vfs_subr.c revision 96034
1/* 2 * Copyright (c) 1989, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 39 * $FreeBSD: head/sys/kern/vfs_subr.c 96034 2002-05-04 17:10:15Z dillon $ 40 */ 41 42/* 43 * External virtual filesystem routines 44 */ 45#include "opt_ddb.h" 46#include "opt_ffs.h" 47 48#include <sys/param.h> 49#include <sys/systm.h> 50#include <sys/bio.h> 51#include <sys/buf.h> 52#include <sys/conf.h> 53#include <sys/eventhandler.h> 54#include <sys/fcntl.h> 55#include <sys/kernel.h> 56#include <sys/kthread.h> 57#include <sys/malloc.h> 58#include <sys/mount.h> 59#include <sys/namei.h> 60#include <sys/stat.h> 61#include <sys/sysctl.h> 62#include <sys/syslog.h> 63#include <sys/vmmeter.h> 64#include <sys/vnode.h> 65 66#include <vm/vm.h> 67#include <vm/vm_object.h> 68#include <vm/vm_extern.h> 69#include <vm/pmap.h> 70#include <vm/vm_map.h> 71#include <vm/vm_page.h> 72#include <vm/uma.h> 73 74static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure"); 75 76static void addalias(struct vnode *vp, dev_t nvp_rdev); 77static void insmntque(struct vnode *vp, struct mount *mp); 78static void vclean(struct vnode *vp, int flags, struct thread *td); 79static void vlruvp(struct vnode *vp); 80 81/* 82 * Number of vnodes in existence. Increased whenever getnewvnode() 83 * allocates a new vnode, never decreased. 84 */ 85static unsigned long numvnodes; 86 87SYSCTL_LONG(_debug, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); 88 89/* 90 * Conversion tables for conversion from vnode types to inode formats 91 * and back. 92 */ 93enum vtype iftovt_tab[16] = { 94 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 95 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 96}; 97int vttoif_tab[9] = { 98 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 99 S_IFSOCK, S_IFIFO, S_IFMT, 100}; 101 102/* 103 * List of vnodes that are ready for recycling. 104 */ 105static TAILQ_HEAD(freelst, vnode) vnode_free_list; 106 107/* 108 * Minimum number of free vnodes. If there are fewer than this free vnodes, 109 * getnewvnode() will return a newly allocated vnode. 110 */ 111static u_long wantfreevnodes = 25; 112SYSCTL_LONG(_debug, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 113/* Number of vnodes in the free list. */ 114static u_long freevnodes; 115SYSCTL_LONG(_debug, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); 116 117/* 118 * Various variables used for debugging the new implementation of 119 * reassignbuf(). 120 * XXX these are probably of (very) limited utility now. 121 */ 122static int reassignbufcalls; 123SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); 124static int reassignbufloops; 125SYSCTL_INT(_vfs, OID_AUTO, reassignbufloops, CTLFLAG_RW, &reassignbufloops, 0, ""); 126static int reassignbufsortgood; 127SYSCTL_INT(_vfs, OID_AUTO, reassignbufsortgood, CTLFLAG_RW, &reassignbufsortgood, 0, ""); 128static int reassignbufsortbad; 129SYSCTL_INT(_vfs, OID_AUTO, reassignbufsortbad, CTLFLAG_RW, &reassignbufsortbad, 0, ""); 130/* Set to 0 for old insertion-sort based reassignbuf, 1 for modern method. */ 131static int reassignbufmethod = 1; 132SYSCTL_INT(_vfs, OID_AUTO, reassignbufmethod, CTLFLAG_RW, &reassignbufmethod, 0, ""); 133static int nameileafonly; 134SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, ""); 135 136#ifdef ENABLE_VFS_IOOPT 137/* See NOTES for a description of this setting. */ 138int vfs_ioopt; 139SYSCTL_INT(_vfs, OID_AUTO, ioopt, CTLFLAG_RW, &vfs_ioopt, 0, ""); 140#endif 141 142/* List of mounted filesystems. */ 143struct mntlist mountlist = TAILQ_HEAD_INITIALIZER(mountlist); 144 145/* For any iteration/modification of mountlist */ 146struct mtx mountlist_mtx; 147 148/* For any iteration/modification of mnt_vnodelist */ 149struct mtx mntvnode_mtx; 150 151/* 152 * Cache for the mount type id assigned to NFS. This is used for 153 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 154 */ 155int nfs_mount_type = -1; 156 157/* To keep more than one thread at a time from running vfs_getnewfsid */ 158static struct mtx mntid_mtx; 159 160/* For any iteration/modification of vnode_free_list */ 161static struct mtx vnode_free_list_mtx; 162 163/* 164 * For any iteration/modification of dev->si_hlist (linked through 165 * v_specnext) 166 */ 167static struct mtx spechash_mtx; 168 169/* Publicly exported FS */ 170struct nfs_public nfs_pub; 171 172/* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 173static uma_zone_t vnode_zone; 174static uma_zone_t vnodepoll_zone; 175 176/* Set to 1 to print out reclaim of active vnodes */ 177int prtactive; 178 179/* 180 * The workitem queue. 181 * 182 * It is useful to delay writes of file data and filesystem metadata 183 * for tens of seconds so that quickly created and deleted files need 184 * not waste disk bandwidth being created and removed. To realize this, 185 * we append vnodes to a "workitem" queue. When running with a soft 186 * updates implementation, most pending metadata dependencies should 187 * not wait for more than a few seconds. Thus, mounted on block devices 188 * are delayed only about a half the time that file data is delayed. 189 * Similarly, directory updates are more critical, so are only delayed 190 * about a third the time that file data is delayed. Thus, there are 191 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 192 * one each second (driven off the filesystem syncer process). The 193 * syncer_delayno variable indicates the next queue that is to be processed. 194 * Items that need to be processed soon are placed in this queue: 195 * 196 * syncer_workitem_pending[syncer_delayno] 197 * 198 * A delay of fifteen seconds is done by placing the request fifteen 199 * entries later in the queue: 200 * 201 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 202 * 203 */ 204static int syncer_delayno; 205static long syncer_mask; 206LIST_HEAD(synclist, vnode); 207static struct synclist *syncer_workitem_pending; 208 209#define SYNCER_MAXDELAY 32 210static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 211static int syncdelay = 30; /* max time to delay syncing data */ 212static int filedelay = 30; /* time to delay syncing files */ 213SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); 214static int dirdelay = 29; /* time to delay syncing directories */ 215SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); 216static int metadelay = 28; /* time to delay syncing metadata */ 217SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); 218static int rushjob; /* number of slots to run ASAP */ 219static int stat_rush_requests; /* number of times I/O speeded up */ 220SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); 221 222/* 223 * Number of vnodes we want to exist at any one time. This is mostly used 224 * to size hash tables in vnode-related code. It is normally not used in 225 * getnewvnode(), as wantfreevnodes is normally nonzero.) 226 * 227 * XXX desiredvnodes is historical cruft and should not exist. 228 */ 229int desiredvnodes; 230SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 231 &desiredvnodes, 0, "Maximum number of vnodes"); 232static int minvnodes; 233SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 234 &minvnodes, 0, "Minimum number of vnodes"); 235static int vnlru_nowhere; 236SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, &vnlru_nowhere, 0, 237 "Number of times the vnlru process ran without success"); 238 239void 240v_addpollinfo(struct vnode *vp) 241{ 242 vp->v_pollinfo = uma_zalloc(vnodepoll_zone, M_WAITOK); 243 mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 244} 245 246/* 247 * Initialize the vnode management data structures. 248 */ 249static void 250vntblinit(void *dummy __unused) 251{ 252 253 desiredvnodes = maxproc + cnt.v_page_count / 4; 254 minvnodes = desiredvnodes / 4; 255 mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF); 256 mtx_init(&mntvnode_mtx, "mntvnode", NULL, MTX_DEF); 257 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 258 mtx_init(&spechash_mtx, "spechash", NULL, MTX_DEF); 259 TAILQ_INIT(&vnode_free_list); 260 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 261 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 262 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 263 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 264 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 265 /* 266 * Initialize the filesystem syncer. 267 */ 268 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 269 &syncer_mask); 270 syncer_maxdelay = syncer_mask + 1; 271} 272SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) 273 274 275/* 276 * Mark a mount point as busy. Used to synchronize access and to delay 277 * unmounting. Interlock is not released on failure. 278 */ 279int 280vfs_busy(mp, flags, interlkp, td) 281 struct mount *mp; 282 int flags; 283 struct mtx *interlkp; 284 struct thread *td; 285{ 286 int lkflags; 287 288 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 289 if (flags & LK_NOWAIT) 290 return (ENOENT); 291 mp->mnt_kern_flag |= MNTK_MWAIT; 292 /* 293 * Since all busy locks are shared except the exclusive 294 * lock granted when unmounting, the only place that a 295 * wakeup needs to be done is at the release of the 296 * exclusive lock at the end of dounmount. 297 */ 298 msleep((caddr_t)mp, interlkp, PVFS, "vfs_busy", 0); 299 return (ENOENT); 300 } 301 lkflags = LK_SHARED | LK_NOPAUSE; 302 if (interlkp) 303 lkflags |= LK_INTERLOCK; 304 if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td)) 305 panic("vfs_busy: unexpected lock failure"); 306 return (0); 307} 308 309/* 310 * Free a busy filesystem. 311 */ 312void 313vfs_unbusy(mp, td) 314 struct mount *mp; 315 struct thread *td; 316{ 317 318 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); 319} 320 321/* 322 * Lookup a filesystem type, and if found allocate and initialize 323 * a mount structure for it. 324 * 325 * Devname is usually updated by mount(8) after booting. 326 */ 327int 328vfs_rootmountalloc(fstypename, devname, mpp) 329 char *fstypename; 330 char *devname; 331 struct mount **mpp; 332{ 333 struct thread *td = curthread; /* XXX */ 334 struct vfsconf *vfsp; 335 struct mount *mp; 336 337 if (fstypename == NULL) 338 return (ENODEV); 339 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) 340 if (!strcmp(vfsp->vfc_name, fstypename)) 341 break; 342 if (vfsp == NULL) 343 return (ENODEV); 344 mp = malloc((u_long)sizeof(struct mount), M_MOUNT, M_WAITOK | M_ZERO); 345 lockinit(&mp->mnt_lock, PVFS, "vfslock", 0, LK_NOPAUSE); 346 (void)vfs_busy(mp, LK_NOWAIT, 0, td); 347 TAILQ_INIT(&mp->mnt_nvnodelist); 348 TAILQ_INIT(&mp->mnt_reservedvnlist); 349 mp->mnt_vfc = vfsp; 350 mp->mnt_op = vfsp->vfc_vfsops; 351 mp->mnt_flag = MNT_RDONLY; 352 mp->mnt_vnodecovered = NULLVP; 353 vfsp->vfc_refcount++; 354 mp->mnt_iosize_max = DFLTPHYS; 355 mp->mnt_stat.f_type = vfsp->vfc_typenum; 356 mp->mnt_flag |= vfsp->vfc_flags & MNT_VISFLAGMASK; 357 strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN); 358 mp->mnt_stat.f_mntonname[0] = '/'; 359 mp->mnt_stat.f_mntonname[1] = 0; 360 (void) copystr(devname, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, 0); 361 *mpp = mp; 362 return (0); 363} 364 365/* 366 * Find an appropriate filesystem to use for the root. If a filesystem 367 * has not been preselected, walk through the list of known filesystems 368 * trying those that have mountroot routines, and try them until one 369 * works or we have tried them all. 370 */ 371#ifdef notdef /* XXX JH */ 372int 373lite2_vfs_mountroot() 374{ 375 struct vfsconf *vfsp; 376 extern int (*lite2_mountroot)(void); 377 int error; 378 379 if (lite2_mountroot != NULL) 380 return ((*lite2_mountroot)()); 381 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) { 382 if (vfsp->vfc_mountroot == NULL) 383 continue; 384 if ((error = (*vfsp->vfc_mountroot)()) == 0) 385 return (0); 386 printf("%s_mountroot failed: %d\n", vfsp->vfc_name, error); 387 } 388 return (ENODEV); 389} 390#endif 391 392/* 393 * Lookup a mount point by filesystem identifier. 394 */ 395struct mount * 396vfs_getvfs(fsid) 397 fsid_t *fsid; 398{ 399 register struct mount *mp; 400 401 mtx_lock(&mountlist_mtx); 402 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 403 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 404 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 405 mtx_unlock(&mountlist_mtx); 406 return (mp); 407 } 408 } 409 mtx_unlock(&mountlist_mtx); 410 return ((struct mount *) 0); 411} 412 413/* 414 * Get a new unique fsid. Try to make its val[0] unique, since this value 415 * will be used to create fake device numbers for stat(). Also try (but 416 * not so hard) make its val[0] unique mod 2^16, since some emulators only 417 * support 16-bit device numbers. We end up with unique val[0]'s for the 418 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 419 * 420 * Keep in mind that several mounts may be running in parallel. Starting 421 * the search one past where the previous search terminated is both a 422 * micro-optimization and a defense against returning the same fsid to 423 * different mounts. 424 */ 425void 426vfs_getnewfsid(mp) 427 struct mount *mp; 428{ 429 static u_int16_t mntid_base; 430 fsid_t tfsid; 431 int mtype; 432 433 mtx_lock(&mntid_mtx); 434 mtype = mp->mnt_vfc->vfc_typenum; 435 tfsid.val[1] = mtype; 436 mtype = (mtype & 0xFF) << 24; 437 for (;;) { 438 tfsid.val[0] = makeudev(255, 439 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 440 mntid_base++; 441 if (vfs_getvfs(&tfsid) == NULL) 442 break; 443 } 444 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 445 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 446 mtx_unlock(&mntid_mtx); 447} 448 449/* 450 * Knob to control the precision of file timestamps: 451 * 452 * 0 = seconds only; nanoseconds zeroed. 453 * 1 = seconds and nanoseconds, accurate within 1/HZ. 454 * 2 = seconds and nanoseconds, truncated to microseconds. 455 * >=3 = seconds and nanoseconds, maximum precision. 456 */ 457enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 458 459static int timestamp_precision = TSP_SEC; 460SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 461 ×tamp_precision, 0, ""); 462 463/* 464 * Get a current timestamp. 465 */ 466void 467vfs_timestamp(tsp) 468 struct timespec *tsp; 469{ 470 struct timeval tv; 471 472 switch (timestamp_precision) { 473 case TSP_SEC: 474 tsp->tv_sec = time_second; 475 tsp->tv_nsec = 0; 476 break; 477 case TSP_HZ: 478 getnanotime(tsp); 479 break; 480 case TSP_USEC: 481 microtime(&tv); 482 TIMEVAL_TO_TIMESPEC(&tv, tsp); 483 break; 484 case TSP_NSEC: 485 default: 486 nanotime(tsp); 487 break; 488 } 489} 490 491/* 492 * Get a mount option by its name. 493 * 494 * Return 0 if the option was found. 495 * Return ENOENT if the option wasn't found. 496 * If len is a non-NULL pointer and *len 497 * a integer different from 0, then the size 498 * of the option will be compared with *len and 499 * if they doesn't match, EINVAL is returned. 500 * If len is non-NULL and *len == 0, it will 501 * be filled with the length of the option. 502 * Finally, if buf is non-NULL, it will be 503 * filled with the address of the option. 504 */ 505int 506vfs_getopt(opts, name, buf, len) 507 struct vfsoptlist *opts; 508 const char *name; 509 void **buf; 510 int *len; 511{ 512 struct vfsopt *opt; 513 int i; 514 515 i = 0; 516 opt = opts->opt; 517 while (i++ < opts->optcnt) { 518 if (strcmp(name, opt->name) == 0) { 519 if (len != NULL) { 520 if ((*len != 0) && (*len != opt->len)) 521 return (EINVAL); 522 *len = opt->len; 523 } 524 if (buf != NULL) 525 *buf = opt->value; 526 return (0); 527 } 528 opt++; 529 } 530 return (ENOENT); 531} 532 533/* 534 * Find and copy a mount option. 535 * The size of the buffer has to be specified 536 * in len, if it is not big enough, EINVAL is 537 * returned. Returns ENOENT if the option is 538 * not found. Otherwise, the number of bytes 539 * actually copied are put in done if it's 540 * non-NULL and 0 is returned. 541 */ 542int 543vfs_copyopt(opts, name, dest, len, done) 544 struct vfsoptlist *opts; 545 const char *name; 546 void *dest; 547 int len, *done; 548{ 549 struct vfsopt *opt; 550 int i; 551 552 i = 0; 553 opt = opts->opt; 554 while (i++ < opts->optcnt) { 555 if (strcmp(name, opt->name) == 0) { 556 if (len < opt->len) 557 return (EINVAL); 558 bcopy(dest, opt->value, opt->len); 559 if (done != NULL) 560 *done = opt->len; 561 return (0); 562 } 563 opt++; 564 } 565 return (ENOENT); 566} 567 568/* 569 * Set vnode attributes to VNOVAL 570 */ 571void 572vattr_null(vap) 573 register struct vattr *vap; 574{ 575 576 vap->va_type = VNON; 577 vap->va_size = VNOVAL; 578 vap->va_bytes = VNOVAL; 579 vap->va_mode = VNOVAL; 580 vap->va_nlink = VNOVAL; 581 vap->va_uid = VNOVAL; 582 vap->va_gid = VNOVAL; 583 vap->va_fsid = VNOVAL; 584 vap->va_fileid = VNOVAL; 585 vap->va_blocksize = VNOVAL; 586 vap->va_rdev = VNOVAL; 587 vap->va_atime.tv_sec = VNOVAL; 588 vap->va_atime.tv_nsec = VNOVAL; 589 vap->va_mtime.tv_sec = VNOVAL; 590 vap->va_mtime.tv_nsec = VNOVAL; 591 vap->va_ctime.tv_sec = VNOVAL; 592 vap->va_ctime.tv_nsec = VNOVAL; 593 vap->va_flags = VNOVAL; 594 vap->va_gen = VNOVAL; 595 vap->va_vaflags = 0; 596} 597 598/* 599 * This routine is called when we have too many vnodes. It attempts 600 * to free <count> vnodes and will potentially free vnodes that still 601 * have VM backing store (VM backing store is typically the cause 602 * of a vnode blowout so we want to do this). Therefore, this operation 603 * is not considered cheap. 604 * 605 * A number of conditions may prevent a vnode from being reclaimed. 606 * the buffer cache may have references on the vnode, a directory 607 * vnode may still have references due to the namei cache representing 608 * underlying files, or the vnode may be in active use. It is not 609 * desireable to reuse such vnodes. These conditions may cause the 610 * number of vnodes to reach some minimum value regardless of what 611 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 612 */ 613static int 614vlrureclaim(struct mount *mp, int count) 615{ 616 struct vnode *vp; 617 int done; 618 int trigger; 619 int usevnodes; 620 621 /* 622 * Calculate the trigger point, don't allow user 623 * screwups to blow us up. This prevents us from 624 * recycling vnodes with lots of resident pages. We 625 * aren't trying to free memory, we are trying to 626 * free vnodes. 627 */ 628 usevnodes = desiredvnodes; 629 if (usevnodes <= 0) 630 usevnodes = 1; 631 trigger = cnt.v_page_count * 2 / usevnodes; 632 633 done = 0; 634 mtx_lock(&mntvnode_mtx); 635 while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) { 636 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 637 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 638 639 if (vp->v_type != VNON && 640 vp->v_type != VBAD && 641 VMIGHTFREE(vp) && /* critical path opt */ 642 (vp->v_object == NULL || vp->v_object->resident_page_count < trigger) && 643 mtx_trylock(&vp->v_interlock) 644 ) { 645 mtx_unlock(&mntvnode_mtx); 646 if (VMIGHTFREE(vp)) { 647 vgonel(vp, curthread); 648 done++; 649 } else { 650 mtx_unlock(&vp->v_interlock); 651 } 652 mtx_lock(&mntvnode_mtx); 653 } 654 --count; 655 } 656 mtx_unlock(&mntvnode_mtx); 657 return done; 658} 659 660/* 661 * Attempt to recycle vnodes in a context that is always safe to block. 662 * Calling vlrurecycle() from the bowels of file system code has some 663 * interesting deadlock problems. 664 */ 665static struct proc *vnlruproc; 666static int vnlruproc_sig; 667 668static void 669vnlru_proc(void) 670{ 671 struct mount *mp, *nmp; 672 int s; 673 int done; 674 struct proc *p = vnlruproc; 675 struct thread *td = FIRST_THREAD_IN_PROC(p); /* XXXKSE */ 676 677 mtx_lock(&Giant); 678 679 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 680 SHUTDOWN_PRI_FIRST); 681 682 s = splbio(); 683 for (;;) { 684 kthread_suspend_check(p); 685 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { 686 vnlruproc_sig = 0; 687 tsleep(vnlruproc, PVFS, "vlruwt", 0); 688 continue; 689 } 690 done = 0; 691 mtx_lock(&mountlist_mtx); 692 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 693 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 694 nmp = TAILQ_NEXT(mp, mnt_list); 695 continue; 696 } 697 done += vlrureclaim(mp, 10); 698 mtx_lock(&mountlist_mtx); 699 nmp = TAILQ_NEXT(mp, mnt_list); 700 vfs_unbusy(mp, td); 701 } 702 mtx_unlock(&mountlist_mtx); 703 if (done == 0) { 704#if 0 705 /* These messages are temporary debugging aids */ 706 if (vnlru_nowhere < 5) 707 printf("vnlru process getting nowhere..\n"); 708 else if (vnlru_nowhere == 5) 709 printf("vnlru process messages stopped.\n"); 710#endif 711 vnlru_nowhere++; 712 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 713 } 714 } 715 splx(s); 716} 717 718static struct kproc_desc vnlru_kp = { 719 "vnlru", 720 vnlru_proc, 721 &vnlruproc 722}; 723SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 724 725 726/* 727 * Routines having to do with the management of the vnode table. 728 */ 729 730/* 731 * Return the next vnode from the free list. 732 */ 733int 734getnewvnode(tag, mp, vops, vpp) 735 enum vtagtype tag; 736 struct mount *mp; 737 vop_t **vops; 738 struct vnode **vpp; 739{ 740 int s; 741 struct thread *td = curthread; /* XXX */ 742 struct vnode *vp = NULL; 743 struct mount *vnmp; 744 vm_object_t object; 745 746 s = splbio(); 747 /* 748 * Try to reuse vnodes if we hit the max. This situation only 749 * occurs in certain large-memory (2G+) situations. We cannot 750 * attempt to directly reclaim vnodes due to nasty recursion 751 * problems. 752 */ 753 if (vnlruproc_sig == 0 && numvnodes - freevnodes > desiredvnodes) { 754 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 755 wakeup(vnlruproc); 756 } 757 758 /* 759 * Attempt to reuse a vnode already on the free list, allocating 760 * a new vnode if we can't find one or if we have not reached a 761 * good minimum for good LRU performance. 762 */ 763 764 mtx_lock(&vnode_free_list_mtx); 765 766 if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) { 767 int count; 768 769 for (count = 0; count < freevnodes; count++) { 770 vp = TAILQ_FIRST(&vnode_free_list); 771 if (vp == NULL || vp->v_usecount) 772 panic("getnewvnode: free vnode isn't"); 773 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 774 775 /* 776 * Don't recycle if we still have cached pages or if 777 * we cannot get the interlock. 778 */ 779 if ((VOP_GETVOBJECT(vp, &object) == 0 && 780 (object->resident_page_count || 781 object->ref_count)) || 782 !mtx_trylock(&vp->v_interlock)) { 783 TAILQ_INSERT_TAIL(&vnode_free_list, vp, 784 v_freelist); 785 vp = NULL; 786 continue; 787 } 788 if (LIST_FIRST(&vp->v_cache_src)) { 789 /* 790 * note: nameileafonly sysctl is temporary, 791 * for debugging only, and will eventually be 792 * removed. 793 */ 794 if (nameileafonly > 0) { 795 /* 796 * Do not reuse namei-cached directory 797 * vnodes that have cached 798 * subdirectories. 799 */ 800 if (cache_leaf_test(vp) < 0) { 801 mtx_unlock(&vp->v_interlock); 802 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 803 vp = NULL; 804 continue; 805 } 806 } else if (nameileafonly < 0 || 807 vmiodirenable == 0) { 808 /* 809 * Do not reuse namei-cached directory 810 * vnodes if nameileafonly is -1 or 811 * if VMIO backing for directories is 812 * turned off (otherwise we reuse them 813 * too quickly). 814 */ 815 mtx_unlock(&vp->v_interlock); 816 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 817 vp = NULL; 818 continue; 819 } 820 } 821 /* 822 * Skip over it if its filesystem is being suspended. 823 */ 824 if (vn_start_write(vp, &vnmp, V_NOWAIT) == 0) 825 break; 826 mtx_unlock(&vp->v_interlock); 827 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 828 vp = NULL; 829 } 830 } 831 if (vp) { 832 vp->v_flag |= VDOOMED; 833 vp->v_flag &= ~VFREE; 834 freevnodes--; 835 mtx_unlock(&vnode_free_list_mtx); 836 cache_purge(vp); 837 if (vp->v_type != VBAD) { 838 vgonel(vp, td); 839 } else { 840 mtx_unlock(&vp->v_interlock); 841 } 842 vn_finished_write(vnmp); 843 844#ifdef INVARIANTS 845 { 846 int s; 847 848 if (vp->v_data) 849 panic("cleaned vnode isn't"); 850 s = splbio(); 851 if (vp->v_numoutput) 852 panic("Clean vnode has pending I/O's"); 853 splx(s); 854 if (vp->v_writecount != 0) 855 panic("Non-zero write count"); 856 } 857#endif 858 if (vp->v_pollinfo) { 859 mtx_destroy(&vp->v_pollinfo->vpi_lock); 860 uma_zfree(vnodepoll_zone, vp->v_pollinfo); 861 } 862 vp->v_pollinfo = NULL; 863 vp->v_flag = 0; 864 vp->v_lastw = 0; 865 vp->v_lasta = 0; 866 vp->v_cstart = 0; 867 vp->v_clen = 0; 868 vp->v_socket = 0; 869 } else { 870 mtx_unlock(&vnode_free_list_mtx); 871 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK); 872 bzero((char *) vp, sizeof *vp); 873 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 874 vp->v_dd = vp; 875 cache_purge(vp); 876 LIST_INIT(&vp->v_cache_src); 877 TAILQ_INIT(&vp->v_cache_dst); 878 numvnodes++; 879 } 880 881 TAILQ_INIT(&vp->v_cleanblkhd); 882 TAILQ_INIT(&vp->v_dirtyblkhd); 883 vp->v_type = VNON; 884 vp->v_tag = tag; 885 vp->v_op = vops; 886 lockinit(&vp->v_lock, PVFS, "vnlock", VLKTIMEOUT, LK_NOPAUSE); 887 insmntque(vp, mp); 888 *vpp = vp; 889 vp->v_usecount = 1; 890 vp->v_data = 0; 891 892 splx(s); 893 894 vfs_object_create(vp, td, td->td_ucred); 895 896#if 0 897 vnodeallocs++; 898 if (vnodeallocs % vnoderecycleperiod == 0 && 899 freevnodes < vnoderecycleminfreevn && 900 vnoderecyclemintotalvn < numvnodes) { 901 /* Recycle vnodes. */ 902 cache_purgeleafdirs(vnoderecyclenumber); 903 } 904#endif 905 906 return (0); 907} 908 909/* 910 * Move a vnode from one mount queue to another. 911 */ 912static void 913insmntque(vp, mp) 914 register struct vnode *vp; 915 register struct mount *mp; 916{ 917 918 mtx_lock(&mntvnode_mtx); 919 /* 920 * Delete from old mount point vnode list, if on one. 921 */ 922 if (vp->v_mount != NULL) 923 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes); 924 /* 925 * Insert into list of vnodes for the new mount point, if available. 926 */ 927 if ((vp->v_mount = mp) == NULL) { 928 mtx_unlock(&mntvnode_mtx); 929 return; 930 } 931 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 932 mtx_unlock(&mntvnode_mtx); 933} 934 935/* 936 * Update outstanding I/O count and do wakeup if requested. 937 */ 938void 939vwakeup(bp) 940 register struct buf *bp; 941{ 942 register struct vnode *vp; 943 944 bp->b_flags &= ~B_WRITEINPROG; 945 if ((vp = bp->b_vp)) { 946 vp->v_numoutput--; 947 if (vp->v_numoutput < 0) 948 panic("vwakeup: neg numoutput"); 949 if ((vp->v_numoutput == 0) && (vp->v_flag & VBWAIT)) { 950 vp->v_flag &= ~VBWAIT; 951 wakeup((caddr_t) &vp->v_numoutput); 952 } 953 } 954} 955 956/* 957 * Flush out and invalidate all buffers associated with a vnode. 958 * Called with the underlying object locked. 959 */ 960int 961vinvalbuf(vp, flags, cred, td, slpflag, slptimeo) 962 register struct vnode *vp; 963 int flags; 964 struct ucred *cred; 965 struct thread *td; 966 int slpflag, slptimeo; 967{ 968 register struct buf *bp; 969 struct buf *nbp, *blist; 970 int s, error; 971 vm_object_t object; 972 973 GIANT_REQUIRED; 974 975 if (flags & V_SAVE) { 976 s = splbio(); 977 while (vp->v_numoutput) { 978 vp->v_flag |= VBWAIT; 979 error = tsleep((caddr_t)&vp->v_numoutput, 980 slpflag | (PRIBIO + 1), "vinvlbuf", slptimeo); 981 if (error) { 982 splx(s); 983 return (error); 984 } 985 } 986 if (!TAILQ_EMPTY(&vp->v_dirtyblkhd)) { 987 splx(s); 988 if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0) 989 return (error); 990 s = splbio(); 991 if (vp->v_numoutput > 0 || 992 !TAILQ_EMPTY(&vp->v_dirtyblkhd)) 993 panic("vinvalbuf: dirty bufs"); 994 } 995 splx(s); 996 } 997 s = splbio(); 998 for (;;) { 999 blist = TAILQ_FIRST(&vp->v_cleanblkhd); 1000 if (!blist) 1001 blist = TAILQ_FIRST(&vp->v_dirtyblkhd); 1002 if (!blist) 1003 break; 1004 1005 for (bp = blist; bp; bp = nbp) { 1006 nbp = TAILQ_NEXT(bp, b_vnbufs); 1007 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 1008 error = BUF_TIMELOCK(bp, 1009 LK_EXCLUSIVE | LK_SLEEPFAIL, 1010 "vinvalbuf", slpflag, slptimeo); 1011 if (error == ENOLCK) 1012 break; 1013 splx(s); 1014 return (error); 1015 } 1016 /* 1017 * XXX Since there are no node locks for NFS, I 1018 * believe there is a slight chance that a delayed 1019 * write will occur while sleeping just above, so 1020 * check for it. Note that vfs_bio_awrite expects 1021 * buffers to reside on a queue, while BUF_WRITE and 1022 * brelse do not. 1023 */ 1024 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1025 (flags & V_SAVE)) { 1026 1027 if (bp->b_vp == vp) { 1028 if (bp->b_flags & B_CLUSTEROK) { 1029 BUF_UNLOCK(bp); 1030 vfs_bio_awrite(bp); 1031 } else { 1032 bremfree(bp); 1033 bp->b_flags |= B_ASYNC; 1034 BUF_WRITE(bp); 1035 } 1036 } else { 1037 bremfree(bp); 1038 (void) BUF_WRITE(bp); 1039 } 1040 break; 1041 } 1042 bremfree(bp); 1043 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 1044 bp->b_flags &= ~B_ASYNC; 1045 brelse(bp); 1046 } 1047 } 1048 1049 /* 1050 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1051 * have write I/O in-progress but if there is a VM object then the 1052 * VM object can also have read-I/O in-progress. 1053 */ 1054 do { 1055 while (vp->v_numoutput > 0) { 1056 vp->v_flag |= VBWAIT; 1057 tsleep(&vp->v_numoutput, PVM, "vnvlbv", 0); 1058 } 1059 if (VOP_GETVOBJECT(vp, &object) == 0) { 1060 while (object->paging_in_progress) 1061 vm_object_pip_sleep(object, "vnvlbx"); 1062 } 1063 } while (vp->v_numoutput > 0); 1064 1065 splx(s); 1066 1067 /* 1068 * Destroy the copy in the VM cache, too. 1069 */ 1070 mtx_lock(&vp->v_interlock); 1071 if (VOP_GETVOBJECT(vp, &object) == 0) { 1072 vm_object_page_remove(object, 0, 0, 1073 (flags & V_SAVE) ? TRUE : FALSE); 1074 } 1075 mtx_unlock(&vp->v_interlock); 1076 1077 if (!TAILQ_EMPTY(&vp->v_dirtyblkhd) || !TAILQ_EMPTY(&vp->v_cleanblkhd)) 1078 panic("vinvalbuf: flush failed"); 1079 return (0); 1080} 1081 1082/* 1083 * Truncate a file's buffer and pages to a specified length. This 1084 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1085 * sync activity. 1086 */ 1087int 1088vtruncbuf(vp, cred, td, length, blksize) 1089 register struct vnode *vp; 1090 struct ucred *cred; 1091 struct thread *td; 1092 off_t length; 1093 int blksize; 1094{ 1095 register struct buf *bp; 1096 struct buf *nbp; 1097 int s, anyfreed; 1098 int trunclbn; 1099 1100 /* 1101 * Round up to the *next* lbn. 1102 */ 1103 trunclbn = (length + blksize - 1) / blksize; 1104 1105 s = splbio(); 1106restart: 1107 anyfreed = 1; 1108 for (;anyfreed;) { 1109 anyfreed = 0; 1110 for (bp = TAILQ_FIRST(&vp->v_cleanblkhd); bp; bp = nbp) { 1111 nbp = TAILQ_NEXT(bp, b_vnbufs); 1112 if (bp->b_lblkno >= trunclbn) { 1113 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 1114 BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL); 1115 goto restart; 1116 } else { 1117 bremfree(bp); 1118 bp->b_flags |= (B_INVAL | B_RELBUF); 1119 bp->b_flags &= ~B_ASYNC; 1120 brelse(bp); 1121 anyfreed = 1; 1122 } 1123 if (nbp && 1124 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1125 (nbp->b_vp != vp) || 1126 (nbp->b_flags & B_DELWRI))) { 1127 goto restart; 1128 } 1129 } 1130 } 1131 1132 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1133 nbp = TAILQ_NEXT(bp, b_vnbufs); 1134 if (bp->b_lblkno >= trunclbn) { 1135 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 1136 BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL); 1137 goto restart; 1138 } else { 1139 bremfree(bp); 1140 bp->b_flags |= (B_INVAL | B_RELBUF); 1141 bp->b_flags &= ~B_ASYNC; 1142 brelse(bp); 1143 anyfreed = 1; 1144 } 1145 if (nbp && 1146 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1147 (nbp->b_vp != vp) || 1148 (nbp->b_flags & B_DELWRI) == 0)) { 1149 goto restart; 1150 } 1151 } 1152 } 1153 } 1154 1155 if (length > 0) { 1156restartsync: 1157 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1158 nbp = TAILQ_NEXT(bp, b_vnbufs); 1159 if ((bp->b_flags & B_DELWRI) && (bp->b_lblkno < 0)) { 1160 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 1161 BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL); 1162 goto restart; 1163 } else { 1164 bremfree(bp); 1165 if (bp->b_vp == vp) { 1166 bp->b_flags |= B_ASYNC; 1167 } else { 1168 bp->b_flags &= ~B_ASYNC; 1169 } 1170 BUF_WRITE(bp); 1171 } 1172 goto restartsync; 1173 } 1174 1175 } 1176 } 1177 1178 while (vp->v_numoutput > 0) { 1179 vp->v_flag |= VBWAIT; 1180 tsleep(&vp->v_numoutput, PVM, "vbtrunc", 0); 1181 } 1182 1183 splx(s); 1184 1185 vnode_pager_setsize(vp, length); 1186 1187 return (0); 1188} 1189 1190/* 1191 * Associate a buffer with a vnode. 1192 */ 1193void 1194bgetvp(vp, bp) 1195 register struct vnode *vp; 1196 register struct buf *bp; 1197{ 1198 int s; 1199 1200 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 1201 1202 vhold(vp); 1203 bp->b_vp = vp; 1204 bp->b_dev = vn_todev(vp); 1205 /* 1206 * Insert onto list for new vnode. 1207 */ 1208 s = splbio(); 1209 bp->b_xflags |= BX_VNCLEAN; 1210 bp->b_xflags &= ~BX_VNDIRTY; 1211 TAILQ_INSERT_TAIL(&vp->v_cleanblkhd, bp, b_vnbufs); 1212 splx(s); 1213} 1214 1215/* 1216 * Disassociate a buffer from a vnode. 1217 */ 1218void 1219brelvp(bp) 1220 register struct buf *bp; 1221{ 1222 struct vnode *vp; 1223 struct buflists *listheadp; 1224 int s; 1225 1226 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1227 1228 /* 1229 * Delete from old vnode list, if on one. 1230 */ 1231 vp = bp->b_vp; 1232 s = splbio(); 1233 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) { 1234 if (bp->b_xflags & BX_VNDIRTY) 1235 listheadp = &vp->v_dirtyblkhd; 1236 else 1237 listheadp = &vp->v_cleanblkhd; 1238 TAILQ_REMOVE(listheadp, bp, b_vnbufs); 1239 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1240 } 1241 if ((vp->v_flag & VONWORKLST) && TAILQ_EMPTY(&vp->v_dirtyblkhd)) { 1242 vp->v_flag &= ~VONWORKLST; 1243 LIST_REMOVE(vp, v_synclist); 1244 } 1245 splx(s); 1246 bp->b_vp = (struct vnode *) 0; 1247 vdrop(vp); 1248} 1249 1250/* 1251 * Add an item to the syncer work queue. 1252 */ 1253static void 1254vn_syncer_add_to_worklist(struct vnode *vp, int delay) 1255{ 1256 int s, slot; 1257 1258 s = splbio(); 1259 1260 if (vp->v_flag & VONWORKLST) { 1261 LIST_REMOVE(vp, v_synclist); 1262 } 1263 1264 if (delay > syncer_maxdelay - 2) 1265 delay = syncer_maxdelay - 2; 1266 slot = (syncer_delayno + delay) & syncer_mask; 1267 1268 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist); 1269 vp->v_flag |= VONWORKLST; 1270 splx(s); 1271} 1272 1273struct proc *updateproc; 1274static void sched_sync(void); 1275static struct kproc_desc up_kp = { 1276 "syncer", 1277 sched_sync, 1278 &updateproc 1279}; 1280SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 1281 1282/* 1283 * System filesystem synchronizer daemon. 1284 */ 1285void 1286sched_sync(void) 1287{ 1288 struct synclist *slp; 1289 struct vnode *vp; 1290 struct mount *mp; 1291 long starttime; 1292 int s; 1293 struct thread *td = FIRST_THREAD_IN_PROC(updateproc); /* XXXKSE */ 1294 1295 mtx_lock(&Giant); 1296 1297 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, td->td_proc, 1298 SHUTDOWN_PRI_LAST); 1299 1300 for (;;) { 1301 kthread_suspend_check(td->td_proc); 1302 1303 starttime = time_second; 1304 1305 /* 1306 * Push files whose dirty time has expired. Be careful 1307 * of interrupt race on slp queue. 1308 */ 1309 s = splbio(); 1310 slp = &syncer_workitem_pending[syncer_delayno]; 1311 syncer_delayno += 1; 1312 if (syncer_delayno == syncer_maxdelay) 1313 syncer_delayno = 0; 1314 splx(s); 1315 1316 while ((vp = LIST_FIRST(slp)) != NULL) { 1317 if (VOP_ISLOCKED(vp, NULL) == 0 && 1318 vn_start_write(vp, &mp, V_NOWAIT) == 0) { 1319 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 1320 (void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td); 1321 VOP_UNLOCK(vp, 0, td); 1322 vn_finished_write(mp); 1323 } 1324 s = splbio(); 1325 if (LIST_FIRST(slp) == vp) { 1326 /* 1327 * Note: v_tag VT_VFS vps can remain on the 1328 * worklist too with no dirty blocks, but 1329 * since sync_fsync() moves it to a different 1330 * slot we are safe. 1331 */ 1332 if (TAILQ_EMPTY(&vp->v_dirtyblkhd) && 1333 !vn_isdisk(vp, NULL)) 1334 panic("sched_sync: fsync failed vp %p tag %d", vp, vp->v_tag); 1335 /* 1336 * Put us back on the worklist. The worklist 1337 * routine will remove us from our current 1338 * position and then add us back in at a later 1339 * position. 1340 */ 1341 vn_syncer_add_to_worklist(vp, syncdelay); 1342 } 1343 splx(s); 1344 } 1345 1346 /* 1347 * Do soft update processing. 1348 */ 1349#ifdef SOFTUPDATES 1350 softdep_process_worklist(NULL); 1351#endif 1352 1353 /* 1354 * The variable rushjob allows the kernel to speed up the 1355 * processing of the filesystem syncer process. A rushjob 1356 * value of N tells the filesystem syncer to process the next 1357 * N seconds worth of work on its queue ASAP. Currently rushjob 1358 * is used by the soft update code to speed up the filesystem 1359 * syncer process when the incore state is getting so far 1360 * ahead of the disk that the kernel memory pool is being 1361 * threatened with exhaustion. 1362 */ 1363 if (rushjob > 0) { 1364 rushjob -= 1; 1365 continue; 1366 } 1367 /* 1368 * If it has taken us less than a second to process the 1369 * current work, then wait. Otherwise start right over 1370 * again. We can still lose time if any single round 1371 * takes more than two seconds, but it does not really 1372 * matter as we are just trying to generally pace the 1373 * filesystem activity. 1374 */ 1375 if (time_second == starttime) 1376 tsleep(&lbolt, PPAUSE, "syncer", 0); 1377 } 1378} 1379 1380/* 1381 * Request the syncer daemon to speed up its work. 1382 * We never push it to speed up more than half of its 1383 * normal turn time, otherwise it could take over the cpu. 1384 * XXXKSE only one update? 1385 */ 1386int 1387speedup_syncer() 1388{ 1389 1390 mtx_lock_spin(&sched_lock); 1391 if (FIRST_THREAD_IN_PROC(updateproc)->td_wchan == &lbolt) /* XXXKSE */ 1392 setrunnable(FIRST_THREAD_IN_PROC(updateproc)); 1393 mtx_unlock_spin(&sched_lock); 1394 if (rushjob < syncdelay / 2) { 1395 rushjob += 1; 1396 stat_rush_requests += 1; 1397 return (1); 1398 } 1399 return(0); 1400} 1401 1402/* 1403 * Associate a p-buffer with a vnode. 1404 * 1405 * Also sets B_PAGING flag to indicate that vnode is not fully associated 1406 * with the buffer. i.e. the bp has not been linked into the vnode or 1407 * ref-counted. 1408 */ 1409void 1410pbgetvp(vp, bp) 1411 register struct vnode *vp; 1412 register struct buf *bp; 1413{ 1414 1415 KASSERT(bp->b_vp == NULL, ("pbgetvp: not free")); 1416 1417 bp->b_vp = vp; 1418 bp->b_flags |= B_PAGING; 1419 bp->b_dev = vn_todev(vp); 1420} 1421 1422/* 1423 * Disassociate a p-buffer from a vnode. 1424 */ 1425void 1426pbrelvp(bp) 1427 register struct buf *bp; 1428{ 1429 1430 KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL")); 1431 1432 /* XXX REMOVE ME */ 1433 if (TAILQ_NEXT(bp, b_vnbufs) != NULL) { 1434 panic( 1435 "relpbuf(): b_vp was probably reassignbuf()d %p %x", 1436 bp, 1437 (int)bp->b_flags 1438 ); 1439 } 1440 bp->b_vp = (struct vnode *) 0; 1441 bp->b_flags &= ~B_PAGING; 1442} 1443 1444/* 1445 * Change the vnode a pager buffer is associated with. 1446 */ 1447void 1448pbreassignbuf(bp, newvp) 1449 struct buf *bp; 1450 struct vnode *newvp; 1451{ 1452 1453 KASSERT(bp->b_flags & B_PAGING, 1454 ("pbreassignbuf() on non phys bp %p", bp)); 1455 bp->b_vp = newvp; 1456} 1457 1458/* 1459 * Reassign a buffer from one vnode to another. 1460 * Used to assign file specific control information 1461 * (indirect blocks) to the vnode to which they belong. 1462 */ 1463void 1464reassignbuf(bp, newvp) 1465 register struct buf *bp; 1466 register struct vnode *newvp; 1467{ 1468 struct buflists *listheadp; 1469 int delay; 1470 int s; 1471 1472 if (newvp == NULL) { 1473 printf("reassignbuf: NULL"); 1474 return; 1475 } 1476 ++reassignbufcalls; 1477 1478 /* 1479 * B_PAGING flagged buffers cannot be reassigned because their vp 1480 * is not fully linked in. 1481 */ 1482 if (bp->b_flags & B_PAGING) 1483 panic("cannot reassign paging buffer"); 1484 1485 s = splbio(); 1486 /* 1487 * Delete from old vnode list, if on one. 1488 */ 1489 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) { 1490 if (bp->b_xflags & BX_VNDIRTY) 1491 listheadp = &bp->b_vp->v_dirtyblkhd; 1492 else 1493 listheadp = &bp->b_vp->v_cleanblkhd; 1494 TAILQ_REMOVE(listheadp, bp, b_vnbufs); 1495 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1496 if (bp->b_vp != newvp) { 1497 vdrop(bp->b_vp); 1498 bp->b_vp = NULL; /* for clarification */ 1499 } 1500 } 1501 /* 1502 * If dirty, put on list of dirty buffers; otherwise insert onto list 1503 * of clean buffers. 1504 */ 1505 if (bp->b_flags & B_DELWRI) { 1506 struct buf *tbp; 1507 1508 listheadp = &newvp->v_dirtyblkhd; 1509 if ((newvp->v_flag & VONWORKLST) == 0) { 1510 switch (newvp->v_type) { 1511 case VDIR: 1512 delay = dirdelay; 1513 break; 1514 case VCHR: 1515 if (newvp->v_rdev->si_mountpoint != NULL) { 1516 delay = metadelay; 1517 break; 1518 } 1519 /* fall through */ 1520 default: 1521 delay = filedelay; 1522 } 1523 vn_syncer_add_to_worklist(newvp, delay); 1524 } 1525 bp->b_xflags |= BX_VNDIRTY; 1526 tbp = TAILQ_FIRST(listheadp); 1527 if (tbp == NULL || 1528 bp->b_lblkno == 0 || 1529 (bp->b_lblkno > 0 && tbp->b_lblkno < 0) || 1530 (bp->b_lblkno > 0 && bp->b_lblkno < tbp->b_lblkno)) { 1531 TAILQ_INSERT_HEAD(listheadp, bp, b_vnbufs); 1532 ++reassignbufsortgood; 1533 } else if (bp->b_lblkno < 0) { 1534 TAILQ_INSERT_TAIL(listheadp, bp, b_vnbufs); 1535 ++reassignbufsortgood; 1536 } else if (reassignbufmethod == 1) { 1537 /* 1538 * New sorting algorithm, only handle sequential case, 1539 * otherwise append to end (but before metadata) 1540 */ 1541 if ((tbp = gbincore(newvp, bp->b_lblkno - 1)) != NULL && 1542 (tbp->b_xflags & BX_VNDIRTY)) { 1543 /* 1544 * Found the best place to insert the buffer 1545 */ 1546 TAILQ_INSERT_AFTER(listheadp, tbp, bp, b_vnbufs); 1547 ++reassignbufsortgood; 1548 } else { 1549 /* 1550 * Missed, append to end, but before meta-data. 1551 * We know that the head buffer in the list is 1552 * not meta-data due to prior conditionals. 1553 * 1554 * Indirect effects: NFS second stage write 1555 * tends to wind up here, giving maximum 1556 * distance between the unstable write and the 1557 * commit rpc. 1558 */ 1559 tbp = TAILQ_LAST(listheadp, buflists); 1560 while (tbp && tbp->b_lblkno < 0) 1561 tbp = TAILQ_PREV(tbp, buflists, b_vnbufs); 1562 TAILQ_INSERT_AFTER(listheadp, tbp, bp, b_vnbufs); 1563 ++reassignbufsortbad; 1564 } 1565 } else { 1566 /* 1567 * Old sorting algorithm, scan queue and insert 1568 */ 1569 struct buf *ttbp; 1570 while ((ttbp = TAILQ_NEXT(tbp, b_vnbufs)) && 1571 (ttbp->b_lblkno < bp->b_lblkno)) { 1572 ++reassignbufloops; 1573 tbp = ttbp; 1574 } 1575 TAILQ_INSERT_AFTER(listheadp, tbp, bp, b_vnbufs); 1576 } 1577 } else { 1578 bp->b_xflags |= BX_VNCLEAN; 1579 TAILQ_INSERT_TAIL(&newvp->v_cleanblkhd, bp, b_vnbufs); 1580 if ((newvp->v_flag & VONWORKLST) && 1581 TAILQ_EMPTY(&newvp->v_dirtyblkhd)) { 1582 newvp->v_flag &= ~VONWORKLST; 1583 LIST_REMOVE(newvp, v_synclist); 1584 } 1585 } 1586 if (bp->b_vp != newvp) { 1587 bp->b_vp = newvp; 1588 vhold(bp->b_vp); 1589 } 1590 splx(s); 1591} 1592 1593/* 1594 * Create a vnode for a device. 1595 * Used for mounting the root file system. 1596 */ 1597int 1598bdevvp(dev, vpp) 1599 dev_t dev; 1600 struct vnode **vpp; 1601{ 1602 register struct vnode *vp; 1603 struct vnode *nvp; 1604 int error; 1605 1606 if (dev == NODEV) { 1607 *vpp = NULLVP; 1608 return (ENXIO); 1609 } 1610 if (vfinddev(dev, VCHR, vpp)) 1611 return (0); 1612 error = getnewvnode(VT_NON, (struct mount *)0, spec_vnodeop_p, &nvp); 1613 if (error) { 1614 *vpp = NULLVP; 1615 return (error); 1616 } 1617 vp = nvp; 1618 vp->v_type = VCHR; 1619 addalias(vp, dev); 1620 *vpp = vp; 1621 return (0); 1622} 1623 1624/* 1625 * Add vnode to the alias list hung off the dev_t. 1626 * 1627 * The reason for this gunk is that multiple vnodes can reference 1628 * the same physical device, so checking vp->v_usecount to see 1629 * how many users there are is inadequate; the v_usecount for 1630 * the vnodes need to be accumulated. vcount() does that. 1631 */ 1632struct vnode * 1633addaliasu(nvp, nvp_rdev) 1634 struct vnode *nvp; 1635 udev_t nvp_rdev; 1636{ 1637 struct vnode *ovp; 1638 vop_t **ops; 1639 dev_t dev; 1640 1641 if (nvp->v_type == VBLK) 1642 return (nvp); 1643 if (nvp->v_type != VCHR) 1644 panic("addaliasu on non-special vnode"); 1645 dev = udev2dev(nvp_rdev, 0); 1646 /* 1647 * Check to see if we have a bdevvp vnode with no associated 1648 * filesystem. If so, we want to associate the filesystem of 1649 * the new newly instigated vnode with the bdevvp vnode and 1650 * discard the newly created vnode rather than leaving the 1651 * bdevvp vnode lying around with no associated filesystem. 1652 */ 1653 if (vfinddev(dev, nvp->v_type, &ovp) == 0 || ovp->v_data != NULL) { 1654 addalias(nvp, dev); 1655 return (nvp); 1656 } 1657 /* 1658 * Discard unneeded vnode, but save its node specific data. 1659 * Note that if there is a lock, it is carried over in the 1660 * node specific data to the replacement vnode. 1661 */ 1662 vref(ovp); 1663 ovp->v_data = nvp->v_data; 1664 ovp->v_tag = nvp->v_tag; 1665 nvp->v_data = NULL; 1666 lockinit(&ovp->v_lock, PVFS, nvp->v_lock.lk_wmesg, 1667 nvp->v_lock.lk_timo, nvp->v_lock.lk_flags & LK_EXTFLG_MASK); 1668 if (nvp->v_vnlock) 1669 ovp->v_vnlock = &ovp->v_lock; 1670 ops = ovp->v_op; 1671 ovp->v_op = nvp->v_op; 1672 if (VOP_ISLOCKED(nvp, curthread)) { 1673 VOP_UNLOCK(nvp, 0, curthread); 1674 vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread); 1675 } 1676 nvp->v_op = ops; 1677 insmntque(ovp, nvp->v_mount); 1678 vrele(nvp); 1679 vgone(nvp); 1680 return (ovp); 1681} 1682 1683/* This is a local helper function that do the same as addaliasu, but for a 1684 * dev_t instead of an udev_t. */ 1685static void 1686addalias(nvp, dev) 1687 struct vnode *nvp; 1688 dev_t dev; 1689{ 1690 1691 KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode")); 1692 nvp->v_rdev = dev; 1693 mtx_lock(&spechash_mtx); 1694 SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext); 1695 mtx_unlock(&spechash_mtx); 1696} 1697 1698/* 1699 * Grab a particular vnode from the free list, increment its 1700 * reference count and lock it. The vnode lock bit is set if the 1701 * vnode is being eliminated in vgone. The process is awakened 1702 * when the transition is completed, and an error returned to 1703 * indicate that the vnode is no longer usable (possibly having 1704 * been changed to a new file system type). 1705 */ 1706int 1707vget(vp, flags, td) 1708 register struct vnode *vp; 1709 int flags; 1710 struct thread *td; 1711{ 1712 int error; 1713 1714 /* 1715 * If the vnode is in the process of being cleaned out for 1716 * another use, we wait for the cleaning to finish and then 1717 * return failure. Cleaning is determined by checking that 1718 * the VXLOCK flag is set. 1719 */ 1720 if ((flags & LK_INTERLOCK) == 0) 1721 mtx_lock(&vp->v_interlock); 1722 if (vp->v_flag & VXLOCK) { 1723 if (vp->v_vxproc == curthread) { 1724#if 0 1725 /* this can now occur in normal operation */ 1726 log(LOG_INFO, "VXLOCK interlock avoided\n"); 1727#endif 1728 } else { 1729 vp->v_flag |= VXWANT; 1730 msleep((caddr_t)vp, &vp->v_interlock, PINOD | PDROP, 1731 "vget", 0); 1732 return (ENOENT); 1733 } 1734 } 1735 1736 vp->v_usecount++; 1737 1738 if (VSHOULDBUSY(vp)) 1739 vbusy(vp); 1740 if (flags & LK_TYPE_MASK) { 1741 if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { 1742 /* 1743 * must expand vrele here because we do not want 1744 * to call VOP_INACTIVE if the reference count 1745 * drops back to zero since it was never really 1746 * active. We must remove it from the free list 1747 * before sleeping so that multiple processes do 1748 * not try to recycle it. 1749 */ 1750 mtx_lock(&vp->v_interlock); 1751 vp->v_usecount--; 1752 if (VSHOULDFREE(vp)) 1753 vfree(vp); 1754 else 1755 vlruvp(vp); 1756 mtx_unlock(&vp->v_interlock); 1757 } 1758 return (error); 1759 } 1760 mtx_unlock(&vp->v_interlock); 1761 return (0); 1762} 1763 1764/* 1765 * Increase the reference count of a vnode. 1766 */ 1767void 1768vref(struct vnode *vp) 1769{ 1770 mtx_lock(&vp->v_interlock); 1771 vp->v_usecount++; 1772 mtx_unlock(&vp->v_interlock); 1773} 1774 1775/* 1776 * Vnode put/release. 1777 * If count drops to zero, call inactive routine and return to freelist. 1778 */ 1779void 1780vrele(vp) 1781 struct vnode *vp; 1782{ 1783 struct thread *td = curthread; /* XXX */ 1784 1785 KASSERT(vp != NULL, ("vrele: null vp")); 1786 1787 mtx_lock(&vp->v_interlock); 1788 1789 /* Skip this v_writecount check if we're going to panic below. */ 1790 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 1791 ("vrele: missed vn_close")); 1792 1793 if (vp->v_usecount > 1) { 1794 1795 vp->v_usecount--; 1796 mtx_unlock(&vp->v_interlock); 1797 1798 return; 1799 } 1800 1801 if (vp->v_usecount == 1) { 1802 vp->v_usecount--; 1803 /* 1804 * We must call VOP_INACTIVE with the node locked. 1805 * If we are doing a vput, the node is already locked, 1806 * but, in the case of vrele, we must explicitly lock 1807 * the vnode before calling VOP_INACTIVE. 1808 */ 1809 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) 1810 VOP_INACTIVE(vp, td); 1811 if (VSHOULDFREE(vp)) 1812 vfree(vp); 1813 else 1814 vlruvp(vp); 1815 1816 } else { 1817#ifdef DIAGNOSTIC 1818 vprint("vrele: negative ref count", vp); 1819 mtx_unlock(&vp->v_interlock); 1820#endif 1821 panic("vrele: negative ref cnt"); 1822 } 1823} 1824 1825/* 1826 * Release an already locked vnode. This give the same effects as 1827 * unlock+vrele(), but takes less time and avoids releasing and 1828 * re-aquiring the lock (as vrele() aquires the lock internally.) 1829 */ 1830void 1831vput(vp) 1832 struct vnode *vp; 1833{ 1834 struct thread *td = curthread; /* XXX */ 1835 1836 GIANT_REQUIRED; 1837 1838 KASSERT(vp != NULL, ("vput: null vp")); 1839 mtx_lock(&vp->v_interlock); 1840 /* Skip this v_writecount check if we're going to panic below. */ 1841 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 1842 ("vput: missed vn_close")); 1843 1844 if (vp->v_usecount > 1) { 1845 vp->v_usecount--; 1846 VOP_UNLOCK(vp, LK_INTERLOCK, td); 1847 return; 1848 } 1849 1850 if (vp->v_usecount == 1) { 1851 vp->v_usecount--; 1852 /* 1853 * We must call VOP_INACTIVE with the node locked. 1854 * If we are doing a vput, the node is already locked, 1855 * so we just need to release the vnode mutex. 1856 */ 1857 mtx_unlock(&vp->v_interlock); 1858 VOP_INACTIVE(vp, td); 1859 if (VSHOULDFREE(vp)) 1860 vfree(vp); 1861 else 1862 vlruvp(vp); 1863 1864 } else { 1865#ifdef DIAGNOSTIC 1866 vprint("vput: negative ref count", vp); 1867#endif 1868 panic("vput: negative ref cnt"); 1869 } 1870} 1871 1872/* 1873 * Somebody doesn't want the vnode recycled. 1874 */ 1875void 1876vhold(vp) 1877 register struct vnode *vp; 1878{ 1879 int s; 1880 1881 s = splbio(); 1882 vp->v_holdcnt++; 1883 if (VSHOULDBUSY(vp)) 1884 vbusy(vp); 1885 splx(s); 1886} 1887 1888/* 1889 * Note that there is one less who cares about this vnode. vdrop() is the 1890 * opposite of vhold(). 1891 */ 1892void 1893vdrop(vp) 1894 register struct vnode *vp; 1895{ 1896 int s; 1897 1898 s = splbio(); 1899 if (vp->v_holdcnt <= 0) 1900 panic("vdrop: holdcnt"); 1901 vp->v_holdcnt--; 1902 if (VSHOULDFREE(vp)) 1903 vfree(vp); 1904 else 1905 vlruvp(vp); 1906 splx(s); 1907} 1908 1909/* 1910 * Remove any vnodes in the vnode table belonging to mount point mp. 1911 * 1912 * If FORCECLOSE is not specified, there should not be any active ones, 1913 * return error if any are found (nb: this is a user error, not a 1914 * system error). If FORCECLOSE is specified, detach any active vnodes 1915 * that are found. 1916 * 1917 * If WRITECLOSE is set, only flush out regular file vnodes open for 1918 * writing. 1919 * 1920 * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped. 1921 * 1922 * `rootrefs' specifies the base reference count for the root vnode 1923 * of this filesystem. The root vnode is considered busy if its 1924 * v_usecount exceeds this value. On a successful return, vflush() 1925 * will call vrele() on the root vnode exactly rootrefs times. 1926 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 1927 * be zero. 1928 */ 1929#ifdef DIAGNOSTIC 1930static int busyprt = 0; /* print out busy vnodes */ 1931SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 1932#endif 1933 1934int 1935vflush(mp, rootrefs, flags) 1936 struct mount *mp; 1937 int rootrefs; 1938 int flags; 1939{ 1940 struct thread *td = curthread; /* XXX */ 1941 struct vnode *vp, *nvp, *rootvp = NULL; 1942 struct vattr vattr; 1943 int busy = 0, error; 1944 1945 if (rootrefs > 0) { 1946 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 1947 ("vflush: bad args")); 1948 /* 1949 * Get the filesystem root vnode. We can vput() it 1950 * immediately, since with rootrefs > 0, it won't go away. 1951 */ 1952 if ((error = VFS_ROOT(mp, &rootvp)) != 0) 1953 return (error); 1954 vput(rootvp); 1955 } 1956 mtx_lock(&mntvnode_mtx); 1957loop: 1958 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp; vp = nvp) { 1959 /* 1960 * Make sure this vnode wasn't reclaimed in getnewvnode(). 1961 * Start over if it has (it won't be on the list anymore). 1962 */ 1963 if (vp->v_mount != mp) 1964 goto loop; 1965 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 1966 1967 mtx_unlock(&mntvnode_mtx); 1968 mtx_lock(&vp->v_interlock); 1969 /* 1970 * Skip over a vnodes marked VSYSTEM. 1971 */ 1972 if ((flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) { 1973 mtx_unlock(&vp->v_interlock); 1974 mtx_lock(&mntvnode_mtx); 1975 continue; 1976 } 1977 /* 1978 * If WRITECLOSE is set, flush out unlinked but still open 1979 * files (even if open only for reading) and regular file 1980 * vnodes open for writing. 1981 */ 1982 if ((flags & WRITECLOSE) && 1983 (vp->v_type == VNON || 1984 (VOP_GETATTR(vp, &vattr, td->td_ucred, td) == 0 && 1985 vattr.va_nlink > 0)) && 1986 (vp->v_writecount == 0 || vp->v_type != VREG)) { 1987 mtx_unlock(&vp->v_interlock); 1988 mtx_lock(&mntvnode_mtx); 1989 continue; 1990 } 1991 1992 /* 1993 * With v_usecount == 0, all we need to do is clear out the 1994 * vnode data structures and we are done. 1995 */ 1996 if (vp->v_usecount == 0) { 1997 vgonel(vp, td); 1998 mtx_lock(&mntvnode_mtx); 1999 continue; 2000 } 2001 2002 /* 2003 * If FORCECLOSE is set, forcibly close the vnode. For block 2004 * or character devices, revert to an anonymous device. For 2005 * all other files, just kill them. 2006 */ 2007 if (flags & FORCECLOSE) { 2008 if (vp->v_type != VCHR) { 2009 vgonel(vp, td); 2010 } else { 2011 vclean(vp, 0, td); 2012 vp->v_op = spec_vnodeop_p; 2013 insmntque(vp, (struct mount *) 0); 2014 } 2015 mtx_lock(&mntvnode_mtx); 2016 continue; 2017 } 2018#ifdef DIAGNOSTIC 2019 if (busyprt) 2020 vprint("vflush: busy vnode", vp); 2021#endif 2022 mtx_unlock(&vp->v_interlock); 2023 mtx_lock(&mntvnode_mtx); 2024 busy++; 2025 } 2026 mtx_unlock(&mntvnode_mtx); 2027 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2028 /* 2029 * If just the root vnode is busy, and if its refcount 2030 * is equal to `rootrefs', then go ahead and kill it. 2031 */ 2032 mtx_lock(&rootvp->v_interlock); 2033 KASSERT(busy > 0, ("vflush: not busy")); 2034 KASSERT(rootvp->v_usecount >= rootrefs, ("vflush: rootrefs")); 2035 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2036 vgonel(rootvp, td); 2037 busy = 0; 2038 } else 2039 mtx_unlock(&rootvp->v_interlock); 2040 } 2041 if (busy) 2042 return (EBUSY); 2043 for (; rootrefs > 0; rootrefs--) 2044 vrele(rootvp); 2045 return (0); 2046} 2047 2048/* 2049 * This moves a now (likely recyclable) vnode to the end of the 2050 * mountlist. XXX However, it is temporarily disabled until we 2051 * can clean up ffs_sync() and friends, which have loop restart 2052 * conditions which this code causes to operate O(N^2). 2053 */ 2054static void 2055vlruvp(struct vnode *vp) 2056{ 2057#if 0 2058 struct mount *mp; 2059 2060 if ((mp = vp->v_mount) != NULL) { 2061 mtx_lock(&mntvnode_mtx); 2062 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2063 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2064 mtx_unlock(&mntvnode_mtx); 2065 } 2066#endif 2067} 2068 2069/* 2070 * Disassociate the underlying file system from a vnode. 2071 */ 2072static void 2073vclean(vp, flags, td) 2074 struct vnode *vp; 2075 int flags; 2076 struct thread *td; 2077{ 2078 int active; 2079 2080 /* 2081 * Check to see if the vnode is in use. If so we have to reference it 2082 * before we clean it out so that its count cannot fall to zero and 2083 * generate a race against ourselves to recycle it. 2084 */ 2085 if ((active = vp->v_usecount)) 2086 vp->v_usecount++; 2087 2088 /* 2089 * Prevent the vnode from being recycled or brought into use while we 2090 * clean it out. 2091 */ 2092 if (vp->v_flag & VXLOCK) 2093 panic("vclean: deadlock"); 2094 vp->v_flag |= VXLOCK; 2095 vp->v_vxproc = curthread; 2096 /* 2097 * Even if the count is zero, the VOP_INACTIVE routine may still 2098 * have the object locked while it cleans it out. The VOP_LOCK 2099 * ensures that the VOP_INACTIVE routine is done with its work. 2100 * For active vnodes, it ensures that no other activity can 2101 * occur while the underlying object is being cleaned out. 2102 */ 2103 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2104 2105 /* 2106 * Clean out any buffers associated with the vnode. 2107 * If the flush fails, just toss the buffers. 2108 */ 2109 if (flags & DOCLOSE) { 2110 if (TAILQ_FIRST(&vp->v_dirtyblkhd) != NULL) 2111 (void) vn_write_suspend_wait(vp, NULL, V_WAIT); 2112 if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0) 2113 vinvalbuf(vp, 0, NOCRED, td, 0, 0); 2114 } 2115 2116 VOP_DESTROYVOBJECT(vp); 2117 2118 /* 2119 * If purging an active vnode, it must be closed and 2120 * deactivated before being reclaimed. Note that the 2121 * VOP_INACTIVE will unlock the vnode. 2122 */ 2123 if (active) { 2124 if (flags & DOCLOSE) 2125 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2126 VOP_INACTIVE(vp, td); 2127 } else { 2128 /* 2129 * Any other processes trying to obtain this lock must first 2130 * wait for VXLOCK to clear, then call the new lock operation. 2131 */ 2132 VOP_UNLOCK(vp, 0, td); 2133 } 2134 /* 2135 * Reclaim the vnode. 2136 */ 2137 if (VOP_RECLAIM(vp, td)) 2138 panic("vclean: cannot reclaim"); 2139 2140 if (active) { 2141 /* 2142 * Inline copy of vrele() since VOP_INACTIVE 2143 * has already been called. 2144 */ 2145 mtx_lock(&vp->v_interlock); 2146 if (--vp->v_usecount <= 0) { 2147#ifdef DIAGNOSTIC 2148 if (vp->v_usecount < 0 || vp->v_writecount != 0) { 2149 vprint("vclean: bad ref count", vp); 2150 panic("vclean: ref cnt"); 2151 } 2152#endif 2153 vfree(vp); 2154 } 2155 mtx_unlock(&vp->v_interlock); 2156 } 2157 2158 cache_purge(vp); 2159 vp->v_vnlock = NULL; 2160 lockdestroy(&vp->v_lock); 2161 2162 if (VSHOULDFREE(vp)) 2163 vfree(vp); 2164 2165 /* 2166 * Done with purge, notify sleepers of the grim news. 2167 */ 2168 vp->v_op = dead_vnodeop_p; 2169 if (vp->v_pollinfo != NULL) 2170 vn_pollgone(vp); 2171 vp->v_tag = VT_NON; 2172 vp->v_flag &= ~VXLOCK; 2173 vp->v_vxproc = NULL; 2174 if (vp->v_flag & VXWANT) { 2175 vp->v_flag &= ~VXWANT; 2176 wakeup((caddr_t) vp); 2177 } 2178} 2179 2180/* 2181 * Eliminate all activity associated with the requested vnode 2182 * and with all vnodes aliased to the requested vnode. 2183 */ 2184int 2185vop_revoke(ap) 2186 struct vop_revoke_args /* { 2187 struct vnode *a_vp; 2188 int a_flags; 2189 } */ *ap; 2190{ 2191 struct vnode *vp, *vq; 2192 dev_t dev; 2193 2194 KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke")); 2195 2196 vp = ap->a_vp; 2197 /* 2198 * If a vgone (or vclean) is already in progress, 2199 * wait until it is done and return. 2200 */ 2201 if (vp->v_flag & VXLOCK) { 2202 vp->v_flag |= VXWANT; 2203 msleep((caddr_t)vp, &vp->v_interlock, PINOD | PDROP, 2204 "vop_revokeall", 0); 2205 return (0); 2206 } 2207 dev = vp->v_rdev; 2208 for (;;) { 2209 mtx_lock(&spechash_mtx); 2210 vq = SLIST_FIRST(&dev->si_hlist); 2211 mtx_unlock(&spechash_mtx); 2212 if (!vq) 2213 break; 2214 vgone(vq); 2215 } 2216 return (0); 2217} 2218 2219/* 2220 * Recycle an unused vnode to the front of the free list. 2221 * Release the passed interlock if the vnode will be recycled. 2222 */ 2223int 2224vrecycle(vp, inter_lkp, td) 2225 struct vnode *vp; 2226 struct mtx *inter_lkp; 2227 struct thread *td; 2228{ 2229 2230 mtx_lock(&vp->v_interlock); 2231 if (vp->v_usecount == 0) { 2232 if (inter_lkp) { 2233 mtx_unlock(inter_lkp); 2234 } 2235 vgonel(vp, td); 2236 return (1); 2237 } 2238 mtx_unlock(&vp->v_interlock); 2239 return (0); 2240} 2241 2242/* 2243 * Eliminate all activity associated with a vnode 2244 * in preparation for reuse. 2245 */ 2246void 2247vgone(vp) 2248 register struct vnode *vp; 2249{ 2250 struct thread *td = curthread; /* XXX */ 2251 2252 mtx_lock(&vp->v_interlock); 2253 vgonel(vp, td); 2254} 2255 2256/* 2257 * vgone, with the vp interlock held. 2258 */ 2259void 2260vgonel(vp, td) 2261 struct vnode *vp; 2262 struct thread *td; 2263{ 2264 int s; 2265 2266 /* 2267 * If a vgone (or vclean) is already in progress, 2268 * wait until it is done and return. 2269 */ 2270 if (vp->v_flag & VXLOCK) { 2271 vp->v_flag |= VXWANT; 2272 msleep((caddr_t)vp, &vp->v_interlock, PINOD | PDROP, 2273 "vgone", 0); 2274 return; 2275 } 2276 2277 /* 2278 * Clean out the filesystem specific data. 2279 */ 2280 vclean(vp, DOCLOSE, td); 2281 mtx_lock(&vp->v_interlock); 2282 2283 /* 2284 * Delete from old mount point vnode list, if on one. 2285 */ 2286 if (vp->v_mount != NULL) 2287 insmntque(vp, (struct mount *)0); 2288 /* 2289 * If special device, remove it from special device alias list 2290 * if it is on one. 2291 */ 2292 if (vp->v_type == VCHR && vp->v_rdev != NULL && vp->v_rdev != NODEV) { 2293 mtx_lock(&spechash_mtx); 2294 SLIST_REMOVE(&vp->v_rdev->si_hlist, vp, vnode, v_specnext); 2295 freedev(vp->v_rdev); 2296 mtx_unlock(&spechash_mtx); 2297 vp->v_rdev = NULL; 2298 } 2299 2300 /* 2301 * If it is on the freelist and not already at the head, 2302 * move it to the head of the list. The test of the 2303 * VDOOMED flag and the reference count of zero is because 2304 * it will be removed from the free list by getnewvnode, 2305 * but will not have its reference count incremented until 2306 * after calling vgone. If the reference count were 2307 * incremented first, vgone would (incorrectly) try to 2308 * close the previous instance of the underlying object. 2309 */ 2310 if (vp->v_usecount == 0 && !(vp->v_flag & VDOOMED)) { 2311 s = splbio(); 2312 mtx_lock(&vnode_free_list_mtx); 2313 if (vp->v_flag & VFREE) 2314 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2315 else 2316 freevnodes++; 2317 vp->v_flag |= VFREE; 2318 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2319 mtx_unlock(&vnode_free_list_mtx); 2320 splx(s); 2321 } 2322 2323 vp->v_type = VBAD; 2324 mtx_unlock(&vp->v_interlock); 2325} 2326 2327/* 2328 * Lookup a vnode by device number. 2329 */ 2330int 2331vfinddev(dev, type, vpp) 2332 dev_t dev; 2333 enum vtype type; 2334 struct vnode **vpp; 2335{ 2336 struct vnode *vp; 2337 2338 mtx_lock(&spechash_mtx); 2339 SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) { 2340 if (type == vp->v_type) { 2341 *vpp = vp; 2342 mtx_unlock(&spechash_mtx); 2343 return (1); 2344 } 2345 } 2346 mtx_unlock(&spechash_mtx); 2347 return (0); 2348} 2349 2350/* 2351 * Calculate the total number of references to a special device. 2352 */ 2353int 2354vcount(vp) 2355 struct vnode *vp; 2356{ 2357 struct vnode *vq; 2358 int count; 2359 2360 count = 0; 2361 mtx_lock(&spechash_mtx); 2362 SLIST_FOREACH(vq, &vp->v_rdev->si_hlist, v_specnext) 2363 count += vq->v_usecount; 2364 mtx_unlock(&spechash_mtx); 2365 return (count); 2366} 2367 2368/* 2369 * Same as above, but using the dev_t as argument 2370 */ 2371int 2372count_dev(dev) 2373 dev_t dev; 2374{ 2375 struct vnode *vp; 2376 2377 vp = SLIST_FIRST(&dev->si_hlist); 2378 if (vp == NULL) 2379 return (0); 2380 return(vcount(vp)); 2381} 2382 2383/* 2384 * Print out a description of a vnode. 2385 */ 2386static char *typename[] = 2387{"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 2388 2389void 2390vprint(label, vp) 2391 char *label; 2392 struct vnode *vp; 2393{ 2394 char buf[96]; 2395 2396 if (label != NULL) 2397 printf("%s: %p: ", label, (void *)vp); 2398 else 2399 printf("%p: ", (void *)vp); 2400 printf("type %s, usecount %d, writecount %d, refcount %d,", 2401 typename[vp->v_type], vp->v_usecount, vp->v_writecount, 2402 vp->v_holdcnt); 2403 buf[0] = '\0'; 2404 if (vp->v_flag & VROOT) 2405 strcat(buf, "|VROOT"); 2406 if (vp->v_flag & VTEXT) 2407 strcat(buf, "|VTEXT"); 2408 if (vp->v_flag & VSYSTEM) 2409 strcat(buf, "|VSYSTEM"); 2410 if (vp->v_flag & VXLOCK) 2411 strcat(buf, "|VXLOCK"); 2412 if (vp->v_flag & VXWANT) 2413 strcat(buf, "|VXWANT"); 2414 if (vp->v_flag & VBWAIT) 2415 strcat(buf, "|VBWAIT"); 2416 if (vp->v_flag & VDOOMED) 2417 strcat(buf, "|VDOOMED"); 2418 if (vp->v_flag & VFREE) 2419 strcat(buf, "|VFREE"); 2420 if (vp->v_flag & VOBJBUF) 2421 strcat(buf, "|VOBJBUF"); 2422 if (buf[0] != '\0') 2423 printf(" flags (%s)", &buf[1]); 2424 if (vp->v_data == NULL) { 2425 printf("\n"); 2426 } else { 2427 printf("\n\t"); 2428 VOP_PRINT(vp); 2429 } 2430} 2431 2432#ifdef DDB 2433#include <ddb/ddb.h> 2434/* 2435 * List all of the locked vnodes in the system. 2436 * Called when debugging the kernel. 2437 */ 2438DB_SHOW_COMMAND(lockedvnodes, lockedvnodes) 2439{ 2440 struct thread *td = curthread; /* XXX */ 2441 struct mount *mp, *nmp; 2442 struct vnode *vp; 2443 2444 printf("Locked vnodes\n"); 2445 mtx_lock(&mountlist_mtx); 2446 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2447 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 2448 nmp = TAILQ_NEXT(mp, mnt_list); 2449 continue; 2450 } 2451 mtx_lock(&mntvnode_mtx); 2452 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2453 if (VOP_ISLOCKED(vp, NULL)) 2454 vprint((char *)0, vp); 2455 } 2456 mtx_unlock(&mntvnode_mtx); 2457 mtx_lock(&mountlist_mtx); 2458 nmp = TAILQ_NEXT(mp, mnt_list); 2459 vfs_unbusy(mp, td); 2460 } 2461 mtx_unlock(&mountlist_mtx); 2462} 2463#endif 2464 2465/* 2466 * Top level filesystem related information gathering. 2467 */ 2468static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2469 2470static int 2471vfs_sysctl(SYSCTL_HANDLER_ARGS) 2472{ 2473 int *name = (int *)arg1 - 1; /* XXX */ 2474 u_int namelen = arg2 + 1; /* XXX */ 2475 struct vfsconf *vfsp; 2476 2477#if 1 || defined(COMPAT_PRELITE2) 2478 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2479 if (namelen == 1) 2480 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2481#endif 2482 2483 /* XXX the below code does not compile; vfs_sysctl does not exist. */ 2484#ifdef notyet 2485 /* all sysctl names at this level are at least name and field */ 2486 if (namelen < 2) 2487 return (ENOTDIR); /* overloaded */ 2488 if (name[0] != VFS_GENERIC) { 2489 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) 2490 if (vfsp->vfc_typenum == name[0]) 2491 break; 2492 if (vfsp == NULL) 2493 return (EOPNOTSUPP); 2494 return ((*vfsp->vfc_vfsops->vfs_sysctl)(&name[1], namelen - 1, 2495 oldp, oldlenp, newp, newlen, td)); 2496 } 2497#endif 2498 switch (name[1]) { 2499 case VFS_MAXTYPENUM: 2500 if (namelen != 2) 2501 return (ENOTDIR); 2502 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2503 case VFS_CONF: 2504 if (namelen != 3) 2505 return (ENOTDIR); /* overloaded */ 2506 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) 2507 if (vfsp->vfc_typenum == name[2]) 2508 break; 2509 if (vfsp == NULL) 2510 return (EOPNOTSUPP); 2511 return (SYSCTL_OUT(req, vfsp, sizeof *vfsp)); 2512 } 2513 return (EOPNOTSUPP); 2514} 2515 2516SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD, vfs_sysctl, 2517 "Generic filesystem"); 2518 2519#if 1 || defined(COMPAT_PRELITE2) 2520 2521static int 2522sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 2523{ 2524 int error; 2525 struct vfsconf *vfsp; 2526 struct ovfsconf ovfs; 2527 2528 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) { 2529 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 2530 strcpy(ovfs.vfc_name, vfsp->vfc_name); 2531 ovfs.vfc_index = vfsp->vfc_typenum; 2532 ovfs.vfc_refcount = vfsp->vfc_refcount; 2533 ovfs.vfc_flags = vfsp->vfc_flags; 2534 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 2535 if (error) 2536 return error; 2537 } 2538 return 0; 2539} 2540 2541#endif /* 1 || COMPAT_PRELITE2 */ 2542 2543#if COMPILING_LINT 2544#define KINFO_VNODESLOP 10 2545/* 2546 * Dump vnode list (via sysctl). 2547 * Copyout address of vnode followed by vnode. 2548 */ 2549/* ARGSUSED */ 2550static int 2551sysctl_vnode(SYSCTL_HANDLER_ARGS) 2552{ 2553 struct thread *td = curthread; /* XXX */ 2554 struct mount *mp, *nmp; 2555 struct vnode *nvp, *vp; 2556 int error; 2557 2558#define VPTRSZ sizeof (struct vnode *) 2559#define VNODESZ sizeof (struct vnode) 2560 2561 req->lock = 0; 2562 if (!req->oldptr) /* Make an estimate */ 2563 return (SYSCTL_OUT(req, 0, 2564 (numvnodes + KINFO_VNODESLOP) * (VPTRSZ + VNODESZ))); 2565 2566 mtx_lock(&mountlist_mtx); 2567 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2568 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 2569 nmp = TAILQ_NEXT(mp, mnt_list); 2570 continue; 2571 } 2572 mtx_lock(&mntvnode_mtx); 2573again: 2574 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 2575 vp != NULL; 2576 vp = nvp) { 2577 /* 2578 * Check that the vp is still associated with 2579 * this filesystem. RACE: could have been 2580 * recycled onto the same filesystem. 2581 */ 2582 if (vp->v_mount != mp) 2583 goto again; 2584 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 2585 mtx_unlock(&mntvnode_mtx); 2586 if ((error = SYSCTL_OUT(req, &vp, VPTRSZ)) || 2587 (error = SYSCTL_OUT(req, vp, VNODESZ))) 2588 return (error); 2589 mtx_lock(&mntvnode_mtx); 2590 } 2591 mtx_unlock(&mntvnode_mtx); 2592 mtx_lock(&mountlist_mtx); 2593 nmp = TAILQ_NEXT(mp, mnt_list); 2594 vfs_unbusy(mp, td); 2595 } 2596 mtx_unlock(&mountlist_mtx); 2597 2598 return (0); 2599} 2600 2601/* 2602 * XXX 2603 * Exporting the vnode list on large systems causes them to crash. 2604 * Exporting the vnode list on medium systems causes sysctl to coredump. 2605 */ 2606SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 2607 0, 0, sysctl_vnode, "S,vnode", ""); 2608#endif 2609 2610/* 2611 * Check to see if a filesystem is mounted on a block device. 2612 */ 2613int 2614vfs_mountedon(vp) 2615 struct vnode *vp; 2616{ 2617 2618 if (vp->v_rdev->si_mountpoint != NULL) 2619 return (EBUSY); 2620 return (0); 2621} 2622 2623/* 2624 * Unmount all filesystems. The list is traversed in reverse order 2625 * of mounting to avoid dependencies. 2626 */ 2627void 2628vfs_unmountall() 2629{ 2630 struct mount *mp; 2631 struct thread *td; 2632 int error; 2633 2634 if (curthread != NULL) 2635 td = curthread; 2636 else 2637 td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */ 2638 /* 2639 * Since this only runs when rebooting, it is not interlocked. 2640 */ 2641 while(!TAILQ_EMPTY(&mountlist)) { 2642 mp = TAILQ_LAST(&mountlist, mntlist); 2643 error = dounmount(mp, MNT_FORCE, td); 2644 if (error) { 2645 TAILQ_REMOVE(&mountlist, mp, mnt_list); 2646 printf("unmount of %s failed (", 2647 mp->mnt_stat.f_mntonname); 2648 if (error == EBUSY) 2649 printf("BUSY)\n"); 2650 else 2651 printf("%d)\n", error); 2652 } else { 2653 /* The unmount has removed mp from the mountlist */ 2654 } 2655 } 2656} 2657 2658/* 2659 * perform msync on all vnodes under a mount point 2660 * the mount point must be locked. 2661 */ 2662void 2663vfs_msync(struct mount *mp, int flags) 2664{ 2665 struct vnode *vp, *nvp; 2666 struct vm_object *obj; 2667 int tries; 2668 2669 GIANT_REQUIRED; 2670 2671 tries = 5; 2672 mtx_lock(&mntvnode_mtx); 2673loop: 2674 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp != NULL; vp = nvp) { 2675 if (vp->v_mount != mp) { 2676 if (--tries > 0) 2677 goto loop; 2678 break; 2679 } 2680 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 2681 2682 if (vp->v_flag & VXLOCK) /* XXX: what if MNT_WAIT? */ 2683 continue; 2684 2685 if (vp->v_flag & VNOSYNC) /* unlinked, skip it */ 2686 continue; 2687 2688 if ((vp->v_flag & VOBJDIRTY) && 2689 (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { 2690 mtx_unlock(&mntvnode_mtx); 2691 if (!vget(vp, 2692 LK_EXCLUSIVE | LK_RETRY | LK_NOOBJ, curthread)) { 2693 if (VOP_GETVOBJECT(vp, &obj) == 0) { 2694 vm_object_page_clean(obj, 0, 0, 2695 flags == MNT_WAIT ? 2696 OBJPC_SYNC : OBJPC_NOSYNC); 2697 } 2698 vput(vp); 2699 } 2700 mtx_lock(&mntvnode_mtx); 2701 if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) { 2702 if (--tries > 0) 2703 goto loop; 2704 break; 2705 } 2706 } 2707 } 2708 mtx_unlock(&mntvnode_mtx); 2709} 2710 2711/* 2712 * Create the VM object needed for VMIO and mmap support. This 2713 * is done for all VREG files in the system. Some filesystems might 2714 * afford the additional metadata buffering capability of the 2715 * VMIO code by making the device node be VMIO mode also. 2716 * 2717 * vp must be locked when vfs_object_create is called. 2718 */ 2719int 2720vfs_object_create(vp, td, cred) 2721 struct vnode *vp; 2722 struct thread *td; 2723 struct ucred *cred; 2724{ 2725 GIANT_REQUIRED; 2726 return (VOP_CREATEVOBJECT(vp, cred, td)); 2727} 2728 2729/* 2730 * Mark a vnode as free, putting it up for recycling. 2731 */ 2732void 2733vfree(vp) 2734 struct vnode *vp; 2735{ 2736 int s; 2737 2738 s = splbio(); 2739 mtx_lock(&vnode_free_list_mtx); 2740 KASSERT((vp->v_flag & VFREE) == 0, ("vnode already free")); 2741 if (vp->v_flag & VAGE) { 2742 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2743 } else { 2744 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 2745 } 2746 freevnodes++; 2747 mtx_unlock(&vnode_free_list_mtx); 2748 vp->v_flag &= ~VAGE; 2749 vp->v_flag |= VFREE; 2750 splx(s); 2751} 2752 2753/* 2754 * Opposite of vfree() - mark a vnode as in use. 2755 */ 2756void 2757vbusy(vp) 2758 struct vnode *vp; 2759{ 2760 int s; 2761 2762 s = splbio(); 2763 mtx_lock(&vnode_free_list_mtx); 2764 KASSERT((vp->v_flag & VFREE) != 0, ("vnode not free")); 2765 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2766 freevnodes--; 2767 mtx_unlock(&vnode_free_list_mtx); 2768 vp->v_flag &= ~(VFREE|VAGE); 2769 splx(s); 2770} 2771 2772/* 2773 * Record a process's interest in events which might happen to 2774 * a vnode. Because poll uses the historic select-style interface 2775 * internally, this routine serves as both the ``check for any 2776 * pending events'' and the ``record my interest in future events'' 2777 * functions. (These are done together, while the lock is held, 2778 * to avoid race conditions.) 2779 */ 2780int 2781vn_pollrecord(vp, td, events) 2782 struct vnode *vp; 2783 struct thread *td; 2784 short events; 2785{ 2786 2787 if (vp->v_pollinfo == NULL) 2788 v_addpollinfo(vp); 2789 mtx_lock(&vp->v_pollinfo->vpi_lock); 2790 if (vp->v_pollinfo->vpi_revents & events) { 2791 /* 2792 * This leaves events we are not interested 2793 * in available for the other process which 2794 * which presumably had requested them 2795 * (otherwise they would never have been 2796 * recorded). 2797 */ 2798 events &= vp->v_pollinfo->vpi_revents; 2799 vp->v_pollinfo->vpi_revents &= ~events; 2800 2801 mtx_unlock(&vp->v_pollinfo->vpi_lock); 2802 return events; 2803 } 2804 vp->v_pollinfo->vpi_events |= events; 2805 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 2806 mtx_unlock(&vp->v_pollinfo->vpi_lock); 2807 return 0; 2808} 2809 2810/* 2811 * Note the occurrence of an event. If the VN_POLLEVENT macro is used, 2812 * it is possible for us to miss an event due to race conditions, but 2813 * that condition is expected to be rare, so for the moment it is the 2814 * preferred interface. 2815 */ 2816void 2817vn_pollevent(vp, events) 2818 struct vnode *vp; 2819 short events; 2820{ 2821 2822 if (vp->v_pollinfo == NULL) 2823 v_addpollinfo(vp); 2824 mtx_lock(&vp->v_pollinfo->vpi_lock); 2825 if (vp->v_pollinfo->vpi_events & events) { 2826 /* 2827 * We clear vpi_events so that we don't 2828 * call selwakeup() twice if two events are 2829 * posted before the polling process(es) is 2830 * awakened. This also ensures that we take at 2831 * most one selwakeup() if the polling process 2832 * is no longer interested. However, it does 2833 * mean that only one event can be noticed at 2834 * a time. (Perhaps we should only clear those 2835 * event bits which we note?) XXX 2836 */ 2837 vp->v_pollinfo->vpi_events = 0; /* &= ~events ??? */ 2838 vp->v_pollinfo->vpi_revents |= events; 2839 selwakeup(&vp->v_pollinfo->vpi_selinfo); 2840 } 2841 mtx_unlock(&vp->v_pollinfo->vpi_lock); 2842} 2843 2844/* 2845 * Wake up anyone polling on vp because it is being revoked. 2846 * This depends on dead_poll() returning POLLHUP for correct 2847 * behavior. 2848 */ 2849void 2850vn_pollgone(vp) 2851 struct vnode *vp; 2852{ 2853 2854 mtx_lock(&vp->v_pollinfo->vpi_lock); 2855 VN_KNOTE(vp, NOTE_REVOKE); 2856 if (vp->v_pollinfo->vpi_events) { 2857 vp->v_pollinfo->vpi_events = 0; 2858 selwakeup(&vp->v_pollinfo->vpi_selinfo); 2859 } 2860 mtx_unlock(&vp->v_pollinfo->vpi_lock); 2861} 2862 2863 2864 2865/* 2866 * Routine to create and manage a filesystem syncer vnode. 2867 */ 2868#define sync_close ((int (*)(struct vop_close_args *))nullop) 2869static int sync_fsync(struct vop_fsync_args *); 2870static int sync_inactive(struct vop_inactive_args *); 2871static int sync_reclaim(struct vop_reclaim_args *); 2872#define sync_lock ((int (*)(struct vop_lock_args *))vop_nolock) 2873#define sync_unlock ((int (*)(struct vop_unlock_args *))vop_nounlock) 2874static int sync_print(struct vop_print_args *); 2875#define sync_islocked ((int(*)(struct vop_islocked_args *))vop_noislocked) 2876 2877static vop_t **sync_vnodeop_p; 2878static struct vnodeopv_entry_desc sync_vnodeop_entries[] = { 2879 { &vop_default_desc, (vop_t *) vop_eopnotsupp }, 2880 { &vop_close_desc, (vop_t *) sync_close }, /* close */ 2881 { &vop_fsync_desc, (vop_t *) sync_fsync }, /* fsync */ 2882 { &vop_inactive_desc, (vop_t *) sync_inactive }, /* inactive */ 2883 { &vop_reclaim_desc, (vop_t *) sync_reclaim }, /* reclaim */ 2884 { &vop_lock_desc, (vop_t *) sync_lock }, /* lock */ 2885 { &vop_unlock_desc, (vop_t *) sync_unlock }, /* unlock */ 2886 { &vop_print_desc, (vop_t *) sync_print }, /* print */ 2887 { &vop_islocked_desc, (vop_t *) sync_islocked }, /* islocked */ 2888 { NULL, NULL } 2889}; 2890static struct vnodeopv_desc sync_vnodeop_opv_desc = 2891 { &sync_vnodeop_p, sync_vnodeop_entries }; 2892 2893VNODEOP_SET(sync_vnodeop_opv_desc); 2894 2895/* 2896 * Create a new filesystem syncer vnode for the specified mount point. 2897 */ 2898int 2899vfs_allocate_syncvnode(mp) 2900 struct mount *mp; 2901{ 2902 struct vnode *vp; 2903 static long start, incr, next; 2904 int error; 2905 2906 /* Allocate a new vnode */ 2907 if ((error = getnewvnode(VT_VFS, mp, sync_vnodeop_p, &vp)) != 0) { 2908 mp->mnt_syncer = NULL; 2909 return (error); 2910 } 2911 vp->v_type = VNON; 2912 /* 2913 * Place the vnode onto the syncer worklist. We attempt to 2914 * scatter them about on the list so that they will go off 2915 * at evenly distributed times even if all the filesystems 2916 * are mounted at once. 2917 */ 2918 next += incr; 2919 if (next == 0 || next > syncer_maxdelay) { 2920 start /= 2; 2921 incr /= 2; 2922 if (start == 0) { 2923 start = syncer_maxdelay / 2; 2924 incr = syncer_maxdelay; 2925 } 2926 next = start; 2927 } 2928 vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0); 2929 mp->mnt_syncer = vp; 2930 return (0); 2931} 2932 2933/* 2934 * Do a lazy sync of the filesystem. 2935 */ 2936static int 2937sync_fsync(ap) 2938 struct vop_fsync_args /* { 2939 struct vnode *a_vp; 2940 struct ucred *a_cred; 2941 int a_waitfor; 2942 struct thread *a_td; 2943 } */ *ap; 2944{ 2945 struct vnode *syncvp = ap->a_vp; 2946 struct mount *mp = syncvp->v_mount; 2947 struct thread *td = ap->a_td; 2948 int asyncflag; 2949 2950 /* 2951 * We only need to do something if this is a lazy evaluation. 2952 */ 2953 if (ap->a_waitfor != MNT_LAZY) 2954 return (0); 2955 2956 /* 2957 * Move ourselves to the back of the sync list. 2958 */ 2959 vn_syncer_add_to_worklist(syncvp, syncdelay); 2960 2961 /* 2962 * Walk the list of vnodes pushing all that are dirty and 2963 * not already on the sync list. 2964 */ 2965 mtx_lock(&mountlist_mtx); 2966 if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { 2967 mtx_unlock(&mountlist_mtx); 2968 return (0); 2969 } 2970 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 2971 vfs_unbusy(mp, td); 2972 return (0); 2973 } 2974 asyncflag = mp->mnt_flag & MNT_ASYNC; 2975 mp->mnt_flag &= ~MNT_ASYNC; 2976 vfs_msync(mp, MNT_NOWAIT); 2977 VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td); 2978 if (asyncflag) 2979 mp->mnt_flag |= MNT_ASYNC; 2980 vn_finished_write(mp); 2981 vfs_unbusy(mp, td); 2982 return (0); 2983} 2984 2985/* 2986 * The syncer vnode is no referenced. 2987 */ 2988static int 2989sync_inactive(ap) 2990 struct vop_inactive_args /* { 2991 struct vnode *a_vp; 2992 struct thread *a_td; 2993 } */ *ap; 2994{ 2995 2996 vgone(ap->a_vp); 2997 return (0); 2998} 2999 3000/* 3001 * The syncer vnode is no longer needed and is being decommissioned. 3002 * 3003 * Modifications to the worklist must be protected at splbio(). 3004 */ 3005static int 3006sync_reclaim(ap) 3007 struct vop_reclaim_args /* { 3008 struct vnode *a_vp; 3009 } */ *ap; 3010{ 3011 struct vnode *vp = ap->a_vp; 3012 int s; 3013 3014 s = splbio(); 3015 vp->v_mount->mnt_syncer = NULL; 3016 if (vp->v_flag & VONWORKLST) { 3017 LIST_REMOVE(vp, v_synclist); 3018 vp->v_flag &= ~VONWORKLST; 3019 } 3020 splx(s); 3021 3022 return (0); 3023} 3024 3025/* 3026 * Print out a syncer vnode. 3027 */ 3028static int 3029sync_print(ap) 3030 struct vop_print_args /* { 3031 struct vnode *a_vp; 3032 } */ *ap; 3033{ 3034 struct vnode *vp = ap->a_vp; 3035 3036 printf("syncer vnode"); 3037 if (vp->v_vnlock != NULL) 3038 lockmgr_printinfo(vp->v_vnlock); 3039 printf("\n"); 3040 return (0); 3041} 3042 3043/* 3044 * extract the dev_t from a VCHR 3045 */ 3046dev_t 3047vn_todev(vp) 3048 struct vnode *vp; 3049{ 3050 if (vp->v_type != VCHR) 3051 return (NODEV); 3052 return (vp->v_rdev); 3053} 3054 3055/* 3056 * Check if vnode represents a disk device 3057 */ 3058int 3059vn_isdisk(vp, errp) 3060 struct vnode *vp; 3061 int *errp; 3062{ 3063 struct cdevsw *cdevsw; 3064 3065 if (vp->v_type != VCHR) { 3066 if (errp != NULL) 3067 *errp = ENOTBLK; 3068 return (0); 3069 } 3070 if (vp->v_rdev == NULL) { 3071 if (errp != NULL) 3072 *errp = ENXIO; 3073 return (0); 3074 } 3075 cdevsw = devsw(vp->v_rdev); 3076 if (cdevsw == NULL) { 3077 if (errp != NULL) 3078 *errp = ENXIO; 3079 return (0); 3080 } 3081 if (!(cdevsw->d_flags & D_DISK)) { 3082 if (errp != NULL) 3083 *errp = ENOTBLK; 3084 return (0); 3085 } 3086 if (errp != NULL) 3087 *errp = 0; 3088 return (1); 3089} 3090 3091/* 3092 * Free data allocated by namei(); see namei(9) for details. 3093 */ 3094void 3095NDFREE(ndp, flags) 3096 struct nameidata *ndp; 3097 const uint flags; 3098{ 3099 if (!(flags & NDF_NO_FREE_PNBUF) && 3100 (ndp->ni_cnd.cn_flags & HASBUF)) { 3101 uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf); 3102 ndp->ni_cnd.cn_flags &= ~HASBUF; 3103 } 3104 if (!(flags & NDF_NO_DVP_UNLOCK) && 3105 (ndp->ni_cnd.cn_flags & LOCKPARENT) && 3106 ndp->ni_dvp != ndp->ni_vp) 3107 VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread); 3108 if (!(flags & NDF_NO_DVP_RELE) && 3109 (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) { 3110 vrele(ndp->ni_dvp); 3111 ndp->ni_dvp = NULL; 3112 } 3113 if (!(flags & NDF_NO_VP_UNLOCK) && 3114 (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp) 3115 VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread); 3116 if (!(flags & NDF_NO_VP_RELE) && 3117 ndp->ni_vp) { 3118 vrele(ndp->ni_vp); 3119 ndp->ni_vp = NULL; 3120 } 3121 if (!(flags & NDF_NO_STARTDIR_RELE) && 3122 (ndp->ni_cnd.cn_flags & SAVESTART)) { 3123 vrele(ndp->ni_startdir); 3124 ndp->ni_startdir = NULL; 3125 } 3126} 3127 3128/* 3129 * Common file system object access control check routine. Accepts a 3130 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3131 * and optional call-by-reference privused argument allowing vaccess() 3132 * to indicate to the caller whether privilege was used to satisfy the 3133 * request. Returns 0 on success, or an errno on failure. 3134 */ 3135int 3136vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused) 3137 enum vtype type; 3138 mode_t file_mode; 3139 uid_t file_uid; 3140 gid_t file_gid; 3141 mode_t acc_mode; 3142 struct ucred *cred; 3143 int *privused; 3144{ 3145 mode_t dac_granted; 3146#ifdef CAPABILITIES 3147 mode_t cap_granted; 3148#endif 3149 3150 /* 3151 * Look for a normal, non-privileged way to access the file/directory 3152 * as requested. If it exists, go with that. 3153 */ 3154 3155 if (privused != NULL) 3156 *privused = 0; 3157 3158 dac_granted = 0; 3159 3160 /* Check the owner. */ 3161 if (cred->cr_uid == file_uid) { 3162 dac_granted |= VADMIN; 3163 if (file_mode & S_IXUSR) 3164 dac_granted |= VEXEC; 3165 if (file_mode & S_IRUSR) 3166 dac_granted |= VREAD; 3167 if (file_mode & S_IWUSR) 3168 dac_granted |= VWRITE; 3169 3170 if ((acc_mode & dac_granted) == acc_mode) 3171 return (0); 3172 3173 goto privcheck; 3174 } 3175 3176 /* Otherwise, check the groups (first match) */ 3177 if (groupmember(file_gid, cred)) { 3178 if (file_mode & S_IXGRP) 3179 dac_granted |= VEXEC; 3180 if (file_mode & S_IRGRP) 3181 dac_granted |= VREAD; 3182 if (file_mode & S_IWGRP) 3183 dac_granted |= VWRITE; 3184 3185 if ((acc_mode & dac_granted) == acc_mode) 3186 return (0); 3187 3188 goto privcheck; 3189 } 3190 3191 /* Otherwise, check everyone else. */ 3192 if (file_mode & S_IXOTH) 3193 dac_granted |= VEXEC; 3194 if (file_mode & S_IROTH) 3195 dac_granted |= VREAD; 3196 if (file_mode & S_IWOTH) 3197 dac_granted |= VWRITE; 3198 if ((acc_mode & dac_granted) == acc_mode) 3199 return (0); 3200 3201privcheck: 3202 if (!suser_cred(cred, PRISON_ROOT)) { 3203 /* XXX audit: privilege used */ 3204 if (privused != NULL) 3205 *privused = 1; 3206 return (0); 3207 } 3208 3209#ifdef CAPABILITIES 3210 /* 3211 * Build a capability mask to determine if the set of capabilities 3212 * satisfies the requirements when combined with the granted mask 3213 * from above. 3214 * For each capability, if the capability is required, bitwise 3215 * or the request type onto the cap_granted mask. 3216 */ 3217 cap_granted = 0; 3218 3219 if (type == VDIR) { 3220 /* 3221 * For directories, use CAP_DAC_READ_SEARCH to satisfy 3222 * VEXEC requests, instead of CAP_DAC_EXECUTE. 3223 */ 3224 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3225 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) 3226 cap_granted |= VEXEC; 3227 } else { 3228 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3229 !cap_check(cred, NULL, CAP_DAC_EXECUTE, PRISON_ROOT)) 3230 cap_granted |= VEXEC; 3231 } 3232 3233 if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && 3234 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) 3235 cap_granted |= VREAD; 3236 3237 if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3238 !cap_check(cred, NULL, CAP_DAC_WRITE, PRISON_ROOT)) 3239 cap_granted |= VWRITE; 3240 3241 if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3242 !cap_check(cred, NULL, CAP_FOWNER, PRISON_ROOT)) 3243 cap_granted |= VADMIN; 3244 3245 if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { 3246 /* XXX audit: privilege used */ 3247 if (privused != NULL) 3248 *privused = 1; 3249 return (0); 3250 } 3251#endif 3252 3253 return ((acc_mode & VADMIN) ? EPERM : EACCES); 3254} 3255 3256