vfs_subr.c revision 136943
1/* 2 * Copyright (c) 1989, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 4. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 35 */ 36 37/* 38 * External virtual filesystem routines 39 */ 40 41#include <sys/cdefs.h> 42__FBSDID("$FreeBSD: head/sys/kern/vfs_subr.c 136943 2004-10-25 09:14:03Z phk $"); 43 44#include "opt_ddb.h" 45#include "opt_mac.h" 46 47#include <sys/param.h> 48#include <sys/systm.h> 49#include <sys/bio.h> 50#include <sys/buf.h> 51#include <sys/conf.h> 52#include <sys/event.h> 53#include <sys/eventhandler.h> 54#include <sys/extattr.h> 55#include <sys/fcntl.h> 56#include <sys/kdb.h> 57#include <sys/kernel.h> 58#include <sys/kthread.h> 59#include <sys/mac.h> 60#include <sys/malloc.h> 61#include <sys/mount.h> 62#include <sys/namei.h> 63#include <sys/reboot.h> 64#include <sys/sleepqueue.h> 65#include <sys/stat.h> 66#include <sys/sysctl.h> 67#include <sys/syslog.h> 68#include <sys/vmmeter.h> 69#include <sys/vnode.h> 70 71#include <vm/vm.h> 72#include <vm/vm_object.h> 73#include <vm/vm_extern.h> 74#include <vm/pmap.h> 75#include <vm/vm_map.h> 76#include <vm/vm_page.h> 77#include <vm/vm_kern.h> 78#include <vm/uma.h> 79 80static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure"); 81 82static void addalias(struct vnode *vp, struct cdev *nvp_rdev); 83static void delmntque(struct vnode *vp); 84static void insmntque(struct vnode *vp, struct mount *mp); 85static void vclean(struct vnode *vp, int flags, struct thread *td); 86static void vlruvp(struct vnode *vp); 87static int flushbuflist(struct buf *blist, int flags, struct vnode *vp, 88 int slpflag, int slptimeo, int *errorp); 89static void syncer_shutdown(void *arg, int howto); 90static int vtryrecycle(struct vnode *vp); 91static void vx_lock(struct vnode *vp); 92static void vx_unlock(struct vnode *vp); 93static void vgonechrl(struct vnode *vp, struct thread *td); 94 95 96/* 97 * Number of vnodes in existence. Increased whenever getnewvnode() 98 * allocates a new vnode, never decreased. 99 */ 100static unsigned long numvnodes; 101 102SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); 103 104/* 105 * Conversion tables for conversion from vnode types to inode formats 106 * and back. 107 */ 108enum vtype iftovt_tab[16] = { 109 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 110 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 111}; 112int vttoif_tab[9] = { 113 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 114 S_IFSOCK, S_IFIFO, S_IFMT, 115}; 116 117/* 118 * List of vnodes that are ready for recycling. 119 */ 120static TAILQ_HEAD(freelst, vnode) vnode_free_list; 121 122/* 123 * Minimum number of free vnodes. If there are fewer than this free vnodes, 124 * getnewvnode() will return a newly allocated vnode. 125 */ 126static u_long wantfreevnodes = 25; 127SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 128/* Number of vnodes in the free list. */ 129static u_long freevnodes; 130SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); 131 132/* 133 * Various variables used for debugging the new implementation of 134 * reassignbuf(). 135 * XXX these are probably of (very) limited utility now. 136 */ 137static int reassignbufcalls; 138SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); 139static int nameileafonly; 140SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, ""); 141 142/* 143 * Cache for the mount type id assigned to NFS. This is used for 144 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 145 */ 146int nfs_mount_type = -1; 147 148/* To keep more than one thread at a time from running vfs_getnewfsid */ 149static struct mtx mntid_mtx; 150 151/* 152 * Lock for any access to the following: 153 * vnode_free_list 154 * numvnodes 155 * freevnodes 156 */ 157static struct mtx vnode_free_list_mtx; 158 159/* Publicly exported FS */ 160struct nfs_public nfs_pub; 161 162/* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 163static uma_zone_t vnode_zone; 164static uma_zone_t vnodepoll_zone; 165 166/* Set to 1 to print out reclaim of active vnodes */ 167int prtactive; 168 169/* 170 * The workitem queue. 171 * 172 * It is useful to delay writes of file data and filesystem metadata 173 * for tens of seconds so that quickly created and deleted files need 174 * not waste disk bandwidth being created and removed. To realize this, 175 * we append vnodes to a "workitem" queue. When running with a soft 176 * updates implementation, most pending metadata dependencies should 177 * not wait for more than a few seconds. Thus, mounted on block devices 178 * are delayed only about a half the time that file data is delayed. 179 * Similarly, directory updates are more critical, so are only delayed 180 * about a third the time that file data is delayed. Thus, there are 181 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 182 * one each second (driven off the filesystem syncer process). The 183 * syncer_delayno variable indicates the next queue that is to be processed. 184 * Items that need to be processed soon are placed in this queue: 185 * 186 * syncer_workitem_pending[syncer_delayno] 187 * 188 * A delay of fifteen seconds is done by placing the request fifteen 189 * entries later in the queue: 190 * 191 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 192 * 193 */ 194static int syncer_delayno; 195static long syncer_mask; 196LIST_HEAD(synclist, vnode); 197static struct synclist *syncer_workitem_pending; 198/* 199 * The sync_mtx protects: 200 * vp->v_synclist 201 * sync_vnode_count 202 * syncer_delayno 203 * syncer_state 204 * syncer_workitem_pending 205 * syncer_worklist_len 206 * rushjob 207 */ 208static struct mtx sync_mtx; 209 210#define SYNCER_MAXDELAY 32 211static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 212static int syncdelay = 30; /* max time to delay syncing data */ 213static int filedelay = 30; /* time to delay syncing files */ 214SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); 215static int dirdelay = 29; /* time to delay syncing directories */ 216SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); 217static int metadelay = 28; /* time to delay syncing metadata */ 218SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); 219static int rushjob; /* number of slots to run ASAP */ 220static int stat_rush_requests; /* number of times I/O speeded up */ 221SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); 222 223/* 224 * When shutting down the syncer, run it at four times normal speed. 225 */ 226#define SYNCER_SHUTDOWN_SPEEDUP 4 227static int sync_vnode_count; 228static int syncer_worklist_len; 229static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 230 syncer_state; 231 232/* 233 * Number of vnodes we want to exist at any one time. This is mostly used 234 * to size hash tables in vnode-related code. It is normally not used in 235 * getnewvnode(), as wantfreevnodes is normally nonzero.) 236 * 237 * XXX desiredvnodes is historical cruft and should not exist. 238 */ 239int desiredvnodes; 240SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 241 &desiredvnodes, 0, "Maximum number of vnodes"); 242static int minvnodes; 243SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 244 &minvnodes, 0, "Minimum number of vnodes"); 245static int vnlru_nowhere; 246SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 247 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 248 249/* Hook for calling soft updates. */ 250int (*softdep_process_worklist_hook)(struct mount *); 251 252/* 253 * Initialize the vnode management data structures. 254 */ 255#ifndef MAXVNODES_MAX 256#define MAXVNODES_MAX 100000 257#endif 258static void 259vntblinit(void *dummy __unused) 260{ 261 262 /* 263 * Desiredvnodes is a function of the physical memory size and 264 * the kernel's heap size. Specifically, desiredvnodes scales 265 * in proportion to the physical memory size until two fifths 266 * of the kernel's heap size is consumed by vnodes and vm 267 * objects. 268 */ 269 desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size / 270 (5 * (sizeof(struct vm_object) + sizeof(struct vnode)))); 271 if (desiredvnodes > MAXVNODES_MAX) { 272 if (bootverbose) 273 printf("Reducing kern.maxvnodes %d -> %d\n", 274 desiredvnodes, MAXVNODES_MAX); 275 desiredvnodes = MAXVNODES_MAX; 276 } 277 minvnodes = desiredvnodes / 4; 278 mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF); 279 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 280 TAILQ_INIT(&vnode_free_list); 281 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 282 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 283 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 284 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 285 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 286 /* 287 * Initialize the filesystem syncer. 288 */ 289 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 290 &syncer_mask); 291 syncer_maxdelay = syncer_mask + 1; 292 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 293} 294SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) 295 296 297/* 298 * Mark a mount point as busy. Used to synchronize access and to delay 299 * unmounting. Interlock is not released on failure. 300 */ 301int 302vfs_busy(mp, flags, interlkp, td) 303 struct mount *mp; 304 int flags; 305 struct mtx *interlkp; 306 struct thread *td; 307{ 308 int lkflags; 309 310 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 311 if (flags & LK_NOWAIT) 312 return (ENOENT); 313 mp->mnt_kern_flag |= MNTK_MWAIT; 314 /* 315 * Since all busy locks are shared except the exclusive 316 * lock granted when unmounting, the only place that a 317 * wakeup needs to be done is at the release of the 318 * exclusive lock at the end of dounmount. 319 */ 320 msleep(mp, interlkp, PVFS, "vfs_busy", 0); 321 return (ENOENT); 322 } 323 lkflags = LK_SHARED | LK_NOPAUSE; 324 if (interlkp) 325 lkflags |= LK_INTERLOCK; 326 if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td)) 327 panic("vfs_busy: unexpected lock failure"); 328 return (0); 329} 330 331/* 332 * Free a busy filesystem. 333 */ 334void 335vfs_unbusy(mp, td) 336 struct mount *mp; 337 struct thread *td; 338{ 339 340 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); 341} 342 343/* 344 * Lookup a mount point by filesystem identifier. 345 */ 346struct mount * 347vfs_getvfs(fsid) 348 fsid_t *fsid; 349{ 350 register struct mount *mp; 351 352 mtx_lock(&mountlist_mtx); 353 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 354 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 355 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 356 mtx_unlock(&mountlist_mtx); 357 return (mp); 358 } 359 } 360 mtx_unlock(&mountlist_mtx); 361 return ((struct mount *) 0); 362} 363 364/* 365 * Check if a user can access priveledged mount options. 366 */ 367int 368vfs_suser(struct mount *mp, struct thread *td) 369{ 370 int error; 371 372 if ((mp->mnt_flag & MNT_USER) == 0 || 373 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 374 if ((error = suser(td)) != 0) 375 return (error); 376 } 377 return (0); 378} 379 380/* 381 * Get a new unique fsid. Try to make its val[0] unique, since this value 382 * will be used to create fake device numbers for stat(). Also try (but 383 * not so hard) make its val[0] unique mod 2^16, since some emulators only 384 * support 16-bit device numbers. We end up with unique val[0]'s for the 385 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 386 * 387 * Keep in mind that several mounts may be running in parallel. Starting 388 * the search one past where the previous search terminated is both a 389 * micro-optimization and a defense against returning the same fsid to 390 * different mounts. 391 */ 392void 393vfs_getnewfsid(mp) 394 struct mount *mp; 395{ 396 static u_int16_t mntid_base; 397 fsid_t tfsid; 398 int mtype; 399 400 mtx_lock(&mntid_mtx); 401 mtype = mp->mnt_vfc->vfc_typenum; 402 tfsid.val[1] = mtype; 403 mtype = (mtype & 0xFF) << 24; 404 for (;;) { 405 tfsid.val[0] = makedev(255, 406 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 407 mntid_base++; 408 if (vfs_getvfs(&tfsid) == NULL) 409 break; 410 } 411 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 412 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 413 mtx_unlock(&mntid_mtx); 414} 415 416/* 417 * Knob to control the precision of file timestamps: 418 * 419 * 0 = seconds only; nanoseconds zeroed. 420 * 1 = seconds and nanoseconds, accurate within 1/HZ. 421 * 2 = seconds and nanoseconds, truncated to microseconds. 422 * >=3 = seconds and nanoseconds, maximum precision. 423 */ 424enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 425 426static int timestamp_precision = TSP_SEC; 427SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 428 ×tamp_precision, 0, ""); 429 430/* 431 * Get a current timestamp. 432 */ 433void 434vfs_timestamp(tsp) 435 struct timespec *tsp; 436{ 437 struct timeval tv; 438 439 switch (timestamp_precision) { 440 case TSP_SEC: 441 tsp->tv_sec = time_second; 442 tsp->tv_nsec = 0; 443 break; 444 case TSP_HZ: 445 getnanotime(tsp); 446 break; 447 case TSP_USEC: 448 microtime(&tv); 449 TIMEVAL_TO_TIMESPEC(&tv, tsp); 450 break; 451 case TSP_NSEC: 452 default: 453 nanotime(tsp); 454 break; 455 } 456} 457 458/* 459 * Set vnode attributes to VNOVAL 460 */ 461void 462vattr_null(vap) 463 register struct vattr *vap; 464{ 465 466 vap->va_type = VNON; 467 vap->va_size = VNOVAL; 468 vap->va_bytes = VNOVAL; 469 vap->va_mode = VNOVAL; 470 vap->va_nlink = VNOVAL; 471 vap->va_uid = VNOVAL; 472 vap->va_gid = VNOVAL; 473 vap->va_fsid = VNOVAL; 474 vap->va_fileid = VNOVAL; 475 vap->va_blocksize = VNOVAL; 476 vap->va_rdev = VNOVAL; 477 vap->va_atime.tv_sec = VNOVAL; 478 vap->va_atime.tv_nsec = VNOVAL; 479 vap->va_mtime.tv_sec = VNOVAL; 480 vap->va_mtime.tv_nsec = VNOVAL; 481 vap->va_ctime.tv_sec = VNOVAL; 482 vap->va_ctime.tv_nsec = VNOVAL; 483 vap->va_birthtime.tv_sec = VNOVAL; 484 vap->va_birthtime.tv_nsec = VNOVAL; 485 vap->va_flags = VNOVAL; 486 vap->va_gen = VNOVAL; 487 vap->va_vaflags = 0; 488} 489 490/* 491 * This routine is called when we have too many vnodes. It attempts 492 * to free <count> vnodes and will potentially free vnodes that still 493 * have VM backing store (VM backing store is typically the cause 494 * of a vnode blowout so we want to do this). Therefore, this operation 495 * is not considered cheap. 496 * 497 * A number of conditions may prevent a vnode from being reclaimed. 498 * the buffer cache may have references on the vnode, a directory 499 * vnode may still have references due to the namei cache representing 500 * underlying files, or the vnode may be in active use. It is not 501 * desireable to reuse such vnodes. These conditions may cause the 502 * number of vnodes to reach some minimum value regardless of what 503 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 504 */ 505static int 506vlrureclaim(struct mount *mp) 507{ 508 struct vnode *vp; 509 int done; 510 int trigger; 511 int usevnodes; 512 int count; 513 514 /* 515 * Calculate the trigger point, don't allow user 516 * screwups to blow us up. This prevents us from 517 * recycling vnodes with lots of resident pages. We 518 * aren't trying to free memory, we are trying to 519 * free vnodes. 520 */ 521 usevnodes = desiredvnodes; 522 if (usevnodes <= 0) 523 usevnodes = 1; 524 trigger = cnt.v_page_count * 2 / usevnodes; 525 526 done = 0; 527 MNT_ILOCK(mp); 528 count = mp->mnt_nvnodelistsize / 10 + 1; 529 while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) { 530 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 531 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 532 533 if (vp->v_type != VNON && 534 vp->v_type != VBAD && 535 VI_TRYLOCK(vp)) { 536 if (VMIGHTFREE(vp) && /* critical path opt */ 537 (vp->v_object == NULL || 538 vp->v_object->resident_page_count < trigger)) { 539 MNT_IUNLOCK(mp); 540 vgonel(vp, curthread); 541 done++; 542 MNT_ILOCK(mp); 543 } else 544 VI_UNLOCK(vp); 545 } 546 --count; 547 } 548 MNT_IUNLOCK(mp); 549 return done; 550} 551 552/* 553 * Attempt to recycle vnodes in a context that is always safe to block. 554 * Calling vlrurecycle() from the bowels of filesystem code has some 555 * interesting deadlock problems. 556 */ 557static struct proc *vnlruproc; 558static int vnlruproc_sig; 559 560static void 561vnlru_proc(void) 562{ 563 struct mount *mp, *nmp; 564 int done; 565 struct proc *p = vnlruproc; 566 struct thread *td = FIRST_THREAD_IN_PROC(p); 567 568 mtx_lock(&Giant); 569 570 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 571 SHUTDOWN_PRI_FIRST); 572 573 for (;;) { 574 kthread_suspend_check(p); 575 mtx_lock(&vnode_free_list_mtx); 576 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { 577 mtx_unlock(&vnode_free_list_mtx); 578 vnlruproc_sig = 0; 579 wakeup(&vnlruproc_sig); 580 tsleep(vnlruproc, PVFS, "vlruwt", hz); 581 continue; 582 } 583 mtx_unlock(&vnode_free_list_mtx); 584 done = 0; 585 mtx_lock(&mountlist_mtx); 586 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 587 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 588 nmp = TAILQ_NEXT(mp, mnt_list); 589 continue; 590 } 591 done += vlrureclaim(mp); 592 mtx_lock(&mountlist_mtx); 593 nmp = TAILQ_NEXT(mp, mnt_list); 594 vfs_unbusy(mp, td); 595 } 596 mtx_unlock(&mountlist_mtx); 597 if (done == 0) { 598#if 0 599 /* These messages are temporary debugging aids */ 600 if (vnlru_nowhere < 5) 601 printf("vnlru process getting nowhere..\n"); 602 else if (vnlru_nowhere == 5) 603 printf("vnlru process messages stopped.\n"); 604#endif 605 vnlru_nowhere++; 606 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 607 } 608 } 609} 610 611static struct kproc_desc vnlru_kp = { 612 "vnlru", 613 vnlru_proc, 614 &vnlruproc 615}; 616SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 617 618 619/* 620 * Routines having to do with the management of the vnode table. 621 */ 622 623/* 624 * Check to see if a free vnode can be recycled. If it can, 625 * recycle it and return it with the vnode interlock held. 626 */ 627static int 628vtryrecycle(struct vnode *vp) 629{ 630 struct thread *td = curthread; 631 vm_object_t object; 632 struct mount *vnmp; 633 int error; 634 635 /* Don't recycle if we can't get the interlock */ 636 if (!VI_TRYLOCK(vp)) 637 return (EWOULDBLOCK); 638 /* 639 * This vnode may found and locked via some other list, if so we 640 * can't recycle it yet. 641 */ 642 if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 643 return (EWOULDBLOCK); 644 /* 645 * Don't recycle if its filesystem is being suspended. 646 */ 647 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 648 VOP_UNLOCK(vp, 0, td); 649 return (EBUSY); 650 } 651 652 /* 653 * Don't recycle if we still have cached pages. 654 */ 655 if (VOP_GETVOBJECT(vp, &object) == 0) { 656 VM_OBJECT_LOCK(object); 657 if (object->resident_page_count || 658 object->ref_count) { 659 VM_OBJECT_UNLOCK(object); 660 error = EBUSY; 661 goto done; 662 } 663 VM_OBJECT_UNLOCK(object); 664 } 665 if (LIST_FIRST(&vp->v_cache_src)) { 666 /* 667 * note: nameileafonly sysctl is temporary, 668 * for debugging only, and will eventually be 669 * removed. 670 */ 671 if (nameileafonly > 0) { 672 /* 673 * Do not reuse namei-cached directory 674 * vnodes that have cached 675 * subdirectories. 676 */ 677 if (cache_leaf_test(vp) < 0) { 678 error = EISDIR; 679 goto done; 680 } 681 } else if (nameileafonly < 0 || 682 vmiodirenable == 0) { 683 /* 684 * Do not reuse namei-cached directory 685 * vnodes if nameileafonly is -1 or 686 * if VMIO backing for directories is 687 * turned off (otherwise we reuse them 688 * too quickly). 689 */ 690 error = EBUSY; 691 goto done; 692 } 693 } 694 /* 695 * If we got this far, we need to acquire the interlock and see if 696 * anyone picked up this vnode from another list. If not, we will 697 * mark it with XLOCK via vgonel() so that anyone who does find it 698 * will skip over it. 699 */ 700 VI_LOCK(vp); 701 if (VSHOULDBUSY(vp) && (vp->v_iflag & VI_XLOCK) == 0) { 702 VI_UNLOCK(vp); 703 error = EBUSY; 704 goto done; 705 } 706 mtx_lock(&vnode_free_list_mtx); 707 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 708 vp->v_iflag &= ~VI_FREE; 709 mtx_unlock(&vnode_free_list_mtx); 710 vp->v_iflag |= VI_DOOMED; 711 if ((vp->v_type != VBAD) || (vp->v_data != NULL)) { 712 VOP_UNLOCK(vp, 0, td); 713 vgonel(vp, td); 714 VI_LOCK(vp); 715 } else 716 VOP_UNLOCK(vp, 0, td); 717 vn_finished_write(vnmp); 718 return (0); 719done: 720 VOP_UNLOCK(vp, 0, td); 721 vn_finished_write(vnmp); 722 return (error); 723} 724 725/* 726 * Return the next vnode from the free list. 727 */ 728int 729getnewvnode(tag, mp, vops, vpp) 730 const char *tag; 731 struct mount *mp; 732 vop_t **vops; 733 struct vnode **vpp; 734{ 735 struct vnode *vp = NULL; 736 struct vpollinfo *pollinfo = NULL; 737 struct bufobj *bo; 738 739 mtx_lock(&vnode_free_list_mtx); 740 741 /* 742 * Try to reuse vnodes if we hit the max. This situation only 743 * occurs in certain large-memory (2G+) situations. We cannot 744 * attempt to directly reclaim vnodes due to nasty recursion 745 * problems. 746 */ 747 while (numvnodes - freevnodes > desiredvnodes) { 748 if (vnlruproc_sig == 0) { 749 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 750 wakeup(vnlruproc); 751 } 752 mtx_unlock(&vnode_free_list_mtx); 753 tsleep(&vnlruproc_sig, PVFS, "vlruwk", hz); 754 mtx_lock(&vnode_free_list_mtx); 755 } 756 757 /* 758 * Attempt to reuse a vnode already on the free list, allocating 759 * a new vnode if we can't find one or if we have not reached a 760 * good minimum for good LRU performance. 761 */ 762 763 if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) { 764 int error; 765 int count; 766 767 for (count = 0; count < freevnodes; count++) { 768 vp = TAILQ_FIRST(&vnode_free_list); 769 770 KASSERT(vp->v_usecount == 0 && 771 (vp->v_iflag & VI_DOINGINACT) == 0, 772 ("getnewvnode: free vnode isn't")); 773 774 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 775 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 776 mtx_unlock(&vnode_free_list_mtx); 777 error = vtryrecycle(vp); 778 mtx_lock(&vnode_free_list_mtx); 779 if (error == 0) 780 break; 781 vp = NULL; 782 } 783 } 784 if (vp) { 785 freevnodes--; 786 bo = &vp->v_bufobj; 787 mtx_unlock(&vnode_free_list_mtx); 788 789#ifdef INVARIANTS 790 { 791 if (vp->v_data) 792 printf("cleaned vnode isn't, " 793 "address %p, inode %p\n", 794 vp, vp->v_data); 795 if (bo->bo_numoutput) 796 panic("Clean vnode has pending I/O's"); 797 if (vp->v_writecount != 0) 798 panic("Non-zero write count"); 799 } 800#endif 801 if ((pollinfo = vp->v_pollinfo) != NULL) { 802 /* 803 * To avoid lock order reversals, the call to 804 * uma_zfree() must be delayed until the vnode 805 * interlock is released. 806 */ 807 vp->v_pollinfo = NULL; 808 } 809#ifdef MAC 810 mac_destroy_vnode(vp); 811#endif 812 vp->v_iflag = 0; 813 vp->v_vflag = 0; 814 vp->v_lastw = 0; 815 vp->v_lasta = 0; 816 vp->v_cstart = 0; 817 vp->v_clen = 0; 818 vp->v_socket = 0; 819 lockdestroy(vp->v_vnlock); 820 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 821 KASSERT(bo->bo_clean.bv_cnt == 0, ("cleanbufcnt not 0")); 822 KASSERT(bo->bo_clean.bv_root == NULL, ("cleanblkroot not NULL")); 823 KASSERT(bo->bo_dirty.bv_cnt == 0, ("dirtybufcnt not 0")); 824 KASSERT(bo->bo_dirty.bv_root == NULL, ("dirtyblkroot not NULL")); 825 } else { 826 numvnodes++; 827 mtx_unlock(&vnode_free_list_mtx); 828 829 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); 830 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 831 VI_LOCK(vp); 832 vp->v_dd = vp; 833 bo = &vp->v_bufobj; 834 bo->bo_mtx = &vp->v_interlock; 835 vp->v_vnlock = &vp->v_lock; 836 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 837 cache_purge(vp); /* Sets up v_id. */ 838 LIST_INIT(&vp->v_cache_src); 839 TAILQ_INIT(&vp->v_cache_dst); 840 } 841 842 TAILQ_INIT(&bo->bo_clean.bv_hd); 843 TAILQ_INIT(&bo->bo_dirty.bv_hd); 844 bo->bo_ops = &buf_ops_bio; 845 vp->v_type = VNON; 846 vp->v_tag = tag; 847 vp->v_op = vops; 848 *vpp = vp; 849 vp->v_usecount = 1; 850 vp->v_data = 0; 851 vp->v_cachedid = -1; 852 VI_UNLOCK(vp); 853 if (pollinfo != NULL) { 854 knlist_destroy(&pollinfo->vpi_selinfo.si_note); 855 mtx_destroy(&pollinfo->vpi_lock); 856 uma_zfree(vnodepoll_zone, pollinfo); 857 } 858#ifdef MAC 859 mac_init_vnode(vp); 860 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 861 mac_associate_vnode_singlelabel(mp, vp); 862 else if (mp == NULL) 863 printf("NULL mp in getnewvnode()\n"); 864#endif 865 delmntque(vp); 866 if (mp != NULL) 867 insmntque(vp, mp); 868 869 return (0); 870} 871 872/* 873 * Delete from old mount point vnode list, if on one. 874 */ 875static void 876delmntque(struct vnode *vp) 877{ 878 struct mount *mp; 879 880 if (vp->v_mount == NULL) 881 return; 882 mp = vp->v_mount; 883 MNT_ILOCK(mp); 884 vp->v_mount = NULL; 885 KASSERT(mp->mnt_nvnodelistsize > 0, 886 ("bad mount point vnode list size")); 887 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 888 mp->mnt_nvnodelistsize--; 889 MNT_IUNLOCK(mp); 890} 891 892/* 893 * Insert into list of vnodes for the new mount point, if available. 894 */ 895static void 896insmntque(struct vnode *vp, struct mount *mp) 897{ 898 899 vp->v_mount = mp; 900 KASSERT(mp != NULL, ("Don't call insmntque(foo, NULL)")); 901 MNT_ILOCK(vp->v_mount); 902 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 903 mp->mnt_nvnodelistsize++; 904 MNT_IUNLOCK(vp->v_mount); 905} 906 907/* 908 * Flush out and invalidate all buffers associated with a vnode. 909 * Called with the underlying object locked. 910 */ 911int 912vinvalbuf(vp, flags, cred, td, slpflag, slptimeo) 913 struct vnode *vp; 914 int flags; 915 struct ucred *cred; 916 struct thread *td; 917 int slpflag, slptimeo; 918{ 919 struct buf *blist; 920 int error; 921 vm_object_t object; 922 struct bufobj *bo; 923 924 GIANT_REQUIRED; 925 926 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 927 928 VI_LOCK(vp); 929 bo = &vp->v_bufobj; 930 if (flags & V_SAVE) { 931 error = bufobj_wwait(bo, slpflag, slptimeo); 932 if (error) { 933 VI_UNLOCK(vp); 934 return (error); 935 } 936 if (bo->bo_dirty.bv_cnt > 0) { 937 VI_UNLOCK(vp); 938 if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0) 939 return (error); 940 /* 941 * XXX We could save a lock/unlock if this was only 942 * enabled under INVARIANTS 943 */ 944 VI_LOCK(vp); 945 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 946 panic("vinvalbuf: dirty bufs"); 947 } 948 } 949 /* 950 * If you alter this loop please notice that interlock is dropped and 951 * reacquired in flushbuflist. Special care is needed to ensure that 952 * no race conditions occur from this. 953 */ 954 for (error = 0;;) { 955 blist = TAILQ_FIRST(&vp->v_bufobj.bo_clean.bv_hd); 956 if (blist != NULL && 957 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 958 if (error) 959 break; 960 continue; 961 } 962 blist = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd); 963 if (blist != NULL && 964 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 965 if (error) 966 break; 967 continue; 968 } 969 break; 970 } 971 if (error) { 972 VI_UNLOCK(vp); 973 return (error); 974 } 975 976 /* 977 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 978 * have write I/O in-progress but if there is a VM object then the 979 * VM object can also have read-I/O in-progress. 980 */ 981 do { 982 bufobj_wwait(bo, 0, 0); 983 VI_UNLOCK(vp); 984 if (VOP_GETVOBJECT(vp, &object) == 0) { 985 VM_OBJECT_LOCK(object); 986 vm_object_pip_wait(object, "vnvlbx"); 987 VM_OBJECT_UNLOCK(object); 988 } 989 VI_LOCK(vp); 990 } while (bo->bo_numoutput > 0); 991 VI_UNLOCK(vp); 992 993 /* 994 * Destroy the copy in the VM cache, too. 995 */ 996 if (VOP_GETVOBJECT(vp, &object) == 0) { 997 VM_OBJECT_LOCK(object); 998 vm_object_page_remove(object, 0, 0, 999 (flags & V_SAVE) ? TRUE : FALSE); 1000 VM_OBJECT_UNLOCK(object); 1001 } 1002 1003#ifdef INVARIANTS 1004 VI_LOCK(vp); 1005 if ((flags & (V_ALT | V_NORMAL)) == 0 && 1006 (vp->v_bufobj.bo_dirty.bv_cnt > 0 || 1007 vp->v_bufobj.bo_clean.bv_cnt > 0)) 1008 panic("vinvalbuf: flush failed"); 1009 VI_UNLOCK(vp); 1010#endif 1011 return (0); 1012} 1013 1014/* 1015 * Flush out buffers on the specified list. 1016 * 1017 */ 1018static int 1019flushbuflist(blist, flags, vp, slpflag, slptimeo, errorp) 1020 struct buf *blist; 1021 int flags; 1022 struct vnode *vp; 1023 int slpflag, slptimeo; 1024 int *errorp; 1025{ 1026 struct buf *bp, *nbp; 1027 int found, error; 1028 1029 ASSERT_VI_LOCKED(vp, "flushbuflist"); 1030 1031 for (found = 0, bp = blist; bp; bp = nbp) { 1032 nbp = TAILQ_NEXT(bp, b_bobufs); 1033 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1034 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1035 continue; 1036 } 1037 found += 1; 1038 error = BUF_TIMELOCK(bp, 1039 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp), 1040 "flushbuf", slpflag, slptimeo); 1041 if (error) { 1042 if (error != ENOLCK) 1043 *errorp = error; 1044 goto done; 1045 } 1046 /* 1047 * XXX Since there are no node locks for NFS, I 1048 * believe there is a slight chance that a delayed 1049 * write will occur while sleeping just above, so 1050 * check for it. Note that vfs_bio_awrite expects 1051 * buffers to reside on a queue, while bwrite and 1052 * brelse do not. 1053 */ 1054 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1055 (flags & V_SAVE)) { 1056 1057 if (bp->b_vp == vp) { 1058 if (bp->b_flags & B_CLUSTEROK) { 1059 vfs_bio_awrite(bp); 1060 } else { 1061 bremfree(bp); 1062 bp->b_flags |= B_ASYNC; 1063 bwrite(bp); 1064 } 1065 } else { 1066 bremfree(bp); 1067 (void) bwrite(bp); 1068 } 1069 goto done; 1070 } 1071 bremfree(bp); 1072 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 1073 bp->b_flags &= ~B_ASYNC; 1074 brelse(bp); 1075 VI_LOCK(vp); 1076 } 1077 return (found); 1078done: 1079 VI_LOCK(vp); 1080 return (found); 1081} 1082 1083/* 1084 * Truncate a file's buffer and pages to a specified length. This 1085 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1086 * sync activity. 1087 */ 1088int 1089vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, off_t length, int blksize) 1090{ 1091 struct buf *bp, *nbp; 1092 int anyfreed; 1093 int trunclbn; 1094 struct bufobj *bo; 1095 1096 /* 1097 * Round up to the *next* lbn. 1098 */ 1099 trunclbn = (length + blksize - 1) / blksize; 1100 1101 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1102restart: 1103 VI_LOCK(vp); 1104 bo = &vp->v_bufobj; 1105 anyfreed = 1; 1106 for (;anyfreed;) { 1107 anyfreed = 0; 1108 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 1109 if (bp->b_lblkno < trunclbn) 1110 continue; 1111 if (BUF_LOCK(bp, 1112 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1113 VI_MTX(vp)) == ENOLCK) 1114 goto restart; 1115 1116 bremfree(bp); 1117 bp->b_flags |= (B_INVAL | B_RELBUF); 1118 bp->b_flags &= ~B_ASYNC; 1119 brelse(bp); 1120 anyfreed = 1; 1121 1122 if (nbp != NULL && 1123 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1124 (nbp->b_vp != vp) || 1125 (nbp->b_flags & B_DELWRI))) { 1126 goto restart; 1127 } 1128 VI_LOCK(vp); 1129 } 1130 1131 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1132 if (bp->b_lblkno < trunclbn) 1133 continue; 1134 if (BUF_LOCK(bp, 1135 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1136 VI_MTX(vp)) == ENOLCK) 1137 goto restart; 1138 bremfree(bp); 1139 bp->b_flags |= (B_INVAL | B_RELBUF); 1140 bp->b_flags &= ~B_ASYNC; 1141 brelse(bp); 1142 anyfreed = 1; 1143 if (nbp != NULL && 1144 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1145 (nbp->b_vp != vp) || 1146 (nbp->b_flags & B_DELWRI) == 0)) { 1147 goto restart; 1148 } 1149 VI_LOCK(vp); 1150 } 1151 } 1152 1153 if (length > 0) { 1154restartsync: 1155 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1156 if (bp->b_lblkno > 0) 1157 continue; 1158 /* 1159 * Since we hold the vnode lock this should only 1160 * fail if we're racing with the buf daemon. 1161 */ 1162 if (BUF_LOCK(bp, 1163 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1164 VI_MTX(vp)) == ENOLCK) { 1165 goto restart; 1166 } 1167 KASSERT((bp->b_flags & B_DELWRI), 1168 ("buf(%p) on dirty queue without DELWRI", bp)); 1169 1170 bremfree(bp); 1171 bawrite(bp); 1172 VI_LOCK(vp); 1173 goto restartsync; 1174 } 1175 } 1176 1177 bufobj_wwait(bo, 0, 0); 1178 VI_UNLOCK(vp); 1179 vnode_pager_setsize(vp, length); 1180 1181 return (0); 1182} 1183 1184/* 1185 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1186 * a vnode. 1187 * 1188 * NOTE: We have to deal with the special case of a background bitmap 1189 * buffer, a situation where two buffers will have the same logical 1190 * block offset. We want (1) only the foreground buffer to be accessed 1191 * in a lookup and (2) must differentiate between the foreground and 1192 * background buffer in the splay tree algorithm because the splay 1193 * tree cannot normally handle multiple entities with the same 'index'. 1194 * We accomplish this by adding differentiating flags to the splay tree's 1195 * numerical domain. 1196 */ 1197static 1198struct buf * 1199buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1200{ 1201 struct buf dummy; 1202 struct buf *lefttreemax, *righttreemin, *y; 1203 1204 if (root == NULL) 1205 return (NULL); 1206 lefttreemax = righttreemin = &dummy; 1207 for (;;) { 1208 if (lblkno < root->b_lblkno || 1209 (lblkno == root->b_lblkno && 1210 (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1211 if ((y = root->b_left) == NULL) 1212 break; 1213 if (lblkno < y->b_lblkno) { 1214 /* Rotate right. */ 1215 root->b_left = y->b_right; 1216 y->b_right = root; 1217 root = y; 1218 if ((y = root->b_left) == NULL) 1219 break; 1220 } 1221 /* Link into the new root's right tree. */ 1222 righttreemin->b_left = root; 1223 righttreemin = root; 1224 } else if (lblkno > root->b_lblkno || 1225 (lblkno == root->b_lblkno && 1226 (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { 1227 if ((y = root->b_right) == NULL) 1228 break; 1229 if (lblkno > y->b_lblkno) { 1230 /* Rotate left. */ 1231 root->b_right = y->b_left; 1232 y->b_left = root; 1233 root = y; 1234 if ((y = root->b_right) == NULL) 1235 break; 1236 } 1237 /* Link into the new root's left tree. */ 1238 lefttreemax->b_right = root; 1239 lefttreemax = root; 1240 } else { 1241 break; 1242 } 1243 root = y; 1244 } 1245 /* Assemble the new root. */ 1246 lefttreemax->b_right = root->b_left; 1247 righttreemin->b_left = root->b_right; 1248 root->b_left = dummy.b_right; 1249 root->b_right = dummy.b_left; 1250 return (root); 1251} 1252 1253static void 1254buf_vlist_remove(struct buf *bp) 1255{ 1256 struct buf *root; 1257 struct bufv *bv; 1258 1259 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1260 ASSERT_BO_LOCKED(bp->b_bufobj); 1261 if (bp->b_xflags & BX_VNDIRTY) 1262 bv = &bp->b_bufobj->bo_dirty; 1263 else 1264 bv = &bp->b_bufobj->bo_clean; 1265 if (bp != bv->bv_root) { 1266 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1267 KASSERT(root == bp, ("splay lookup failed in remove")); 1268 } 1269 if (bp->b_left == NULL) { 1270 root = bp->b_right; 1271 } else { 1272 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1273 root->b_right = bp->b_right; 1274 } 1275 bv->bv_root = root; 1276 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 1277 bv->bv_cnt--; 1278 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1279} 1280 1281/* 1282 * Add the buffer to the sorted clean or dirty block list using a 1283 * splay tree algorithm. 1284 * 1285 * NOTE: xflags is passed as a constant, optimizing this inline function! 1286 */ 1287static void 1288buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 1289{ 1290 struct buf *root; 1291 struct bufv *bv; 1292 1293 ASSERT_BO_LOCKED(bo); 1294 bp->b_xflags |= xflags; 1295 if (xflags & BX_VNDIRTY) 1296 bv = &bo->bo_dirty; 1297 else 1298 bv = &bo->bo_clean; 1299 1300 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1301 if (root == NULL) { 1302 bp->b_left = NULL; 1303 bp->b_right = NULL; 1304 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 1305 } else if (bp->b_lblkno < root->b_lblkno || 1306 (bp->b_lblkno == root->b_lblkno && 1307 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1308 bp->b_left = root->b_left; 1309 bp->b_right = root; 1310 root->b_left = NULL; 1311 TAILQ_INSERT_BEFORE(root, bp, b_bobufs); 1312 } else { 1313 bp->b_right = root->b_right; 1314 bp->b_left = root; 1315 root->b_right = NULL; 1316 TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs); 1317 } 1318 bv->bv_cnt++; 1319 bv->bv_root = bp; 1320} 1321 1322/* 1323 * Lookup a buffer using the splay tree. Note that we specifically avoid 1324 * shadow buffers used in background bitmap writes. 1325 * 1326 * This code isn't quite efficient as it could be because we are maintaining 1327 * two sorted lists and do not know which list the block resides in. 1328 * 1329 * During a "make buildworld" the desired buffer is found at one of 1330 * the roots more than 60% of the time. Thus, checking both roots 1331 * before performing either splay eliminates unnecessary splays on the 1332 * first tree splayed. 1333 */ 1334struct buf * 1335gbincore(struct bufobj *bo, daddr_t lblkno) 1336{ 1337 struct buf *bp; 1338 1339 GIANT_REQUIRED; 1340 1341 ASSERT_BO_LOCKED(bo); 1342 if ((bp = bo->bo_clean.bv_root) != NULL && 1343 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1344 return (bp); 1345 if ((bp = bo->bo_dirty.bv_root) != NULL && 1346 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1347 return (bp); 1348 if ((bp = bo->bo_clean.bv_root) != NULL) { 1349 bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp); 1350 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1351 return (bp); 1352 } 1353 if ((bp = bo->bo_dirty.bv_root) != NULL) { 1354 bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp); 1355 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1356 return (bp); 1357 } 1358 return (NULL); 1359} 1360 1361/* 1362 * Associate a buffer with a vnode. 1363 */ 1364void 1365bgetvp(struct vnode *vp, struct buf *bp) 1366{ 1367 1368 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 1369 1370 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1371 ("bgetvp: bp already attached! %p", bp)); 1372 1373 ASSERT_VI_LOCKED(vp, "bgetvp"); 1374 vholdl(vp); 1375 bp->b_vp = vp; 1376 bp->b_bufobj = &vp->v_bufobj; 1377 bp->b_dev = vn_todev(vp); 1378 /* 1379 * Insert onto list for new vnode. 1380 */ 1381 buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN); 1382} 1383 1384/* 1385 * Disassociate a buffer from a vnode. 1386 */ 1387void 1388brelvp(struct buf *bp) 1389{ 1390 struct bufobj *bo; 1391 struct vnode *vp; 1392 1393 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1394 1395 /* 1396 * Delete from old vnode list, if on one. 1397 */ 1398 vp = bp->b_vp; 1399 VI_LOCK(vp); 1400 bo = bp->b_bufobj; 1401 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1402 buf_vlist_remove(bp); 1403 if ((vp->v_iflag & VI_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1404 vp->v_iflag &= ~VI_ONWORKLST; 1405 mtx_lock(&sync_mtx); 1406 LIST_REMOVE(vp, v_synclist); 1407 syncer_worklist_len--; 1408 mtx_unlock(&sync_mtx); 1409 } 1410 vdropl(vp); 1411 bp->b_vp = NULL; 1412 bp->b_bufobj = NULL; 1413 VI_UNLOCK(vp); 1414} 1415 1416/* 1417 * Add an item to the syncer work queue. 1418 */ 1419static void 1420vn_syncer_add_to_worklist(struct vnode *vp, int delay) 1421{ 1422 int slot; 1423 1424 ASSERT_VI_LOCKED(vp, "vn_syncer_add_to_worklist"); 1425 1426 mtx_lock(&sync_mtx); 1427 if (vp->v_iflag & VI_ONWORKLST) 1428 LIST_REMOVE(vp, v_synclist); 1429 else { 1430 vp->v_iflag |= VI_ONWORKLST; 1431 syncer_worklist_len++; 1432 } 1433 1434 if (delay > syncer_maxdelay - 2) 1435 delay = syncer_maxdelay - 2; 1436 slot = (syncer_delayno + delay) & syncer_mask; 1437 1438 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist); 1439 mtx_unlock(&sync_mtx); 1440} 1441 1442static int 1443sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 1444{ 1445 int error, len; 1446 1447 mtx_lock(&sync_mtx); 1448 len = syncer_worklist_len - sync_vnode_count; 1449 mtx_unlock(&sync_mtx); 1450 error = SYSCTL_OUT(req, &len, sizeof(len)); 1451 return (error); 1452} 1453 1454SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0, 1455 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 1456 1457struct proc *updateproc; 1458static void sched_sync(void); 1459static struct kproc_desc up_kp = { 1460 "syncer", 1461 sched_sync, 1462 &updateproc 1463}; 1464SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 1465 1466/* 1467 * System filesystem synchronizer daemon. 1468 */ 1469static void 1470sched_sync(void) 1471{ 1472 struct synclist *next; 1473 struct synclist *slp; 1474 struct vnode *vp; 1475 struct mount *mp; 1476 long starttime; 1477 struct thread *td = FIRST_THREAD_IN_PROC(updateproc); 1478 static int dummychan; 1479 int last_work_seen; 1480 int net_worklist_len; 1481 int syncer_final_iter; 1482 int first_printf; 1483 1484 mtx_lock(&Giant); 1485 last_work_seen = 0; 1486 syncer_final_iter = 0; 1487 first_printf = 1; 1488 syncer_state = SYNCER_RUNNING; 1489 starttime = time_second; 1490 1491 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 1492 SHUTDOWN_PRI_LAST); 1493 1494 for (;;) { 1495 mtx_lock(&sync_mtx); 1496 if (syncer_state == SYNCER_FINAL_DELAY && 1497 syncer_final_iter == 0) { 1498 mtx_unlock(&sync_mtx); 1499 kthread_suspend_check(td->td_proc); 1500 mtx_lock(&sync_mtx); 1501 } 1502 net_worklist_len = syncer_worklist_len - sync_vnode_count; 1503 if (syncer_state != SYNCER_RUNNING && 1504 starttime != time_second) { 1505 if (first_printf) { 1506 printf("\nSyncing disks, vnodes remaining..."); 1507 first_printf = 0; 1508 } 1509 printf("%d ", net_worklist_len); 1510 } 1511 starttime = time_second; 1512 1513 /* 1514 * Push files whose dirty time has expired. Be careful 1515 * of interrupt race on slp queue. 1516 * 1517 * Skip over empty worklist slots when shutting down. 1518 */ 1519 do { 1520 slp = &syncer_workitem_pending[syncer_delayno]; 1521 syncer_delayno += 1; 1522 if (syncer_delayno == syncer_maxdelay) 1523 syncer_delayno = 0; 1524 next = &syncer_workitem_pending[syncer_delayno]; 1525 /* 1526 * If the worklist has wrapped since the 1527 * it was emptied of all but syncer vnodes, 1528 * switch to the FINAL_DELAY state and run 1529 * for one more second. 1530 */ 1531 if (syncer_state == SYNCER_SHUTTING_DOWN && 1532 net_worklist_len == 0 && 1533 last_work_seen == syncer_delayno) { 1534 syncer_state = SYNCER_FINAL_DELAY; 1535 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 1536 } 1537 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 1538 syncer_worklist_len > 0); 1539 1540 /* 1541 * Keep track of the last time there was anything 1542 * on the worklist other than syncer vnodes. 1543 * Return to the SHUTTING_DOWN state if any 1544 * new work appears. 1545 */ 1546 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 1547 last_work_seen = syncer_delayno; 1548 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 1549 syncer_state = SYNCER_SHUTTING_DOWN; 1550 while ((vp = LIST_FIRST(slp)) != NULL) { 1551 if (VOP_ISLOCKED(vp, NULL) != 0 || 1552 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1553 LIST_REMOVE(vp, v_synclist); 1554 LIST_INSERT_HEAD(next, vp, v_synclist); 1555 continue; 1556 } 1557 if (VI_TRYLOCK(vp) == 0) { 1558 LIST_REMOVE(vp, v_synclist); 1559 LIST_INSERT_HEAD(next, vp, v_synclist); 1560 vn_finished_write(mp); 1561 continue; 1562 } 1563 /* 1564 * We use vhold in case the vnode does not 1565 * successfully sync. vhold prevents the vnode from 1566 * going away when we unlock the sync_mtx so that 1567 * we can acquire the vnode interlock. 1568 */ 1569 vholdl(vp); 1570 mtx_unlock(&sync_mtx); 1571 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, td); 1572 (void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td); 1573 VOP_UNLOCK(vp, 0, td); 1574 vn_finished_write(mp); 1575 VI_LOCK(vp); 1576 if ((vp->v_iflag & VI_ONWORKLST) != 0) { 1577 /* 1578 * Put us back on the worklist. The worklist 1579 * routine will remove us from our current 1580 * position and then add us back in at a later 1581 * position. 1582 */ 1583 vn_syncer_add_to_worklist(vp, syncdelay); 1584 } 1585 vdropl(vp); 1586 VI_UNLOCK(vp); 1587 mtx_lock(&sync_mtx); 1588 } 1589 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 1590 syncer_final_iter--; 1591 mtx_unlock(&sync_mtx); 1592 1593 /* 1594 * Do soft update processing. 1595 */ 1596 if (softdep_process_worklist_hook != NULL) 1597 (*softdep_process_worklist_hook)(NULL); 1598 1599 /* 1600 * The variable rushjob allows the kernel to speed up the 1601 * processing of the filesystem syncer process. A rushjob 1602 * value of N tells the filesystem syncer to process the next 1603 * N seconds worth of work on its queue ASAP. Currently rushjob 1604 * is used by the soft update code to speed up the filesystem 1605 * syncer process when the incore state is getting so far 1606 * ahead of the disk that the kernel memory pool is being 1607 * threatened with exhaustion. 1608 */ 1609 mtx_lock(&sync_mtx); 1610 if (rushjob > 0) { 1611 rushjob -= 1; 1612 mtx_unlock(&sync_mtx); 1613 continue; 1614 } 1615 mtx_unlock(&sync_mtx); 1616 /* 1617 * Just sleep for a short period if time between 1618 * iterations when shutting down to allow some I/O 1619 * to happen. 1620 * 1621 * If it has taken us less than a second to process the 1622 * current work, then wait. Otherwise start right over 1623 * again. We can still lose time if any single round 1624 * takes more than two seconds, but it does not really 1625 * matter as we are just trying to generally pace the 1626 * filesystem activity. 1627 */ 1628 if (syncer_state != SYNCER_RUNNING) 1629 tsleep(&dummychan, PPAUSE, "syncfnl", 1630 hz / SYNCER_SHUTDOWN_SPEEDUP); 1631 else if (time_second == starttime) 1632 tsleep(&lbolt, PPAUSE, "syncer", 0); 1633 } 1634} 1635 1636/* 1637 * Request the syncer daemon to speed up its work. 1638 * We never push it to speed up more than half of its 1639 * normal turn time, otherwise it could take over the cpu. 1640 */ 1641int 1642speedup_syncer() 1643{ 1644 struct thread *td; 1645 int ret = 0; 1646 1647 td = FIRST_THREAD_IN_PROC(updateproc); 1648 sleepq_remove(td, &lbolt); 1649 mtx_lock(&sync_mtx); 1650 if (rushjob < syncdelay / 2) { 1651 rushjob += 1; 1652 stat_rush_requests += 1; 1653 ret = 1; 1654 } 1655 mtx_unlock(&sync_mtx); 1656 return (ret); 1657} 1658 1659/* 1660 * Tell the syncer to speed up its work and run though its work 1661 * list several times, then tell it to shut down. 1662 */ 1663static void 1664syncer_shutdown(void *arg, int howto) 1665{ 1666 struct thread *td; 1667 1668 if (howto & RB_NOSYNC) 1669 return; 1670 td = FIRST_THREAD_IN_PROC(updateproc); 1671 sleepq_remove(td, &lbolt); 1672 mtx_lock(&sync_mtx); 1673 syncer_state = SYNCER_SHUTTING_DOWN; 1674 rushjob = 0; 1675 mtx_unlock(&sync_mtx); 1676 kproc_shutdown(arg, howto); 1677} 1678 1679/* 1680 * Associate a p-buffer with a vnode. 1681 * 1682 * Also sets B_PAGING flag to indicate that vnode is not fully associated 1683 * with the buffer. i.e. the bp has not been linked into the vnode or 1684 * ref-counted. 1685 */ 1686void 1687pbgetvp(vp, bp) 1688 register struct vnode *vp; 1689 register struct buf *bp; 1690{ 1691 1692 KASSERT(bp->b_vp == NULL, ("pbgetvp: not free")); 1693 1694 bp->b_vp = vp; 1695 bp->b_flags |= B_PAGING; 1696 bp->b_dev = vn_todev(vp); 1697 bp->b_bufobj = &vp->v_bufobj; 1698} 1699 1700/* 1701 * Disassociate a p-buffer from a vnode. 1702 */ 1703void 1704pbrelvp(bp) 1705 register struct buf *bp; 1706{ 1707 1708 KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL")); 1709 KASSERT(bp->b_bufobj != NULL, ("pbrelvp: NULL bufobj")); 1710 1711 /* XXX REMOVE ME */ 1712 BO_LOCK(bp->b_bufobj); 1713 if (TAILQ_NEXT(bp, b_bobufs) != NULL) { 1714 panic( 1715 "relpbuf(): b_vp was probably reassignbuf()d %p %x", 1716 bp, 1717 (int)bp->b_flags 1718 ); 1719 } 1720 BO_UNLOCK(bp->b_bufobj); 1721 bp->b_vp = NULL; 1722 bp->b_bufobj = NULL; 1723 bp->b_flags &= ~B_PAGING; 1724} 1725 1726/* 1727 * Reassign a buffer from one vnode to another. 1728 * Used to assign file specific control information 1729 * (indirect blocks) to the vnode to which they belong. 1730 */ 1731void 1732reassignbuf(struct buf *bp) 1733{ 1734 struct vnode *vp; 1735 struct bufobj *bo; 1736 int delay; 1737 1738 vp = bp->b_vp; 1739 bo = bp->b_bufobj; 1740 ++reassignbufcalls; 1741 1742 /* 1743 * B_PAGING flagged buffers cannot be reassigned because their vp 1744 * is not fully linked in. 1745 */ 1746 if (bp->b_flags & B_PAGING) 1747 panic("cannot reassign paging buffer"); 1748 1749 /* 1750 * Delete from old vnode list, if on one. 1751 */ 1752 VI_LOCK(vp); 1753 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1754 buf_vlist_remove(bp); 1755 /* 1756 * If dirty, put on list of dirty buffers; otherwise insert onto list 1757 * of clean buffers. 1758 */ 1759 if (bp->b_flags & B_DELWRI) { 1760 if ((vp->v_iflag & VI_ONWORKLST) == 0) { 1761 switch (vp->v_type) { 1762 case VDIR: 1763 delay = dirdelay; 1764 break; 1765 case VCHR: 1766 delay = metadelay; 1767 break; 1768 default: 1769 delay = filedelay; 1770 } 1771 vn_syncer_add_to_worklist(vp, delay); 1772 } 1773 buf_vlist_add(bp, bo, BX_VNDIRTY); 1774 } else { 1775 buf_vlist_add(bp, bo, BX_VNCLEAN); 1776 1777 if ((vp->v_iflag & VI_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1778 mtx_lock(&sync_mtx); 1779 LIST_REMOVE(vp, v_synclist); 1780 syncer_worklist_len--; 1781 mtx_unlock(&sync_mtx); 1782 vp->v_iflag &= ~VI_ONWORKLST; 1783 } 1784 } 1785 VI_UNLOCK(vp); 1786} 1787 1788/* 1789 * Create a vnode for a device. 1790 * Used for mounting the root filesystem. 1791 */ 1792int 1793bdevvp(dev, vpp) 1794 struct cdev *dev; 1795 struct vnode **vpp; 1796{ 1797 register struct vnode *vp; 1798 struct vnode *nvp; 1799 int error; 1800 1801 if (dev == NULL) { 1802 *vpp = NULLVP; 1803 return (ENXIO); 1804 } 1805 if (vfinddev(dev, vpp)) 1806 return (0); 1807 1808 error = getnewvnode("none", (struct mount *)0, devfs_specop_p, &nvp); 1809 if (error) { 1810 *vpp = NULLVP; 1811 return (error); 1812 } 1813 vp = nvp; 1814 vp->v_type = VCHR; 1815 addalias(vp, dev); 1816 *vpp = vp; 1817 return (0); 1818} 1819 1820static void 1821v_incr_usecount(struct vnode *vp, int delta) 1822{ 1823 1824 vp->v_usecount += delta; 1825 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1826 dev_lock(); 1827 vp->v_rdev->si_usecount += delta; 1828 dev_unlock(); 1829 } 1830} 1831 1832/* 1833 * Add vnode to the alias list hung off the struct cdev *. 1834 * 1835 * The reason for this gunk is that multiple vnodes can reference 1836 * the same physical device, so checking vp->v_usecount to see 1837 * how many users there are is inadequate; the v_usecount for 1838 * the vnodes need to be accumulated. vcount() does that. 1839 */ 1840struct vnode * 1841addaliasu(nvp, nvp_rdev) 1842 struct vnode *nvp; 1843 dev_t nvp_rdev; 1844{ 1845 struct vnode *ovp; 1846 vop_t **ops; 1847 struct cdev *dev; 1848 1849 if (nvp->v_type == VBLK) 1850 return (nvp); 1851 if (nvp->v_type != VCHR) 1852 panic("addaliasu on non-special vnode"); 1853 dev = findcdev(nvp_rdev); 1854 if (dev == NULL) 1855 return (nvp); 1856 /* 1857 * Check to see if we have a bdevvp vnode with no associated 1858 * filesystem. If so, we want to associate the filesystem of 1859 * the new newly instigated vnode with the bdevvp vnode and 1860 * discard the newly created vnode rather than leaving the 1861 * bdevvp vnode lying around with no associated filesystem. 1862 */ 1863 if (vfinddev(dev, &ovp) == 0 || ovp->v_data != NULL) { 1864 addalias(nvp, dev); 1865 return (nvp); 1866 } 1867 /* 1868 * Discard unneeded vnode, but save its node specific data. 1869 * Note that if there is a lock, it is carried over in the 1870 * node specific data to the replacement vnode. 1871 */ 1872 vref(ovp); 1873 ovp->v_data = nvp->v_data; 1874 ovp->v_tag = nvp->v_tag; 1875 nvp->v_data = NULL; 1876 lockdestroy(ovp->v_vnlock); 1877 lockinit(ovp->v_vnlock, PVFS, nvp->v_vnlock->lk_wmesg, 1878 nvp->v_vnlock->lk_timo, nvp->v_vnlock->lk_flags & LK_EXTFLG_MASK); 1879 ops = ovp->v_op; 1880 ovp->v_op = nvp->v_op; 1881 if (VOP_ISLOCKED(nvp, curthread)) { 1882 VOP_UNLOCK(nvp, 0, curthread); 1883 vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread); 1884 } 1885 nvp->v_op = ops; 1886 delmntque(ovp); 1887 insmntque(ovp, nvp->v_mount); 1888 vrele(nvp); 1889 vgone(nvp); 1890 return (ovp); 1891} 1892 1893/* This is a local helper function that do the same as addaliasu, but for a 1894 * struct cdev *instead of an dev_t. */ 1895static void 1896addalias(nvp, dev) 1897 struct vnode *nvp; 1898 struct cdev *dev; 1899{ 1900 1901 KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode")); 1902 VI_LOCK(nvp); 1903 dev_lock(); 1904 dev->si_refcount++; 1905 nvp->v_rdev = dev; 1906 SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext); 1907 dev->si_usecount += nvp->v_usecount; 1908 dev_unlock(); 1909 VI_UNLOCK(nvp); 1910} 1911 1912/* 1913 * Grab a particular vnode from the free list, increment its 1914 * reference count and lock it. The vnode lock bit is set if the 1915 * vnode is being eliminated in vgone. The process is awakened 1916 * when the transition is completed, and an error returned to 1917 * indicate that the vnode is no longer usable (possibly having 1918 * been changed to a new filesystem type). 1919 */ 1920int 1921vget(vp, flags, td) 1922 register struct vnode *vp; 1923 int flags; 1924 struct thread *td; 1925{ 1926 int error; 1927 1928 /* 1929 * If the vnode is in the process of being cleaned out for 1930 * another use, we wait for the cleaning to finish and then 1931 * return failure. Cleaning is determined by checking that 1932 * the VI_XLOCK flag is set. 1933 */ 1934 if ((flags & LK_INTERLOCK) == 0) 1935 VI_LOCK(vp); 1936 if (vp->v_iflag & VI_XLOCK && vp->v_vxthread != curthread) { 1937 if ((flags & LK_NOWAIT) == 0) { 1938 vp->v_iflag |= VI_XWANT; 1939 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0); 1940 return (ENOENT); 1941 } 1942 VI_UNLOCK(vp); 1943 return (EBUSY); 1944 } 1945 1946 v_incr_usecount(vp, 1); 1947 1948 if (VSHOULDBUSY(vp)) 1949 vbusy(vp); 1950 if (flags & LK_TYPE_MASK) { 1951 if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { 1952 /* 1953 * must expand vrele here because we do not want 1954 * to call VOP_INACTIVE if the reference count 1955 * drops back to zero since it was never really 1956 * active. We must remove it from the free list 1957 * before sleeping so that multiple processes do 1958 * not try to recycle it. 1959 */ 1960 VI_LOCK(vp); 1961 v_incr_usecount(vp, -1); 1962 if (VSHOULDFREE(vp)) 1963 vfree(vp); 1964 else 1965 vlruvp(vp); 1966 VI_UNLOCK(vp); 1967 } 1968 return (error); 1969 } 1970 VI_UNLOCK(vp); 1971 return (0); 1972} 1973 1974/* 1975 * Increase the reference count of a vnode. 1976 */ 1977void 1978vref(struct vnode *vp) 1979{ 1980 1981 VI_LOCK(vp); 1982 v_incr_usecount(vp, 1); 1983 VI_UNLOCK(vp); 1984} 1985 1986/* 1987 * Return reference count of a vnode. 1988 * 1989 * The results of this call are only guaranteed when some mechanism other 1990 * than the VI lock is used to stop other processes from gaining references 1991 * to the vnode. This may be the case if the caller holds the only reference. 1992 * This is also useful when stale data is acceptable as race conditions may 1993 * be accounted for by some other means. 1994 */ 1995int 1996vrefcnt(struct vnode *vp) 1997{ 1998 int usecnt; 1999 2000 VI_LOCK(vp); 2001 usecnt = vp->v_usecount; 2002 VI_UNLOCK(vp); 2003 2004 return (usecnt); 2005} 2006 2007 2008/* 2009 * Vnode put/release. 2010 * If count drops to zero, call inactive routine and return to freelist. 2011 */ 2012void 2013vrele(vp) 2014 struct vnode *vp; 2015{ 2016 struct thread *td = curthread; /* XXX */ 2017 2018 GIANT_REQUIRED; 2019 2020 KASSERT(vp != NULL, ("vrele: null vp")); 2021 2022 VI_LOCK(vp); 2023 2024 /* Skip this v_writecount check if we're going to panic below. */ 2025 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2026 ("vrele: missed vn_close")); 2027 2028 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2029 vp->v_usecount == 1)) { 2030 v_incr_usecount(vp, -1); 2031 VI_UNLOCK(vp); 2032 2033 return; 2034 } 2035 2036 if (vp->v_usecount == 1) { 2037 v_incr_usecount(vp, -1); 2038 /* 2039 * We must call VOP_INACTIVE with the node locked. Mark 2040 * as VI_DOINGINACT to avoid recursion. 2041 */ 2042 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) { 2043 VI_LOCK(vp); 2044 vp->v_iflag |= VI_DOINGINACT; 2045 VI_UNLOCK(vp); 2046 VOP_INACTIVE(vp, td); 2047 VI_LOCK(vp); 2048 KASSERT(vp->v_iflag & VI_DOINGINACT, 2049 ("vrele: lost VI_DOINGINACT")); 2050 vp->v_iflag &= ~VI_DOINGINACT; 2051 } else 2052 VI_LOCK(vp); 2053 if (VSHOULDFREE(vp)) 2054 vfree(vp); 2055 else 2056 vlruvp(vp); 2057 VI_UNLOCK(vp); 2058 2059 } else { 2060#ifdef DIAGNOSTIC 2061 vprint("vrele: negative ref count", vp); 2062#endif 2063 VI_UNLOCK(vp); 2064 panic("vrele: negative ref cnt"); 2065 } 2066} 2067 2068/* 2069 * Release an already locked vnode. This give the same effects as 2070 * unlock+vrele(), but takes less time and avoids releasing and 2071 * re-aquiring the lock (as vrele() aquires the lock internally.) 2072 */ 2073void 2074vput(vp) 2075 struct vnode *vp; 2076{ 2077 struct thread *td = curthread; /* XXX */ 2078 2079 GIANT_REQUIRED; 2080 2081 KASSERT(vp != NULL, ("vput: null vp")); 2082 VI_LOCK(vp); 2083 /* Skip this v_writecount check if we're going to panic below. */ 2084 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2085 ("vput: missed vn_close")); 2086 2087 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2088 vp->v_usecount == 1)) { 2089 v_incr_usecount(vp, -1); 2090 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2091 return; 2092 } 2093 2094 if (vp->v_usecount == 1) { 2095 v_incr_usecount(vp, -1); 2096 /* 2097 * We must call VOP_INACTIVE with the node locked, so 2098 * we just need to release the vnode mutex. Mark as 2099 * as VI_DOINGINACT to avoid recursion. 2100 */ 2101 vp->v_iflag |= VI_DOINGINACT; 2102 VI_UNLOCK(vp); 2103 VOP_INACTIVE(vp, td); 2104 VI_LOCK(vp); 2105 KASSERT(vp->v_iflag & VI_DOINGINACT, 2106 ("vput: lost VI_DOINGINACT")); 2107 vp->v_iflag &= ~VI_DOINGINACT; 2108 if (VSHOULDFREE(vp)) 2109 vfree(vp); 2110 else 2111 vlruvp(vp); 2112 VI_UNLOCK(vp); 2113 2114 } else { 2115#ifdef DIAGNOSTIC 2116 vprint("vput: negative ref count", vp); 2117#endif 2118 panic("vput: negative ref cnt"); 2119 } 2120} 2121 2122/* 2123 * Somebody doesn't want the vnode recycled. 2124 */ 2125void 2126vhold(struct vnode *vp) 2127{ 2128 2129 VI_LOCK(vp); 2130 vholdl(vp); 2131 VI_UNLOCK(vp); 2132} 2133 2134void 2135vholdl(struct vnode *vp) 2136{ 2137 2138 vp->v_holdcnt++; 2139 if (VSHOULDBUSY(vp)) 2140 vbusy(vp); 2141} 2142 2143/* 2144 * Note that there is one less who cares about this vnode. vdrop() is the 2145 * opposite of vhold(). 2146 */ 2147void 2148vdrop(struct vnode *vp) 2149{ 2150 2151 VI_LOCK(vp); 2152 vdropl(vp); 2153 VI_UNLOCK(vp); 2154} 2155 2156void 2157vdropl(vp) 2158 register struct vnode *vp; 2159{ 2160 2161 if (vp->v_holdcnt <= 0) 2162 panic("vdrop: holdcnt"); 2163 vp->v_holdcnt--; 2164 if (VSHOULDFREE(vp)) 2165 vfree(vp); 2166 else 2167 vlruvp(vp); 2168} 2169 2170/* 2171 * Remove any vnodes in the vnode table belonging to mount point mp. 2172 * 2173 * If FORCECLOSE is not specified, there should not be any active ones, 2174 * return error if any are found (nb: this is a user error, not a 2175 * system error). If FORCECLOSE is specified, detach any active vnodes 2176 * that are found. 2177 * 2178 * If WRITECLOSE is set, only flush out regular file vnodes open for 2179 * writing. 2180 * 2181 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2182 * 2183 * `rootrefs' specifies the base reference count for the root vnode 2184 * of this filesystem. The root vnode is considered busy if its 2185 * v_usecount exceeds this value. On a successful return, vflush(, td) 2186 * will call vrele() on the root vnode exactly rootrefs times. 2187 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2188 * be zero. 2189 */ 2190#ifdef DIAGNOSTIC 2191static int busyprt = 0; /* print out busy vnodes */ 2192SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 2193#endif 2194 2195int 2196vflush(mp, rootrefs, flags, td) 2197 struct mount *mp; 2198 int rootrefs; 2199 int flags; 2200 struct thread *td; 2201{ 2202 struct vnode *vp, *nvp, *rootvp = NULL; 2203 struct vattr vattr; 2204 int busy = 0, error; 2205 2206 if (rootrefs > 0) { 2207 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2208 ("vflush: bad args")); 2209 /* 2210 * Get the filesystem root vnode. We can vput() it 2211 * immediately, since with rootrefs > 0, it won't go away. 2212 */ 2213 if ((error = VFS_ROOT(mp, &rootvp, td)) != 0) 2214 return (error); 2215 vput(rootvp); 2216 2217 } 2218 MNT_ILOCK(mp); 2219loop: 2220 MNT_VNODE_FOREACH(vp, mp, nvp) { 2221 2222 VI_LOCK(vp); 2223 MNT_IUNLOCK(mp); 2224 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td); 2225 if (error) { 2226 MNT_ILOCK(mp); 2227 goto loop; 2228 } 2229 /* 2230 * Skip over a vnodes marked VV_SYSTEM. 2231 */ 2232 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2233 VOP_UNLOCK(vp, 0, td); 2234 MNT_ILOCK(mp); 2235 continue; 2236 } 2237 /* 2238 * If WRITECLOSE is set, flush out unlinked but still open 2239 * files (even if open only for reading) and regular file 2240 * vnodes open for writing. 2241 */ 2242 if (flags & WRITECLOSE) { 2243 error = VOP_GETATTR(vp, &vattr, td->td_ucred, td); 2244 VI_LOCK(vp); 2245 2246 if ((vp->v_type == VNON || 2247 (error == 0 && vattr.va_nlink > 0)) && 2248 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2249 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2250 MNT_ILOCK(mp); 2251 continue; 2252 } 2253 } else 2254 VI_LOCK(vp); 2255 2256 VOP_UNLOCK(vp, 0, td); 2257 2258 /* 2259 * With v_usecount == 0, all we need to do is clear out the 2260 * vnode data structures and we are done. 2261 */ 2262 if (vp->v_usecount == 0) { 2263 vgonel(vp, td); 2264 MNT_ILOCK(mp); 2265 continue; 2266 } 2267 2268 /* 2269 * If FORCECLOSE is set, forcibly close the vnode. For block 2270 * or character devices, revert to an anonymous device. For 2271 * all other files, just kill them. 2272 */ 2273 if (flags & FORCECLOSE) { 2274 if (vp->v_type != VCHR) 2275 vgonel(vp, td); 2276 else 2277 vgonechrl(vp, td); 2278 MNT_ILOCK(mp); 2279 continue; 2280 } 2281#ifdef DIAGNOSTIC 2282 if (busyprt) 2283 vprint("vflush: busy vnode", vp); 2284#endif 2285 VI_UNLOCK(vp); 2286 MNT_ILOCK(mp); 2287 busy++; 2288 } 2289 MNT_IUNLOCK(mp); 2290 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2291 /* 2292 * If just the root vnode is busy, and if its refcount 2293 * is equal to `rootrefs', then go ahead and kill it. 2294 */ 2295 VI_LOCK(rootvp); 2296 KASSERT(busy > 0, ("vflush: not busy")); 2297 KASSERT(rootvp->v_usecount >= rootrefs, 2298 ("vflush: usecount %d < rootrefs %d", 2299 rootvp->v_usecount, rootrefs)); 2300 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2301 vgonel(rootvp, td); 2302 busy = 0; 2303 } else 2304 VI_UNLOCK(rootvp); 2305 } 2306 if (busy) 2307 return (EBUSY); 2308 for (; rootrefs > 0; rootrefs--) 2309 vrele(rootvp); 2310 return (0); 2311} 2312 2313/* 2314 * This moves a now (likely recyclable) vnode to the end of the 2315 * mountlist. XXX However, it is temporarily disabled until we 2316 * can clean up ffs_sync() and friends, which have loop restart 2317 * conditions which this code causes to operate O(N^2). 2318 */ 2319static void 2320vlruvp(struct vnode *vp) 2321{ 2322#if 0 2323 struct mount *mp; 2324 2325 if ((mp = vp->v_mount) != NULL) { 2326 MNT_ILOCK(mp); 2327 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2328 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2329 MNT_IUNLOCK(mp); 2330 } 2331#endif 2332} 2333 2334static void 2335vx_lock(struct vnode *vp) 2336{ 2337 2338 ASSERT_VI_LOCKED(vp, "vx_lock"); 2339 2340 /* 2341 * Prevent the vnode from being recycled or brought into use while we 2342 * clean it out. 2343 */ 2344 if (vp->v_iflag & VI_XLOCK) 2345 panic("vclean: deadlock"); 2346 vp->v_iflag |= VI_XLOCK; 2347 vp->v_vxthread = curthread; 2348} 2349 2350static void 2351vx_unlock(struct vnode *vp) 2352{ 2353 ASSERT_VI_LOCKED(vp, "vx_unlock"); 2354 vp->v_iflag &= ~VI_XLOCK; 2355 vp->v_vxthread = NULL; 2356 if (vp->v_iflag & VI_XWANT) { 2357 vp->v_iflag &= ~VI_XWANT; 2358 wakeup(vp); 2359 } 2360} 2361 2362/* 2363 * Disassociate the underlying filesystem from a vnode. 2364 */ 2365static void 2366vclean(vp, flags, td) 2367 struct vnode *vp; 2368 int flags; 2369 struct thread *td; 2370{ 2371 int active; 2372 2373 ASSERT_VI_LOCKED(vp, "vclean"); 2374 /* 2375 * Check to see if the vnode is in use. If so we have to reference it 2376 * before we clean it out so that its count cannot fall to zero and 2377 * generate a race against ourselves to recycle it. 2378 */ 2379 if ((active = vp->v_usecount)) 2380 v_incr_usecount(vp, 1); 2381 2382 /* 2383 * Even if the count is zero, the VOP_INACTIVE routine may still 2384 * have the object locked while it cleans it out. The VOP_LOCK 2385 * ensures that the VOP_INACTIVE routine is done with its work. 2386 * For active vnodes, it ensures that no other activity can 2387 * occur while the underlying object is being cleaned out. 2388 */ 2389 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2390 2391 /* 2392 * Clean out any buffers associated with the vnode. 2393 * If the flush fails, just toss the buffers. 2394 */ 2395 if (flags & DOCLOSE) { 2396 struct buf *bp; 2397 bp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd); 2398 if (bp != NULL) 2399 (void) vn_write_suspend_wait(vp, NULL, V_WAIT); 2400 if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0) 2401 vinvalbuf(vp, 0, NOCRED, td, 0, 0); 2402 } 2403 2404 VOP_DESTROYVOBJECT(vp); 2405 2406 /* 2407 * Any other processes trying to obtain this lock must first 2408 * wait for VXLOCK to clear, then call the new lock operation. 2409 */ 2410 VOP_UNLOCK(vp, 0, td); 2411 2412 /* 2413 * If purging an active vnode, it must be closed and 2414 * deactivated before being reclaimed. Note that the 2415 * VOP_INACTIVE will unlock the vnode. 2416 */ 2417 if (active) { 2418 if (flags & DOCLOSE) 2419 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2420 VI_LOCK(vp); 2421 if ((vp->v_iflag & VI_DOINGINACT) == 0) { 2422 vp->v_iflag |= VI_DOINGINACT; 2423 VI_UNLOCK(vp); 2424 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 2425 panic("vclean: cannot relock."); 2426 VOP_INACTIVE(vp, td); 2427 VI_LOCK(vp); 2428 KASSERT(vp->v_iflag & VI_DOINGINACT, 2429 ("vclean: lost VI_DOINGINACT")); 2430 vp->v_iflag &= ~VI_DOINGINACT; 2431 } 2432 VI_UNLOCK(vp); 2433 } 2434 /* 2435 * Reclaim the vnode. 2436 */ 2437 if (VOP_RECLAIM(vp, td)) 2438 panic("vclean: cannot reclaim"); 2439 2440 if (active) { 2441 /* 2442 * Inline copy of vrele() since VOP_INACTIVE 2443 * has already been called. 2444 */ 2445 VI_LOCK(vp); 2446 v_incr_usecount(vp, -1); 2447 if (vp->v_usecount <= 0) { 2448#ifdef INVARIANTS 2449 if (vp->v_usecount < 0 || vp->v_writecount != 0) { 2450 vprint("vclean: bad ref count", vp); 2451 panic("vclean: ref cnt"); 2452 } 2453#endif 2454 if (VSHOULDFREE(vp)) 2455 vfree(vp); 2456 } 2457 VI_UNLOCK(vp); 2458 } 2459 /* 2460 * Delete from old mount point vnode list. 2461 */ 2462 delmntque(vp); 2463 cache_purge(vp); 2464 VI_LOCK(vp); 2465 if (VSHOULDFREE(vp)) 2466 vfree(vp); 2467 2468 /* 2469 * Done with purge, reset to the standard lock and 2470 * notify sleepers of the grim news. 2471 */ 2472 vp->v_vnlock = &vp->v_lock; 2473 vp->v_op = dead_vnodeop_p; 2474 if (vp->v_pollinfo != NULL) 2475 vn_pollgone(vp); 2476 vp->v_tag = "none"; 2477} 2478 2479/* 2480 * Eliminate all activity associated with the requested vnode 2481 * and with all vnodes aliased to the requested vnode. 2482 */ 2483int 2484vop_revoke(ap) 2485 struct vop_revoke_args /* { 2486 struct vnode *a_vp; 2487 int a_flags; 2488 } */ *ap; 2489{ 2490 struct vnode *vp, *vq; 2491 struct cdev *dev; 2492 2493 KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke")); 2494 vp = ap->a_vp; 2495 KASSERT((vp->v_type == VCHR), ("vop_revoke: not VCHR")); 2496 2497 VI_LOCK(vp); 2498 /* 2499 * If a vgone (or vclean) is already in progress, 2500 * wait until it is done and return. 2501 */ 2502 if (vp->v_iflag & VI_XLOCK) { 2503 vp->v_iflag |= VI_XWANT; 2504 msleep(vp, VI_MTX(vp), PINOD | PDROP, 2505 "vop_revokeall", 0); 2506 return (0); 2507 } 2508 VI_UNLOCK(vp); 2509 dev = vp->v_rdev; 2510 for (;;) { 2511 dev_lock(); 2512 vq = SLIST_FIRST(&dev->si_hlist); 2513 dev_unlock(); 2514 if (vq == NULL) 2515 break; 2516 vgone(vq); 2517 } 2518 return (0); 2519} 2520 2521/* 2522 * Recycle an unused vnode to the front of the free list. 2523 * Release the passed interlock if the vnode will be recycled. 2524 */ 2525int 2526vrecycle(vp, inter_lkp, td) 2527 struct vnode *vp; 2528 struct mtx *inter_lkp; 2529 struct thread *td; 2530{ 2531 2532 VI_LOCK(vp); 2533 if (vp->v_usecount == 0) { 2534 if (inter_lkp) { 2535 mtx_unlock(inter_lkp); 2536 } 2537 vgonel(vp, td); 2538 return (1); 2539 } 2540 VI_UNLOCK(vp); 2541 return (0); 2542} 2543 2544/* 2545 * Eliminate all activity associated with a vnode 2546 * in preparation for reuse. 2547 */ 2548void 2549vgone(vp) 2550 register struct vnode *vp; 2551{ 2552 struct thread *td = curthread; /* XXX */ 2553 2554 VI_LOCK(vp); 2555 vgonel(vp, td); 2556} 2557 2558/* 2559 * Disassociate a character device from the its underlying filesystem and 2560 * attach it to spec. This is for use when the chr device is still active 2561 * and the filesystem is going away. 2562 */ 2563static void 2564vgonechrl(struct vnode *vp, struct thread *td) 2565{ 2566 ASSERT_VI_LOCKED(vp, "vgonechrl"); 2567 vx_lock(vp); 2568 /* 2569 * This is a custom version of vclean() which does not tearm down 2570 * the bufs or vm objects held by this vnode. This allows filesystems 2571 * to continue using devices which were discovered via another 2572 * filesystem that has been unmounted. 2573 */ 2574 if (vp->v_usecount != 0) { 2575 v_incr_usecount(vp, 1); 2576 /* 2577 * Ensure that no other activity can occur while the 2578 * underlying object is being cleaned out. 2579 */ 2580 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2581 /* 2582 * Any other processes trying to obtain this lock must first 2583 * wait for VXLOCK to clear, then call the new lock operation. 2584 */ 2585 VOP_UNLOCK(vp, 0, td); 2586 vp->v_vnlock = &vp->v_lock; 2587 vp->v_tag = "orphanchr"; 2588 vp->v_op = devfs_specop_p; 2589 delmntque(vp); 2590 cache_purge(vp); 2591 vrele(vp); 2592 VI_LOCK(vp); 2593 } else 2594 vclean(vp, 0, td); 2595 vp->v_op = devfs_specop_p; 2596 vx_unlock(vp); 2597 VI_UNLOCK(vp); 2598} 2599 2600/* 2601 * vgone, with the vp interlock held. 2602 */ 2603void 2604vgonel(vp, td) 2605 struct vnode *vp; 2606 struct thread *td; 2607{ 2608 /* 2609 * If a vgone (or vclean) is already in progress, 2610 * wait until it is done and return. 2611 */ 2612 ASSERT_VI_LOCKED(vp, "vgonel"); 2613 if (vp->v_iflag & VI_XLOCK) { 2614 vp->v_iflag |= VI_XWANT; 2615 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0); 2616 return; 2617 } 2618 vx_lock(vp); 2619 2620 /* 2621 * Clean out the filesystem specific data. 2622 */ 2623 vclean(vp, DOCLOSE, td); 2624 VI_UNLOCK(vp); 2625 2626 /* 2627 * If special device, remove it from special device alias list 2628 * if it is on one. 2629 */ 2630 VI_LOCK(vp); 2631 if (vp->v_type == VCHR && vp->v_rdev != NULL) 2632 dev_rel(vp); 2633 2634 /* 2635 * If it is on the freelist and not already at the head, 2636 * move it to the head of the list. The test of the 2637 * VDOOMED flag and the reference count of zero is because 2638 * it will be removed from the free list by getnewvnode, 2639 * but will not have its reference count incremented until 2640 * after calling vgone. If the reference count were 2641 * incremented first, vgone would (incorrectly) try to 2642 * close the previous instance of the underlying object. 2643 */ 2644 if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) { 2645 mtx_lock(&vnode_free_list_mtx); 2646 if (vp->v_iflag & VI_FREE) { 2647 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2648 } else { 2649 vp->v_iflag |= VI_FREE; 2650 freevnodes++; 2651 } 2652 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2653 mtx_unlock(&vnode_free_list_mtx); 2654 } 2655 2656 vp->v_type = VBAD; 2657 vx_unlock(vp); 2658 VI_UNLOCK(vp); 2659} 2660 2661/* 2662 * Lookup a vnode by device number. 2663 */ 2664int 2665vfinddev(dev, vpp) 2666 struct cdev *dev; 2667 struct vnode **vpp; 2668{ 2669 struct vnode *vp; 2670 2671 dev_lock(); 2672 SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) { 2673 *vpp = vp; 2674 dev_unlock(); 2675 return (1); 2676 } 2677 dev_unlock(); 2678 return (0); 2679} 2680 2681/* 2682 * Calculate the total number of references to a special device. 2683 */ 2684int 2685vcount(vp) 2686 struct vnode *vp; 2687{ 2688 int count; 2689 2690 dev_lock(); 2691 count = vp->v_rdev->si_usecount; 2692 dev_unlock(); 2693 return (count); 2694} 2695 2696/* 2697 * Same as above, but using the struct cdev *as argument 2698 */ 2699int 2700count_dev(dev) 2701 struct cdev *dev; 2702{ 2703 int count; 2704 2705 dev_lock(); 2706 count = dev->si_usecount; 2707 dev_unlock(); 2708 return(count); 2709} 2710 2711/* 2712 * Print out a description of a vnode. 2713 */ 2714static char *typename[] = 2715{"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 2716 2717void 2718vprint(label, vp) 2719 char *label; 2720 struct vnode *vp; 2721{ 2722 char buf[96]; 2723 2724 if (label != NULL) 2725 printf("%s: %p: ", label, (void *)vp); 2726 else 2727 printf("%p: ", (void *)vp); 2728 printf("tag %s, type %s, usecount %d, writecount %d, refcount %d,", 2729 vp->v_tag, typename[vp->v_type], vp->v_usecount, 2730 vp->v_writecount, vp->v_holdcnt); 2731 buf[0] = '\0'; 2732 if (vp->v_vflag & VV_ROOT) 2733 strcat(buf, "|VV_ROOT"); 2734 if (vp->v_vflag & VV_TEXT) 2735 strcat(buf, "|VV_TEXT"); 2736 if (vp->v_vflag & VV_SYSTEM) 2737 strcat(buf, "|VV_SYSTEM"); 2738 if (vp->v_iflag & VI_XLOCK) 2739 strcat(buf, "|VI_XLOCK"); 2740 if (vp->v_iflag & VI_XWANT) 2741 strcat(buf, "|VI_XWANT"); 2742 if (vp->v_iflag & VI_DOOMED) 2743 strcat(buf, "|VI_DOOMED"); 2744 if (vp->v_iflag & VI_FREE) 2745 strcat(buf, "|VI_FREE"); 2746 if (vp->v_vflag & VV_OBJBUF) 2747 strcat(buf, "|VV_OBJBUF"); 2748 if (buf[0] != '\0') 2749 printf(" flags (%s),", &buf[1]); 2750 lockmgr_printinfo(vp->v_vnlock); 2751 printf("\n"); 2752 if (vp->v_data != NULL) 2753 VOP_PRINT(vp); 2754} 2755 2756#ifdef DDB 2757#include <ddb/ddb.h> 2758/* 2759 * List all of the locked vnodes in the system. 2760 * Called when debugging the kernel. 2761 */ 2762DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 2763{ 2764 struct mount *mp, *nmp; 2765 struct vnode *vp; 2766 2767 /* 2768 * Note: because this is DDB, we can't obey the locking semantics 2769 * for these structures, which means we could catch an inconsistent 2770 * state and dereference a nasty pointer. Not much to be done 2771 * about that. 2772 */ 2773 printf("Locked vnodes\n"); 2774 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2775 nmp = TAILQ_NEXT(mp, mnt_list); 2776 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2777 if (VOP_ISLOCKED(vp, NULL)) 2778 vprint(NULL, vp); 2779 } 2780 nmp = TAILQ_NEXT(mp, mnt_list); 2781 } 2782} 2783#endif 2784 2785/* 2786 * Fill in a struct xvfsconf based on a struct vfsconf. 2787 */ 2788static void 2789vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) 2790{ 2791 2792 strcpy(xvfsp->vfc_name, vfsp->vfc_name); 2793 xvfsp->vfc_typenum = vfsp->vfc_typenum; 2794 xvfsp->vfc_refcount = vfsp->vfc_refcount; 2795 xvfsp->vfc_flags = vfsp->vfc_flags; 2796 /* 2797 * These are unused in userland, we keep them 2798 * to not break binary compatibility. 2799 */ 2800 xvfsp->vfc_vfsops = NULL; 2801 xvfsp->vfc_next = NULL; 2802} 2803 2804/* 2805 * Top level filesystem related information gathering. 2806 */ 2807static int 2808sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 2809{ 2810 struct vfsconf *vfsp; 2811 struct xvfsconf xvfsp; 2812 int error; 2813 2814 error = 0; 2815 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2816 vfsconf2x(vfsp, &xvfsp); 2817 error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp); 2818 if (error) 2819 break; 2820 } 2821 return (error); 2822} 2823 2824SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, 2825 "S,xvfsconf", "List of all configured filesystems"); 2826 2827#ifndef BURN_BRIDGES 2828static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2829 2830static int 2831vfs_sysctl(SYSCTL_HANDLER_ARGS) 2832{ 2833 int *name = (int *)arg1 - 1; /* XXX */ 2834 u_int namelen = arg2 + 1; /* XXX */ 2835 struct vfsconf *vfsp; 2836 struct xvfsconf xvfsp; 2837 2838 printf("WARNING: userland calling deprecated sysctl, " 2839 "please rebuild world\n"); 2840 2841#if 1 || defined(COMPAT_PRELITE2) 2842 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2843 if (namelen == 1) 2844 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2845#endif 2846 2847 switch (name[1]) { 2848 case VFS_MAXTYPENUM: 2849 if (namelen != 2) 2850 return (ENOTDIR); 2851 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2852 case VFS_CONF: 2853 if (namelen != 3) 2854 return (ENOTDIR); /* overloaded */ 2855 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) 2856 if (vfsp->vfc_typenum == name[2]) 2857 break; 2858 if (vfsp == NULL) 2859 return (EOPNOTSUPP); 2860 vfsconf2x(vfsp, &xvfsp); 2861 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 2862 } 2863 return (EOPNOTSUPP); 2864} 2865 2866SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl, 2867 "Generic filesystem"); 2868 2869#if 1 || defined(COMPAT_PRELITE2) 2870 2871static int 2872sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 2873{ 2874 int error; 2875 struct vfsconf *vfsp; 2876 struct ovfsconf ovfs; 2877 2878 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2879 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 2880 strcpy(ovfs.vfc_name, vfsp->vfc_name); 2881 ovfs.vfc_index = vfsp->vfc_typenum; 2882 ovfs.vfc_refcount = vfsp->vfc_refcount; 2883 ovfs.vfc_flags = vfsp->vfc_flags; 2884 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 2885 if (error) 2886 return error; 2887 } 2888 return 0; 2889} 2890 2891#endif /* 1 || COMPAT_PRELITE2 */ 2892#endif /* !BURN_BRIDGES */ 2893 2894#define KINFO_VNODESLOP 10 2895#ifdef notyet 2896/* 2897 * Dump vnode list (via sysctl). 2898 */ 2899/* ARGSUSED */ 2900static int 2901sysctl_vnode(SYSCTL_HANDLER_ARGS) 2902{ 2903 struct xvnode *xvn; 2904 struct thread *td = req->td; 2905 struct mount *mp; 2906 struct vnode *vp; 2907 int error, len, n; 2908 2909 /* 2910 * Stale numvnodes access is not fatal here. 2911 */ 2912 req->lock = 0; 2913 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 2914 if (!req->oldptr) 2915 /* Make an estimate */ 2916 return (SYSCTL_OUT(req, 0, len)); 2917 2918 error = sysctl_wire_old_buffer(req, 0); 2919 if (error != 0) 2920 return (error); 2921 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 2922 n = 0; 2923 mtx_lock(&mountlist_mtx); 2924 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 2925 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 2926 continue; 2927 MNT_ILOCK(mp); 2928 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2929 if (n == len) 2930 break; 2931 vref(vp); 2932 xvn[n].xv_size = sizeof *xvn; 2933 xvn[n].xv_vnode = vp; 2934#define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 2935 XV_COPY(usecount); 2936 XV_COPY(writecount); 2937 XV_COPY(holdcnt); 2938 XV_COPY(id); 2939 XV_COPY(mount); 2940 XV_COPY(numoutput); 2941 XV_COPY(type); 2942#undef XV_COPY 2943 xvn[n].xv_flag = vp->v_vflag; 2944 2945 switch (vp->v_type) { 2946 case VREG: 2947 case VDIR: 2948 case VLNK: 2949 xvn[n].xv_dev = vp->v_cachedfs; 2950 xvn[n].xv_ino = vp->v_cachedid; 2951 break; 2952 case VBLK: 2953 case VCHR: 2954 if (vp->v_rdev == NULL) { 2955 vrele(vp); 2956 continue; 2957 } 2958 xvn[n].xv_dev = dev2udev(vp->v_rdev); 2959 break; 2960 case VSOCK: 2961 xvn[n].xv_socket = vp->v_socket; 2962 break; 2963 case VFIFO: 2964 xvn[n].xv_fifo = vp->v_fifoinfo; 2965 break; 2966 case VNON: 2967 case VBAD: 2968 default: 2969 /* shouldn't happen? */ 2970 vrele(vp); 2971 continue; 2972 } 2973 vrele(vp); 2974 ++n; 2975 } 2976 MNT_IUNLOCK(mp); 2977 mtx_lock(&mountlist_mtx); 2978 vfs_unbusy(mp, td); 2979 if (n == len) 2980 break; 2981 } 2982 mtx_unlock(&mountlist_mtx); 2983 2984 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 2985 free(xvn, M_TEMP); 2986 return (error); 2987} 2988 2989SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 2990 0, 0, sysctl_vnode, "S,xvnode", ""); 2991#endif 2992 2993/* 2994 * Check to see if a filesystem is mounted on a block device. 2995 */ 2996int 2997vfs_mountedon(vp) 2998 struct vnode *vp; 2999{ 3000 3001 if (vp->v_rdev->si_mountpoint != NULL) 3002 return (EBUSY); 3003 return (0); 3004} 3005 3006/* 3007 * Unmount all filesystems. The list is traversed in reverse order 3008 * of mounting to avoid dependencies. 3009 */ 3010void 3011vfs_unmountall() 3012{ 3013 struct mount *mp; 3014 struct thread *td; 3015 int error; 3016 3017 if (curthread != NULL) 3018 td = curthread; 3019 else 3020 td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */ 3021 /* 3022 * Since this only runs when rebooting, it is not interlocked. 3023 */ 3024 while(!TAILQ_EMPTY(&mountlist)) { 3025 mp = TAILQ_LAST(&mountlist, mntlist); 3026 error = dounmount(mp, MNT_FORCE, td); 3027 if (error) { 3028 TAILQ_REMOVE(&mountlist, mp, mnt_list); 3029 printf("unmount of %s failed (", 3030 mp->mnt_stat.f_mntonname); 3031 if (error == EBUSY) 3032 printf("BUSY)\n"); 3033 else 3034 printf("%d)\n", error); 3035 } else { 3036 /* The unmount has removed mp from the mountlist */ 3037 } 3038 } 3039} 3040 3041/* 3042 * perform msync on all vnodes under a mount point 3043 * the mount point must be locked. 3044 */ 3045void 3046vfs_msync(struct mount *mp, int flags) 3047{ 3048 struct vnode *vp, *nvp; 3049 struct vm_object *obj; 3050 int tries; 3051 3052 GIANT_REQUIRED; 3053 3054 tries = 5; 3055 MNT_ILOCK(mp); 3056loop: 3057 TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) { 3058 if (vp->v_mount != mp) { 3059 if (--tries > 0) 3060 goto loop; 3061 break; 3062 } 3063 3064 VI_LOCK(vp); 3065 if (vp->v_iflag & VI_XLOCK) { 3066 VI_UNLOCK(vp); 3067 continue; 3068 } 3069 3070 if ((vp->v_iflag & VI_OBJDIRTY) && 3071 (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { 3072 MNT_IUNLOCK(mp); 3073 if (!vget(vp, 3074 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 3075 curthread)) { 3076 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 3077 vput(vp); 3078 MNT_ILOCK(mp); 3079 continue; 3080 } 3081 3082 if (VOP_GETVOBJECT(vp, &obj) == 0) { 3083 VM_OBJECT_LOCK(obj); 3084 vm_object_page_clean(obj, 0, 0, 3085 flags == MNT_WAIT ? 3086 OBJPC_SYNC : OBJPC_NOSYNC); 3087 VM_OBJECT_UNLOCK(obj); 3088 } 3089 vput(vp); 3090 } 3091 MNT_ILOCK(mp); 3092 if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) { 3093 if (--tries > 0) 3094 goto loop; 3095 break; 3096 } 3097 } else 3098 VI_UNLOCK(vp); 3099 } 3100 MNT_IUNLOCK(mp); 3101} 3102 3103/* 3104 * Create the VM object needed for VMIO and mmap support. This 3105 * is done for all VREG files in the system. Some filesystems might 3106 * afford the additional metadata buffering capability of the 3107 * VMIO code by making the device node be VMIO mode also. 3108 * 3109 * vp must be locked when vfs_object_create is called. 3110 */ 3111int 3112vfs_object_create(struct vnode *vp, struct thread *td, struct ucred *cred) 3113{ 3114 3115 GIANT_REQUIRED; 3116 return (VOP_CREATEVOBJECT(vp, cred, td)); 3117} 3118 3119/* 3120 * Mark a vnode as free, putting it up for recycling. 3121 */ 3122void 3123vfree(struct vnode *vp) 3124{ 3125 3126 ASSERT_VI_LOCKED(vp, "vfree"); 3127 mtx_lock(&vnode_free_list_mtx); 3128 KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free")); 3129 if (vp->v_iflag & VI_AGE) { 3130 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 3131 } else { 3132 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 3133 } 3134 freevnodes++; 3135 mtx_unlock(&vnode_free_list_mtx); 3136 vp->v_iflag &= ~VI_AGE; 3137 vp->v_iflag |= VI_FREE; 3138} 3139 3140/* 3141 * Opposite of vfree() - mark a vnode as in use. 3142 */ 3143void 3144vbusy(struct vnode *vp) 3145{ 3146 3147 ASSERT_VI_LOCKED(vp, "vbusy"); 3148 KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free")); 3149 3150 mtx_lock(&vnode_free_list_mtx); 3151 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 3152 freevnodes--; 3153 mtx_unlock(&vnode_free_list_mtx); 3154 3155 vp->v_iflag &= ~(VI_FREE|VI_AGE); 3156} 3157 3158/* 3159 * Initalize per-vnode helper structure to hold poll-related state. 3160 */ 3161void 3162v_addpollinfo(struct vnode *vp) 3163{ 3164 struct vpollinfo *vi; 3165 3166 vi = uma_zalloc(vnodepoll_zone, M_WAITOK); 3167 if (vp->v_pollinfo != NULL) { 3168 uma_zfree(vnodepoll_zone, vi); 3169 return; 3170 } 3171 vp->v_pollinfo = vi; 3172 mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 3173 knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, 3174 &vp->v_pollinfo->vpi_lock); 3175} 3176 3177/* 3178 * Record a process's interest in events which might happen to 3179 * a vnode. Because poll uses the historic select-style interface 3180 * internally, this routine serves as both the ``check for any 3181 * pending events'' and the ``record my interest in future events'' 3182 * functions. (These are done together, while the lock is held, 3183 * to avoid race conditions.) 3184 */ 3185int 3186vn_pollrecord(vp, td, events) 3187 struct vnode *vp; 3188 struct thread *td; 3189 short events; 3190{ 3191 3192 if (vp->v_pollinfo == NULL) 3193 v_addpollinfo(vp); 3194 mtx_lock(&vp->v_pollinfo->vpi_lock); 3195 if (vp->v_pollinfo->vpi_revents & events) { 3196 /* 3197 * This leaves events we are not interested 3198 * in available for the other process which 3199 * which presumably had requested them 3200 * (otherwise they would never have been 3201 * recorded). 3202 */ 3203 events &= vp->v_pollinfo->vpi_revents; 3204 vp->v_pollinfo->vpi_revents &= ~events; 3205 3206 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3207 return events; 3208 } 3209 vp->v_pollinfo->vpi_events |= events; 3210 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 3211 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3212 return 0; 3213} 3214 3215/* 3216 * Note the occurrence of an event. If the VN_POLLEVENT macro is used, 3217 * it is possible for us to miss an event due to race conditions, but 3218 * that condition is expected to be rare, so for the moment it is the 3219 * preferred interface. 3220 */ 3221void 3222vn_pollevent(vp, events) 3223 struct vnode *vp; 3224 short events; 3225{ 3226 3227 if (vp->v_pollinfo == NULL) 3228 v_addpollinfo(vp); 3229 mtx_lock(&vp->v_pollinfo->vpi_lock); 3230 if (vp->v_pollinfo->vpi_events & events) { 3231 /* 3232 * We clear vpi_events so that we don't 3233 * call selwakeup() twice if two events are 3234 * posted before the polling process(es) is 3235 * awakened. This also ensures that we take at 3236 * most one selwakeup() if the polling process 3237 * is no longer interested. However, it does 3238 * mean that only one event can be noticed at 3239 * a time. (Perhaps we should only clear those 3240 * event bits which we note?) XXX 3241 */ 3242 vp->v_pollinfo->vpi_events = 0; /* &= ~events ??? */ 3243 vp->v_pollinfo->vpi_revents |= events; 3244 selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO); 3245 } 3246 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3247} 3248 3249/* 3250 * Wake up anyone polling on vp because it is being revoked. 3251 * This depends on dead_poll() returning POLLHUP for correct 3252 * behavior. 3253 */ 3254void 3255vn_pollgone(vp) 3256 struct vnode *vp; 3257{ 3258 3259 mtx_lock(&vp->v_pollinfo->vpi_lock); 3260 VN_KNOTE_LOCKED(vp, NOTE_REVOKE); 3261 if (vp->v_pollinfo->vpi_events) { 3262 vp->v_pollinfo->vpi_events = 0; 3263 selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO); 3264 } 3265 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3266} 3267 3268 3269 3270/* 3271 * Routine to create and manage a filesystem syncer vnode. 3272 */ 3273#define sync_close ((int (*)(struct vop_close_args *))nullop) 3274static int sync_fsync(struct vop_fsync_args *); 3275static int sync_inactive(struct vop_inactive_args *); 3276static int sync_reclaim(struct vop_reclaim_args *); 3277 3278static vop_t **sync_vnodeop_p; 3279static struct vnodeopv_entry_desc sync_vnodeop_entries[] = { 3280 { &vop_default_desc, (vop_t *) vop_eopnotsupp }, 3281 { &vop_close_desc, (vop_t *) sync_close }, /* close */ 3282 { &vop_fsync_desc, (vop_t *) sync_fsync }, /* fsync */ 3283 { &vop_inactive_desc, (vop_t *) sync_inactive }, /* inactive */ 3284 { &vop_reclaim_desc, (vop_t *) sync_reclaim }, /* reclaim */ 3285 { &vop_lock_desc, (vop_t *) vop_stdlock }, /* lock */ 3286 { &vop_unlock_desc, (vop_t *) vop_stdunlock }, /* unlock */ 3287 { &vop_islocked_desc, (vop_t *) vop_stdislocked }, /* islocked */ 3288 { NULL, NULL } 3289}; 3290static struct vnodeopv_desc sync_vnodeop_opv_desc = 3291 { &sync_vnodeop_p, sync_vnodeop_entries }; 3292 3293VNODEOP_SET(sync_vnodeop_opv_desc); 3294 3295/* 3296 * Create a new filesystem syncer vnode for the specified mount point. 3297 */ 3298int 3299vfs_allocate_syncvnode(mp) 3300 struct mount *mp; 3301{ 3302 struct vnode *vp; 3303 static long start, incr, next; 3304 int error; 3305 3306 /* Allocate a new vnode */ 3307 if ((error = getnewvnode("syncer", mp, sync_vnodeop_p, &vp)) != 0) { 3308 mp->mnt_syncer = NULL; 3309 return (error); 3310 } 3311 vp->v_type = VNON; 3312 /* 3313 * Place the vnode onto the syncer worklist. We attempt to 3314 * scatter them about on the list so that they will go off 3315 * at evenly distributed times even if all the filesystems 3316 * are mounted at once. 3317 */ 3318 next += incr; 3319 if (next == 0 || next > syncer_maxdelay) { 3320 start /= 2; 3321 incr /= 2; 3322 if (start == 0) { 3323 start = syncer_maxdelay / 2; 3324 incr = syncer_maxdelay; 3325 } 3326 next = start; 3327 } 3328 VI_LOCK(vp); 3329 vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0); 3330 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 3331 mtx_lock(&sync_mtx); 3332 sync_vnode_count++; 3333 mtx_unlock(&sync_mtx); 3334 VI_UNLOCK(vp); 3335 mp->mnt_syncer = vp; 3336 return (0); 3337} 3338 3339/* 3340 * Do a lazy sync of the filesystem. 3341 */ 3342static int 3343sync_fsync(ap) 3344 struct vop_fsync_args /* { 3345 struct vnode *a_vp; 3346 struct ucred *a_cred; 3347 int a_waitfor; 3348 struct thread *a_td; 3349 } */ *ap; 3350{ 3351 struct vnode *syncvp = ap->a_vp; 3352 struct mount *mp = syncvp->v_mount; 3353 struct thread *td = ap->a_td; 3354 int error, asyncflag; 3355 3356 /* 3357 * We only need to do something if this is a lazy evaluation. 3358 */ 3359 if (ap->a_waitfor != MNT_LAZY) 3360 return (0); 3361 3362 /* 3363 * Move ourselves to the back of the sync list. 3364 */ 3365 VI_LOCK(syncvp); 3366 vn_syncer_add_to_worklist(syncvp, syncdelay); 3367 VI_UNLOCK(syncvp); 3368 3369 /* 3370 * Walk the list of vnodes pushing all that are dirty and 3371 * not already on the sync list. 3372 */ 3373 mtx_lock(&mountlist_mtx); 3374 if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { 3375 mtx_unlock(&mountlist_mtx); 3376 return (0); 3377 } 3378 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 3379 vfs_unbusy(mp, td); 3380 return (0); 3381 } 3382 asyncflag = mp->mnt_flag & MNT_ASYNC; 3383 mp->mnt_flag &= ~MNT_ASYNC; 3384 vfs_msync(mp, MNT_NOWAIT); 3385 error = VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td); 3386 if (asyncflag) 3387 mp->mnt_flag |= MNT_ASYNC; 3388 vn_finished_write(mp); 3389 vfs_unbusy(mp, td); 3390 return (error); 3391} 3392 3393/* 3394 * The syncer vnode is no referenced. 3395 */ 3396static int 3397sync_inactive(ap) 3398 struct vop_inactive_args /* { 3399 struct vnode *a_vp; 3400 struct thread *a_td; 3401 } */ *ap; 3402{ 3403 3404 VOP_UNLOCK(ap->a_vp, 0, ap->a_td); 3405 vgone(ap->a_vp); 3406 return (0); 3407} 3408 3409/* 3410 * The syncer vnode is no longer needed and is being decommissioned. 3411 * 3412 * Modifications to the worklist must be protected by sync_mtx. 3413 */ 3414static int 3415sync_reclaim(ap) 3416 struct vop_reclaim_args /* { 3417 struct vnode *a_vp; 3418 } */ *ap; 3419{ 3420 struct vnode *vp = ap->a_vp; 3421 3422 VI_LOCK(vp); 3423 vp->v_mount->mnt_syncer = NULL; 3424 if (vp->v_iflag & VI_ONWORKLST) { 3425 mtx_lock(&sync_mtx); 3426 LIST_REMOVE(vp, v_synclist); 3427 syncer_worklist_len--; 3428 sync_vnode_count--; 3429 mtx_unlock(&sync_mtx); 3430 vp->v_iflag &= ~VI_ONWORKLST; 3431 } 3432 VI_UNLOCK(vp); 3433 3434 return (0); 3435} 3436 3437/* 3438 * extract the struct cdev *from a VCHR 3439 */ 3440struct cdev * 3441vn_todev(vp) 3442 struct vnode *vp; 3443{ 3444 3445 if (vp->v_type != VCHR) 3446 return (NULL); 3447 return (vp->v_rdev); 3448} 3449 3450/* 3451 * Check if vnode represents a disk device 3452 */ 3453int 3454vn_isdisk(vp, errp) 3455 struct vnode *vp; 3456 int *errp; 3457{ 3458 int error; 3459 3460 error = 0; 3461 dev_lock(); 3462 if (vp->v_type != VCHR) 3463 error = ENOTBLK; 3464 else if (vp->v_rdev == NULL) 3465 error = ENXIO; 3466 else if (vp->v_rdev->si_devsw == NULL) 3467 error = ENXIO; 3468 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 3469 error = ENOTBLK; 3470 dev_unlock(); 3471 if (errp != NULL) 3472 *errp = error; 3473 return (error == 0); 3474} 3475 3476/* 3477 * Free data allocated by namei(); see namei(9) for details. 3478 */ 3479void 3480NDFREE(ndp, flags) 3481 struct nameidata *ndp; 3482 const u_int flags; 3483{ 3484 3485 if (!(flags & NDF_NO_FREE_PNBUF) && 3486 (ndp->ni_cnd.cn_flags & HASBUF)) { 3487 uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf); 3488 ndp->ni_cnd.cn_flags &= ~HASBUF; 3489 } 3490 if (!(flags & NDF_NO_DVP_UNLOCK) && 3491 (ndp->ni_cnd.cn_flags & LOCKPARENT) && 3492 ndp->ni_dvp != ndp->ni_vp) 3493 VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread); 3494 if (!(flags & NDF_NO_DVP_RELE) && 3495 (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) { 3496 vrele(ndp->ni_dvp); 3497 ndp->ni_dvp = NULL; 3498 } 3499 if (!(flags & NDF_NO_VP_UNLOCK) && 3500 (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp) 3501 VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread); 3502 if (!(flags & NDF_NO_VP_RELE) && 3503 ndp->ni_vp) { 3504 vrele(ndp->ni_vp); 3505 ndp->ni_vp = NULL; 3506 } 3507 if (!(flags & NDF_NO_STARTDIR_RELE) && 3508 (ndp->ni_cnd.cn_flags & SAVESTART)) { 3509 vrele(ndp->ni_startdir); 3510 ndp->ni_startdir = NULL; 3511 } 3512} 3513 3514/* 3515 * Common filesystem object access control check routine. Accepts a 3516 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3517 * and optional call-by-reference privused argument allowing vaccess() 3518 * to indicate to the caller whether privilege was used to satisfy the 3519 * request (obsoleted). Returns 0 on success, or an errno on failure. 3520 */ 3521int 3522vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused) 3523 enum vtype type; 3524 mode_t file_mode; 3525 uid_t file_uid; 3526 gid_t file_gid; 3527 mode_t acc_mode; 3528 struct ucred *cred; 3529 int *privused; 3530{ 3531 mode_t dac_granted; 3532#ifdef CAPABILITIES 3533 mode_t cap_granted; 3534#endif 3535 3536 /* 3537 * Look for a normal, non-privileged way to access the file/directory 3538 * as requested. If it exists, go with that. 3539 */ 3540 3541 if (privused != NULL) 3542 *privused = 0; 3543 3544 dac_granted = 0; 3545 3546 /* Check the owner. */ 3547 if (cred->cr_uid == file_uid) { 3548 dac_granted |= VADMIN; 3549 if (file_mode & S_IXUSR) 3550 dac_granted |= VEXEC; 3551 if (file_mode & S_IRUSR) 3552 dac_granted |= VREAD; 3553 if (file_mode & S_IWUSR) 3554 dac_granted |= (VWRITE | VAPPEND); 3555 3556 if ((acc_mode & dac_granted) == acc_mode) 3557 return (0); 3558 3559 goto privcheck; 3560 } 3561 3562 /* Otherwise, check the groups (first match) */ 3563 if (groupmember(file_gid, cred)) { 3564 if (file_mode & S_IXGRP) 3565 dac_granted |= VEXEC; 3566 if (file_mode & S_IRGRP) 3567 dac_granted |= VREAD; 3568 if (file_mode & S_IWGRP) 3569 dac_granted |= (VWRITE | VAPPEND); 3570 3571 if ((acc_mode & dac_granted) == acc_mode) 3572 return (0); 3573 3574 goto privcheck; 3575 } 3576 3577 /* Otherwise, check everyone else. */ 3578 if (file_mode & S_IXOTH) 3579 dac_granted |= VEXEC; 3580 if (file_mode & S_IROTH) 3581 dac_granted |= VREAD; 3582 if (file_mode & S_IWOTH) 3583 dac_granted |= (VWRITE | VAPPEND); 3584 if ((acc_mode & dac_granted) == acc_mode) 3585 return (0); 3586 3587privcheck: 3588 if (!suser_cred(cred, SUSER_ALLOWJAIL)) { 3589 /* XXX audit: privilege used */ 3590 if (privused != NULL) 3591 *privused = 1; 3592 return (0); 3593 } 3594 3595#ifdef CAPABILITIES 3596 /* 3597 * Build a capability mask to determine if the set of capabilities 3598 * satisfies the requirements when combined with the granted mask 3599 * from above. 3600 * For each capability, if the capability is required, bitwise 3601 * or the request type onto the cap_granted mask. 3602 */ 3603 cap_granted = 0; 3604 3605 if (type == VDIR) { 3606 /* 3607 * For directories, use CAP_DAC_READ_SEARCH to satisfy 3608 * VEXEC requests, instead of CAP_DAC_EXECUTE. 3609 */ 3610 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3611 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL)) 3612 cap_granted |= VEXEC; 3613 } else { 3614 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3615 !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL)) 3616 cap_granted |= VEXEC; 3617 } 3618 3619 if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && 3620 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL)) 3621 cap_granted |= VREAD; 3622 3623 if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3624 !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL)) 3625 cap_granted |= (VWRITE | VAPPEND); 3626 3627 if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3628 !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL)) 3629 cap_granted |= VADMIN; 3630 3631 if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { 3632 /* XXX audit: privilege used */ 3633 if (privused != NULL) 3634 *privused = 1; 3635 return (0); 3636 } 3637#endif 3638 3639 return ((acc_mode & VADMIN) ? EPERM : EACCES); 3640} 3641 3642/* 3643 * Credential check based on process requesting service, and per-attribute 3644 * permissions. 3645 */ 3646int 3647extattr_check_cred(struct vnode *vp, int attrnamespace, 3648 struct ucred *cred, struct thread *td, int access) 3649{ 3650 3651 /* 3652 * Kernel-invoked always succeeds. 3653 */ 3654 if (cred == NOCRED) 3655 return (0); 3656 3657 /* 3658 * Do not allow privileged processes in jail to directly 3659 * manipulate system attributes. 3660 * 3661 * XXX What capability should apply here? 3662 * Probably CAP_SYS_SETFFLAG. 3663 */ 3664 switch (attrnamespace) { 3665 case EXTATTR_NAMESPACE_SYSTEM: 3666 /* Potentially should be: return (EPERM); */ 3667 return (suser_cred(cred, 0)); 3668 case EXTATTR_NAMESPACE_USER: 3669 return (VOP_ACCESS(vp, access, cred, td)); 3670 default: 3671 return (EPERM); 3672 } 3673} 3674 3675#ifdef DEBUG_VFS_LOCKS 3676/* 3677 * This only exists to supress warnings from unlocked specfs accesses. It is 3678 * no longer ok to have an unlocked VFS. 3679 */ 3680#define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD) 3681 3682int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 3683SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, ""); 3684 3685int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 3686SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, ""); 3687 3688int vfs_badlock_print = 1; /* Print lock violations. */ 3689SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, ""); 3690 3691#ifdef KDB 3692int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 3693SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, ""); 3694#endif 3695 3696static void 3697vfs_badlock(const char *msg, const char *str, struct vnode *vp) 3698{ 3699 3700#ifdef KDB 3701 if (vfs_badlock_backtrace) 3702 kdb_backtrace(); 3703#endif 3704 if (vfs_badlock_print) 3705 printf("%s: %p %s\n", str, (void *)vp, msg); 3706 if (vfs_badlock_ddb) 3707 kdb_enter("lock violation"); 3708} 3709 3710void 3711assert_vi_locked(struct vnode *vp, const char *str) 3712{ 3713 3714 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 3715 vfs_badlock("interlock is not locked but should be", str, vp); 3716} 3717 3718void 3719assert_vi_unlocked(struct vnode *vp, const char *str) 3720{ 3721 3722 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 3723 vfs_badlock("interlock is locked but should not be", str, vp); 3724} 3725 3726void 3727assert_vop_locked(struct vnode *vp, const char *str) 3728{ 3729 3730 if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0) 3731 vfs_badlock("is not locked but should be", str, vp); 3732} 3733 3734void 3735assert_vop_unlocked(struct vnode *vp, const char *str) 3736{ 3737 3738 if (vp && !IGNORE_LOCK(vp) && 3739 VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE) 3740 vfs_badlock("is locked but should not be", str, vp); 3741} 3742 3743#if 0 3744void 3745assert_vop_elocked(struct vnode *vp, const char *str) 3746{ 3747 3748 if (vp && !IGNORE_LOCK(vp) && 3749 VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE) 3750 vfs_badlock("is not exclusive locked but should be", str, vp); 3751} 3752 3753void 3754assert_vop_elocked_other(struct vnode *vp, const char *str) 3755{ 3756 3757 if (vp && !IGNORE_LOCK(vp) && 3758 VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER) 3759 vfs_badlock("is not exclusive locked by another thread", 3760 str, vp); 3761} 3762 3763void 3764assert_vop_slocked(struct vnode *vp, const char *str) 3765{ 3766 3767 if (vp && !IGNORE_LOCK(vp) && 3768 VOP_ISLOCKED(vp, curthread) != LK_SHARED) 3769 vfs_badlock("is not locked shared but should be", str, vp); 3770} 3771#endif /* 0 */ 3772 3773void 3774vop_rename_pre(void *ap) 3775{ 3776 struct vop_rename_args *a = ap; 3777 3778 if (a->a_tvp) 3779 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 3780 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 3781 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 3782 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 3783 3784 /* Check the source (from). */ 3785 if (a->a_tdvp != a->a_fdvp) 3786 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 3787 if (a->a_tvp != a->a_fvp) 3788 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked"); 3789 3790 /* Check the target. */ 3791 if (a->a_tvp) 3792 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 3793 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 3794} 3795 3796void 3797vop_strategy_pre(void *ap) 3798{ 3799 struct vop_strategy_args *a; 3800 struct buf *bp; 3801 3802 a = ap; 3803 bp = a->a_bp; 3804 3805 /* 3806 * Cluster ops lock their component buffers but not the IO container. 3807 */ 3808 if ((bp->b_flags & B_CLUSTER) != 0) 3809 return; 3810 3811 if (BUF_REFCNT(bp) < 1) { 3812 if (vfs_badlock_print) 3813 printf( 3814 "VOP_STRATEGY: bp is not locked but should be\n"); 3815 if (vfs_badlock_ddb) 3816 kdb_enter("lock violation"); 3817 } 3818} 3819 3820void 3821vop_lookup_pre(void *ap) 3822{ 3823 struct vop_lookup_args *a; 3824 struct vnode *dvp; 3825 3826 a = ap; 3827 dvp = a->a_dvp; 3828 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3829 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 3830} 3831 3832void 3833vop_lookup_post(void *ap, int rc) 3834{ 3835 struct vop_lookup_args *a; 3836 struct componentname *cnp; 3837 struct vnode *dvp; 3838 struct vnode *vp; 3839 int flags; 3840 3841 a = ap; 3842 dvp = a->a_dvp; 3843 cnp = a->a_cnp; 3844 vp = *(a->a_vpp); 3845 flags = cnp->cn_flags; 3846 3847 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3848 3849 /* 3850 * If this is the last path component for this lookup and LOCKPARENT 3851 * is set, OR if there is an error the directory has to be locked. 3852 */ 3853 if ((flags & LOCKPARENT) && (flags & ISLASTCN)) 3854 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)"); 3855 else if (rc != 0) 3856 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)"); 3857 else if (dvp != vp) 3858 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)"); 3859 if (flags & PDIRUNLOCK) 3860 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)"); 3861} 3862 3863void 3864vop_lock_pre(void *ap) 3865{ 3866 struct vop_lock_args *a = ap; 3867 3868 if ((a->a_flags & LK_INTERLOCK) == 0) 3869 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3870 else 3871 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 3872} 3873 3874void 3875vop_lock_post(void *ap, int rc) 3876{ 3877 struct vop_lock_args *a = ap; 3878 3879 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3880 if (rc == 0) 3881 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 3882} 3883 3884void 3885vop_unlock_pre(void *ap) 3886{ 3887 struct vop_unlock_args *a = ap; 3888 3889 if (a->a_flags & LK_INTERLOCK) 3890 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 3891 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 3892} 3893 3894void 3895vop_unlock_post(void *ap, int rc) 3896{ 3897 struct vop_unlock_args *a = ap; 3898 3899 if (a->a_flags & LK_INTERLOCK) 3900 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 3901} 3902#endif /* DEBUG_VFS_LOCKS */ 3903 3904static struct knlist fs_knlist; 3905 3906static void 3907vfs_event_init(void *arg) 3908{ 3909 knlist_init(&fs_knlist, NULL); 3910} 3911/* XXX - correct order? */ 3912SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 3913 3914void 3915vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused) 3916{ 3917 3918 KNOTE_UNLOCKED(&fs_knlist, event); 3919} 3920 3921static int filt_fsattach(struct knote *kn); 3922static void filt_fsdetach(struct knote *kn); 3923static int filt_fsevent(struct knote *kn, long hint); 3924 3925struct filterops fs_filtops = 3926 { 0, filt_fsattach, filt_fsdetach, filt_fsevent }; 3927 3928static int 3929filt_fsattach(struct knote *kn) 3930{ 3931 3932 kn->kn_flags |= EV_CLEAR; 3933 knlist_add(&fs_knlist, kn, 0); 3934 return (0); 3935} 3936 3937static void 3938filt_fsdetach(struct knote *kn) 3939{ 3940 3941 knlist_remove(&fs_knlist, kn, 0); 3942} 3943 3944static int 3945filt_fsevent(struct knote *kn, long hint) 3946{ 3947 3948 kn->kn_fflags |= hint; 3949 return (kn->kn_fflags != 0); 3950} 3951 3952static int 3953sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 3954{ 3955 struct vfsidctl vc; 3956 int error; 3957 struct mount *mp; 3958 3959 error = SYSCTL_IN(req, &vc, sizeof(vc)); 3960 if (error) 3961 return (error); 3962 if (vc.vc_vers != VFS_CTL_VERS1) 3963 return (EINVAL); 3964 mp = vfs_getvfs(&vc.vc_fsid); 3965 if (mp == NULL) 3966 return (ENOENT); 3967 /* ensure that a specific sysctl goes to the right filesystem. */ 3968 if (strcmp(vc.vc_fstypename, "*") != 0 && 3969 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 3970 return (EINVAL); 3971 } 3972 VCTLTOREQ(&vc, req); 3973 return (VFS_SYSCTL(mp, vc.vc_op, req)); 3974} 3975 3976SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR, 3977 NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid"); 3978 3979/* 3980 * Function to initialize a va_filerev field sensibly. 3981 * XXX: Wouldn't a random number make a lot more sense ?? 3982 */ 3983u_quad_t 3984init_va_filerev(void) 3985{ 3986 struct bintime bt; 3987 3988 getbinuptime(&bt); 3989 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 3990} 3991