrefs")); 2454 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2455 vgonel(rootvp, td); 2456 busy = 0; 2457 } else 2458 VI_UNLOCK(rootvp); 2459 } 2460 if (busy) 2461 return (EBUSY); 2462 for (; rootrefs > 0; rootrefs--) 2463 vrele(rootvp); 2464 return (0); 2465} 2466 2467/* 2468 * This moves a now (likely recyclable) vnode to the end of the 2469 * mountlist. XXX However, it is temporarily disabled until we 2470 * can clean up ffs_sync() and friends, which have loop restart 2471 * conditions which this code causes to operate O(N^2). 2472 */ 2473static void 2474vlruvp(struct vnode *vp) 2475{ 2476#if 0 2477 struct mount *mp; 2478 2479 if ((mp = vp->v_mount) != NULL) { 2480 mtx_lock(&mntvnode_mtx); 2481 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2482 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2483 mtx_unlock(&mntvnode_mtx); 2484 } 2485#endif 2486} 2487 2488/* 2489 * Disassociate the underlying filesystem from a vnode. 2490 */ 2491static void 2492vclean(vp, flags, td) 2493 struct vnode *vp; 2494 int flags; 2495 struct thread *td; 2496{ 2497 int active; 2498 2499 ASSERT_VI_LOCKED(vp, "vclean"); 2500 /* 2501 * Check to see if the vnode is in use. If so we have to reference it 2502 * before we clean it out so that its count cannot fall to zero and 2503 * generate a race against ourselves to recycle it. 2504 */ 2505 if ((active = vp->v_usecount)) 2506 v_incr_usecount(vp, 1); 2507 2508 /* 2509 * Prevent the vnode from being recycled or brought into use while we 2510 * clean it out. 2511 */ 2512 if (vp->v_iflag & VI_XLOCK) 2513 panic("vclean: deadlock"); 2514 vp->v_iflag |= VI_XLOCK; 2515 vp->v_vxproc = curthread; 2516 /* 2517 * Even if the count is zero, the VOP_INACTIVE routine may still 2518 * have the object locked while it cleans it out. The VOP_LOCK 2519 * ensures that the VOP_INACTIVE routine is done with its work. 2520 * For active vnodes, it ensures that no other activity can 2521 * occur while the underlying object is being cleaned out. 2522 */ 2523 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2524 2525 /* 2526 * Clean out any buffers associated with the vnode. 2527 * If the flush fails, just toss the buffers. 2528 */ 2529 if (flags & DOCLOSE) { 2530 struct buf *bp; 2531 VI_LOCK(vp); 2532 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 2533 VI_UNLOCK(vp); 2534 if (bp != NULL) 2535 (void) vn_write_suspend_wait(vp, NULL, V_WAIT); 2536 if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0) 2537 vinvalbuf(vp, 0, NOCRED, td, 0, 0); 2538 } 2539 2540 VOP_DESTROYVOBJECT(vp); 2541 2542 /* 2543 * Any other processes trying to obtain this lock must first 2544 * wait for VXLOCK to clear, then call the new lock operation. 2545 */ 2546 VOP_UNLOCK(vp, 0, td); 2547 2548 /* 2549 * If purging an active vnode, it must be closed and 2550 * deactivated before being reclaimed. Note that the 2551 * VOP_INACTIVE will unlock the vnode. 2552 */ 2553 if (active) { 2554 if (flags & DOCLOSE) 2555 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2556 VI_LOCK(vp); 2557 if ((vp->v_iflag & VI_DOINGINACT) == 0) { 2558 vp->v_iflag |= VI_DOINGINACT; 2559 VI_UNLOCK(vp); 2560 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 2561 panic("vclean: cannot relock."); 2562 VOP_INACTIVE(vp, td); 2563 VI_LOCK(vp); 2564 KASSERT(vp->v_iflag & VI_DOINGINACT, 2565 ("vclean: lost VI_DOINGINACT")); 2566 vp->v_iflag &= ~VI_DOINGINACT; 2567 } 2568 VI_UNLOCK(vp); 2569 } 2570 2571 /* 2572 * Reclaim the vnode. 2573 */ 2574 if (VOP_RECLAIM(vp, td)) 2575 panic("vclean: cannot reclaim"); 2576 2577 if (active) { 2578 /* 2579 * Inline copy of vrele() since VOP_INACTIVE 2580 * has already been called. 2581 */ 2582 VI_LOCK(vp); 2583 v_incr_usecount(vp, -1); 2584 if (vp->v_usecount <= 0) { 2585#ifdef DIAGNOSTIC 2586 if (vp->v_usecount < 0 || vp->v_writecount != 0) { 2587 vprint("vclean: bad ref count", vp); 2588 panic("vclean: ref cnt"); 2589 } 2590#endif 2591 vfree(vp); 2592 } 2593 VI_UNLOCK(vp); 2594 } 2595 2596 cache_purge(vp); 2597 VI_LOCK(vp); 2598 if (VSHOULDFREE(vp)) 2599 vfree(vp); 2600 2601 /* 2602 * Done with purge, reset to the standard lock and 2603 * notify sleepers of the grim news. 2604 */ 2605 vp->v_vnlock = &vp->v_lock; 2606 vp->v_op = dead_vnodeop_p; 2607 if (vp->v_pollinfo != NULL) 2608 vn_pollgone(vp); 2609 vp->v_tag = "none"; 2610 vp->v_iflag &= ~VI_XLOCK; 2611 vp->v_vxproc = NULL; 2612 if (vp->v_iflag & VI_XWANT) { 2613 vp->v_iflag &= ~VI_XWANT; 2614 wakeup(vp); 2615 } 2616} 2617 2618/* 2619 * Eliminate all activity associated with the requested vnode 2620 * and with all vnodes aliased to the requested vnode. 2621 */ 2622int 2623vop_revoke(ap) 2624 struct vop_revoke_args /* { 2625 struct vnode *a_vp; 2626 int a_flags; 2627 } */ *ap; 2628{ 2629 struct vnode *vp, *vq; 2630 dev_t dev; 2631 2632 KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke")); 2633 vp = ap->a_vp; 2634 KASSERT((vp->v_type == VCHR), ("vop_revoke: not VCHR")); 2635 2636 VI_LOCK(vp); 2637 /* 2638 * If a vgone (or vclean) is already in progress, 2639 * wait until it is done and return. 2640 */ 2641 if (vp->v_iflag & VI_XLOCK) { 2642 vp->v_iflag |= VI_XWANT; 2643 msleep(vp, VI_MTX(vp), PINOD | PDROP, 2644 "vop_revokeall", 0); 2645 return (0); 2646 } 2647 VI_UNLOCK(vp); 2648 dev = vp->v_rdev; 2649 for (;;) { 2650 mtx_lock(&spechash_mtx); 2651 vq = SLIST_FIRST(&dev->si_hlist); 2652 mtx_unlock(&spechash_mtx); 2653 if (!vq) 2654 break; 2655 vgone(vq); 2656 } 2657 return (0); 2658} 2659 2660/* 2661 * Recycle an unused vnode to the front of the free list. 2662 * Release the passed interlock if the vnode will be recycled. 2663 */ 2664int 2665vrecycle(vp, inter_lkp, td) 2666 struct vnode *vp; 2667 struct mtx *inter_lkp; 2668 struct thread *td; 2669{ 2670 2671 VI_LOCK(vp); 2672 if (vp->v_usecount == 0) { 2673 if (inter_lkp) { 2674 mtx_unlock(inter_lkp); 2675 } 2676 vgonel(vp, td); 2677 return (1); 2678 } 2679 VI_UNLOCK(vp); 2680 return (0); 2681} 2682 2683/* 2684 * Eliminate all activity associated with a vnode 2685 * in preparation for reuse. 2686 */ 2687void 2688vgone(vp) 2689 register struct vnode *vp; 2690{ 2691 struct thread *td = curthread; /* XXX */ 2692 2693 VI_LOCK(vp); 2694 vgonel(vp, td); 2695} 2696 2697/* 2698 * vgone, with the vp interlock held. 2699 */ 2700void 2701vgonel(vp, td) 2702 struct vnode *vp; 2703 struct thread *td; 2704{ 2705 int s; 2706 2707 /* 2708 * If a vgone (or vclean) is already in progress, 2709 * wait until it is done and return. 2710 */ 2711 ASSERT_VI_LOCKED(vp, "vgonel"); 2712 if (vp->v_iflag & VI_XLOCK) { 2713 vp->v_iflag |= VI_XWANT; 2714 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0); 2715 return; 2716 } 2717 2718 /* 2719 * Clean out the filesystem specific data. 2720 */ 2721 vclean(vp, DOCLOSE, td); 2722 VI_UNLOCK(vp); 2723 2724 /* 2725 * Delete from old mount point vnode list, if on one. 2726 */ 2727 if (vp->v_mount != NULL) 2728 insmntque(vp, (struct mount *)0); 2729 /* 2730 * If special device, remove it from special device alias list 2731 * if it is on one. 2732 */ 2733 if (vp->v_type == VCHR && vp->v_rdev != NULL && vp->v_rdev != NODEV) { 2734 VI_LOCK(vp); 2735 mtx_lock(&spechash_mtx); 2736 SLIST_REMOVE(&vp->v_rdev->si_hlist, vp, vnode, v_specnext); 2737 vp->v_rdev->si_usecount -= vp->v_usecount; 2738 mtx_unlock(&spechash_mtx); 2739 VI_UNLOCK(vp); 2740 vp->v_rdev = NULL; 2741 } 2742 2743 /* 2744 * If it is on the freelist and not already at the head, 2745 * move it to the head of the list. The test of the 2746 * VDOOMED flag and the reference count of zero is because 2747 * it will be removed from the free list by getnewvnode, 2748 * but will not have its reference count incremented until 2749 * after calling vgone. If the reference count were 2750 * incremented first, vgone would (incorrectly) try to 2751 * close the previous instance of the underlying object. 2752 */ 2753 VI_LOCK(vp); 2754 if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) { 2755 s = splbio(); 2756 mtx_lock(&vnode_free_list_mtx); 2757 if (vp->v_iflag & VI_FREE) { 2758 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2759 } else { 2760 vp->v_iflag |= VI_FREE; 2761 freevnodes++; 2762 } 2763 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2764 mtx_unlock(&vnode_free_list_mtx); 2765 splx(s); 2766 } 2767 2768 vp->v_type = VBAD; 2769 VI_UNLOCK(vp); 2770} 2771 2772/* 2773 * Lookup a vnode by device number. 2774 */ 2775int 2776vfinddev(dev, type, vpp) 2777 dev_t dev; 2778 enum vtype type; 2779 struct vnode **vpp; 2780{ 2781 struct vnode *vp; 2782 2783 mtx_lock(&spechash_mtx); 2784 SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) { 2785 if (type == vp->v_type) { 2786 *vpp = vp; 2787 mtx_unlock(&spechash_mtx); 2788 return (1); 2789 } 2790 } 2791 mtx_unlock(&spechash_mtx); 2792 return (0); 2793} 2794 2795/* 2796 * Calculate the total number of references to a special device. 2797 */ 2798int 2799vcount(vp) 2800 struct vnode *vp; 2801{ 2802 int count; 2803 2804 mtx_lock(&spechash_mtx); 2805 count = vp->v_rdev->si_usecount; 2806 mtx_unlock(&spechash_mtx); 2807 return (count); 2808} 2809 2810/* 2811 * Same as above, but using the dev_t as argument 2812 */ 2813int 2814count_dev(dev) 2815 dev_t dev; 2816{ 2817 struct vnode *vp; 2818 2819 vp = SLIST_FIRST(&dev->si_hlist); 2820 if (vp == NULL) 2821 return (0); 2822 return(vcount(vp)); 2823} 2824 2825/* 2826 * Print out a description of a vnode. 2827 */ 2828static char *typename[] = 2829{"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 2830 2831void 2832vprint(label, vp) 2833 char *label; 2834 struct vnode *vp; 2835{ 2836 char buf[96]; 2837 2838 if (label != NULL) 2839 printf("%s: %p: ", label, (void *)vp); 2840 else 2841 printf("%p: ", (void *)vp); 2842 printf("tag %s, type %s, usecount %d, writecount %d, refcount %d,", 2843 vp->v_tag, typename[vp->v_type], vp->v_usecount, 2844 vp->v_writecount, vp->v_holdcnt); 2845 buf[0] = '\0'; 2846 if (vp->v_vflag & VV_ROOT) 2847 strcat(buf, "|VV_ROOT"); 2848 if (vp->v_vflag & VV_TEXT) 2849 strcat(buf, "|VV_TEXT"); 2850 if (vp->v_vflag & VV_SYSTEM) 2851 strcat(buf, "|VV_SYSTEM"); 2852 if (vp->v_iflag & VI_XLOCK) 2853 strcat(buf, "|VI_XLOCK"); 2854 if (vp->v_iflag & VI_XWANT) 2855 strcat(buf, "|VI_XWANT"); 2856 if (vp->v_iflag & VI_BWAIT) 2857 strcat(buf, "|VI_BWAIT"); 2858 if (vp->v_iflag & VI_DOOMED) 2859 strcat(buf, "|VI_DOOMED"); 2860 if (vp->v_iflag & VI_FREE) 2861 strcat(buf, "|VI_FREE"); 2862 if (vp->v_vflag & VV_OBJBUF) 2863 strcat(buf, "|VV_OBJBUF"); 2864 if (buf[0] != '\0') 2865 printf(" flags (%s),", &buf[1]); 2866 lockmgr_printinfo(vp->v_vnlock); 2867 printf("\n"); 2868 if (vp->v_data != NULL) 2869 VOP_PRINT(vp); 2870} 2871 2872#ifdef DDB 2873#include <ddb/ddb.h> 2874/* 2875 * List all of the locked vnodes in the system. 2876 * Called when debugging the kernel. 2877 */ 2878DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 2879{ 2880 struct thread *td = curthread; /* XXX */ 2881 struct mount *mp, *nmp; 2882 struct vnode *vp; 2883 2884 printf("Locked vnodes\n"); 2885 mtx_lock(&mountlist_mtx); 2886 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2887 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 2888 nmp = TAILQ_NEXT(mp, mnt_list); 2889 continue; 2890 } 2891 mtx_lock(&mntvnode_mtx); 2892 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2893 if (VOP_ISLOCKED(vp, NULL)) 2894 vprint(NULL, vp); 2895 } 2896 mtx_unlock(&mntvnode_mtx); 2897 mtx_lock(&mountlist_mtx); 2898 nmp = TAILQ_NEXT(mp, mnt_list); 2899 vfs_unbusy(mp, td); 2900 } 2901 mtx_unlock(&mountlist_mtx); 2902} 2903#endif 2904 2905/* 2906 * Fill in a struct xvfsconf based on a struct vfsconf. 2907 */ 2908static void 2909vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) 2910{ 2911 2912 strcpy(xvfsp->vfc_name, vfsp->vfc_name); 2913 xvfsp->vfc_typenum = vfsp->vfc_typenum; 2914 xvfsp->vfc_refcount = vfsp->vfc_refcount; 2915 xvfsp->vfc_flags = vfsp->vfc_flags; 2916 /* 2917 * These are unused in userland, we keep them 2918 * to not break binary compatibility. 2919 */ 2920 xvfsp->vfc_vfsops = NULL; 2921 xvfsp->vfc_next = NULL; 2922} 2923 2924static int 2925sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 2926{ 2927 struct vfsconf *vfsp; 2928 struct xvfsconf *xvfsp; 2929 int cnt, error, i; 2930 2931 cnt = 0; 2932 for (vfsp = vfsconf; vfsp != NULL; vfsp = vfsp->vfc_next) 2933 cnt++; 2934 xvfsp = malloc(sizeof(struct xvfsconf) * cnt, M_TEMP, M_WAITOK); 2935 /* 2936 * Handle the race that we will have here when struct vfsconf 2937 * will be locked down by using both cnt and checking vfc_next 2938 * against NULL to determine the end of the loop. The race will 2939 * happen because we will have to unlock before calling malloc(). 2940 * We are protected by Giant for now. 2941 */ 2942 i = 0; 2943 for (vfsp = vfsconf; vfsp != NULL && i < cnt; vfsp = vfsp->vfc_next) { 2944 vfsconf2x(vfsp, xvfsp + i); 2945 i++; 2946 } 2947 error = SYSCTL_OUT(req, xvfsp, sizeof(struct xvfsconf) * i); 2948 free(xvfsp, M_TEMP); 2949 return (error); 2950} 2951 2952SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, 2953 "S,xvfsconf", "List of all configured filesystems"); 2954 2955/* 2956 * Top level filesystem related information gathering. 2957 */ 2958static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2959 2960static int 2961vfs_sysctl(SYSCTL_HANDLER_ARGS) 2962{ 2963 int *name = (int *)arg1 - 1; /* XXX */ 2964 u_int namelen = arg2 + 1; /* XXX */ 2965 struct vfsconf *vfsp; 2966 struct xvfsconf xvfsp; 2967 2968 printf("WARNING: userland calling deprecated sysctl, " 2969 "please rebuild world\n"); 2970 2971#if 1 || defined(COMPAT_PRELITE2) 2972 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2973 if (namelen == 1) 2974 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2975#endif 2976 2977 switch (name[1]) { 2978 case VFS_MAXTYPENUM: 2979 if (namelen != 2) 2980 return (ENOTDIR); 2981 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2982 case VFS_CONF: 2983 if (namelen != 3) 2984 return (ENOTDIR); /* overloaded */ 2985 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) 2986 if (vfsp->vfc_typenum == name[2]) 2987 break; 2988 if (vfsp == NULL) 2989 return (EOPNOTSUPP); 2990 vfsconf2x(vfsp, &xvfsp); 2991 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 2992 } 2993 return (EOPNOTSUPP); 2994} 2995 2996SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl, 2997 "Generic filesystem"); 2998 2999#if 1 || defined(COMPAT_PRELITE2) 3000 3001static int 3002sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 3003{ 3004 int error; 3005 struct vfsconf *vfsp; 3006 struct ovfsconf ovfs; 3007 3008 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) { 3009 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 3010 strcpy(ovfs.vfc_name, vfsp->vfc_name); 3011 ovfs.vfc_index = vfsp->vfc_typenum; 3012 ovfs.vfc_refcount = vfsp->vfc_refcount; 3013 ovfs.vfc_flags = vfsp->vfc_flags; 3014 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 3015 if (error) 3016 return error; 3017 } 3018 return 0; 3019} 3020 3021#endif /* 1 || COMPAT_PRELITE2 */ 3022 3023#define KINFO_VNODESLOP 10 3024#ifdef notyet 3025/* 3026 * Dump vnode list (via sysctl). 3027 */ 3028/* ARGSUSED */ 3029static int 3030sysctl_vnode(SYSCTL_HANDLER_ARGS) 3031{ 3032 struct xvnode *xvn; 3033 struct thread *td = req->td; 3034 struct mount *mp; 3035 struct vnode *vp; 3036 int error, len, n; 3037 3038 /* 3039 * Stale numvnodes access is not fatal here. 3040 */ 3041 req->lock = 0; 3042 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 3043 if (!req->oldptr) 3044 /* Make an estimate */ 3045 return (SYSCTL_OUT(req, 0, len)); 3046 3047 sysctl_wire_old_buffer(req, 0); 3048 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 3049 n = 0; 3050 mtx_lock(&mountlist_mtx); 3051 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3052 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 3053 continue; 3054 mtx_lock(&mntvnode_mtx); 3055 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3056 if (n == len) 3057 break; 3058 vref(vp); 3059 xvn[n].xv_size = sizeof *xvn; 3060 xvn[n].xv_vnode = vp; 3061#define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 3062 XV_COPY(usecount); 3063 XV_COPY(writecount); 3064 XV_COPY(holdcnt); 3065 XV_COPY(id); 3066 XV_COPY(mount); 3067 XV_COPY(numoutput); 3068 XV_COPY(type); 3069#undef XV_COPY 3070 xvn[n].xv_flag = vp->v_vflag; 3071 3072 switch (vp->v_type) { 3073 case VREG: 3074 case VDIR: 3075 case VLNK: 3076 xvn[n].xv_dev = vp->v_cachedfs; 3077 xvn[n].xv_ino = vp->v_cachedid; 3078 break; 3079 case VBLK: 3080 case VCHR: 3081 if (vp->v_rdev == NULL) { 3082 vrele(vp); 3083 continue; 3084 } 3085 xvn[n].xv_dev = dev2udev(vp->v_rdev); 3086 break; 3087 case VSOCK: 3088 xvn[n].xv_socket = vp->v_socket; 3089 break; 3090 case VFIFO: 3091 xvn[n].xv_fifo = vp->v_fifoinfo; 3092 break; 3093 case VNON: 3094 case VBAD: 3095 default: 3096 /* shouldn't happen? */ 3097 vrele(vp); 3098 continue; 3099 } 3100 vrele(vp); 3101 ++n; 3102 } 3103 mtx_unlock(&mntvnode_mtx); 3104 mtx_lock(&mountlist_mtx); 3105 vfs_unbusy(mp, td); 3106 if (n == len) 3107 break; 3108 } 3109 mtx_unlock(&mountlist_mtx); 3110 3111 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 3112 free(xvn, M_TEMP); 3113 return (error); 3114} 3115 3116SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 3117 0, 0, sysctl_vnode, "S,xvnode", ""); 3118#endif 3119 3120/* 3121 * Check to see if a filesystem is mounted on a block device. 3122 */ 3123int 3124vfs_mountedon(vp) 3125 struct vnode *vp; 3126{ 3127 3128 if (vp->v_rdev->si_mountpoint != NULL) 3129 return (EBUSY); 3130 return (0); 3131} 3132 3133/* 3134 * Unmount all filesystems. The list is traversed in reverse order 3135 * of mounting to avoid dependencies. 3136 */ 3137void 3138vfs_unmountall() 3139{ 3140 struct mount *mp; 3141 struct thread *td; 3142 int error; 3143 3144 if (curthread != NULL) 3145 td = curthread; 3146 else 3147 td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */ 3148 /* 3149 * Since this only runs when rebooting, it is not interlocked. 3150 */ 3151 while(!TAILQ_EMPTY(&mountlist)) { 3152 mp = TAILQ_LAST(&mountlist, mntlist); 3153 error = dounmount(mp, MNT_FORCE, td); 3154 if (error) { 3155 TAILQ_REMOVE(&mountlist, mp, mnt_list); 3156 printf("unmount of %s failed (", 3157 mp->mnt_stat.f_mntonname); 3158 if (error == EBUSY) 3159 printf("BUSY)\n"); 3160 else 3161 printf("%d)\n", error); 3162 } else { 3163 /* The unmount has removed mp from the mountlist */ 3164 } 3165 } 3166} 3167 3168/* 3169 * perform msync on all vnodes under a mount point 3170 * the mount point must be locked. 3171 */ 3172void 3173vfs_msync(struct mount *mp, int flags) 3174{ 3175 struct vnode *vp, *nvp; 3176 struct vm_object *obj; 3177 int tries; 3178 3179 GIANT_REQUIRED; 3180 3181 tries = 5; 3182 mtx_lock(&mntvnode_mtx); 3183loop: 3184 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp != NULL; vp = nvp) { 3185 if (vp->v_mount != mp) { 3186 if (--tries > 0) 3187 goto loop; 3188 break; 3189 } 3190 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 3191 3192 VI_LOCK(vp); 3193 if (vp->v_iflag & VI_XLOCK) { /* XXX: what if MNT_WAIT? */ 3194 VI_UNLOCK(vp); 3195 continue; 3196 } 3197 3198 if ((vp->v_iflag & VI_OBJDIRTY) && 3199 (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { 3200 mtx_unlock(&mntvnode_mtx); 3201 if (!vget(vp, 3202 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 3203 curthread)) { 3204 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 3205 vput(vp); 3206 mtx_lock(&mntvnode_mtx); 3207 continue; 3208 } 3209 3210 if (VOP_GETVOBJECT(vp, &obj) == 0) { 3211 VM_OBJECT_LOCK(obj); 3212 vm_object_page_clean(obj, 0, 0, 3213 flags == MNT_WAIT ? 3214 OBJPC_SYNC : OBJPC_NOSYNC); 3215 VM_OBJECT_UNLOCK(obj); 3216 } 3217 vput(vp); 3218 } 3219 mtx_lock(&mntvnode_mtx); 3220 if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) { 3221 if (--tries > 0) 3222 goto loop; 3223 break; 3224 } 3225 } else 3226 VI_UNLOCK(vp); 3227 } 3228 mtx_unlock(&mntvnode_mtx); 3229} 3230 3231/* 3232 * Create the VM object needed for VMIO and mmap support. This 3233 * is done for all VREG files in the system. Some filesystems might 3234 * afford the additional metadata buffering capability of the 3235 * VMIO code by making the device node be VMIO mode also. 3236 * 3237 * vp must be locked when vfs_object_create is called. 3238 */ 3239int 3240vfs_object_create(vp, td, cred) 3241 struct vnode *vp; 3242 struct thread *td; 3243 struct ucred *cred; 3244{ 3245 GIANT_REQUIRED; 3246 return (VOP_CREATEVOBJECT(vp, cred, td)); 3247} 3248 3249/* 3250 * Mark a vnode as free, putting it up for recycling. 3251 */ 3252void 3253vfree(vp) 3254 struct vnode *vp; 3255{ 3256 int s; 3257 3258 ASSERT_VI_LOCKED(vp, "vfree"); 3259 s = splbio(); 3260 mtx_lock(&vnode_free_list_mtx); 3261 KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free")); 3262 if (vp->v_iflag & VI_AGE) { 3263 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 3264 } else { 3265 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 3266 } 3267 freevnodes++; 3268 mtx_unlock(&vnode_free_list_mtx); 3269 vp->v_iflag &= ~VI_AGE; 3270 vp->v_iflag |= VI_FREE; 3271 splx(s); 3272} 3273 3274/* 3275 * Opposite of vfree() - mark a vnode as in use. 3276 */ 3277void 3278vbusy(vp) 3279 struct vnode *vp; 3280{ 3281 int s; 3282 3283 s = splbio(); 3284 ASSERT_VI_LOCKED(vp, "vbusy"); 3285 KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free")); 3286 3287 mtx_lock(&vnode_free_list_mtx); 3288 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 3289 freevnodes--; 3290 mtx_unlock(&vnode_free_list_mtx); 3291 3292 vp->v_iflag &= ~(VI_FREE|VI_AGE); 3293 splx(s); 3294} 3295 3296/* 3297 * Record a process's interest in events which might happen to 3298 * a vnode. Because poll uses the historic select-style interface 3299 * internally, this routine serves as both the ``check for any 3300 * pending events'' and the ``record my interest in future events'' 3301 * functions. (These are done together, while the lock is held, 3302 * to avoid race conditions.) 3303 */ 3304int 3305vn_pollrecord(vp, td, events) 3306 struct vnode *vp; 3307 struct thread *td; 3308 short events; 3309{ 3310 3311 if (vp->v_pollinfo == NULL) 3312 v_addpollinfo(vp); 3313 mtx_lock(&vp->v_pollinfo->vpi_lock); 3314 if (vp->v_pollinfo->vpi_revents & events) { 3315 /* 3316 * This leaves events we are not interested 3317 * in available for the other process which 3318 * which presumably had requested them 3319 * (otherwise they would never have been 3320 * recorded). 3321 */ 3322 events &= vp->v_pollinfo->vpi_revents; 3323 vp->v_pollinfo->vpi_revents &= ~events; 3324 3325 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3326 return events; 3327 } 3328 vp->v_pollinfo->vpi_events |= events; 3329 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 3330 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3331 return 0; 3332} 3333 3334/* 3335 * Note the occurrence of an event. If the VN_POLLEVENT macro is used, 3336 * it is possible for us to miss an event due to race conditions, but 3337 * that condition is expected to be rare, so for the moment it is the 3338 * preferred interface. 3339 */ 3340void 3341vn_pollevent(vp, events) 3342 struct vnode *vp; 3343 short events; 3344{ 3345 3346 if (vp->v_pollinfo == NULL) 3347 v_addpollinfo(vp); 3348 mtx_lock(&vp->v_pollinfo->vpi_lock); 3349 if (vp->v_pollinfo->vpi_events & events) { 3350 /* 3351 * We clear vpi_events so that we don't 3352 * call selwakeup() twice if two events are 3353 * posted before the polling process(es) is 3354 * awakened. This also ensures that we take at 3355 * most one selwakeup() if the polling process 3356 * is no longer interested. However, it does 3357 * mean that only one event can be noticed at 3358 * a time. (Perhaps we should only clear those 3359 * event bits which we note?) XXX 3360 */ 3361 vp->v_pollinfo->vpi_events = 0; /* &= ~events ??? */ 3362 vp->v_pollinfo->vpi_revents |= events; 3363 selwakeup(&vp->v_pollinfo->vpi_selinfo); 3364 } 3365 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3366} 3367 3368/* 3369 * Wake up anyone polling on vp because it is being revoked. 3370 * This depends on dead_poll() returning POLLHUP for correct 3371 * behavior. 3372 */ 3373void 3374vn_pollgone(vp) 3375 struct vnode *vp; 3376{ 3377 3378 mtx_lock(&vp->v_pollinfo->vpi_lock); 3379 VN_KNOTE(vp, NOTE_REVOKE); 3380 if (vp->v_pollinfo->vpi_events) { 3381 vp->v_pollinfo->vpi_events = 0; 3382 selwakeup(&vp->v_pollinfo->vpi_selinfo); 3383 } 3384 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3385} 3386 3387 3388 3389/* 3390 * Routine to create and manage a filesystem syncer vnode. 3391 */ 3392#define sync_close ((int (*)(struct vop_close_args *))nullop) 3393static int sync_fsync(struct vop_fsync_args *); 3394static int sync_inactive(struct vop_inactive_args *); 3395static int sync_reclaim(struct vop_reclaim_args *); 3396 3397static vop_t **sync_vnodeop_p; 3398static struct vnodeopv_entry_desc sync_vnodeop_entries[] = { 3399 { &vop_default_desc, (vop_t *) vop_eopnotsupp }, 3400 { &vop_close_desc, (vop_t *) sync_close }, /* close */ 3401 { &vop_fsync_desc, (vop_t *) sync_fsync }, /* fsync */ 3402 { &vop_inactive_desc, (vop_t *) sync_inactive }, /* inactive */ 3403 { &vop_reclaim_desc, (vop_t *) sync_reclaim }, /* reclaim */ 3404 { &vop_lock_desc, (vop_t *) vop_stdlock }, /* lock */ 3405 { &vop_unlock_desc, (vop_t *) vop_stdunlock }, /* unlock */ 3406 { &vop_islocked_desc, (vop_t *) vop_stdislocked }, /* islocked */ 3407 { NULL, NULL } 3408}; 3409static struct vnodeopv_desc sync_vnodeop_opv_desc = 3410 { &sync_vnodeop_p, sync_vnodeop_entries }; 3411 3412VNODEOP_SET(sync_vnodeop_opv_desc); 3413 3414/* 3415 * Create a new filesystem syncer vnode for the specified mount point. 3416 */ 3417int 3418vfs_allocate_syncvnode(mp) 3419 struct mount *mp; 3420{ 3421 struct vnode *vp; 3422 static long start, incr, next; 3423 int error; 3424 3425 /* Allocate a new vnode */ 3426 if ((error = getnewvnode("syncer", mp, sync_vnodeop_p, &vp)) != 0) { 3427 mp->mnt_syncer = NULL; 3428 return (error); 3429 } 3430 vp->v_type = VNON; 3431 /* 3432 * Place the vnode onto the syncer worklist. We attempt to 3433 * scatter them about on the list so that they will go off 3434 * at evenly distributed times even if all the filesystems 3435 * are mounted at once. 3436 */ 3437 next += incr; 3438 if (next == 0 || next > syncer_maxdelay) { 3439 start /= 2; 3440 incr /= 2; 3441 if (start == 0) { 3442 start = syncer_maxdelay / 2; 3443 incr = syncer_maxdelay; 3444 } 3445 next = start; 3446 } 3447 VI_LOCK(vp); 3448 vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0); 3449 VI_UNLOCK(vp); 3450 mp->mnt_syncer = vp; 3451 return (0); 3452} 3453 3454/* 3455 * Do a lazy sync of the filesystem. 3456 */ 3457static int 3458sync_fsync(ap) 3459 struct vop_fsync_args /* { 3460 struct vnode *a_vp; 3461 struct ucred *a_cred; 3462 int a_waitfor; 3463 struct thread *a_td; 3464 } */ *ap; 3465{ 3466 struct vnode *syncvp = ap->a_vp; 3467 struct mount *mp = syncvp->v_mount; 3468 struct thread *td = ap->a_td; 3469 int error, asyncflag; 3470 3471 /* 3472 * We only need to do something if this is a lazy evaluation. 3473 */ 3474 if (ap->a_waitfor != MNT_LAZY) 3475 return (0); 3476 3477 /* 3478 * Move ourselves to the back of the sync list. 3479 */ 3480 VI_LOCK(syncvp); 3481 vn_syncer_add_to_worklist(syncvp, syncdelay); 3482 VI_UNLOCK(syncvp); 3483 3484 /* 3485 * Walk the list of vnodes pushing all that are dirty and 3486 * not already on the sync list. 3487 */ 3488 mtx_lock(&mountlist_mtx); 3489 if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { 3490 mtx_unlock(&mountlist_mtx); 3491 return (0); 3492 } 3493 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 3494 vfs_unbusy(mp, td); 3495 return (0); 3496 } 3497 asyncflag = mp->mnt_flag & MNT_ASYNC; 3498 mp->mnt_flag &= ~MNT_ASYNC; 3499 vfs_msync(mp, MNT_NOWAIT); 3500 error = VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td); 3501 if (asyncflag) 3502 mp->mnt_flag |= MNT_ASYNC; 3503 vn_finished_write(mp); 3504 vfs_unbusy(mp, td); 3505 return (error); 3506} 3507 3508/* 3509 * The syncer vnode is no referenced. 3510 */ 3511static int 3512sync_inactive(ap) 3513 struct vop_inactive_args /* { 3514 struct vnode *a_vp; 3515 struct thread *a_td; 3516 } */ *ap; 3517{ 3518 3519 VOP_UNLOCK(ap->a_vp, 0, ap->a_td); 3520 vgone(ap->a_vp); 3521 return (0); 3522} 3523 3524/* 3525 * The syncer vnode is no longer needed and is being decommissioned. 3526 * 3527 * Modifications to the worklist must be protected at splbio(). 3528 */ 3529static int 3530sync_reclaim(ap) 3531 struct vop_reclaim_args /* { 3532 struct vnode *a_vp; 3533 } */ *ap; 3534{ 3535 struct vnode *vp = ap->a_vp; 3536 int s; 3537 3538 s = splbio(); 3539 vp->v_mount->mnt_syncer = NULL; 3540 VI_LOCK(vp); 3541 if (vp->v_iflag & VI_ONWORKLST) { 3542 mtx_lock(&sync_mtx); 3543 LIST_REMOVE(vp, v_synclist); 3544 mtx_unlock(&sync_mtx); 3545 vp->v_iflag &= ~VI_ONWORKLST; 3546 } 3547 VI_UNLOCK(vp); 3548 splx(s); 3549 3550 return (0); 3551} 3552 3553/* 3554 * extract the dev_t from a VCHR 3555 */ 3556dev_t 3557vn_todev(vp) 3558 struct vnode *vp; 3559{ 3560 if (vp->v_type != VCHR) 3561 return (NODEV); 3562 return (vp->v_rdev); 3563} 3564 3565/* 3566 * Check if vnode represents a disk device 3567 */ 3568int 3569vn_isdisk(vp, errp) 3570 struct vnode *vp; 3571 int *errp; 3572{ 3573 struct cdevsw *cdevsw; 3574 3575 if (vp->v_type != VCHR) { 3576 if (errp != NULL) 3577 *errp = ENOTBLK; 3578 return (0); 3579 } 3580 if (vp->v_rdev == NULL) { 3581 if (errp != NULL) 3582 *errp = ENXIO; 3583 return (0); 3584 } 3585 cdevsw = devsw(vp->v_rdev); 3586 if (cdevsw == NULL) { 3587 if (errp != NULL) 3588 *errp = ENXIO; 3589 return (0); 3590 } 3591 if (!(cdevsw->d_flags & D_DISK)) { 3592 if (errp != NULL) 3593 *errp = ENOTBLK; 3594 return (0); 3595 } 3596 if (errp != NULL) 3597 *errp = 0; 3598 return (1); 3599} 3600 3601/* 3602 * Free data allocated by namei(); see namei(9) for details. 3603 */ 3604void 3605NDFREE(ndp, flags) 3606 struct nameidata *ndp; 3607 const uint flags; 3608{ 3609 if (!(flags & NDF_NO_FREE_PNBUF) && 3610 (ndp->ni_cnd.cn_flags & HASBUF)) { 3611 uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf); 3612 ndp->ni_cnd.cn_flags &= ~HASBUF; 3613 } 3614 if (!(flags & NDF_NO_DVP_UNLOCK) && 3615 (ndp->ni_cnd.cn_flags & LOCKPARENT) && 3616 ndp->ni_dvp != ndp->ni_vp) 3617 VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread); 3618 if (!(flags & NDF_NO_DVP_RELE) && 3619 (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) { 3620 vrele(ndp->ni_dvp); 3621 ndp->ni_dvp = NULL; 3622 } 3623 if (!(flags & NDF_NO_VP_UNLOCK) && 3624 (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp) 3625 VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread); 3626 if (!(flags & NDF_NO_VP_RELE) && 3627 ndp->ni_vp) { 3628 vrele(ndp->ni_vp); 3629 ndp->ni_vp = NULL; 3630 } 3631 if (!(flags & NDF_NO_STARTDIR_RELE) && 3632 (ndp->ni_cnd.cn_flags & SAVESTART)) { 3633 vrele(ndp->ni_startdir); 3634 ndp->ni_startdir = NULL; 3635 } 3636} 3637 3638/* 3639 * Common filesystem object access control check routine. Accepts a 3640 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3641 * and optional call-by-reference privused argument allowing vaccess() 3642 * to indicate to the caller whether privilege was used to satisfy the 3643 * request (obsoleted). Returns 0 on success, or an errno on failure. 3644 */ 3645int 3646vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused) 3647 enum vtype type; 3648 mode_t file_mode; 3649 uid_t file_uid; 3650 gid_t file_gid; 3651 mode_t acc_mode; 3652 struct ucred *cred; 3653 int *privused; 3654{ 3655 mode_t dac_granted; 3656#ifdef CAPABILITIES 3657 mode_t cap_granted; 3658#endif 3659 3660 /* 3661 * Look for a normal, non-privileged way to access the file/directory 3662 * as requested. If it exists, go with that. 3663 */ 3664 3665 if (privused != NULL) 3666 *privused = 0; 3667 3668 dac_granted = 0; 3669 3670 /* Check the owner. */ 3671 if (cred->cr_uid == file_uid) { 3672 dac_granted |= VADMIN; 3673 if (file_mode & S_IXUSR) 3674 dac_granted |= VEXEC; 3675 if (file_mode & S_IRUSR) 3676 dac_granted |= VREAD; 3677 if (file_mode & S_IWUSR) 3678 dac_granted |= (VWRITE | VAPPEND); 3679 3680 if ((acc_mode & dac_granted) == acc_mode) 3681 return (0); 3682 3683 goto privcheck; 3684 } 3685 3686 /* Otherwise, check the groups (first match) */ 3687 if (groupmember(file_gid, cred)) { 3688 if (file_mode & S_IXGRP) 3689 dac_granted |= VEXEC; 3690 if (file_mode & S_IRGRP) 3691 dac_granted |= VREAD; 3692 if (file_mode & S_IWGRP) 3693 dac_granted |= (VWRITE | VAPPEND); 3694 3695 if ((acc_mode & dac_granted) == acc_mode) 3696 return (0); 3697 3698 goto privcheck; 3699 } 3700 3701 /* Otherwise, check everyone else. */ 3702 if (file_mode & S_IXOTH) 3703 dac_granted |= VEXEC; 3704 if (file_mode & S_IROTH) 3705 dac_granted |= VREAD; 3706 if (file_mode & S_IWOTH) 3707 dac_granted |= (VWRITE | VAPPEND); 3708 if ((acc_mode & dac_granted) == acc_mode) 3709 return (0); 3710 3711privcheck: 3712 if (!suser_cred(cred, PRISON_ROOT)) { 3713 /* XXX audit: privilege used */ 3714 if (privused != NULL) 3715 *privused = 1; 3716 return (0); 3717 } 3718 3719#ifdef CAPABILITIES 3720 /* 3721 * Build a capability mask to determine if the set of capabilities 3722 * satisfies the requirements when combined with the granted mask 3723 * from above. 3724 * For each capability, if the capability is required, bitwise 3725 * or the request type onto the cap_granted mask. 3726 */ 3727 cap_granted = 0; 3728 3729 if (type == VDIR) { 3730 /* 3731 * For directories, use CAP_DAC_READ_SEARCH to satisfy 3732 * VEXEC requests, instead of CAP_DAC_EXECUTE. 3733 */ 3734 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3735 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) 3736 cap_granted |= VEXEC; 3737 } else { 3738 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3739 !cap_check(cred, NULL, CAP_DAC_EXECUTE, PRISON_ROOT)) 3740 cap_granted |= VEXEC; 3741 } 3742 3743 if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && 3744 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) 3745 cap_granted |= VREAD; 3746 3747 if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3748 !cap_check(cred, NULL, CAP_DAC_WRITE, PRISON_ROOT)) 3749 cap_granted |= (VWRITE | VAPPEND); 3750 3751 if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3752 !cap_check(cred, NULL, CAP_FOWNER, PRISON_ROOT)) 3753 cap_granted |= VADMIN; 3754 3755 if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { 3756 /* XXX audit: privilege used */ 3757 if (privused != NULL) 3758 *privused = 1; 3759 return (0); 3760 } 3761#endif 3762 3763 return ((acc_mode & VADMIN) ? EPERM : EACCES); 3764} 3765 3766/* 3767 * Credential check based on process requesting service, and per-attribute 3768 * permissions. 3769 */ 3770int 3771extattr_check_cred(struct vnode *vp, int attrnamespace, 3772 struct ucred *cred, struct thread *td, int access) 3773{ 3774 3775 /* 3776 * Kernel-invoked always succeeds. 3777 */ 3778 if (cred == NOCRED) 3779 return (0); 3780 3781 /* 3782 * Do not allow privileged processes in jail to directly 3783 * manipulate system attributes. 3784 * 3785 * XXX What capability should apply here? 3786 * Probably CAP_SYS_SETFFLAG. 3787 */ 3788 switch (attrnamespace) { 3789 case EXTATTR_NAMESPACE_SYSTEM: 3790 /* Potentially should be: return (EPERM); */ 3791 return (suser_cred(cred, 0)); 3792 case EXTATTR_NAMESPACE_USER: 3793 return (VOP_ACCESS(vp, access, cred, td)); 3794 default: 3795 return (EPERM); 3796 } 3797} 3798