vfs_subr.c revision 120746
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 */ 40 41/* 42 * External virtual filesystem routines 43 */ 44 45#include <sys/cdefs.h> 46__FBSDID("$FreeBSD: head/sys/kern/vfs_subr.c 120746 2003-10-04 15:10:40Z jeff $"); 47 48#include "opt_ddb.h" 49#include "opt_mac.h" 50 51#include <sys/param.h> 52#include <sys/systm.h> 53#include <sys/bio.h> 54#include <sys/buf.h> 55#include <sys/conf.h> 56#include <sys/eventhandler.h> 57#include <sys/extattr.h> 58#include <sys/fcntl.h> 59#include <sys/kernel.h> 60#include <sys/kthread.h> 61#include <sys/mac.h> 62#include <sys/malloc.h> 63#include <sys/mount.h> 64#include <sys/namei.h> 65#include <sys/stat.h> 66#include <sys/sysctl.h> 67#include <sys/syslog.h> 68#include <sys/vmmeter.h> 69#include <sys/vnode.h> 70 71#include <vm/vm.h> 72#include <vm/vm_object.h> 73#include <vm/vm_extern.h> 74#include <vm/pmap.h> 75#include <vm/vm_map.h> 76#include <vm/vm_page.h> 77#include <vm/vm_kern.h> 78#include <vm/uma.h> 79 80static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure"); 81 82static void addalias(struct vnode *vp, dev_t nvp_rdev); 83static void insmntque(struct vnode *vp, struct mount *mp); 84static void vclean(struct vnode *vp, int flags, struct thread *td); 85static void vlruvp(struct vnode *vp); 86static int flushbuflist(struct buf *blist, int flags, struct vnode *vp, 87 int slpflag, int slptimeo, int *errorp); 88static int vcanrecycle(struct vnode *vp, struct mount **vnmpp); 89 90 91/* 92 * Number of vnodes in existence. Increased whenever getnewvnode() 93 * allocates a new vnode, never decreased. 94 */ 95static unsigned long numvnodes; 96 97SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); 98 99/* 100 * Conversion tables for conversion from vnode types to inode formats 101 * and back. 102 */ 103enum vtype iftovt_tab[16] = { 104 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 105 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 106}; 107int vttoif_tab[9] = { 108 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 109 S_IFSOCK, S_IFIFO, S_IFMT, 110}; 111 112/* 113 * List of vnodes that are ready for recycling. 114 */ 115static TAILQ_HEAD(freelst, vnode) vnode_free_list; 116 117/* 118 * Minimum number of free vnodes. If there are fewer than this free vnodes, 119 * getnewvnode() will return a newly allocated vnode. 120 */ 121static u_long wantfreevnodes = 25; 122SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 123/* Number of vnodes in the free list. */ 124static u_long freevnodes; 125SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); 126 127/* 128 * Various variables used for debugging the new implementation of 129 * reassignbuf(). 130 * XXX these are probably of (very) limited utility now. 131 */ 132static int reassignbufcalls; 133SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); 134static int nameileafonly; 135SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, ""); 136 137/* 138 * Cache for the mount type id assigned to NFS. This is used for 139 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 140 */ 141int nfs_mount_type = -1; 142 143/* To keep more than one thread at a time from running vfs_getnewfsid */ 144static struct mtx mntid_mtx; 145 146/* 147 * Lock for any access to the following: 148 * vnode_free_list 149 * numvnodes 150 * freevnodes 151 */ 152static struct mtx vnode_free_list_mtx; 153 154/* 155 * For any iteration/modification of dev->si_hlist (linked through 156 * v_specnext) 157 */ 158static struct mtx spechash_mtx; 159 160/* Publicly exported FS */ 161struct nfs_public nfs_pub; 162 163/* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 164static uma_zone_t vnode_zone; 165static uma_zone_t vnodepoll_zone; 166 167/* Set to 1 to print out reclaim of active vnodes */ 168int prtactive; 169 170/* 171 * The workitem queue. 172 * 173 * It is useful to delay writes of file data and filesystem metadata 174 * for tens of seconds so that quickly created and deleted files need 175 * not waste disk bandwidth being created and removed. To realize this, 176 * we append vnodes to a "workitem" queue. When running with a soft 177 * updates implementation, most pending metadata dependencies should 178 * not wait for more than a few seconds. Thus, mounted on block devices 179 * are delayed only about a half the time that file data is delayed. 180 * Similarly, directory updates are more critical, so are only delayed 181 * about a third the time that file data is delayed. Thus, there are 182 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 183 * one each second (driven off the filesystem syncer process). The 184 * syncer_delayno variable indicates the next queue that is to be processed. 185 * Items that need to be processed soon are placed in this queue: 186 * 187 * syncer_workitem_pending[syncer_delayno] 188 * 189 * A delay of fifteen seconds is done by placing the request fifteen 190 * entries later in the queue: 191 * 192 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 193 * 194 */ 195static int syncer_delayno; 196static long syncer_mask; 197LIST_HEAD(synclist, vnode); 198static struct synclist *syncer_workitem_pending; 199/* 200 * The sync_mtx protects: 201 * vp->v_synclist 202 * syncer_delayno 203 * syncer_workitem_pending 204 * rushjob 205 */ 206static struct mtx sync_mtx; 207 208#define SYNCER_MAXDELAY 32 209static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 210static int syncdelay = 30; /* max time to delay syncing data */ 211static int filedelay = 30; /* time to delay syncing files */ 212SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); 213static int dirdelay = 29; /* time to delay syncing directories */ 214SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); 215static int metadelay = 28; /* time to delay syncing metadata */ 216SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); 217static int rushjob; /* number of slots to run ASAP */ 218static int stat_rush_requests; /* number of times I/O speeded up */ 219SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); 220 221/* 222 * Number of vnodes we want to exist at any one time. This is mostly used 223 * to size hash tables in vnode-related code. It is normally not used in 224 * getnewvnode(), as wantfreevnodes is normally nonzero.) 225 * 226 * XXX desiredvnodes is historical cruft and should not exist. 227 */ 228int desiredvnodes; 229SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 230 &desiredvnodes, 0, "Maximum number of vnodes"); 231static int minvnodes; 232SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 233 &minvnodes, 0, "Minimum number of vnodes"); 234static int vnlru_nowhere; 235SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, &vnlru_nowhere, 0, 236 "Number of times the vnlru process ran without success"); 237 238/* Hook for calling soft updates */ 239int (*softdep_process_worklist_hook)(struct mount *); 240 241/* 242 * This only exists to supress warnings from unlocked specfs accesses. It is 243 * no longer ok to have an unlocked VFS. 244 */ 245#define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD) 246 247/* Print lock violations */ 248int vfs_badlock_print = 1; 249 250/* Panic on violation */ 251int vfs_badlock_panic = 1; 252 253/* Check for interlock across VOPs */ 254int vfs_badlock_mutex = 1; 255 256static void 257vfs_badlock(char *msg, char *str, struct vnode *vp) 258{ 259 if (vfs_badlock_print) 260 printf("%s: %p %s\n", str, vp, msg); 261 if (vfs_badlock_panic) 262 Debugger("Lock violation.\n"); 263} 264 265void 266assert_vi_unlocked(struct vnode *vp, char *str) 267{ 268 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 269 vfs_badlock("interlock is locked but should not be", str, vp); 270} 271 272void 273assert_vi_locked(struct vnode *vp, char *str) 274{ 275 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 276 vfs_badlock("interlock is not locked but should be", str, vp); 277} 278 279void 280assert_vop_locked(struct vnode *vp, char *str) 281{ 282 if (vp && !IGNORE_LOCK(vp) && !VOP_ISLOCKED(vp, NULL)) 283 vfs_badlock("is not locked but should be", str, vp); 284} 285 286void 287assert_vop_unlocked(struct vnode *vp, char *str) 288{ 289 if (vp && !IGNORE_LOCK(vp) && 290 VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE) 291 vfs_badlock("is locked but should not be", str, vp); 292} 293 294void 295assert_vop_elocked(struct vnode *vp, char *str) 296{ 297 if (vp && !IGNORE_LOCK(vp) && 298 VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE) 299 vfs_badlock("is not exclusive locked but should be", str, vp); 300} 301 302void 303assert_vop_elocked_other(struct vnode *vp, char *str) 304{ 305 if (vp && !IGNORE_LOCK(vp) && 306 VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER) 307 vfs_badlock("is not exclusive locked by another thread", 308 str, vp); 309} 310 311void 312assert_vop_slocked(struct vnode *vp, char *str) 313{ 314 if (vp && !IGNORE_LOCK(vp) && 315 VOP_ISLOCKED(vp, curthread) != LK_SHARED) 316 vfs_badlock("is not locked shared but should be", str, vp); 317} 318 319void 320vop_rename_pre(void *ap) 321{ 322 struct vop_rename_args *a = ap; 323 324 if (a->a_tvp) 325 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 326 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 327 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 328 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 329 330 /* Check the source (from) */ 331 if (a->a_tdvp != a->a_fdvp) 332 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked.\n"); 333 if (a->a_tvp != a->a_fvp) 334 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked.\n"); 335 336 /* Check the target */ 337 if (a->a_tvp) 338 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked.\n"); 339 340 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked.\n"); 341} 342 343void 344vop_strategy_pre(void *ap) 345{ 346 struct vop_strategy_args *a = ap; 347 struct buf *bp; 348 349 bp = a->a_bp; 350 351 /* 352 * Cluster ops lock their component buffers but not the IO container. 353 */ 354 if ((bp->b_flags & B_CLUSTER) != 0) 355 return; 356 357 if (BUF_REFCNT(bp) < 1) { 358 if (vfs_badlock_print) 359 printf("VOP_STRATEGY: bp is not locked but should be.\n"); 360 if (vfs_badlock_panic) 361 Debugger("Lock violation.\n"); 362 } 363} 364 365void 366vop_lookup_pre(void *ap) 367{ 368 struct vop_lookup_args *a = ap; 369 struct vnode *dvp; 370 371 dvp = a->a_dvp; 372 373 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 374 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 375} 376 377void 378vop_lookup_post(void *ap, int rc) 379{ 380 struct vop_lookup_args *a = ap; 381 struct componentname *cnp; 382 struct vnode *dvp; 383 struct vnode *vp; 384 int flags; 385 386 dvp = a->a_dvp; 387 cnp = a->a_cnp; 388 vp = *(a->a_vpp); 389 flags = cnp->cn_flags; 390 391 392 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 393 /* 394 * If this is the last path component for this lookup and LOCPARENT 395 * is set, OR if there is an error the directory has to be locked. 396 */ 397 if ((flags & LOCKPARENT) && (flags & ISLASTCN)) 398 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)"); 399 else if (rc != 0) 400 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)"); 401 else if (dvp != vp) 402 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)"); 403 404 if (flags & PDIRUNLOCK) 405 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)"); 406} 407 408void 409vop_unlock_pre(void *ap) 410{ 411 struct vop_unlock_args *a = ap; 412 413 if (a->a_flags & LK_INTERLOCK) 414 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 415 416 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 417} 418 419void 420vop_unlock_post(void *ap, int rc) 421{ 422 struct vop_unlock_args *a = ap; 423 424 if (a->a_flags & LK_INTERLOCK) 425 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 426} 427 428void 429vop_lock_pre(void *ap) 430{ 431 struct vop_lock_args *a = ap; 432 433 if ((a->a_flags & LK_INTERLOCK) == 0) 434 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 435 else 436 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 437} 438 439void 440vop_lock_post(void *ap, int rc) 441{ 442 struct vop_lock_args *a; 443 444 a = ap; 445 446 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 447 if (rc == 0) 448 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 449} 450 451void 452v_addpollinfo(struct vnode *vp) 453{ 454 vp->v_pollinfo = uma_zalloc(vnodepoll_zone, M_WAITOK); 455 mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 456} 457 458/* 459 * Initialize the vnode management data structures. 460 */ 461static void 462vntblinit(void *dummy __unused) 463{ 464 465 /* 466 * Desiredvnodes is a function of the physical memory size and 467 * the kernel's heap size. Specifically, desiredvnodes scales 468 * in proportion to the physical memory size until two fifths 469 * of the kernel's heap size is consumed by vnodes and vm 470 * objects. 471 */ 472 desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size / 473 (5 * (sizeof(struct vm_object) + sizeof(struct vnode)))); 474 minvnodes = desiredvnodes / 4; 475 mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF); 476 mtx_init(&mntvnode_mtx, "mntvnode", NULL, MTX_DEF); 477 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 478 mtx_init(&spechash_mtx, "spechash", NULL, MTX_DEF); 479 TAILQ_INIT(&vnode_free_list); 480 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 481 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 482 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 483 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 484 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 485 /* 486 * Initialize the filesystem syncer. 487 */ 488 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 489 &syncer_mask); 490 syncer_maxdelay = syncer_mask + 1; 491 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 492} 493SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) 494 495 496/* 497 * Mark a mount point as busy. Used to synchronize access and to delay 498 * unmounting. Interlock is not released on failure. 499 */ 500int 501vfs_busy(mp, flags, interlkp, td) 502 struct mount *mp; 503 int flags; 504 struct mtx *interlkp; 505 struct thread *td; 506{ 507 int lkflags; 508 509 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 510 if (flags & LK_NOWAIT) 511 return (ENOENT); 512 mp->mnt_kern_flag |= MNTK_MWAIT; 513 /* 514 * Since all busy locks are shared except the exclusive 515 * lock granted when unmounting, the only place that a 516 * wakeup needs to be done is at the release of the 517 * exclusive lock at the end of dounmount. 518 */ 519 msleep(mp, interlkp, PVFS, "vfs_busy", 0); 520 return (ENOENT); 521 } 522 lkflags = LK_SHARED | LK_NOPAUSE; 523 if (interlkp) 524 lkflags |= LK_INTERLOCK; 525 if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td)) 526 panic("vfs_busy: unexpected lock failure"); 527 return (0); 528} 529 530/* 531 * Free a busy filesystem. 532 */ 533void 534vfs_unbusy(mp, td) 535 struct mount *mp; 536 struct thread *td; 537{ 538 539 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); 540} 541 542/* 543 * Lookup a mount point by filesystem identifier. 544 */ 545struct mount * 546vfs_getvfs(fsid) 547 fsid_t *fsid; 548{ 549 register struct mount *mp; 550 551 mtx_lock(&mountlist_mtx); 552 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 553 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 554 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 555 mtx_unlock(&mountlist_mtx); 556 return (mp); 557 } 558 } 559 mtx_unlock(&mountlist_mtx); 560 return ((struct mount *) 0); 561} 562 563/* 564 * Get a new unique fsid. Try to make its val[0] unique, since this value 565 * will be used to create fake device numbers for stat(). Also try (but 566 * not so hard) make its val[0] unique mod 2^16, since some emulators only 567 * support 16-bit device numbers. We end up with unique val[0]'s for the 568 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 569 * 570 * Keep in mind that several mounts may be running in parallel. Starting 571 * the search one past where the previous search terminated is both a 572 * micro-optimization and a defense against returning the same fsid to 573 * different mounts. 574 */ 575void 576vfs_getnewfsid(mp) 577 struct mount *mp; 578{ 579 static u_int16_t mntid_base; 580 fsid_t tfsid; 581 int mtype; 582 583 mtx_lock(&mntid_mtx); 584 mtype = mp->mnt_vfc->vfc_typenum; 585 tfsid.val[1] = mtype; 586 mtype = (mtype & 0xFF) << 24; 587 for (;;) { 588 tfsid.val[0] = makeudev(255, 589 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 590 mntid_base++; 591 if (vfs_getvfs(&tfsid) == NULL) 592 break; 593 } 594 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 595 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 596 mtx_unlock(&mntid_mtx); 597} 598 599/* 600 * Knob to control the precision of file timestamps: 601 * 602 * 0 = seconds only; nanoseconds zeroed. 603 * 1 = seconds and nanoseconds, accurate within 1/HZ. 604 * 2 = seconds and nanoseconds, truncated to microseconds. 605 * >=3 = seconds and nanoseconds, maximum precision. 606 */ 607enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 608 609static int timestamp_precision = TSP_SEC; 610SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 611 ×tamp_precision, 0, ""); 612 613/* 614 * Get a current timestamp. 615 */ 616void 617vfs_timestamp(tsp) 618 struct timespec *tsp; 619{ 620 struct timeval tv; 621 622 switch (timestamp_precision) { 623 case TSP_SEC: 624 tsp->tv_sec = time_second; 625 tsp->tv_nsec = 0; 626 break; 627 case TSP_HZ: 628 getnanotime(tsp); 629 break; 630 case TSP_USEC: 631 microtime(&tv); 632 TIMEVAL_TO_TIMESPEC(&tv, tsp); 633 break; 634 case TSP_NSEC: 635 default: 636 nanotime(tsp); 637 break; 638 } 639} 640 641/* 642 * Set vnode attributes to VNOVAL 643 */ 644void 645vattr_null(vap) 646 register struct vattr *vap; 647{ 648 649 vap->va_type = VNON; 650 vap->va_size = VNOVAL; 651 vap->va_bytes = VNOVAL; 652 vap->va_mode = VNOVAL; 653 vap->va_nlink = VNOVAL; 654 vap->va_uid = VNOVAL; 655 vap->va_gid = VNOVAL; 656 vap->va_fsid = VNOVAL; 657 vap->va_fileid = VNOVAL; 658 vap->va_blocksize = VNOVAL; 659 vap->va_rdev = VNOVAL; 660 vap->va_atime.tv_sec = VNOVAL; 661 vap->va_atime.tv_nsec = VNOVAL; 662 vap->va_mtime.tv_sec = VNOVAL; 663 vap->va_mtime.tv_nsec = VNOVAL; 664 vap->va_ctime.tv_sec = VNOVAL; 665 vap->va_ctime.tv_nsec = VNOVAL; 666 vap->va_birthtime.tv_sec = VNOVAL; 667 vap->va_birthtime.tv_nsec = VNOVAL; 668 vap->va_flags = VNOVAL; 669 vap->va_gen = VNOVAL; 670 vap->va_vaflags = 0; 671} 672 673/* 674 * This routine is called when we have too many vnodes. It attempts 675 * to free <count> vnodes and will potentially free vnodes that still 676 * have VM backing store (VM backing store is typically the cause 677 * of a vnode blowout so we want to do this). Therefore, this operation 678 * is not considered cheap. 679 * 680 * A number of conditions may prevent a vnode from being reclaimed. 681 * the buffer cache may have references on the vnode, a directory 682 * vnode may still have references due to the namei cache representing 683 * underlying files, or the vnode may be in active use. It is not 684 * desireable to reuse such vnodes. These conditions may cause the 685 * number of vnodes to reach some minimum value regardless of what 686 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 687 */ 688static int 689vlrureclaim(struct mount *mp) 690{ 691 struct vnode *vp; 692 int done; 693 int trigger; 694 int usevnodes; 695 int count; 696 697 /* 698 * Calculate the trigger point, don't allow user 699 * screwups to blow us up. This prevents us from 700 * recycling vnodes with lots of resident pages. We 701 * aren't trying to free memory, we are trying to 702 * free vnodes. 703 */ 704 usevnodes = desiredvnodes; 705 if (usevnodes <= 0) 706 usevnodes = 1; 707 trigger = cnt.v_page_count * 2 / usevnodes; 708 709 done = 0; 710 mtx_lock(&mntvnode_mtx); 711 count = mp->mnt_nvnodelistsize / 10 + 1; 712 while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) { 713 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 714 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 715 716 if (vp->v_type != VNON && 717 vp->v_type != VBAD && 718 VI_TRYLOCK(vp)) { 719 if (VMIGHTFREE(vp) && /* critical path opt */ 720 (vp->v_object == NULL || 721 vp->v_object->resident_page_count < trigger)) { 722 mtx_unlock(&mntvnode_mtx); 723 vgonel(vp, curthread); 724 done++; 725 mtx_lock(&mntvnode_mtx); 726 } else 727 VI_UNLOCK(vp); 728 } 729 --count; 730 } 731 mtx_unlock(&mntvnode_mtx); 732 return done; 733} 734 735/* 736 * Attempt to recycle vnodes in a context that is always safe to block. 737 * Calling vlrurecycle() from the bowels of filesystem code has some 738 * interesting deadlock problems. 739 */ 740static struct proc *vnlruproc; 741static int vnlruproc_sig; 742 743static void 744vnlru_proc(void) 745{ 746 struct mount *mp, *nmp; 747 int done; 748 struct proc *p = vnlruproc; 749 struct thread *td = FIRST_THREAD_IN_PROC(p); /* XXXKSE */ 750 751 mtx_lock(&Giant); 752 753 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 754 SHUTDOWN_PRI_FIRST); 755 756 for (;;) { 757 kthread_suspend_check(p); 758 mtx_lock(&vnode_free_list_mtx); 759 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { 760 mtx_unlock(&vnode_free_list_mtx); 761 vnlruproc_sig = 0; 762 wakeup(&vnlruproc_sig); 763 tsleep(vnlruproc, PVFS, "vlruwt", hz); 764 continue; 765 } 766 mtx_unlock(&vnode_free_list_mtx); 767 done = 0; 768 mtx_lock(&mountlist_mtx); 769 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 770 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 771 nmp = TAILQ_NEXT(mp, mnt_list); 772 continue; 773 } 774 done += vlrureclaim(mp); 775 mtx_lock(&mountlist_mtx); 776 nmp = TAILQ_NEXT(mp, mnt_list); 777 vfs_unbusy(mp, td); 778 } 779 mtx_unlock(&mountlist_mtx); 780 if (done == 0) { 781#if 0 782 /* These messages are temporary debugging aids */ 783 if (vnlru_nowhere < 5) 784 printf("vnlru process getting nowhere..\n"); 785 else if (vnlru_nowhere == 5) 786 printf("vnlru process messages stopped.\n"); 787#endif 788 vnlru_nowhere++; 789 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 790 } 791 } 792} 793 794static struct kproc_desc vnlru_kp = { 795 "vnlru", 796 vnlru_proc, 797 &vnlruproc 798}; 799SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 800 801 802/* 803 * Routines having to do with the management of the vnode table. 804 */ 805 806/* 807 * Check to see if a free vnode can be recycled. If it can, 808 * return it locked with the vn lock, but not interlock. Also 809 * get the vn_start_write lock. Otherwise indicate the error. 810 */ 811static int 812vcanrecycle(struct vnode *vp, struct mount **vnmpp) 813{ 814 struct thread *td = curthread; 815 vm_object_t object; 816 int error; 817 818 /* Don't recycle if we can't get the interlock */ 819 if (!VI_TRYLOCK(vp)) 820 return (EWOULDBLOCK); 821 /* 822 * This vnode may found and locked via some other list, if so we 823 * can't recycle it yet. 824 */ 825 if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 826 return (EWOULDBLOCK); 827 /* 828 * Don't recycle if its filesystem is being suspended. 829 */ 830 if (vn_start_write(vp, vnmpp, V_NOWAIT) != 0) { 831 error = EBUSY; 832 goto done; 833 } 834 835 /* 836 * Don't recycle if we still have cached pages. 837 */ 838 if (VOP_GETVOBJECT(vp, &object) == 0) { 839 VM_OBJECT_LOCK(object); 840 if (object->resident_page_count || 841 object->ref_count) { 842 VM_OBJECT_UNLOCK(object); 843 error = EBUSY; 844 goto done; 845 } 846 VM_OBJECT_UNLOCK(object); 847 } 848 if (LIST_FIRST(&vp->v_cache_src)) { 849 /* 850 * note: nameileafonly sysctl is temporary, 851 * for debugging only, and will eventually be 852 * removed. 853 */ 854 if (nameileafonly > 0) { 855 /* 856 * Do not reuse namei-cached directory 857 * vnodes that have cached 858 * subdirectories. 859 */ 860 if (cache_leaf_test(vp) < 0) { 861 error = EISDIR; 862 goto done; 863 } 864 } else if (nameileafonly < 0 || 865 vmiodirenable == 0) { 866 /* 867 * Do not reuse namei-cached directory 868 * vnodes if nameileafonly is -1 or 869 * if VMIO backing for directories is 870 * turned off (otherwise we reuse them 871 * too quickly). 872 */ 873 error = EBUSY; 874 goto done; 875 } 876 } 877 return (0); 878done: 879 VOP_UNLOCK(vp, 0, td); 880 return (error); 881} 882 883/* 884 * Return the next vnode from the free list. 885 */ 886int 887getnewvnode(tag, mp, vops, vpp) 888 const char *tag; 889 struct mount *mp; 890 vop_t **vops; 891 struct vnode **vpp; 892{ 893 struct thread *td = curthread; /* XXX */ 894 struct vnode *vp = NULL; 895 struct vpollinfo *pollinfo = NULL; 896 struct mount *vnmp; 897 898 mtx_lock(&vnode_free_list_mtx); 899 900 /* 901 * Try to reuse vnodes if we hit the max. This situation only 902 * occurs in certain large-memory (2G+) situations. We cannot 903 * attempt to directly reclaim vnodes due to nasty recursion 904 * problems. 905 */ 906 while (numvnodes - freevnodes > desiredvnodes) { 907 if (vnlruproc_sig == 0) { 908 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 909 wakeup(vnlruproc); 910 } 911 mtx_unlock(&vnode_free_list_mtx); 912 tsleep(&vnlruproc_sig, PVFS, "vlruwk", hz); 913 mtx_lock(&vnode_free_list_mtx); 914 } 915 916 /* 917 * Attempt to reuse a vnode already on the free list, allocating 918 * a new vnode if we can't find one or if we have not reached a 919 * good minimum for good LRU performance. 920 */ 921 922 if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) { 923 int error; 924 int count; 925 926 for (count = 0; count < freevnodes; count++) { 927 vp = TAILQ_FIRST(&vnode_free_list); 928 929 KASSERT(vp->v_usecount == 0 && 930 (vp->v_iflag & VI_DOINGINACT) == 0, 931 ("getnewvnode: free vnode isn't")); 932 933 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 934 /* 935 * We have to drop the free list mtx to avoid lock 936 * order reversals with interlock. 937 */ 938 mtx_unlock(&vnode_free_list_mtx); 939 error = vcanrecycle(vp, &vnmp); 940 mtx_lock(&vnode_free_list_mtx); 941 if (error == 0) 942 break; 943 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 944 vp = NULL; 945 } 946 } 947 if (vp) { 948 freevnodes--; 949 mtx_unlock(&vnode_free_list_mtx); 950 951 cache_purge(vp); 952 VI_LOCK(vp); 953 vp->v_iflag |= VI_DOOMED; 954 vp->v_iflag &= ~VI_FREE; 955 if (vp->v_type != VBAD) { 956 VOP_UNLOCK(vp, 0, td); 957 vgonel(vp, td); 958 VI_LOCK(vp); 959 } else { 960 VOP_UNLOCK(vp, 0, td); 961 } 962 vn_finished_write(vnmp); 963 964#ifdef INVARIANTS 965 { 966 if (vp->v_data) 967 panic("cleaned vnode isn't"); 968 if (vp->v_numoutput) 969 panic("Clean vnode has pending I/O's"); 970 if (vp->v_writecount != 0) 971 panic("Non-zero write count"); 972 } 973#endif 974 if ((pollinfo = vp->v_pollinfo) != NULL) { 975 /* 976 * To avoid lock order reversals, the call to 977 * uma_zfree() must be delayed until the vnode 978 * interlock is released. 979 */ 980 vp->v_pollinfo = NULL; 981 } 982#ifdef MAC 983 mac_destroy_vnode(vp); 984#endif 985 vp->v_iflag = 0; 986 vp->v_vflag = 0; 987 vp->v_lastw = 0; 988 vp->v_lasta = 0; 989 vp->v_cstart = 0; 990 vp->v_clen = 0; 991 vp->v_socket = 0; 992 lockdestroy(vp->v_vnlock); 993 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 994 KASSERT(vp->v_cleanbufcnt == 0, ("cleanbufcnt not 0")); 995 KASSERT(vp->v_cleanblkroot == NULL, ("cleanblkroot not NULL")); 996 KASSERT(vp->v_dirtybufcnt == 0, ("dirtybufcnt not 0")); 997 KASSERT(vp->v_dirtyblkroot == NULL, ("dirtyblkroot not NULL")); 998 } else { 999 numvnodes++; 1000 mtx_unlock(&vnode_free_list_mtx); 1001 1002 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); 1003 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 1004 VI_LOCK(vp); 1005 vp->v_dd = vp; 1006 vp->v_vnlock = &vp->v_lock; 1007 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 1008 cache_purge(vp); /* Sets up v_id. */ 1009 LIST_INIT(&vp->v_cache_src); 1010 TAILQ_INIT(&vp->v_cache_dst); 1011 } 1012 1013 TAILQ_INIT(&vp->v_cleanblkhd); 1014 TAILQ_INIT(&vp->v_dirtyblkhd); 1015 vp->v_type = VNON; 1016 vp->v_tag = tag; 1017 vp->v_op = vops; 1018 *vpp = vp; 1019 vp->v_usecount = 1; 1020 vp->v_data = 0; 1021 vp->v_cachedid = -1; 1022 VI_UNLOCK(vp); 1023 if (pollinfo != NULL) { 1024 mtx_destroy(&pollinfo->vpi_lock); 1025 uma_zfree(vnodepoll_zone, pollinfo); 1026 } 1027#ifdef MAC 1028 mac_init_vnode(vp); 1029 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 1030 mac_associate_vnode_singlelabel(mp, vp); 1031#endif 1032 insmntque(vp, mp); 1033 1034 return (0); 1035} 1036 1037/* 1038 * Move a vnode from one mount queue to another. 1039 */ 1040static void 1041insmntque(vp, mp) 1042 register struct vnode *vp; 1043 register struct mount *mp; 1044{ 1045 1046 mtx_lock(&mntvnode_mtx); 1047 /* 1048 * Delete from old mount point vnode list, if on one. 1049 */ 1050 if (vp->v_mount != NULL) { 1051 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0, 1052 ("bad mount point vnode list size")); 1053 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes); 1054 vp->v_mount->mnt_nvnodelistsize--; 1055 } 1056 /* 1057 * Insert into list of vnodes for the new mount point, if available. 1058 */ 1059 if ((vp->v_mount = mp) != NULL) { 1060 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1061 mp->mnt_nvnodelistsize++; 1062 } 1063 mtx_unlock(&mntvnode_mtx); 1064} 1065 1066/* 1067 * Update outstanding I/O count and do wakeup if requested. 1068 */ 1069void 1070vwakeup(bp) 1071 register struct buf *bp; 1072{ 1073 register struct vnode *vp; 1074 1075 bp->b_flags &= ~B_WRITEINPROG; 1076 if ((vp = bp->b_vp)) { 1077 VI_LOCK(vp); 1078 vp->v_numoutput--; 1079 if (vp->v_numoutput < 0) 1080 panic("vwakeup: neg numoutput"); 1081 if ((vp->v_numoutput == 0) && (vp->v_iflag & VI_BWAIT)) { 1082 vp->v_iflag &= ~VI_BWAIT; 1083 wakeup(&vp->v_numoutput); 1084 } 1085 VI_UNLOCK(vp); 1086 } 1087} 1088 1089/* 1090 * Flush out and invalidate all buffers associated with a vnode. 1091 * Called with the underlying object locked. 1092 */ 1093int 1094vinvalbuf(vp, flags, cred, td, slpflag, slptimeo) 1095 struct vnode *vp; 1096 int flags; 1097 struct ucred *cred; 1098 struct thread *td; 1099 int slpflag, slptimeo; 1100{ 1101 struct buf *blist; 1102 int error; 1103 vm_object_t object; 1104 1105 GIANT_REQUIRED; 1106 1107 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1108 1109 VI_LOCK(vp); 1110 if (flags & V_SAVE) { 1111 while (vp->v_numoutput) { 1112 vp->v_iflag |= VI_BWAIT; 1113 error = msleep(&vp->v_numoutput, VI_MTX(vp), 1114 slpflag | (PRIBIO + 1), "vinvlbuf", slptimeo); 1115 if (error) { 1116 VI_UNLOCK(vp); 1117 return (error); 1118 } 1119 } 1120 if (!TAILQ_EMPTY(&vp->v_dirtyblkhd)) { 1121 VI_UNLOCK(vp); 1122 if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0) 1123 return (error); 1124 /* 1125 * XXX We could save a lock/unlock if this was only 1126 * enabled under INVARIANTS 1127 */ 1128 VI_LOCK(vp); 1129 if (vp->v_numoutput > 0 || 1130 !TAILQ_EMPTY(&vp->v_dirtyblkhd)) 1131 panic("vinvalbuf: dirty bufs"); 1132 } 1133 } 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 VI_UNLOCK(vp); 1156 return (error); 1157 } 1158 1159 /* 1160 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1161 * have write I/O in-progress but if there is a VM object then the 1162 * VM object can also have read-I/O in-progress. 1163 */ 1164 do { 1165 while (vp->v_numoutput > 0) { 1166 vp->v_iflag |= VI_BWAIT; 1167 msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vnvlbv", 0); 1168 } 1169 VI_UNLOCK(vp); 1170 if (VOP_GETVOBJECT(vp, &object) == 0) { 1171 VM_OBJECT_LOCK(object); 1172 vm_object_pip_wait(object, "vnvlbx"); 1173 VM_OBJECT_UNLOCK(object); 1174 } 1175 VI_LOCK(vp); 1176 } while (vp->v_numoutput > 0); 1177 VI_UNLOCK(vp); 1178 1179 /* 1180 * Destroy the copy in the VM cache, too. 1181 */ 1182 if (VOP_GETVOBJECT(vp, &object) == 0) { 1183 VM_OBJECT_LOCK(object); 1184 vm_object_page_remove(object, 0, 0, 1185 (flags & V_SAVE) ? TRUE : FALSE); 1186 VM_OBJECT_UNLOCK(object); 1187 } 1188 1189#ifdef INVARIANTS 1190 VI_LOCK(vp); 1191 if ((flags & (V_ALT | V_NORMAL)) == 0 && 1192 (!TAILQ_EMPTY(&vp->v_dirtyblkhd) || 1193 !TAILQ_EMPTY(&vp->v_cleanblkhd))) 1194 panic("vinvalbuf: flush failed"); 1195 VI_UNLOCK(vp); 1196#endif 1197 return (0); 1198} 1199 1200/* 1201 * Flush out buffers on the specified list. 1202 * 1203 */ 1204static int 1205flushbuflist(blist, flags, vp, slpflag, slptimeo, errorp) 1206 struct buf *blist; 1207 int flags; 1208 struct vnode *vp; 1209 int slpflag, slptimeo; 1210 int *errorp; 1211{ 1212 struct buf *bp, *nbp; 1213 int found, error; 1214 1215 ASSERT_VI_LOCKED(vp, "flushbuflist"); 1216 1217 for (found = 0, bp = blist; bp; bp = nbp) { 1218 nbp = TAILQ_NEXT(bp, b_vnbufs); 1219 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1220 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1221 continue; 1222 } 1223 found += 1; 1224 error = BUF_TIMELOCK(bp, 1225 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp), 1226 "flushbuf", slpflag, slptimeo); 1227 if (error) { 1228 if (error != ENOLCK) 1229 *errorp = error; 1230 goto done; 1231 } 1232 /* 1233 * XXX Since there are no node locks for NFS, I 1234 * believe there is a slight chance that a delayed 1235 * write will occur while sleeping just above, so 1236 * check for it. Note that vfs_bio_awrite expects 1237 * buffers to reside on a queue, while BUF_WRITE and 1238 * brelse do not. 1239 */ 1240 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1241 (flags & V_SAVE)) { 1242 1243 if (bp->b_vp == vp) { 1244 if (bp->b_flags & B_CLUSTEROK) { 1245 vfs_bio_awrite(bp); 1246 } else { 1247 bremfree(bp); 1248 bp->b_flags |= B_ASYNC; 1249 BUF_WRITE(bp); 1250 } 1251 } else { 1252 bremfree(bp); 1253 (void) BUF_WRITE(bp); 1254 } 1255 goto done; 1256 } 1257 bremfree(bp); 1258 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 1259 bp->b_flags &= ~B_ASYNC; 1260 brelse(bp); 1261 VI_LOCK(vp); 1262 } 1263 return (found); 1264done: 1265 VI_LOCK(vp); 1266 return (found); 1267} 1268 1269/* 1270 * Truncate a file's buffer and pages to a specified length. This 1271 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1272 * sync activity. 1273 */ 1274int 1275vtruncbuf(vp, cred, td, length, blksize) 1276 register struct vnode *vp; 1277 struct ucred *cred; 1278 struct thread *td; 1279 off_t length; 1280 int blksize; 1281{ 1282 register struct buf *bp; 1283 struct buf *nbp; 1284 int anyfreed; 1285 int trunclbn; 1286 1287 /* 1288 * Round up to the *next* lbn. 1289 */ 1290 trunclbn = (length + blksize - 1) / blksize; 1291 1292 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1293restart: 1294 VI_LOCK(vp); 1295 anyfreed = 1; 1296 for (;anyfreed;) { 1297 anyfreed = 0; 1298 for (bp = TAILQ_FIRST(&vp->v_cleanblkhd); bp; bp = nbp) { 1299 nbp = TAILQ_NEXT(bp, b_vnbufs); 1300 if (bp->b_lblkno >= trunclbn) { 1301 if (BUF_LOCK(bp, 1302 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1303 VI_MTX(vp)) == ENOLCK) 1304 goto restart; 1305 1306 bremfree(bp); 1307 bp->b_flags |= (B_INVAL | B_RELBUF); 1308 bp->b_flags &= ~B_ASYNC; 1309 brelse(bp); 1310 anyfreed = 1; 1311 1312 if (nbp && 1313 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1314 (nbp->b_vp != vp) || 1315 (nbp->b_flags & B_DELWRI))) { 1316 goto restart; 1317 } 1318 VI_LOCK(vp); 1319 } 1320 } 1321 1322 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1323 nbp = TAILQ_NEXT(bp, b_vnbufs); 1324 if (bp->b_lblkno >= trunclbn) { 1325 if (BUF_LOCK(bp, 1326 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1327 VI_MTX(vp)) == ENOLCK) 1328 goto restart; 1329 bremfree(bp); 1330 bp->b_flags |= (B_INVAL | B_RELBUF); 1331 bp->b_flags &= ~B_ASYNC; 1332 brelse(bp); 1333 anyfreed = 1; 1334 if (nbp && 1335 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1336 (nbp->b_vp != vp) || 1337 (nbp->b_flags & B_DELWRI) == 0)) { 1338 goto restart; 1339 } 1340 VI_LOCK(vp); 1341 } 1342 } 1343 } 1344 1345 if (length > 0) { 1346restartsync: 1347 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1348 nbp = TAILQ_NEXT(bp, b_vnbufs); 1349 if (bp->b_lblkno > 0) 1350 continue; 1351 /* 1352 * Since we hold the vnode lock this should only 1353 * fail if we're racing with the buf daemon. 1354 */ 1355 if (BUF_LOCK(bp, 1356 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1357 VI_MTX(vp)) == ENOLCK) { 1358 goto restart; 1359 } 1360 KASSERT((bp->b_flags & B_DELWRI), 1361 ("buf(%p) on dirty queue without DELWRI.", bp)); 1362 1363 bremfree(bp); 1364 bawrite(bp); 1365 VI_LOCK(vp); 1366 goto restartsync; 1367 } 1368 } 1369 1370 while (vp->v_numoutput > 0) { 1371 vp->v_iflag |= VI_BWAIT; 1372 msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vbtrunc", 0); 1373 } 1374 VI_UNLOCK(vp); 1375 vnode_pager_setsize(vp, length); 1376 1377 return (0); 1378} 1379 1380/* 1381 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1382 * a vnode. 1383 * 1384 * NOTE: We have to deal with the special case of a background bitmap 1385 * buffer, a situation where two buffers will have the same logical 1386 * block offset. We want (1) only the foreground buffer to be accessed 1387 * in a lookup and (2) must differentiate between the foreground and 1388 * background buffer in the splay tree algorithm because the splay 1389 * tree cannot normally handle multiple entities with the same 'index'. 1390 * We accomplish this by adding differentiating flags to the splay tree's 1391 * numerical domain. 1392 */ 1393static 1394struct buf * 1395buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1396{ 1397 struct buf dummy; 1398 struct buf *lefttreemax, *righttreemin, *y; 1399 1400 if (root == NULL) 1401 return (NULL); 1402 lefttreemax = righttreemin = &dummy; 1403 for (;;) { 1404 if (lblkno < root->b_lblkno || 1405 (lblkno == root->b_lblkno && 1406 (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1407 if ((y = root->b_left) == NULL) 1408 break; 1409 if (lblkno < y->b_lblkno) { 1410 /* Rotate right. */ 1411 root->b_left = y->b_right; 1412 y->b_right = root; 1413 root = y; 1414 if ((y = root->b_left) == NULL) 1415 break; 1416 } 1417 /* Link into the new root's right tree. */ 1418 righttreemin->b_left = root; 1419 righttreemin = root; 1420 } else if (lblkno > root->b_lblkno || 1421 (lblkno == root->b_lblkno && 1422 (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { 1423 if ((y = root->b_right) == NULL) 1424 break; 1425 if (lblkno > y->b_lblkno) { 1426 /* Rotate left. */ 1427 root->b_right = y->b_left; 1428 y->b_left = root; 1429 root = y; 1430 if ((y = root->b_right) == NULL) 1431 break; 1432 } 1433 /* Link into the new root's left tree. */ 1434 lefttreemax->b_right = root; 1435 lefttreemax = root; 1436 } else { 1437 break; 1438 } 1439 root = y; 1440 } 1441 /* Assemble the new root. */ 1442 lefttreemax->b_right = root->b_left; 1443 righttreemin->b_left = root->b_right; 1444 root->b_left = dummy.b_right; 1445 root->b_right = dummy.b_left; 1446 return (root); 1447} 1448 1449static 1450void 1451buf_vlist_remove(struct buf *bp) 1452{ 1453 struct vnode *vp = bp->b_vp; 1454 struct buf *root; 1455 1456 ASSERT_VI_LOCKED(vp, "buf_vlist_remove"); 1457 if (bp->b_xflags & BX_VNDIRTY) { 1458 if (bp != vp->v_dirtyblkroot) { 1459 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot); 1460 KASSERT(root == bp, ("splay lookup failed during dirty remove")); 1461 } 1462 if (bp->b_left == NULL) { 1463 root = bp->b_right; 1464 } else { 1465 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1466 root->b_right = bp->b_right; 1467 } 1468 vp->v_dirtyblkroot = root; 1469 TAILQ_REMOVE(&vp->v_dirtyblkhd, bp, b_vnbufs); 1470 vp->v_dirtybufcnt--; 1471 } else { 1472 /* KASSERT(bp->b_xflags & BX_VNCLEAN, ("bp wasn't clean")); */ 1473 if (bp != vp->v_cleanblkroot) { 1474 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot); 1475 KASSERT(root == bp, ("splay lookup failed during clean remove")); 1476 } 1477 if (bp->b_left == NULL) { 1478 root = bp->b_right; 1479 } else { 1480 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1481 root->b_right = bp->b_right; 1482 } 1483 vp->v_cleanblkroot = root; 1484 TAILQ_REMOVE(&vp->v_cleanblkhd, bp, b_vnbufs); 1485 vp->v_cleanbufcnt--; 1486 } 1487 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1488} 1489 1490/* 1491 * Add the buffer to the sorted clean or dirty block list using a 1492 * splay tree algorithm. 1493 * 1494 * NOTE: xflags is passed as a constant, optimizing this inline function! 1495 */ 1496static 1497void 1498buf_vlist_add(struct buf *bp, struct vnode *vp, b_xflags_t xflags) 1499{ 1500 struct buf *root; 1501 1502 ASSERT_VI_LOCKED(vp, "buf_vlist_add"); 1503 bp->b_xflags |= xflags; 1504 if (xflags & BX_VNDIRTY) { 1505 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot); 1506 if (root == NULL) { 1507 bp->b_left = NULL; 1508 bp->b_right = NULL; 1509 TAILQ_INSERT_TAIL(&vp->v_dirtyblkhd, bp, b_vnbufs); 1510 } else if (bp->b_lblkno < root->b_lblkno || 1511 (bp->b_lblkno == root->b_lblkno && 1512 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1513 bp->b_left = root->b_left; 1514 bp->b_right = root; 1515 root->b_left = NULL; 1516 TAILQ_INSERT_BEFORE(root, bp, b_vnbufs); 1517 } else { 1518 bp->b_right = root->b_right; 1519 bp->b_left = root; 1520 root->b_right = NULL; 1521 TAILQ_INSERT_AFTER(&vp->v_dirtyblkhd, 1522 root, bp, b_vnbufs); 1523 } 1524 vp->v_dirtybufcnt++; 1525 vp->v_dirtyblkroot = bp; 1526 } else { 1527 /* KASSERT(xflags & BX_VNCLEAN, ("xflags not clean")); */ 1528 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot); 1529 if (root == NULL) { 1530 bp->b_left = NULL; 1531 bp->b_right = NULL; 1532 TAILQ_INSERT_TAIL(&vp->v_cleanblkhd, bp, b_vnbufs); 1533 } else if (bp->b_lblkno < root->b_lblkno || 1534 (bp->b_lblkno == root->b_lblkno && 1535 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1536 bp->b_left = root->b_left; 1537 bp->b_right = root; 1538 root->b_left = NULL; 1539 TAILQ_INSERT_BEFORE(root, bp, b_vnbufs); 1540 } else { 1541 bp->b_right = root->b_right; 1542 bp->b_left = root; 1543 root->b_right = NULL; 1544 TAILQ_INSERT_AFTER(&vp->v_cleanblkhd, 1545 root, bp, b_vnbufs); 1546 } 1547 vp->v_cleanbufcnt++; 1548 vp->v_cleanblkroot = bp; 1549 } 1550} 1551 1552/* 1553 * Lookup a buffer using the splay tree. Note that we specifically avoid 1554 * shadow buffers used in background bitmap writes. 1555 * 1556 * This code isn't quite efficient as it could be because we are maintaining 1557 * two sorted lists and do not know which list the block resides in. 1558 * 1559 * During a "make buildworld" the desired buffer is found at one of 1560 * the roots more than 60% of the time. Thus, checking both roots 1561 * before performing either splay eliminates unnecessary splays on the 1562 * first tree splayed. 1563 */ 1564struct buf * 1565gbincore(struct vnode *vp, daddr_t lblkno) 1566{ 1567 struct buf *bp; 1568 1569 GIANT_REQUIRED; 1570 1571 ASSERT_VI_LOCKED(vp, "gbincore"); 1572 if ((bp = vp->v_cleanblkroot) != NULL && 1573 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1574 return (bp); 1575 if ((bp = vp->v_dirtyblkroot) != NULL && 1576 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1577 return (bp); 1578 if ((bp = vp->v_cleanblkroot) != NULL) { 1579 vp->v_cleanblkroot = bp = buf_splay(lblkno, 0, bp); 1580 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1581 return (bp); 1582 } 1583 if ((bp = vp->v_dirtyblkroot) != NULL) { 1584 vp->v_dirtyblkroot = bp = buf_splay(lblkno, 0, bp); 1585 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1586 return (bp); 1587 } 1588 return (NULL); 1589} 1590 1591/* 1592 * Associate a buffer with a vnode. 1593 */ 1594void 1595bgetvp(vp, bp) 1596 register struct vnode *vp; 1597 register struct buf *bp; 1598{ 1599 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 1600 1601 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1602 ("bgetvp: bp already attached! %p", bp)); 1603 1604 ASSERT_VI_LOCKED(vp, "bgetvp"); 1605 vholdl(vp); 1606 bp->b_vp = vp; 1607 bp->b_dev = vn_todev(vp); 1608 /* 1609 * Insert onto list for new vnode. 1610 */ 1611 buf_vlist_add(bp, vp, BX_VNCLEAN); 1612} 1613 1614/* 1615 * Disassociate a buffer from a vnode. 1616 */ 1617void 1618brelvp(bp) 1619 register struct buf *bp; 1620{ 1621 struct vnode *vp; 1622 1623 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1624 1625 /* 1626 * Delete from old vnode list, if on one. 1627 */ 1628 vp = bp->b_vp; 1629 VI_LOCK(vp); 1630 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1631 buf_vlist_remove(bp); 1632 if ((vp->v_iflag & VI_ONWORKLST) && TAILQ_EMPTY(&vp->v_dirtyblkhd)) { 1633 vp->v_iflag &= ~VI_ONWORKLST; 1634 mtx_lock(&sync_mtx); 1635 LIST_REMOVE(vp, v_synclist); 1636 mtx_unlock(&sync_mtx); 1637 } 1638 vdropl(vp); 1639 bp->b_vp = (struct vnode *) 0; 1640 if (bp->b_object) 1641 bp->b_object = NULL; 1642 VI_UNLOCK(vp); 1643} 1644 1645/* 1646 * Add an item to the syncer work queue. 1647 */ 1648static void 1649vn_syncer_add_to_worklist(struct vnode *vp, int delay) 1650{ 1651 int slot; 1652 1653 ASSERT_VI_LOCKED(vp, "vn_syncer_add_to_worklist"); 1654 1655 mtx_lock(&sync_mtx); 1656 if (vp->v_iflag & VI_ONWORKLST) 1657 LIST_REMOVE(vp, v_synclist); 1658 else 1659 vp->v_iflag |= VI_ONWORKLST; 1660 1661 if (delay > syncer_maxdelay - 2) 1662 delay = syncer_maxdelay - 2; 1663 slot = (syncer_delayno + delay) & syncer_mask; 1664 1665 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist); 1666 mtx_unlock(&sync_mtx); 1667} 1668 1669struct proc *updateproc; 1670static void sched_sync(void); 1671static struct kproc_desc up_kp = { 1672 "syncer", 1673 sched_sync, 1674 &updateproc 1675}; 1676SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 1677 1678/* 1679 * System filesystem synchronizer daemon. 1680 */ 1681static void 1682sched_sync(void) 1683{ 1684 struct synclist *slp; 1685 struct vnode *vp; 1686 struct mount *mp; 1687 long starttime; 1688 struct thread *td = FIRST_THREAD_IN_PROC(updateproc); /* XXXKSE */ 1689 1690 mtx_lock(&Giant); 1691 1692 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, td->td_proc, 1693 SHUTDOWN_PRI_LAST); 1694 1695 for (;;) { 1696 kthread_suspend_check(td->td_proc); 1697 1698 starttime = time_second; 1699 1700 /* 1701 * Push files whose dirty time has expired. Be careful 1702 * of interrupt race on slp queue. 1703 */ 1704 mtx_lock(&sync_mtx); 1705 slp = &syncer_workitem_pending[syncer_delayno]; 1706 syncer_delayno += 1; 1707 if (syncer_delayno == syncer_maxdelay) 1708 syncer_delayno = 0; 1709 1710 while ((vp = LIST_FIRST(slp)) != NULL) { 1711 /* 1712 * XXX we have no guarantees about this vnodes 1713 * identity due to a lack of interlock. 1714 */ 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 mtx_lock(&sync_mtx); 1724 if (LIST_FIRST(slp) == vp) { 1725 mtx_unlock(&sync_mtx); 1726 /* 1727 * Note: VFS vnodes can remain on the 1728 * worklist too with no dirty blocks, but 1729 * since sync_fsync() moves it to a different 1730 * slot we are safe. 1731 */ 1732 VI_LOCK(vp); 1733 if (TAILQ_EMPTY(&vp->v_dirtyblkhd) && 1734 !vn_isdisk(vp, NULL)) { 1735 panic("sched_sync: fsync failed " 1736 "vp %p tag %s", vp, vp->v_tag); 1737 } 1738 /* 1739 * Put us back on the worklist. The worklist 1740 * routine will remove us from our current 1741 * position and then add us back in at a later 1742 * position. 1743 */ 1744 vn_syncer_add_to_worklist(vp, syncdelay); 1745 VI_UNLOCK(vp); 1746 mtx_lock(&sync_mtx); 1747 } 1748 } 1749 mtx_unlock(&sync_mtx); 1750 1751 /* 1752 * Do soft update processing. 1753 */ 1754 if (softdep_process_worklist_hook != NULL) 1755 (*softdep_process_worklist_hook)(NULL); 1756 1757 /* 1758 * The variable rushjob allows the kernel to speed up the 1759 * processing of the filesystem syncer process. A rushjob 1760 * value of N tells the filesystem syncer to process the next 1761 * N seconds worth of work on its queue ASAP. Currently rushjob 1762 * is used by the soft update code to speed up the filesystem 1763 * syncer process when the incore state is getting so far 1764 * ahead of the disk that the kernel memory pool is being 1765 * threatened with exhaustion. 1766 */ 1767 mtx_lock(&sync_mtx); 1768 if (rushjob > 0) { 1769 rushjob -= 1; 1770 mtx_unlock(&sync_mtx); 1771 continue; 1772 } 1773 mtx_unlock(&sync_mtx); 1774 /* 1775 * If it has taken us less than a second to process the 1776 * current work, then wait. Otherwise start right over 1777 * again. We can still lose time if any single round 1778 * takes more than two seconds, but it does not really 1779 * matter as we are just trying to generally pace the 1780 * filesystem activity. 1781 */ 1782 if (time_second == starttime) 1783 tsleep(&lbolt, PPAUSE, "syncer", 0); 1784 } 1785} 1786 1787/* 1788 * Request the syncer daemon to speed up its work. 1789 * We never push it to speed up more than half of its 1790 * normal turn time, otherwise it could take over the cpu. 1791 * XXXKSE only one update? 1792 */ 1793int 1794speedup_syncer() 1795{ 1796 struct thread *td; 1797 int ret = 0; 1798 1799 td = FIRST_THREAD_IN_PROC(updateproc); 1800 mtx_lock_spin(&sched_lock); 1801 if (td->td_wchan == &lbolt) { 1802 unsleep(td); 1803 TD_CLR_SLEEPING(td); 1804 setrunnable(td); 1805 } 1806 mtx_unlock_spin(&sched_lock); 1807 mtx_lock(&sync_mtx); 1808 if (rushjob < syncdelay / 2) { 1809 rushjob += 1; 1810 stat_rush_requests += 1; 1811 ret = 1; 1812 } 1813 mtx_unlock(&sync_mtx); 1814 return (ret); 1815} 1816 1817/* 1818 * Associate a p-buffer with a vnode. 1819 * 1820 * Also sets B_PAGING flag to indicate that vnode is not fully associated 1821 * with the buffer. i.e. the bp has not been linked into the vnode or 1822 * ref-counted. 1823 */ 1824void 1825pbgetvp(vp, bp) 1826 register struct vnode *vp; 1827 register struct buf *bp; 1828{ 1829 1830 KASSERT(bp->b_vp == NULL, ("pbgetvp: not free")); 1831 1832 bp->b_vp = vp; 1833 bp->b_flags |= B_PAGING; 1834 bp->b_dev = vn_todev(vp); 1835} 1836 1837/* 1838 * Disassociate a p-buffer from a vnode. 1839 */ 1840void 1841pbrelvp(bp) 1842 register struct buf *bp; 1843{ 1844 1845 KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL")); 1846 1847 /* XXX REMOVE ME */ 1848 VI_LOCK(bp->b_vp); 1849 if (TAILQ_NEXT(bp, b_vnbufs) != NULL) { 1850 panic( 1851 "relpbuf(): b_vp was probably reassignbuf()d %p %x", 1852 bp, 1853 (int)bp->b_flags 1854 ); 1855 } 1856 VI_UNLOCK(bp->b_vp); 1857 bp->b_vp = (struct vnode *) 0; 1858 bp->b_flags &= ~B_PAGING; 1859} 1860 1861/* 1862 * Reassign a buffer from one vnode to another. 1863 * Used to assign file specific control information 1864 * (indirect blocks) to the vnode to which they belong. 1865 */ 1866void 1867reassignbuf(bp, newvp) 1868 register struct buf *bp; 1869 register struct vnode *newvp; 1870{ 1871 struct vnode *vp; 1872 int delay; 1873 1874 if (newvp == NULL) { 1875 printf("reassignbuf: NULL"); 1876 return; 1877 } 1878 vp = bp->b_vp; 1879 ++reassignbufcalls; 1880 1881 /* 1882 * B_PAGING flagged buffers cannot be reassigned because their vp 1883 * is not fully linked in. 1884 */ 1885 if (bp->b_flags & B_PAGING) 1886 panic("cannot reassign paging buffer"); 1887 1888 /* 1889 * Delete from old vnode list, if on one. 1890 */ 1891 VI_LOCK(vp); 1892 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) { 1893 buf_vlist_remove(bp); 1894 if (vp != newvp) { 1895 vdropl(bp->b_vp); 1896 bp->b_vp = NULL; /* for clarification */ 1897 } 1898 } 1899 if (vp != newvp) { 1900 VI_UNLOCK(vp); 1901 VI_LOCK(newvp); 1902 } 1903 /* 1904 * If dirty, put on list of dirty buffers; otherwise insert onto list 1905 * of clean buffers. 1906 */ 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} 1942 1943/* 1944 * Create a vnode for a device. 1945 * Used for mounting the root filesystem. 1946 */ 1947int 1948bdevvp(dev, vpp) 1949 dev_t dev; 1950 struct vnode **vpp; 1951{ 1952 register struct vnode *vp; 1953 struct vnode *nvp; 1954 int error; 1955 1956 if (dev == NODEV) { 1957 *vpp = NULLVP; 1958 return (ENXIO); 1959 } 1960 if (vfinddev(dev, VCHR, vpp)) 1961 return (0); 1962 error = getnewvnode("none", (struct mount *)0, spec_vnodeop_p, &nvp); 1963 if (error) { 1964 *vpp = NULLVP; 1965 return (error); 1966 } 1967 vp = nvp; 1968 vp->v_type = VCHR; 1969 addalias(vp, dev); 1970 *vpp = vp; 1971 return (0); 1972} 1973 1974static void 1975v_incr_usecount(struct vnode *vp, int delta) 1976{ 1977 vp->v_usecount += delta; 1978 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1979 mtx_lock(&spechash_mtx); 1980 vp->v_rdev->si_usecount += delta; 1981 mtx_unlock(&spechash_mtx); 1982 } 1983} 1984 1985/* 1986 * Add vnode to the alias list hung off the dev_t. 1987 * 1988 * The reason for this gunk is that multiple vnodes can reference 1989 * the same physical device, so checking vp->v_usecount to see 1990 * how many users there are is inadequate; the v_usecount for 1991 * the vnodes need to be accumulated. vcount() does that. 1992 */ 1993struct vnode * 1994addaliasu(nvp, nvp_rdev) 1995 struct vnode *nvp; 1996 udev_t nvp_rdev; 1997{ 1998 struct vnode *ovp; 1999 vop_t **ops; 2000 dev_t dev; 2001 2002 if (nvp->v_type == VBLK) 2003 return (nvp); 2004 if (nvp->v_type != VCHR) 2005 panic("addaliasu on non-special vnode"); 2006 dev = udev2dev(nvp_rdev, 0); 2007 /* 2008 * Check to see if we have a bdevvp vnode with no associated 2009 * filesystem. If so, we want to associate the filesystem of 2010 * the new newly instigated vnode with the bdevvp vnode and 2011 * discard the newly created vnode rather than leaving the 2012 * bdevvp vnode lying around with no associated filesystem. 2013 */ 2014 if (vfinddev(dev, nvp->v_type, &ovp) == 0 || ovp->v_data != NULL) { 2015 addalias(nvp, dev); 2016 return (nvp); 2017 } 2018 /* 2019 * Discard unneeded vnode, but save its node specific data. 2020 * Note that if there is a lock, it is carried over in the 2021 * node specific data to the replacement vnode. 2022 */ 2023 vref(ovp); 2024 ovp->v_data = nvp->v_data; 2025 ovp->v_tag = nvp->v_tag; 2026 nvp->v_data = NULL; 2027 lockdestroy(ovp->v_vnlock); 2028 lockinit(ovp->v_vnlock, PVFS, nvp->v_vnlock->lk_wmesg, 2029 nvp->v_vnlock->lk_timo, nvp->v_vnlock->lk_flags & LK_EXTFLG_MASK); 2030 ops = ovp->v_op; 2031 ovp->v_op = nvp->v_op; 2032 if (VOP_ISLOCKED(nvp, curthread)) { 2033 VOP_UNLOCK(nvp, 0, curthread); 2034 vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread); 2035 } 2036 nvp->v_op = ops; 2037 insmntque(ovp, nvp->v_mount); 2038 vrele(nvp); 2039 vgone(nvp); 2040 return (ovp); 2041} 2042 2043/* This is a local helper function that do the same as addaliasu, but for a 2044 * dev_t instead of an udev_t. */ 2045static void 2046addalias(nvp, dev) 2047 struct vnode *nvp; 2048 dev_t dev; 2049{ 2050 2051 KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode")); 2052 nvp->v_rdev = dev; 2053 VI_LOCK(nvp); 2054 mtx_lock(&spechash_mtx); 2055 SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext); 2056 dev->si_usecount += nvp->v_usecount; 2057 mtx_unlock(&spechash_mtx); 2058 VI_UNLOCK(nvp); 2059} 2060 2061/* 2062 * Grab a particular vnode from the free list, increment its 2063 * reference count and lock it. The vnode lock bit is set if the 2064 * vnode is being eliminated in vgone. The process is awakened 2065 * when the transition is completed, and an error returned to 2066 * indicate that the vnode is no longer usable (possibly having 2067 * been changed to a new filesystem type). 2068 */ 2069int 2070vget(vp, flags, td) 2071 register struct vnode *vp; 2072 int flags; 2073 struct thread *td; 2074{ 2075 int error; 2076 2077 /* 2078 * If the vnode is in the process of being cleaned out for 2079 * another use, we wait for the cleaning to finish and then 2080 * return failure. Cleaning is determined by checking that 2081 * the VI_XLOCK flag is set. 2082 */ 2083 if ((flags & LK_INTERLOCK) == 0) 2084 VI_LOCK(vp); 2085 if (vp->v_iflag & VI_XLOCK && vp->v_vxproc != curthread) { 2086 if ((flags & LK_NOWAIT) == 0) { 2087 vp->v_iflag |= VI_XWANT; 2088 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0); 2089 } 2090 return (ENOENT); 2091 } 2092 2093 v_incr_usecount(vp, 1); 2094 2095 if (VSHOULDBUSY(vp)) 2096 vbusy(vp); 2097 if (flags & LK_TYPE_MASK) { 2098 if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { 2099 /* 2100 * must expand vrele here because we do not want 2101 * to call VOP_INACTIVE if the reference count 2102 * drops back to zero since it was never really 2103 * active. We must remove it from the free list 2104 * before sleeping so that multiple processes do 2105 * not try to recycle it. 2106 */ 2107 VI_LOCK(vp); 2108 v_incr_usecount(vp, -1); 2109 if (VSHOULDFREE(vp)) 2110 vfree(vp); 2111 else 2112 vlruvp(vp); 2113 VI_UNLOCK(vp); 2114 } 2115 return (error); 2116 } 2117 VI_UNLOCK(vp); 2118 return (0); 2119} 2120 2121/* 2122 * Increase the reference count of a vnode. 2123 */ 2124void 2125vref(struct vnode *vp) 2126{ 2127 VI_LOCK(vp); 2128 v_incr_usecount(vp, 1); 2129 VI_UNLOCK(vp); 2130} 2131 2132/* 2133 * Return reference count of a vnode. 2134 * 2135 * The results of this call are only guaranteed when some mechanism other 2136 * than the VI lock is used to stop other processes from gaining references 2137 * to the vnode. This may be the case if the caller holds the only reference. 2138 * This is also useful when stale data is acceptable as race conditions may 2139 * be accounted for by some other means. 2140 */ 2141int 2142vrefcnt(struct vnode *vp) 2143{ 2144 int usecnt; 2145 2146 VI_LOCK(vp); 2147 usecnt = vp->v_usecount; 2148 VI_UNLOCK(vp); 2149 2150 return (usecnt); 2151} 2152 2153 2154/* 2155 * Vnode put/release. 2156 * If count drops to zero, call inactive routine and return to freelist. 2157 */ 2158void 2159vrele(vp) 2160 struct vnode *vp; 2161{ 2162 struct thread *td = curthread; /* XXX */ 2163 2164 KASSERT(vp != NULL, ("vrele: null vp")); 2165 2166 VI_LOCK(vp); 2167 2168 /* Skip this v_writecount check if we're going to panic below. */ 2169 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2170 ("vrele: missed vn_close")); 2171 2172 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2173 vp->v_usecount == 1)) { 2174 v_incr_usecount(vp, -1); 2175 VI_UNLOCK(vp); 2176 2177 return; 2178 } 2179 2180 if (vp->v_usecount == 1) { 2181 v_incr_usecount(vp, -1); 2182 /* 2183 * We must call VOP_INACTIVE with the node locked. Mark 2184 * as VI_DOINGINACT to avoid recursion. 2185 */ 2186 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) { 2187 VI_LOCK(vp); 2188 vp->v_iflag |= VI_DOINGINACT; 2189 VI_UNLOCK(vp); 2190 VOP_INACTIVE(vp, td); 2191 VI_LOCK(vp); 2192 KASSERT(vp->v_iflag & VI_DOINGINACT, 2193 ("vrele: lost VI_DOINGINACT")); 2194 vp->v_iflag &= ~VI_DOINGINACT; 2195 } else 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 vp->v_holdcnt++; 2282 if (VSHOULDBUSY(vp)) 2283 vbusy(vp); 2284} 2285 2286/* 2287 * Note that there is one less who cares about this vnode. vdrop() is the 2288 * opposite of vhold(). 2289 */ 2290void 2291vdrop(struct vnode *vp) 2292{ 2293 VI_LOCK(vp); 2294 vdropl(vp); 2295 VI_UNLOCK(vp); 2296} 2297 2298void 2299vdropl(vp) 2300 register struct vnode *vp; 2301{ 2302 if (vp->v_holdcnt <= 0) 2303 panic("vdrop: holdcnt"); 2304 vp->v_holdcnt--; 2305 if (VSHOULDFREE(vp)) 2306 vfree(vp); 2307 else 2308 vlruvp(vp); 2309} 2310 2311/* 2312 * Remove any vnodes in the vnode table belonging to mount point mp. 2313 * 2314 * If FORCECLOSE is not specified, there should not be any active ones, 2315 * return error if any are found (nb: this is a user error, not a 2316 * system error). If FORCECLOSE is specified, detach any active vnodes 2317 * that are found. 2318 * 2319 * If WRITECLOSE is set, only flush out regular file vnodes open for 2320 * writing. 2321 * 2322 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2323 * 2324 * `rootrefs' specifies the base reference count for the root vnode 2325 * of this filesystem. The root vnode is considered busy if its 2326 * v_usecount exceeds this value. On a successful return, vflush() 2327 * will call vrele() on the root vnode exactly rootrefs times. 2328 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2329 * be zero. 2330 */ 2331#ifdef DIAGNOSTIC 2332static int busyprt = 0; /* print out busy vnodes */ 2333SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 2334#endif 2335 2336int 2337vflush(mp, rootrefs, flags) 2338 struct mount *mp; 2339 int rootrefs; 2340 int flags; 2341{ 2342 struct thread *td = curthread; /* XXX */ 2343 struct vnode *vp, *nvp, *rootvp = NULL; 2344 struct vattr vattr; 2345 int busy = 0, error; 2346 2347 if (rootrefs > 0) { 2348 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2349 ("vflush: bad args")); 2350 /* 2351 * Get the filesystem root vnode. We can vput() it 2352 * immediately, since with rootrefs > 0, it won't go away. 2353 */ 2354 if ((error = VFS_ROOT(mp, &rootvp)) != 0) 2355 return (error); 2356 vput(rootvp); 2357 2358 } 2359 mtx_lock(&mntvnode_mtx); 2360loop: 2361 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp; vp = nvp) { 2362 /* 2363 * Make sure this vnode wasn't reclaimed in getnewvnode(). 2364 * Start over if it has (it won't be on the list anymore). 2365 */ 2366 if (vp->v_mount != mp) 2367 goto loop; 2368 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 2369 2370 VI_LOCK(vp); 2371 mtx_unlock(&mntvnode_mtx); 2372 /* 2373 * XXX Does not check vn_lock error. Should restart loop if 2374 * error == ENOENT. 2375 */ 2376 vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_RETRY, td); 2377 /* 2378 * This vnode could have been reclaimed while we were 2379 * waiting for the lock since we are not holding a 2380 * reference. 2381 * Start over if the vnode was reclaimed. 2382 */ 2383 if (vp->v_mount != mp) { 2384 VOP_UNLOCK(vp, 0, td); 2385 mtx_lock(&mntvnode_mtx); 2386 goto loop; 2387 } 2388 /* 2389 * Skip over a vnodes marked VV_SYSTEM. 2390 */ 2391 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2392 VOP_UNLOCK(vp, 0, td); 2393 mtx_lock(&mntvnode_mtx); 2394 continue; 2395 } 2396 /* 2397 * If WRITECLOSE is set, flush out unlinked but still open 2398 * files (even if open only for reading) and regular file 2399 * vnodes open for writing. 2400 */ 2401 if (flags & WRITECLOSE) { 2402 error = VOP_GETATTR(vp, &vattr, td->td_ucred, td); 2403 VI_LOCK(vp); 2404 2405 if ((vp->v_type == VNON || 2406 (error == 0 && vattr.va_nlink > 0)) && 2407 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2408 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2409 mtx_lock(&mntvnode_mtx); 2410 continue; 2411 } 2412 } else 2413 VI_LOCK(vp); 2414 2415 VOP_UNLOCK(vp, 0, td); 2416 2417 /* 2418 * With v_usecount == 0, all we need to do is clear out the 2419 * vnode data structures and we are done. 2420 */ 2421 if (vp->v_usecount == 0) { 2422 vgonel(vp, td); 2423 mtx_lock(&mntvnode_mtx); 2424 continue; 2425 } 2426 2427 /* 2428 * If FORCECLOSE is set, forcibly close the vnode. For block 2429 * or character devices, revert to an anonymous device. For 2430 * all other files, just kill them. 2431 */ 2432 if (flags & FORCECLOSE) { 2433 if (vp->v_type != VCHR) { 2434 vgonel(vp, td); 2435 } else { 2436 vclean(vp, 0, td); 2437 VI_UNLOCK(vp); 2438 vp->v_op = spec_vnodeop_p; 2439 insmntque(vp, (struct mount *) 0); 2440 } 2441 mtx_lock(&mntvnode_mtx); 2442 continue; 2443 } 2444#ifdef DIAGNOSTIC 2445 if (busyprt) 2446 vprint("vflush: busy vnode", vp); 2447#endif 2448 VI_UNLOCK(vp); 2449 mtx_lock(&mntvnode_mtx); 2450 busy++; 2451 } 2452 mtx_unlock(&mntvnode_mtx); 2453 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2454 /* 2455 * If just the root vnode is busy, and if its refcount 2456 * is equal to `rootrefs', then go ahead and kill it. 2457 */ 2458 VI_LOCK(rootvp); 2459 KASSERT(busy > 0, ("vflush: not busy")); 2460 KASSERT(rootvp->v_usecount >= rootrefs, ("vflush: rootrefs")); 2461 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2462 vgonel(rootvp, td); 2463 busy = 0; 2464 } else 2465 VI_UNLOCK(rootvp); 2466 } 2467 if (busy) 2468 return (EBUSY); 2469 for (; rootrefs > 0; rootrefs--) 2470 vrele(rootvp); 2471 return (0); 2472} 2473 2474/* 2475 * This moves a now (likely recyclable) vnode to the end of the 2476 * mountlist. XXX However, it is temporarily disabled until we 2477 * can clean up ffs_sync() and friends, which have loop restart 2478 * conditions which this code causes to operate O(N^2). 2479 */ 2480static void 2481vlruvp(struct vnode *vp) 2482{ 2483#if 0 2484 struct mount *mp; 2485 2486 if ((mp = vp->v_mount) != NULL) { 2487 mtx_lock(&mntvnode_mtx); 2488 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2489 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2490 mtx_unlock(&mntvnode_mtx); 2491 } 2492#endif 2493} 2494 2495/* 2496 * Disassociate the underlying filesystem from a vnode. 2497 */ 2498static void 2499vclean(vp, flags, td) 2500 struct vnode *vp; 2501 int flags; 2502 struct thread *td; 2503{ 2504 int active; 2505 2506 ASSERT_VI_LOCKED(vp, "vclean"); 2507 /* 2508 * Check to see if the vnode is in use. If so we have to reference it 2509 * before we clean it out so that its count cannot fall to zero and 2510 * generate a race against ourselves to recycle it. 2511 */ 2512 if ((active = vp->v_usecount)) 2513 v_incr_usecount(vp, 1); 2514 2515 /* 2516 * Prevent the vnode from being recycled or brought into use while we 2517 * clean it out. 2518 */ 2519 if (vp->v_iflag & VI_XLOCK) 2520 panic("vclean: deadlock"); 2521 vp->v_iflag |= VI_XLOCK; 2522 vp->v_vxproc = curthread; 2523 /* 2524 * Even if the count is zero, the VOP_INACTIVE routine may still 2525 * have the object locked while it cleans it out. The VOP_LOCK 2526 * ensures that the VOP_INACTIVE routine is done with its work. 2527 * For active vnodes, it ensures that no other activity can 2528 * occur while the underlying object is being cleaned out. 2529 */ 2530 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2531 2532 /* 2533 * Clean out any buffers associated with the vnode. 2534 * If the flush fails, just toss the buffers. 2535 */ 2536 if (flags & DOCLOSE) { 2537 struct buf *bp; 2538 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 2539 if (bp != NULL) 2540 (void) vn_write_suspend_wait(vp, NULL, V_WAIT); 2541 if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0) 2542 vinvalbuf(vp, 0, NOCRED, td, 0, 0); 2543 } 2544 2545 VOP_DESTROYVOBJECT(vp); 2546 2547 /* 2548 * Any other processes trying to obtain this lock must first 2549 * wait for VXLOCK to clear, then call the new lock operation. 2550 */ 2551 VOP_UNLOCK(vp, 0, td); 2552 2553 /* 2554 * If purging an active vnode, it must be closed and 2555 * deactivated before being reclaimed. Note that the 2556 * VOP_INACTIVE will unlock the vnode. 2557 */ 2558 if (active) { 2559 if (flags & DOCLOSE) 2560 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2561 VI_LOCK(vp); 2562 if ((vp->v_iflag & VI_DOINGINACT) == 0) { 2563 vp->v_iflag |= VI_DOINGINACT; 2564 VI_UNLOCK(vp); 2565 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 2566 panic("vclean: cannot relock."); 2567 VOP_INACTIVE(vp, td); 2568 VI_LOCK(vp); 2569 KASSERT(vp->v_iflag & VI_DOINGINACT, 2570 ("vclean: lost VI_DOINGINACT")); 2571 vp->v_iflag &= ~VI_DOINGINACT; 2572 } 2573 VI_UNLOCK(vp); 2574 } 2575 2576 /* 2577 * Reclaim the vnode. 2578 */ 2579 if (VOP_RECLAIM(vp, td)) 2580 panic("vclean: cannot reclaim"); 2581 2582 if (active) { 2583 /* 2584 * Inline copy of vrele() since VOP_INACTIVE 2585 * has already been called. 2586 */ 2587 VI_LOCK(vp); 2588 v_incr_usecount(vp, -1); 2589 if (vp->v_usecount <= 0) { 2590#ifdef INVARIANTS 2591 if (vp->v_usecount < 0 || vp->v_writecount != 0) { 2592 vprint("vclean: bad ref count", vp); 2593 panic("vclean: ref cnt"); 2594 } 2595#endif 2596 if (VSHOULDFREE(vp)) 2597 vfree(vp); 2598 } 2599 VI_UNLOCK(vp); 2600 } 2601 2602 cache_purge(vp); 2603 VI_LOCK(vp); 2604 if (VSHOULDFREE(vp)) 2605 vfree(vp); 2606 2607 /* 2608 * Done with purge, reset to the standard lock and 2609 * notify sleepers of the grim news. 2610 */ 2611 vp->v_vnlock = &vp->v_lock; 2612 vp->v_op = dead_vnodeop_p; 2613 if (vp->v_pollinfo != NULL) 2614 vn_pollgone(vp); 2615 vp->v_tag = "none"; 2616 vp->v_iflag &= ~VI_XLOCK; 2617 vp->v_vxproc = NULL; 2618 if (vp->v_iflag & VI_XWANT) { 2619 vp->v_iflag &= ~VI_XWANT; 2620 wakeup(vp); 2621 } 2622} 2623 2624/* 2625 * Eliminate all activity associated with the requested vnode 2626 * and with all vnodes aliased to the requested vnode. 2627 */ 2628int 2629vop_revoke(ap) 2630 struct vop_revoke_args /* { 2631 struct vnode *a_vp; 2632 int a_flags; 2633 } */ *ap; 2634{ 2635 struct vnode *vp, *vq; 2636 dev_t dev; 2637 2638 KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke")); 2639 vp = ap->a_vp; 2640 KASSERT((vp->v_type == VCHR), ("vop_revoke: not VCHR")); 2641 2642 VI_LOCK(vp); 2643 /* 2644 * If a vgone (or vclean) is already in progress, 2645 * wait until it is done and return. 2646 */ 2647 if (vp->v_iflag & VI_XLOCK) { 2648 vp->v_iflag |= VI_XWANT; 2649 msleep(vp, VI_MTX(vp), PINOD | PDROP, 2650 "vop_revokeall", 0); 2651 return (0); 2652 } 2653 VI_UNLOCK(vp); 2654 dev = vp->v_rdev; 2655 for (;;) { 2656 mtx_lock(&spechash_mtx); 2657 vq = SLIST_FIRST(&dev->si_hlist); 2658 mtx_unlock(&spechash_mtx); 2659 if (!vq) 2660 break; 2661 vgone(vq); 2662 } 2663 return (0); 2664} 2665 2666/* 2667 * Recycle an unused vnode to the front of the free list. 2668 * Release the passed interlock if the vnode will be recycled. 2669 */ 2670int 2671vrecycle(vp, inter_lkp, td) 2672 struct vnode *vp; 2673 struct mtx *inter_lkp; 2674 struct thread *td; 2675{ 2676 2677 VI_LOCK(vp); 2678 if (vp->v_usecount == 0) { 2679 if (inter_lkp) { 2680 mtx_unlock(inter_lkp); 2681 } 2682 vgonel(vp, td); 2683 return (1); 2684 } 2685 VI_UNLOCK(vp); 2686 return (0); 2687} 2688 2689/* 2690 * Eliminate all activity associated with a vnode 2691 * in preparation for reuse. 2692 */ 2693void 2694vgone(vp) 2695 register struct vnode *vp; 2696{ 2697 struct thread *td = curthread; /* XXX */ 2698 2699 VI_LOCK(vp); 2700 vgonel(vp, td); 2701} 2702 2703/* 2704 * vgone, with the vp interlock held. 2705 */ 2706void 2707vgonel(vp, td) 2708 struct vnode *vp; 2709 struct thread *td; 2710{ 2711 /* 2712 * If a vgone (or vclean) is already in progress, 2713 * wait until it is done and return. 2714 */ 2715 ASSERT_VI_LOCKED(vp, "vgonel"); 2716 if (vp->v_iflag & VI_XLOCK) { 2717 vp->v_iflag |= VI_XWANT; 2718 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0); 2719 return; 2720 } 2721 2722 /* 2723 * Clean out the filesystem specific data. 2724 */ 2725 vclean(vp, DOCLOSE, td); 2726 VI_UNLOCK(vp); 2727 2728 /* 2729 * Delete from old mount point vnode list, if on one. 2730 */ 2731 if (vp->v_mount != NULL) 2732 insmntque(vp, (struct mount *)0); 2733 /* 2734 * If special device, remove it from special device alias list 2735 * if it is on one. 2736 */ 2737 VI_LOCK(vp); 2738 if (vp->v_type == VCHR && vp->v_rdev != NULL && vp->v_rdev != NODEV) { 2739 mtx_lock(&spechash_mtx); 2740 SLIST_REMOVE(&vp->v_rdev->si_hlist, vp, vnode, v_specnext); 2741 vp->v_rdev->si_usecount -= vp->v_usecount; 2742 mtx_unlock(&spechash_mtx); 2743 vp->v_rdev = NULL; 2744 } 2745 2746 /* 2747 * If it is on the freelist and not already at the head, 2748 * move it to the head of the list. The test of the 2749 * VDOOMED flag and the reference count of zero is because 2750 * it will be removed from the free list by getnewvnode, 2751 * but will not have its reference count incremented until 2752 * after calling vgone. If the reference count were 2753 * incremented first, vgone would (incorrectly) try to 2754 * close the previous instance of the underlying object. 2755 */ 2756 if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) { 2757 mtx_lock(&vnode_free_list_mtx); 2758 if (vp->v_iflag & VI_FREE) { 2759 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2760 } else { 2761 vp->v_iflag |= VI_FREE; 2762 freevnodes++; 2763 } 2764 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2765 mtx_unlock(&vnode_free_list_mtx); 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 mount *mp, *nmp; 2881 struct vnode *vp; 2882 2883 /* 2884 * Note: because this is DDB, we can't obey the locking semantics 2885 * for these structures, which means we could catch an inconsistent 2886 * state and dereference a nasty pointer. Not much to be done 2887 * about that. 2888 */ 2889 printf("Locked vnodes\n"); 2890 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2891 nmp = TAILQ_NEXT(mp, mnt_list); 2892 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2893 if (VOP_ISLOCKED(vp, NULL)) 2894 vprint(NULL, vp); 2895 } 2896 nmp = TAILQ_NEXT(mp, mnt_list); 2897 } 2898} 2899#endif 2900 2901/* 2902 * Fill in a struct xvfsconf based on a struct vfsconf. 2903 */ 2904static void 2905vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) 2906{ 2907 2908 strcpy(xvfsp->vfc_name, vfsp->vfc_name); 2909 xvfsp->vfc_typenum = vfsp->vfc_typenum; 2910 xvfsp->vfc_refcount = vfsp->vfc_refcount; 2911 xvfsp->vfc_flags = vfsp->vfc_flags; 2912 /* 2913 * These are unused in userland, we keep them 2914 * to not break binary compatibility. 2915 */ 2916 xvfsp->vfc_vfsops = NULL; 2917 xvfsp->vfc_next = NULL; 2918} 2919 2920static int 2921sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 2922{ 2923 struct vfsconf *vfsp; 2924 struct xvfsconf *xvfsp; 2925 int cnt, error, i; 2926 2927 cnt = 0; 2928 for (vfsp = vfsconf; vfsp != NULL; vfsp = vfsp->vfc_next) 2929 cnt++; 2930 xvfsp = malloc(sizeof(struct xvfsconf) * cnt, M_TEMP, M_WAITOK); 2931 /* 2932 * Handle the race that we will have here when struct vfsconf 2933 * will be locked down by using both cnt and checking vfc_next 2934 * against NULL to determine the end of the loop. The race will 2935 * happen because we will have to unlock before calling malloc(). 2936 * We are protected by Giant for now. 2937 */ 2938 i = 0; 2939 for (vfsp = vfsconf; vfsp != NULL && i < cnt; vfsp = vfsp->vfc_next) { 2940 vfsconf2x(vfsp, xvfsp + i); 2941 i++; 2942 } 2943 error = SYSCTL_OUT(req, xvfsp, sizeof(struct xvfsconf) * i); 2944 free(xvfsp, M_TEMP); 2945 return (error); 2946} 2947 2948SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, 2949 "S,xvfsconf", "List of all configured filesystems"); 2950 2951/* 2952 * Top level filesystem related information gathering. 2953 */ 2954static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2955 2956static int 2957vfs_sysctl(SYSCTL_HANDLER_ARGS) 2958{ 2959 int *name = (int *)arg1 - 1; /* XXX */ 2960 u_int namelen = arg2 + 1; /* XXX */ 2961 struct vfsconf *vfsp; 2962 struct xvfsconf xvfsp; 2963 2964 printf("WARNING: userland calling deprecated sysctl, " 2965 "please rebuild world\n"); 2966 2967#if 1 || defined(COMPAT_PRELITE2) 2968 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2969 if (namelen == 1) 2970 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2971#endif 2972 2973 switch (name[1]) { 2974 case VFS_MAXTYPENUM: 2975 if (namelen != 2) 2976 return (ENOTDIR); 2977 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2978 case VFS_CONF: 2979 if (namelen != 3) 2980 return (ENOTDIR); /* overloaded */ 2981 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) 2982 if (vfsp->vfc_typenum == name[2]) 2983 break; 2984 if (vfsp == NULL) 2985 return (EOPNOTSUPP); 2986 vfsconf2x(vfsp, &xvfsp); 2987 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 2988 } 2989 return (EOPNOTSUPP); 2990} 2991 2992SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl, 2993 "Generic filesystem"); 2994 2995#if 1 || defined(COMPAT_PRELITE2) 2996 2997static int 2998sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 2999{ 3000 int error; 3001 struct vfsconf *vfsp; 3002 struct ovfsconf ovfs; 3003 3004 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) { 3005 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 3006 strcpy(ovfs.vfc_name, vfsp->vfc_name); 3007 ovfs.vfc_index = vfsp->vfc_typenum; 3008 ovfs.vfc_refcount = vfsp->vfc_refcount; 3009 ovfs.vfc_flags = vfsp->vfc_flags; 3010 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 3011 if (error) 3012 return error; 3013 } 3014 return 0; 3015} 3016 3017#endif /* 1 || COMPAT_PRELITE2 */ 3018 3019#define KINFO_VNODESLOP 10 3020#ifdef notyet 3021/* 3022 * Dump vnode list (via sysctl). 3023 */ 3024/* ARGSUSED */ 3025static int 3026sysctl_vnode(SYSCTL_HANDLER_ARGS) 3027{ 3028 struct xvnode *xvn; 3029 struct thread *td = req->td; 3030 struct mount *mp; 3031 struct vnode *vp; 3032 int error, len, n; 3033 3034 /* 3035 * Stale numvnodes access is not fatal here. 3036 */ 3037 req->lock = 0; 3038 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 3039 if (!req->oldptr) 3040 /* Make an estimate */ 3041 return (SYSCTL_OUT(req, 0, len)); 3042 3043 sysctl_wire_old_buffer(req, 0); 3044 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 3045 n = 0; 3046 mtx_lock(&mountlist_mtx); 3047 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3048 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 3049 continue; 3050 mtx_lock(&mntvnode_mtx); 3051 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3052 if (n == len) 3053 break; 3054 vref(vp); 3055 xvn[n].xv_size = sizeof *xvn; 3056 xvn[n].xv_vnode = vp; 3057#define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 3058 XV_COPY(usecount); 3059 XV_COPY(writecount); 3060 XV_COPY(holdcnt); 3061 XV_COPY(id); 3062 XV_COPY(mount); 3063 XV_COPY(numoutput); 3064 XV_COPY(type); 3065#undef XV_COPY 3066 xvn[n].xv_flag = vp->v_vflag; 3067 3068 switch (vp->v_type) { 3069 case VREG: 3070 case VDIR: 3071 case VLNK: 3072 xvn[n].xv_dev = vp->v_cachedfs; 3073 xvn[n].xv_ino = vp->v_cachedid; 3074 break; 3075 case VBLK: 3076 case VCHR: 3077 if (vp->v_rdev == NULL) { 3078 vrele(vp); 3079 continue; 3080 } 3081 xvn[n].xv_dev = dev2udev(vp->v_rdev); 3082 break; 3083 case VSOCK: 3084 xvn[n].xv_socket = vp->v_socket; 3085 break; 3086 case VFIFO: 3087 xvn[n].xv_fifo = vp->v_fifoinfo; 3088 break; 3089 case VNON: 3090 case VBAD: 3091 default: 3092 /* shouldn't happen? */ 3093 vrele(vp); 3094 continue; 3095 } 3096 vrele(vp); 3097 ++n; 3098 } 3099 mtx_unlock(&mntvnode_mtx); 3100 mtx_lock(&mountlist_mtx); 3101 vfs_unbusy(mp, td); 3102 if (n == len) 3103 break; 3104 } 3105 mtx_unlock(&mountlist_mtx); 3106 3107 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 3108 free(xvn, M_TEMP); 3109 return (error); 3110} 3111 3112SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 3113 0, 0, sysctl_vnode, "S,xvnode", ""); 3114#endif 3115 3116/* 3117 * Check to see if a filesystem is mounted on a block device. 3118 */ 3119int 3120vfs_mountedon(vp) 3121 struct vnode *vp; 3122{ 3123 3124 if (vp->v_rdev->si_mountpoint != NULL) 3125 return (EBUSY); 3126 return (0); 3127} 3128 3129/* 3130 * Unmount all filesystems. The list is traversed in reverse order 3131 * of mounting to avoid dependencies. 3132 */ 3133void 3134vfs_unmountall() 3135{ 3136 struct mount *mp; 3137 struct thread *td; 3138 int error; 3139 3140 if (curthread != NULL) 3141 td = curthread; 3142 else 3143 td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */ 3144 /* 3145 * Since this only runs when rebooting, it is not interlocked. 3146 */ 3147 while(!TAILQ_EMPTY(&mountlist)) { 3148 mp = TAILQ_LAST(&mountlist, mntlist); 3149 error = dounmount(mp, MNT_FORCE, td); 3150 if (error) { 3151 TAILQ_REMOVE(&mountlist, mp, mnt_list); 3152 printf("unmount of %s failed (", 3153 mp->mnt_stat.f_mntonname); 3154 if (error == EBUSY) 3155 printf("BUSY)\n"); 3156 else 3157 printf("%d)\n", error); 3158 } else { 3159 /* The unmount has removed mp from the mountlist */ 3160 } 3161 } 3162} 3163 3164/* 3165 * perform msync on all vnodes under a mount point 3166 * the mount point must be locked. 3167 */ 3168void 3169vfs_msync(struct mount *mp, int flags) 3170{ 3171 struct vnode *vp, *nvp; 3172 struct vm_object *obj; 3173 int tries; 3174 3175 GIANT_REQUIRED; 3176 3177 tries = 5; 3178 mtx_lock(&mntvnode_mtx); 3179loop: 3180 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp != NULL; vp = nvp) { 3181 if (vp->v_mount != mp) { 3182 if (--tries > 0) 3183 goto loop; 3184 break; 3185 } 3186 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 3187 3188 VI_LOCK(vp); 3189 if (vp->v_iflag & VI_XLOCK) { /* XXX: what if MNT_WAIT? */ 3190 VI_UNLOCK(vp); 3191 continue; 3192 } 3193 3194 if ((vp->v_iflag & VI_OBJDIRTY) && 3195 (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { 3196 mtx_unlock(&mntvnode_mtx); 3197 if (!vget(vp, 3198 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 3199 curthread)) { 3200 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 3201 vput(vp); 3202 mtx_lock(&mntvnode_mtx); 3203 continue; 3204 } 3205 3206 if (VOP_GETVOBJECT(vp, &obj) == 0) { 3207 VM_OBJECT_LOCK(obj); 3208 vm_object_page_clean(obj, 0, 0, 3209 flags == MNT_WAIT ? 3210 OBJPC_SYNC : OBJPC_NOSYNC); 3211 VM_OBJECT_UNLOCK(obj); 3212 } 3213 vput(vp); 3214 } 3215 mtx_lock(&mntvnode_mtx); 3216 if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) { 3217 if (--tries > 0) 3218 goto loop; 3219 break; 3220 } 3221 } else 3222 VI_UNLOCK(vp); 3223 } 3224 mtx_unlock(&mntvnode_mtx); 3225} 3226 3227/* 3228 * Create the VM object needed for VMIO and mmap support. This 3229 * is done for all VREG files in the system. Some filesystems might 3230 * afford the additional metadata buffering capability of the 3231 * VMIO code by making the device node be VMIO mode also. 3232 * 3233 * vp must be locked when vfs_object_create is called. 3234 */ 3235int 3236vfs_object_create(vp, td, cred) 3237 struct vnode *vp; 3238 struct thread *td; 3239 struct ucred *cred; 3240{ 3241 GIANT_REQUIRED; 3242 return (VOP_CREATEVOBJECT(vp, cred, td)); 3243} 3244 3245/* 3246 * Mark a vnode as free, putting it up for recycling. 3247 */ 3248void 3249vfree(vp) 3250 struct vnode *vp; 3251{ 3252 ASSERT_VI_LOCKED(vp, "vfree"); 3253 mtx_lock(&vnode_free_list_mtx); 3254 KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free")); 3255 if (vp->v_iflag & VI_AGE) { 3256 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 3257 } else { 3258 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 3259 } 3260 freevnodes++; 3261 mtx_unlock(&vnode_free_list_mtx); 3262 vp->v_iflag &= ~VI_AGE; 3263 vp->v_iflag |= VI_FREE; 3264} 3265 3266/* 3267 * Opposite of vfree() - mark a vnode as in use. 3268 */ 3269void 3270vbusy(vp) 3271 struct vnode *vp; 3272{ 3273 ASSERT_VI_LOCKED(vp, "vbusy"); 3274 KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free")); 3275 3276 mtx_lock(&vnode_free_list_mtx); 3277 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 3278 freevnodes--; 3279 mtx_unlock(&vnode_free_list_mtx); 3280 3281 vp->v_iflag &= ~(VI_FREE|VI_AGE); 3282} 3283 3284/* 3285 * Record a process's interest in events which might happen to 3286 * a vnode. Because poll uses the historic select-style interface 3287 * internally, this routine serves as both the ``check for any 3288 * pending events'' and the ``record my interest in future events'' 3289 * functions. (These are done together, while the lock is held, 3290 * to avoid race conditions.) 3291 */ 3292int 3293vn_pollrecord(vp, td, events) 3294 struct vnode *vp; 3295 struct thread *td; 3296 short events; 3297{ 3298 3299 if (vp->v_pollinfo == NULL) 3300 v_addpollinfo(vp); 3301 mtx_lock(&vp->v_pollinfo->vpi_lock); 3302 if (vp->v_pollinfo->vpi_revents & events) { 3303 /* 3304 * This leaves events we are not interested 3305 * in available for the other process which 3306 * which presumably had requested them 3307 * (otherwise they would never have been 3308 * recorded). 3309 */ 3310 events &= vp->v_pollinfo->vpi_revents; 3311 vp->v_pollinfo->vpi_revents &= ~events; 3312 3313 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3314 return events; 3315 } 3316 vp->v_pollinfo->vpi_events |= events; 3317 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 3318 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3319 return 0; 3320} 3321 3322/* 3323 * Note the occurrence of an event. If the VN_POLLEVENT macro is used, 3324 * it is possible for us to miss an event due to race conditions, but 3325 * that condition is expected to be rare, so for the moment it is the 3326 * preferred interface. 3327 */ 3328void 3329vn_pollevent(vp, events) 3330 struct vnode *vp; 3331 short events; 3332{ 3333 3334 if (vp->v_pollinfo == NULL) 3335 v_addpollinfo(vp); 3336 mtx_lock(&vp->v_pollinfo->vpi_lock); 3337 if (vp->v_pollinfo->vpi_events & events) { 3338 /* 3339 * We clear vpi_events so that we don't 3340 * call selwakeup() twice if two events are 3341 * posted before the polling process(es) is 3342 * awakened. This also ensures that we take at 3343 * most one selwakeup() if the polling process 3344 * is no longer interested. However, it does 3345 * mean that only one event can be noticed at 3346 * a time. (Perhaps we should only clear those 3347 * event bits which we note?) XXX 3348 */ 3349 vp->v_pollinfo->vpi_events = 0; /* &= ~events ??? */ 3350 vp->v_pollinfo->vpi_revents |= events; 3351 selwakeup(&vp->v_pollinfo->vpi_selinfo); 3352 } 3353 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3354} 3355 3356/* 3357 * Wake up anyone polling on vp because it is being revoked. 3358 * This depends on dead_poll() returning POLLHUP for correct 3359 * behavior. 3360 */ 3361void 3362vn_pollgone(vp) 3363 struct vnode *vp; 3364{ 3365 3366 mtx_lock(&vp->v_pollinfo->vpi_lock); 3367 VN_KNOTE(vp, NOTE_REVOKE); 3368 if (vp->v_pollinfo->vpi_events) { 3369 vp->v_pollinfo->vpi_events = 0; 3370 selwakeup(&vp->v_pollinfo->vpi_selinfo); 3371 } 3372 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3373} 3374 3375 3376 3377/* 3378 * Routine to create and manage a filesystem syncer vnode. 3379 */ 3380#define sync_close ((int (*)(struct vop_close_args *))nullop) 3381static int sync_fsync(struct vop_fsync_args *); 3382static int sync_inactive(struct vop_inactive_args *); 3383static int sync_reclaim(struct vop_reclaim_args *); 3384 3385static vop_t **sync_vnodeop_p; 3386static struct vnodeopv_entry_desc sync_vnodeop_entries[] = { 3387 { &vop_default_desc, (vop_t *) vop_eopnotsupp }, 3388 { &vop_close_desc, (vop_t *) sync_close }, /* close */ 3389 { &vop_fsync_desc, (vop_t *) sync_fsync }, /* fsync */ 3390 { &vop_inactive_desc, (vop_t *) sync_inactive }, /* inactive */ 3391 { &vop_reclaim_desc, (vop_t *) sync_reclaim }, /* reclaim */ 3392 { &vop_lock_desc, (vop_t *) vop_stdlock }, /* lock */ 3393 { &vop_unlock_desc, (vop_t *) vop_stdunlock }, /* unlock */ 3394 { &vop_islocked_desc, (vop_t *) vop_stdislocked }, /* islocked */ 3395 { NULL, NULL } 3396}; 3397static struct vnodeopv_desc sync_vnodeop_opv_desc = 3398 { &sync_vnodeop_p, sync_vnodeop_entries }; 3399 3400VNODEOP_SET(sync_vnodeop_opv_desc); 3401 3402/* 3403 * Create a new filesystem syncer vnode for the specified mount point. 3404 */ 3405int 3406vfs_allocate_syncvnode(mp) 3407 struct mount *mp; 3408{ 3409 struct vnode *vp; 3410 static long start, incr, next; 3411 int error; 3412 3413 /* Allocate a new vnode */ 3414 if ((error = getnewvnode("syncer", mp, sync_vnodeop_p, &vp)) != 0) { 3415 mp->mnt_syncer = NULL; 3416 return (error); 3417 } 3418 vp->v_type = VNON; 3419 /* 3420 * Place the vnode onto the syncer worklist. We attempt to 3421 * scatter them about on the list so that they will go off 3422 * at evenly distributed times even if all the filesystems 3423 * are mounted at once. 3424 */ 3425 next += incr; 3426 if (next == 0 || next > syncer_maxdelay) { 3427 start /= 2; 3428 incr /= 2; 3429 if (start == 0) { 3430 start = syncer_maxdelay / 2; 3431 incr = syncer_maxdelay; 3432 } 3433 next = start; 3434 } 3435 VI_LOCK(vp); 3436 vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0); 3437 VI_UNLOCK(vp); 3438 mp->mnt_syncer = vp; 3439 return (0); 3440} 3441 3442/* 3443 * Do a lazy sync of the filesystem. 3444 */ 3445static int 3446sync_fsync(ap) 3447 struct vop_fsync_args /* { 3448 struct vnode *a_vp; 3449 struct ucred *a_cred; 3450 int a_waitfor; 3451 struct thread *a_td; 3452 } */ *ap; 3453{ 3454 struct vnode *syncvp = ap->a_vp; 3455 struct mount *mp = syncvp->v_mount; 3456 struct thread *td = ap->a_td; 3457 int error, asyncflag; 3458 3459 /* 3460 * We only need to do something if this is a lazy evaluation. 3461 */ 3462 if (ap->a_waitfor != MNT_LAZY) 3463 return (0); 3464 3465 /* 3466 * Move ourselves to the back of the sync list. 3467 */ 3468 VI_LOCK(syncvp); 3469 vn_syncer_add_to_worklist(syncvp, syncdelay); 3470 VI_UNLOCK(syncvp); 3471 3472 /* 3473 * Walk the list of vnodes pushing all that are dirty and 3474 * not already on the sync list. 3475 */ 3476 mtx_lock(&mountlist_mtx); 3477 if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { 3478 mtx_unlock(&mountlist_mtx); 3479 return (0); 3480 } 3481 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 3482 vfs_unbusy(mp, td); 3483 return (0); 3484 } 3485 asyncflag = mp->mnt_flag & MNT_ASYNC; 3486 mp->mnt_flag &= ~MNT_ASYNC; 3487 vfs_msync(mp, MNT_NOWAIT); 3488 error = VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td); 3489 if (asyncflag) 3490 mp->mnt_flag |= MNT_ASYNC; 3491 vn_finished_write(mp); 3492 vfs_unbusy(mp, td); 3493 return (error); 3494} 3495 3496/* 3497 * The syncer vnode is no referenced. 3498 */ 3499static int 3500sync_inactive(ap) 3501 struct vop_inactive_args /* { 3502 struct vnode *a_vp; 3503 struct thread *a_td; 3504 } */ *ap; 3505{ 3506 3507 VOP_UNLOCK(ap->a_vp, 0, ap->a_td); 3508 vgone(ap->a_vp); 3509 return (0); 3510} 3511 3512/* 3513 * The syncer vnode is no longer needed and is being decommissioned. 3514 * 3515 * Modifications to the worklist must be protected by sync_mtx. 3516 */ 3517static int 3518sync_reclaim(ap) 3519 struct vop_reclaim_args /* { 3520 struct vnode *a_vp; 3521 } */ *ap; 3522{ 3523 struct vnode *vp = ap->a_vp; 3524 3525 VI_LOCK(vp); 3526 vp->v_mount->mnt_syncer = NULL; 3527 if (vp->v_iflag & VI_ONWORKLST) { 3528 mtx_lock(&sync_mtx); 3529 LIST_REMOVE(vp, v_synclist); 3530 mtx_unlock(&sync_mtx); 3531 vp->v_iflag &= ~VI_ONWORKLST; 3532 } 3533 VI_UNLOCK(vp); 3534 3535 return (0); 3536} 3537 3538/* 3539 * extract the dev_t from a VCHR 3540 */ 3541dev_t 3542vn_todev(vp) 3543 struct vnode *vp; 3544{ 3545 if (vp->v_type != VCHR) 3546 return (NODEV); 3547 return (vp->v_rdev); 3548} 3549 3550/* 3551 * Check if vnode represents a disk device 3552 */ 3553int 3554vn_isdisk(vp, errp) 3555 struct vnode *vp; 3556 int *errp; 3557{ 3558 struct cdevsw *cdevsw; 3559 3560 if (vp->v_type != VCHR) { 3561 if (errp != NULL) 3562 *errp = ENOTBLK; 3563 return (0); 3564 } 3565 if (vp->v_rdev == NULL) { 3566 if (errp != NULL) 3567 *errp = ENXIO; 3568 return (0); 3569 } 3570 cdevsw = devsw(vp->v_rdev); 3571 if (cdevsw == NULL) { 3572 if (errp != NULL) 3573 *errp = ENXIO; 3574 return (0); 3575 } 3576 if (!(cdevsw->d_flags & D_DISK)) { 3577 if (errp != NULL) 3578 *errp = ENOTBLK; 3579 return (0); 3580 } 3581 if (errp != NULL) 3582 *errp = 0; 3583 return (1); 3584} 3585 3586/* 3587 * Free data allocated by namei(); see namei(9) for details. 3588 */ 3589void 3590NDFREE(ndp, flags) 3591 struct nameidata *ndp; 3592 const u_int flags; 3593{ 3594 if (!(flags & NDF_NO_FREE_PNBUF) && 3595 (ndp->ni_cnd.cn_flags & HASBUF)) { 3596 uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf); 3597 ndp->ni_cnd.cn_flags &= ~HASBUF; 3598 } 3599 if (!(flags & NDF_NO_DVP_UNLOCK) && 3600 (ndp->ni_cnd.cn_flags & LOCKPARENT) && 3601 ndp->ni_dvp != ndp->ni_vp) 3602 VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread); 3603 if (!(flags & NDF_NO_DVP_RELE) && 3604 (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) { 3605 vrele(ndp->ni_dvp); 3606 ndp->ni_dvp = NULL; 3607 } 3608 if (!(flags & NDF_NO_VP_UNLOCK) && 3609 (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp) 3610 VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread); 3611 if (!(flags & NDF_NO_VP_RELE) && 3612 ndp->ni_vp) { 3613 vrele(ndp->ni_vp); 3614 ndp->ni_vp = NULL; 3615 } 3616 if (!(flags & NDF_NO_STARTDIR_RELE) && 3617 (ndp->ni_cnd.cn_flags & SAVESTART)) { 3618 vrele(ndp->ni_startdir); 3619 ndp->ni_startdir = NULL; 3620 } 3621} 3622 3623/* 3624 * Common filesystem object access control check routine. Accepts a 3625 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3626 * and optional call-by-reference privused argument allowing vaccess() 3627 * to indicate to the caller whether privilege was used to satisfy the 3628 * request (obsoleted). Returns 0 on success, or an errno on failure. 3629 */ 3630int 3631vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused) 3632 enum vtype type; 3633 mode_t file_mode; 3634 uid_t file_uid; 3635 gid_t file_gid; 3636 mode_t acc_mode; 3637 struct ucred *cred; 3638 int *privused; 3639{ 3640 mode_t dac_granted; 3641#ifdef CAPABILITIES 3642 mode_t cap_granted; 3643#endif 3644 3645 /* 3646 * Look for a normal, non-privileged way to access the file/directory 3647 * as requested. If it exists, go with that. 3648 */ 3649 3650 if (privused != NULL) 3651 *privused = 0; 3652 3653 dac_granted = 0; 3654 3655 /* Check the owner. */ 3656 if (cred->cr_uid == file_uid) { 3657 dac_granted |= VADMIN; 3658 if (file_mode & S_IXUSR) 3659 dac_granted |= VEXEC; 3660 if (file_mode & S_IRUSR) 3661 dac_granted |= VREAD; 3662 if (file_mode & S_IWUSR) 3663 dac_granted |= (VWRITE | VAPPEND); 3664 3665 if ((acc_mode & dac_granted) == acc_mode) 3666 return (0); 3667 3668 goto privcheck; 3669 } 3670 3671 /* Otherwise, check the groups (first match) */ 3672 if (groupmember(file_gid, cred)) { 3673 if (file_mode & S_IXGRP) 3674 dac_granted |= VEXEC; 3675 if (file_mode & S_IRGRP) 3676 dac_granted |= VREAD; 3677 if (file_mode & S_IWGRP) 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 everyone else. */ 3687 if (file_mode & S_IXOTH) 3688 dac_granted |= VEXEC; 3689 if (file_mode & S_IROTH) 3690 dac_granted |= VREAD; 3691 if (file_mode & S_IWOTH) 3692 dac_granted |= (VWRITE | VAPPEND); 3693 if ((acc_mode & dac_granted) == acc_mode) 3694 return (0); 3695 3696privcheck: 3697 if (!suser_cred(cred, PRISON_ROOT)) { 3698 /* XXX audit: privilege used */ 3699 if (privused != NULL) 3700 *privused = 1; 3701 return (0); 3702 } 3703 3704#ifdef CAPABILITIES 3705 /* 3706 * Build a capability mask to determine if the set of capabilities 3707 * satisfies the requirements when combined with the granted mask 3708 * from above. 3709 * For each capability, if the capability is required, bitwise 3710 * or the request type onto the cap_granted mask. 3711 */ 3712 cap_granted = 0; 3713 3714 if (type == VDIR) { 3715 /* 3716 * For directories, use CAP_DAC_READ_SEARCH to satisfy 3717 * VEXEC requests, instead of CAP_DAC_EXECUTE. 3718 */ 3719 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3720 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) 3721 cap_granted |= VEXEC; 3722 } else { 3723 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3724 !cap_check(cred, NULL, CAP_DAC_EXECUTE, PRISON_ROOT)) 3725 cap_granted |= VEXEC; 3726 } 3727 3728 if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && 3729 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) 3730 cap_granted |= VREAD; 3731 3732 if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3733 !cap_check(cred, NULL, CAP_DAC_WRITE, PRISON_ROOT)) 3734 cap_granted |= (VWRITE | VAPPEND); 3735 3736 if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3737 !cap_check(cred, NULL, CAP_FOWNER, PRISON_ROOT)) 3738 cap_granted |= VADMIN; 3739 3740 if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { 3741 /* XXX audit: privilege used */ 3742 if (privused != NULL) 3743 *privused = 1; 3744 return (0); 3745 } 3746#endif 3747 3748 return ((acc_mode & VADMIN) ? EPERM : EACCES); 3749} 3750 3751/* 3752 * Credential check based on process requesting service, and per-attribute 3753 * permissions. 3754 */ 3755int 3756extattr_check_cred(struct vnode *vp, int attrnamespace, 3757 struct ucred *cred, struct thread *td, int access) 3758{ 3759 3760 /* 3761 * Kernel-invoked always succeeds. 3762 */ 3763 if (cred == NOCRED) 3764 return (0); 3765 3766 /* 3767 * Do not allow privileged processes in jail to directly 3768 * manipulate system attributes. 3769 * 3770 * XXX What capability should apply here? 3771 * Probably CAP_SYS_SETFFLAG. 3772 */ 3773 switch (attrnamespace) { 3774 case EXTATTR_NAMESPACE_SYSTEM: 3775 /* Potentially should be: return (EPERM); */ 3776 return (suser_cred(cred, 0)); 3777 case EXTATTR_NAMESPACE_USER: 3778 return (VOP_ACCESS(vp, access, cred, td)); 3779 default: 3780 return (EPERM); 3781 } 3782} 3783