vfs_subr.c revision 144900
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 144900 2005-04-11 09:28:32Z jeff $"); 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 <machine/stdarg.h> 72 73#include <vm/vm.h> 74#include <vm/vm_object.h> 75#include <vm/vm_extern.h> 76#include <vm/pmap.h> 77#include <vm/vm_map.h> 78#include <vm/vm_page.h> 79#include <vm/vm_kern.h> 80#include <vm/uma.h> 81 82static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure"); 83 84static void delmntque(struct vnode *vp); 85static void insmntque(struct vnode *vp, struct mount *mp); 86static void vlruvp(struct vnode *vp); 87static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, 88 int slpflag, int slptimeo); 89static void syncer_shutdown(void *arg, int howto); 90static int vtryrecycle(struct vnode *vp); 91static void vbusy(struct vnode *vp); 92static void vdropl(struct vnode *vp); 93static void vinactive(struct vnode *, struct thread *); 94static void v_incr_usecount(struct vnode *, int); 95static void vfree(struct vnode *); 96static void vnlru_free(int); 97static void vdestroy(struct vnode *); 98 99/* 100 * Enable Giant pushdown based on whether or not the vm is mpsafe in this 101 * build. Without mpsafevm the buffer cache can not run Giant free. 102 */ 103#if defined(__alpha__) || defined(__amd64__) || defined(__i386__) 104int mpsafe_vfs = 1; 105#else 106int mpsafe_vfs; 107#endif 108TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs); 109SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0, 110 "MPSAFE VFS"); 111 112/* 113 * Number of vnodes in existence. Increased whenever getnewvnode() 114 * allocates a new vnode, never decreased. 115 */ 116static unsigned long numvnodes; 117 118SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); 119 120/* 121 * Conversion tables for conversion from vnode types to inode formats 122 * and back. 123 */ 124enum vtype iftovt_tab[16] = { 125 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 126 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 127}; 128int vttoif_tab[9] = { 129 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 130 S_IFSOCK, S_IFIFO, S_IFMT, 131}; 132 133/* 134 * List of vnodes that are ready for recycling. 135 */ 136static TAILQ_HEAD(freelst, vnode) vnode_free_list; 137 138/* 139 * Free vnode target. Free vnodes may simply be files which have been stat'd 140 * but not read. This is somewhat common, and a small cache of such files 141 * should be kept to avoid recreation costs. 142 */ 143static u_long wantfreevnodes; 144SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 145/* Number of vnodes in the free list. */ 146static u_long freevnodes; 147SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); 148 149/* 150 * Various variables used for debugging the new implementation of 151 * reassignbuf(). 152 * XXX these are probably of (very) limited utility now. 153 */ 154static int reassignbufcalls; 155SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); 156 157/* 158 * Cache for the mount type id assigned to NFS. This is used for 159 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 160 */ 161int nfs_mount_type = -1; 162 163/* To keep more than one thread at a time from running vfs_getnewfsid */ 164static struct mtx mntid_mtx; 165 166/* 167 * Lock for any access to the following: 168 * vnode_free_list 169 * numvnodes 170 * freevnodes 171 */ 172static struct mtx vnode_free_list_mtx; 173 174/* Publicly exported FS */ 175struct nfs_public nfs_pub; 176 177/* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 178static uma_zone_t vnode_zone; 179static uma_zone_t vnodepoll_zone; 180 181/* Set to 1 to print out reclaim of active vnodes */ 182int prtactive; 183 184/* 185 * The workitem queue. 186 * 187 * It is useful to delay writes of file data and filesystem metadata 188 * for tens of seconds so that quickly created and deleted files need 189 * not waste disk bandwidth being created and removed. To realize this, 190 * we append vnodes to a "workitem" queue. When running with a soft 191 * updates implementation, most pending metadata dependencies should 192 * not wait for more than a few seconds. Thus, mounted on block devices 193 * are delayed only about a half the time that file data is delayed. 194 * Similarly, directory updates are more critical, so are only delayed 195 * about a third the time that file data is delayed. Thus, there are 196 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 197 * one each second (driven off the filesystem syncer process). The 198 * syncer_delayno variable indicates the next queue that is to be processed. 199 * Items that need to be processed soon are placed in this queue: 200 * 201 * syncer_workitem_pending[syncer_delayno] 202 * 203 * A delay of fifteen seconds is done by placing the request fifteen 204 * entries later in the queue: 205 * 206 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 207 * 208 */ 209static int syncer_delayno; 210static long syncer_mask; 211LIST_HEAD(synclist, bufobj); 212static struct synclist *syncer_workitem_pending; 213/* 214 * The sync_mtx protects: 215 * bo->bo_synclist 216 * sync_vnode_count 217 * syncer_delayno 218 * syncer_state 219 * syncer_workitem_pending 220 * syncer_worklist_len 221 * rushjob 222 */ 223static struct mtx sync_mtx; 224 225#define SYNCER_MAXDELAY 32 226static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 227static int syncdelay = 30; /* max time to delay syncing data */ 228static int filedelay = 30; /* time to delay syncing files */ 229SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); 230static int dirdelay = 29; /* time to delay syncing directories */ 231SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); 232static int metadelay = 28; /* time to delay syncing metadata */ 233SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); 234static int rushjob; /* number of slots to run ASAP */ 235static int stat_rush_requests; /* number of times I/O speeded up */ 236SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); 237 238/* 239 * When shutting down the syncer, run it at four times normal speed. 240 */ 241#define SYNCER_SHUTDOWN_SPEEDUP 4 242static int sync_vnode_count; 243static int syncer_worklist_len; 244static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 245 syncer_state; 246 247/* 248 * Number of vnodes we want to exist at any one time. This is mostly used 249 * to size hash tables in vnode-related code. It is normally not used in 250 * getnewvnode(), as wantfreevnodes is normally nonzero.) 251 * 252 * XXX desiredvnodes is historical cruft and should not exist. 253 */ 254int desiredvnodes; 255SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 256 &desiredvnodes, 0, "Maximum number of vnodes"); 257SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 258 &wantfreevnodes, 0, "Minimum number of vnodes (legacy)"); 259static int vnlru_nowhere; 260SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 261 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 262 263/* Hook for calling soft updates. */ 264int (*softdep_process_worklist_hook)(struct mount *); 265 266/* 267 * Macros to control when a vnode is freed and recycled. All require 268 * the vnode interlock. 269 */ 270#define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt) 271#define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt) 272#define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt) 273 274 275/* 276 * Initialize the vnode management data structures. 277 */ 278#ifndef MAXVNODES_MAX 279#define MAXVNODES_MAX 100000 280#endif 281static void 282vntblinit(void *dummy __unused) 283{ 284 285 /* 286 * Desiredvnodes is a function of the physical memory size and 287 * the kernel's heap size. Specifically, desiredvnodes scales 288 * in proportion to the physical memory size until two fifths 289 * of the kernel's heap size is consumed by vnodes and vm 290 * objects. 291 */ 292 desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size / 293 (5 * (sizeof(struct vm_object) + sizeof(struct vnode)))); 294 if (desiredvnodes > MAXVNODES_MAX) { 295 if (bootverbose) 296 printf("Reducing kern.maxvnodes %d -> %d\n", 297 desiredvnodes, MAXVNODES_MAX); 298 desiredvnodes = MAXVNODES_MAX; 299 } 300 wantfreevnodes = desiredvnodes / 4; 301 mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF); 302 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 303 TAILQ_INIT(&vnode_free_list); 304 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 305 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 306 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 307 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 308 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 309 /* 310 * Initialize the filesystem syncer. 311 */ 312 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 313 &syncer_mask); 314 syncer_maxdelay = syncer_mask + 1; 315 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 316} 317SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) 318 319 320/* 321 * Mark a mount point as busy. Used to synchronize access and to delay 322 * unmounting. Interlock is not released on failure. 323 */ 324int 325vfs_busy(mp, flags, interlkp, td) 326 struct mount *mp; 327 int flags; 328 struct mtx *interlkp; 329 struct thread *td; 330{ 331 int lkflags; 332 333 MNT_ILOCK(mp); 334 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 335 if (flags & LK_NOWAIT) { 336 MNT_IUNLOCK(mp); 337 return (ENOENT); 338 } 339 if (interlkp) 340 mtx_unlock(interlkp); 341 mp->mnt_kern_flag |= MNTK_MWAIT; 342 /* 343 * Since all busy locks are shared except the exclusive 344 * lock granted when unmounting, the only place that a 345 * wakeup needs to be done is at the release of the 346 * exclusive lock at the end of dounmount. 347 */ 348 msleep(mp, MNT_MTX(mp), PVFS|PDROP, "vfs_busy", 0); 349 if (interlkp) 350 mtx_lock(interlkp); 351 return (ENOENT); 352 } 353 if (interlkp) 354 mtx_unlock(interlkp); 355 lkflags = LK_SHARED | LK_INTERLOCK; 356 if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td)) 357 panic("vfs_busy: unexpected lock failure"); 358 return (0); 359} 360 361/* 362 * Free a busy filesystem. 363 */ 364void 365vfs_unbusy(mp, td) 366 struct mount *mp; 367 struct thread *td; 368{ 369 370 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); 371} 372 373/* 374 * Lookup a mount point by filesystem identifier. 375 */ 376struct mount * 377vfs_getvfs(fsid) 378 fsid_t *fsid; 379{ 380 struct mount *mp; 381 382 mtx_lock(&mountlist_mtx); 383 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 384 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 385 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 386 mtx_unlock(&mountlist_mtx); 387 return (mp); 388 } 389 } 390 mtx_unlock(&mountlist_mtx); 391 return ((struct mount *) 0); 392} 393 394/* 395 * Check if a user can access priveledged mount options. 396 */ 397int 398vfs_suser(struct mount *mp, struct thread *td) 399{ 400 int error; 401 402 if ((mp->mnt_flag & MNT_USER) == 0 || 403 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 404 if ((error = suser(td)) != 0) 405 return (error); 406 } 407 return (0); 408} 409 410/* 411 * Get a new unique fsid. Try to make its val[0] unique, since this value 412 * will be used to create fake device numbers for stat(). Also try (but 413 * not so hard) make its val[0] unique mod 2^16, since some emulators only 414 * support 16-bit device numbers. We end up with unique val[0]'s for the 415 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 416 * 417 * Keep in mind that several mounts may be running in parallel. Starting 418 * the search one past where the previous search terminated is both a 419 * micro-optimization and a defense against returning the same fsid to 420 * different mounts. 421 */ 422void 423vfs_getnewfsid(mp) 424 struct mount *mp; 425{ 426 static u_int16_t mntid_base; 427 fsid_t tfsid; 428 int mtype; 429 430 mtx_lock(&mntid_mtx); 431 mtype = mp->mnt_vfc->vfc_typenum; 432 tfsid.val[1] = mtype; 433 mtype = (mtype & 0xFF) << 24; 434 for (;;) { 435 tfsid.val[0] = makedev(255, 436 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 437 mntid_base++; 438 if (vfs_getvfs(&tfsid) == NULL) 439 break; 440 } 441 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 442 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 443 mtx_unlock(&mntid_mtx); 444} 445 446/* 447 * Knob to control the precision of file timestamps: 448 * 449 * 0 = seconds only; nanoseconds zeroed. 450 * 1 = seconds and nanoseconds, accurate within 1/HZ. 451 * 2 = seconds and nanoseconds, truncated to microseconds. 452 * >=3 = seconds and nanoseconds, maximum precision. 453 */ 454enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 455 456static int timestamp_precision = TSP_SEC; 457SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 458 ×tamp_precision, 0, ""); 459 460/* 461 * Get a current timestamp. 462 */ 463void 464vfs_timestamp(tsp) 465 struct timespec *tsp; 466{ 467 struct timeval tv; 468 469 switch (timestamp_precision) { 470 case TSP_SEC: 471 tsp->tv_sec = time_second; 472 tsp->tv_nsec = 0; 473 break; 474 case TSP_HZ: 475 getnanotime(tsp); 476 break; 477 case TSP_USEC: 478 microtime(&tv); 479 TIMEVAL_TO_TIMESPEC(&tv, tsp); 480 break; 481 case TSP_NSEC: 482 default: 483 nanotime(tsp); 484 break; 485 } 486} 487 488/* 489 * Set vnode attributes to VNOVAL 490 */ 491void 492vattr_null(vap) 493 struct vattr *vap; 494{ 495 496 vap->va_type = VNON; 497 vap->va_size = VNOVAL; 498 vap->va_bytes = VNOVAL; 499 vap->va_mode = VNOVAL; 500 vap->va_nlink = VNOVAL; 501 vap->va_uid = VNOVAL; 502 vap->va_gid = VNOVAL; 503 vap->va_fsid = VNOVAL; 504 vap->va_fileid = VNOVAL; 505 vap->va_blocksize = VNOVAL; 506 vap->va_rdev = VNOVAL; 507 vap->va_atime.tv_sec = VNOVAL; 508 vap->va_atime.tv_nsec = VNOVAL; 509 vap->va_mtime.tv_sec = VNOVAL; 510 vap->va_mtime.tv_nsec = VNOVAL; 511 vap->va_ctime.tv_sec = VNOVAL; 512 vap->va_ctime.tv_nsec = VNOVAL; 513 vap->va_birthtime.tv_sec = VNOVAL; 514 vap->va_birthtime.tv_nsec = VNOVAL; 515 vap->va_flags = VNOVAL; 516 vap->va_gen = VNOVAL; 517 vap->va_vaflags = 0; 518} 519 520/* 521 * This routine is called when we have too many vnodes. It attempts 522 * to free <count> vnodes and will potentially free vnodes that still 523 * have VM backing store (VM backing store is typically the cause 524 * of a vnode blowout so we want to do this). Therefore, this operation 525 * is not considered cheap. 526 * 527 * A number of conditions may prevent a vnode from being reclaimed. 528 * the buffer cache may have references on the vnode, a directory 529 * vnode may still have references due to the namei cache representing 530 * underlying files, or the vnode may be in active use. It is not 531 * desireable to reuse such vnodes. These conditions may cause the 532 * number of vnodes to reach some minimum value regardless of what 533 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 534 */ 535static int 536vlrureclaim(struct mount *mp) 537{ 538 struct vnode *vp; 539 int done; 540 int trigger; 541 int usevnodes; 542 int count; 543 544 /* 545 * Calculate the trigger point, don't allow user 546 * screwups to blow us up. This prevents us from 547 * recycling vnodes with lots of resident pages. We 548 * aren't trying to free memory, we are trying to 549 * free vnodes. 550 */ 551 usevnodes = desiredvnodes; 552 if (usevnodes <= 0) 553 usevnodes = 1; 554 trigger = cnt.v_page_count * 2 / usevnodes; 555 556 done = 0; 557 vn_start_write(NULL, &mp, V_WAIT); 558 MNT_ILOCK(mp); 559 count = mp->mnt_nvnodelistsize / 10 + 1; 560 while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) { 561 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 562 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 563 564 if (vp->v_type != VNON && 565 vp->v_type != VBAD && 566 VI_TRYLOCK(vp)) { 567 /* critical path opt */ 568 if (LIST_EMPTY(&(vp)->v_cache_src) && 569 !(vp)->v_usecount && 570 (vp->v_object == NULL || 571 vp->v_object->resident_page_count < trigger)) { 572 struct thread *td; 573 574 td = curthread; 575 MNT_IUNLOCK(mp); 576 VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE, td); 577 if ((vp->v_iflag & VI_DOOMED) == 0) 578 vgone(vp); 579 VOP_UNLOCK(vp, 0, td); 580 done++; 581 MNT_ILOCK(mp); 582 } else 583 VI_UNLOCK(vp); 584 } 585 --count; 586 } 587 MNT_IUNLOCK(mp); 588 vn_finished_write(mp); 589 return done; 590} 591 592/* 593 * Attempt to keep the free list at wantfreevnodes length. 594 */ 595static void 596vnlru_free(int count) 597{ 598 struct vnode *vp; 599 600 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 601 for (; count > 0; count--) { 602 vp = TAILQ_FIRST(&vnode_free_list); 603 /* 604 * The list can be modified while the free_list_mtx 605 * has been dropped and vp could be NULL here. 606 */ 607 if (!vp) 608 break; 609 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 610 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 611 /* 612 * Don't recycle if we can't get the interlock. 613 */ 614 if (!VI_TRYLOCK(vp)) 615 continue; 616 if (!VCANRECYCLE(vp)) { 617 VI_UNLOCK(vp); 618 continue; 619 } 620 /* 621 * We assume success to avoid having to relock the frelist 622 * in the common case, simply restore counts on failure. 623 */ 624 freevnodes--; 625 numvnodes--; 626 mtx_unlock(&vnode_free_list_mtx); 627 if (vtryrecycle(vp) != 0) { 628 mtx_lock(&vnode_free_list_mtx); 629 freevnodes++; 630 numvnodes++; 631 continue; 632 } 633 vdestroy(vp); 634 mtx_lock(&vnode_free_list_mtx); 635 } 636} 637/* 638 * Attempt to recycle vnodes in a context that is always safe to block. 639 * Calling vlrurecycle() from the bowels of filesystem code has some 640 * interesting deadlock problems. 641 */ 642static struct proc *vnlruproc; 643static int vnlruproc_sig; 644 645static void 646vnlru_proc(void) 647{ 648 struct mount *mp, *nmp; 649 int done; 650 struct proc *p = vnlruproc; 651 struct thread *td = FIRST_THREAD_IN_PROC(p); 652 653 mtx_lock(&Giant); 654 655 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 656 SHUTDOWN_PRI_FIRST); 657 658 for (;;) { 659 kthread_suspend_check(p); 660 mtx_lock(&vnode_free_list_mtx); 661 if (freevnodes > wantfreevnodes) 662 vnlru_free(freevnodes - wantfreevnodes); 663 if (numvnodes <= desiredvnodes * 9 / 10) { 664 vnlruproc_sig = 0; 665 wakeup(&vnlruproc_sig); 666 msleep(vnlruproc, &vnode_free_list_mtx, 667 PVFS|PDROP, "vlruwt", hz); 668 continue; 669 } 670 mtx_unlock(&vnode_free_list_mtx); 671 done = 0; 672 mtx_lock(&mountlist_mtx); 673 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 674 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 675 nmp = TAILQ_NEXT(mp, mnt_list); 676 continue; 677 } 678 done += vlrureclaim(mp); 679 mtx_lock(&mountlist_mtx); 680 nmp = TAILQ_NEXT(mp, mnt_list); 681 vfs_unbusy(mp, td); 682 } 683 mtx_unlock(&mountlist_mtx); 684 if (done == 0) { 685#if 0 686 /* These messages are temporary debugging aids */ 687 if (vnlru_nowhere < 5) 688 printf("vnlru process getting nowhere..\n"); 689 else if (vnlru_nowhere == 5) 690 printf("vnlru process messages stopped.\n"); 691#endif 692 vnlru_nowhere++; 693 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 694 } 695 } 696} 697 698static struct kproc_desc vnlru_kp = { 699 "vnlru", 700 vnlru_proc, 701 &vnlruproc 702}; 703SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 704 705/* 706 * Routines having to do with the management of the vnode table. 707 */ 708 709static void 710vdestroy(struct vnode *vp) 711{ 712 struct bufobj *bo; 713 714 bo = &vp->v_bufobj; 715 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 716 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 717 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 718 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 719 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 720 VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL")); 721 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 722 VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL")); 723#ifdef MAC 724 mac_destroy_vnode(vp); 725#endif 726 if (vp->v_pollinfo != NULL) { 727 knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note); 728 mtx_destroy(&vp->v_pollinfo->vpi_lock); 729 uma_zfree(vnodepoll_zone, vp->v_pollinfo); 730 } 731 lockdestroy(vp->v_vnlock); 732 mtx_destroy(&vp->v_interlock); 733 uma_zfree(vnode_zone, vp); 734} 735 736/* 737 * Check to see if a free vnode can be recycled. If it can, 738 * recycle it and return it with the vnode interlock held. 739 */ 740static int 741vtryrecycle(struct vnode *vp) 742{ 743 struct thread *td = curthread; 744 struct mount *vnmp; 745 int error; 746 747 ASSERT_VI_LOCKED(vp, "vtryrecycle"); 748 error = 0; 749 /* 750 * This vnode may found and locked via some other list, if so we 751 * can't recycle it yet. 752 */ 753 if (VOP_LOCK(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 754 return (EWOULDBLOCK); 755 /* 756 * Don't recycle if its filesystem is being suspended. 757 */ 758 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 759 VOP_UNLOCK(vp, 0, td); 760 return (EBUSY); 761 } 762 /* 763 * If we got this far, we need to acquire the interlock and see if 764 * anyone picked up this vnode from another list. If not, we will 765 * mark it with DOOMED via vgonel() so that anyone who does find it 766 * will skip over it. 767 */ 768 VI_LOCK(vp); 769 if (!VCANRECYCLE(vp)) { 770 VI_UNLOCK(vp); 771 error = EBUSY; 772 goto done; 773 } 774 mtx_lock(&vnode_free_list_mtx); 775 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 776 vp->v_iflag &= ~VI_FREE; 777 mtx_unlock(&vnode_free_list_mtx); 778 if ((vp->v_iflag & VI_DOOMED) == 0) { 779 vp->v_iflag |= VI_DOOMED; 780 vgonel(vp, td); 781 VI_LOCK(vp); 782 } 783 /* 784 * If someone ref'd the vnode while we were cleaning, we have to 785 * free it once the last ref is dropped. 786 */ 787 if (vp->v_holdcnt) 788 error = EBUSY; 789 VI_UNLOCK(vp); 790done: 791 VOP_UNLOCK(vp, 0, td); 792 vn_finished_write(vnmp); 793 return (error); 794} 795 796/* 797 * Return the next vnode from the free list. 798 */ 799int 800getnewvnode(tag, mp, vops, vpp) 801 const char *tag; 802 struct mount *mp; 803 struct vop_vector *vops; 804 struct vnode **vpp; 805{ 806 struct vnode *vp = NULL; 807 struct bufobj *bo; 808 809 mtx_lock(&vnode_free_list_mtx); 810 /* 811 * Lend our context to reclaim vnodes if they've exceeded the max. 812 */ 813 if (freevnodes > wantfreevnodes) 814 vnlru_free(1); 815 /* 816 * Wait for available vnodes. 817 */ 818 if (numvnodes > desiredvnodes) { 819 if (vnlruproc_sig == 0) { 820 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 821 wakeup(vnlruproc); 822 } 823 msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS, 824 "vlruwk", hz); 825 if (numvnodes > desiredvnodes) { 826 mtx_unlock(&vnode_free_list_mtx); 827 return (ENFILE); 828 } 829 } 830 numvnodes++; 831 mtx_unlock(&vnode_free_list_mtx); 832 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); 833 /* 834 * Setup locks. 835 */ 836 vp->v_vnlock = &vp->v_lock; 837 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 838 /* 839 * By default, don't allow shared locks unless filesystems 840 * opt-in. 841 */ 842 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE); 843 /* 844 * Initialize bufobj. 845 */ 846 bo = &vp->v_bufobj; 847 bo->__bo_vnode = vp; 848 bo->bo_mtx = &vp->v_interlock; 849 bo->bo_ops = &buf_ops_bio; 850 bo->bo_private = vp; 851 TAILQ_INIT(&bo->bo_clean.bv_hd); 852 TAILQ_INIT(&bo->bo_dirty.bv_hd); 853 /* 854 * Initialize namecache. 855 */ 856 LIST_INIT(&vp->v_cache_src); 857 TAILQ_INIT(&vp->v_cache_dst); 858 /* 859 * Finalize various vnode identity bits. 860 */ 861 vp->v_type = VNON; 862 vp->v_tag = tag; 863 vp->v_op = vops; 864 v_incr_usecount(vp, 1); 865 vp->v_data = 0; 866#ifdef MAC 867 mac_init_vnode(vp); 868 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 869 mac_associate_vnode_singlelabel(mp, vp); 870 else if (mp == NULL) 871 printf("NULL mp in getnewvnode()\n"); 872#endif 873 delmntque(vp); 874 if (mp != NULL) { 875 insmntque(vp, mp); 876 bo->bo_bsize = mp->mnt_stat.f_iosize; 877 } 878 879 *vpp = vp; 880 return (0); 881} 882 883/* 884 * Delete from old mount point vnode list, if on one. 885 */ 886static void 887delmntque(struct vnode *vp) 888{ 889 struct mount *mp; 890 891 if (vp->v_mount == NULL) 892 return; 893 mp = vp->v_mount; 894 MNT_ILOCK(mp); 895 vp->v_mount = NULL; 896 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 897 ("bad mount point vnode list size")); 898 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 899 mp->mnt_nvnodelistsize--; 900 MNT_IUNLOCK(mp); 901} 902 903/* 904 * Insert into list of vnodes for the new mount point, if available. 905 */ 906static void 907insmntque(struct vnode *vp, struct mount *mp) 908{ 909 910 vp->v_mount = mp; 911 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 912 MNT_ILOCK(vp->v_mount); 913 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 914 mp->mnt_nvnodelistsize++; 915 MNT_IUNLOCK(vp->v_mount); 916} 917 918/* 919 * Flush out and invalidate all buffers associated with a bufobj 920 * Called with the underlying object locked. 921 */ 922int 923bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag, int slptimeo) 924{ 925 int error; 926 927 BO_LOCK(bo); 928 if (flags & V_SAVE) { 929 error = bufobj_wwait(bo, slpflag, slptimeo); 930 if (error) { 931 BO_UNLOCK(bo); 932 return (error); 933 } 934 if (bo->bo_dirty.bv_cnt > 0) { 935 BO_UNLOCK(bo); 936 if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0) 937 return (error); 938 /* 939 * XXX We could save a lock/unlock if this was only 940 * enabled under INVARIANTS 941 */ 942 BO_LOCK(bo); 943 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 944 panic("vinvalbuf: dirty bufs"); 945 } 946 } 947 /* 948 * If you alter this loop please notice that interlock is dropped and 949 * reacquired in flushbuflist. Special care is needed to ensure that 950 * no race conditions occur from this. 951 */ 952 do { 953 error = flushbuflist(&bo->bo_clean, 954 flags, bo, slpflag, slptimeo); 955 if (error == 0) 956 error = flushbuflist(&bo->bo_dirty, 957 flags, bo, slpflag, slptimeo); 958 if (error != 0 && error != EAGAIN) { 959 BO_UNLOCK(bo); 960 return (error); 961 } 962 } while (error != 0); 963 964 /* 965 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 966 * have write I/O in-progress but if there is a VM object then the 967 * VM object can also have read-I/O in-progress. 968 */ 969 do { 970 bufobj_wwait(bo, 0, 0); 971 BO_UNLOCK(bo); 972 if (bo->bo_object != NULL) { 973 VM_OBJECT_LOCK(bo->bo_object); 974 vm_object_pip_wait(bo->bo_object, "bovlbx"); 975 VM_OBJECT_UNLOCK(bo->bo_object); 976 } 977 BO_LOCK(bo); 978 } while (bo->bo_numoutput > 0); 979 BO_UNLOCK(bo); 980 981 /* 982 * Destroy the copy in the VM cache, too. 983 */ 984 if (bo->bo_object != NULL) { 985 VM_OBJECT_LOCK(bo->bo_object); 986 vm_object_page_remove(bo->bo_object, 0, 0, 987 (flags & V_SAVE) ? TRUE : FALSE); 988 VM_OBJECT_UNLOCK(bo->bo_object); 989 } 990 991#ifdef INVARIANTS 992 BO_LOCK(bo); 993 if ((flags & (V_ALT | V_NORMAL)) == 0 && 994 (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0)) 995 panic("vinvalbuf: flush failed"); 996 BO_UNLOCK(bo); 997#endif 998 return (0); 999} 1000 1001/* 1002 * Flush out and invalidate all buffers associated with a vnode. 1003 * Called with the underlying object locked. 1004 */ 1005int 1006vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag, int slptimeo) 1007{ 1008 1009 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1010 return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo)); 1011} 1012 1013/* 1014 * Flush out buffers on the specified list. 1015 * 1016 */ 1017static int 1018flushbuflist(bufv, flags, bo, slpflag, slptimeo) 1019 struct bufv *bufv; 1020 int flags; 1021 struct bufobj *bo; 1022 int slpflag, slptimeo; 1023{ 1024 struct buf *bp, *nbp; 1025 int retval, error; 1026 1027 ASSERT_BO_LOCKED(bo); 1028 1029 retval = 0; 1030 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 1031 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1032 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1033 continue; 1034 } 1035 retval = EAGAIN; 1036 error = BUF_TIMELOCK(bp, 1037 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo), 1038 "flushbuf", slpflag, slptimeo); 1039 if (error) { 1040 BO_LOCK(bo); 1041 return (error != ENOLCK ? error : EAGAIN); 1042 } 1043 KASSERT(bp->b_bufobj == bo, 1044 ("wrong b_bufobj %p should be %p", bp->b_bufobj, bo)); 1045 if (bp->b_bufobj != bo) { /* XXX: necessary ? */ 1046 BUF_UNLOCK(bp); 1047 BO_LOCK(bo); 1048 return (EAGAIN); 1049 } 1050 /* 1051 * XXX Since there are no node locks for NFS, I 1052 * believe there is a slight chance that a delayed 1053 * write will occur while sleeping just above, so 1054 * check for it. 1055 */ 1056 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1057 (flags & V_SAVE)) { 1058 bremfree(bp); 1059 bp->b_flags |= B_ASYNC; 1060 bwrite(bp); 1061 BO_LOCK(bo); 1062 return (EAGAIN); /* XXX: why not loop ? */ 1063 } 1064 bremfree(bp); 1065 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 1066 bp->b_flags &= ~B_ASYNC; 1067 brelse(bp); 1068 BO_LOCK(bo); 1069 } 1070 return (retval); 1071} 1072 1073/* 1074 * Truncate a file's buffer and pages to a specified length. This 1075 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1076 * sync activity. 1077 */ 1078int 1079vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, off_t length, int blksize) 1080{ 1081 struct buf *bp, *nbp; 1082 int anyfreed; 1083 int trunclbn; 1084 struct bufobj *bo; 1085 1086 /* 1087 * Round up to the *next* lbn. 1088 */ 1089 trunclbn = (length + blksize - 1) / blksize; 1090 1091 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1092restart: 1093 VI_LOCK(vp); 1094 bo = &vp->v_bufobj; 1095 anyfreed = 1; 1096 for (;anyfreed;) { 1097 anyfreed = 0; 1098 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 1099 if (bp->b_lblkno < trunclbn) 1100 continue; 1101 if (BUF_LOCK(bp, 1102 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1103 VI_MTX(vp)) == ENOLCK) 1104 goto restart; 1105 1106 bremfree(bp); 1107 bp->b_flags |= (B_INVAL | B_RELBUF); 1108 bp->b_flags &= ~B_ASYNC; 1109 brelse(bp); 1110 anyfreed = 1; 1111 1112 if (nbp != NULL && 1113 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1114 (nbp->b_vp != vp) || 1115 (nbp->b_flags & B_DELWRI))) { 1116 goto restart; 1117 } 1118 VI_LOCK(vp); 1119 } 1120 1121 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1122 if (bp->b_lblkno < trunclbn) 1123 continue; 1124 if (BUF_LOCK(bp, 1125 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1126 VI_MTX(vp)) == ENOLCK) 1127 goto restart; 1128 bremfree(bp); 1129 bp->b_flags |= (B_INVAL | B_RELBUF); 1130 bp->b_flags &= ~B_ASYNC; 1131 brelse(bp); 1132 anyfreed = 1; 1133 if (nbp != NULL && 1134 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1135 (nbp->b_vp != vp) || 1136 (nbp->b_flags & B_DELWRI) == 0)) { 1137 goto restart; 1138 } 1139 VI_LOCK(vp); 1140 } 1141 } 1142 1143 if (length > 0) { 1144restartsync: 1145 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1146 if (bp->b_lblkno > 0) 1147 continue; 1148 /* 1149 * Since we hold the vnode lock this should only 1150 * fail if we're racing with the buf daemon. 1151 */ 1152 if (BUF_LOCK(bp, 1153 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1154 VI_MTX(vp)) == ENOLCK) { 1155 goto restart; 1156 } 1157 VNASSERT((bp->b_flags & B_DELWRI), vp, 1158 ("buf(%p) on dirty queue without DELWRI", bp)); 1159 1160 bremfree(bp); 1161 bawrite(bp); 1162 VI_LOCK(vp); 1163 goto restartsync; 1164 } 1165 } 1166 1167 bufobj_wwait(bo, 0, 0); 1168 VI_UNLOCK(vp); 1169 vnode_pager_setsize(vp, length); 1170 1171 return (0); 1172} 1173 1174/* 1175 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1176 * a vnode. 1177 * 1178 * NOTE: We have to deal with the special case of a background bitmap 1179 * buffer, a situation where two buffers will have the same logical 1180 * block offset. We want (1) only the foreground buffer to be accessed 1181 * in a lookup and (2) must differentiate between the foreground and 1182 * background buffer in the splay tree algorithm because the splay 1183 * tree cannot normally handle multiple entities with the same 'index'. 1184 * We accomplish this by adding differentiating flags to the splay tree's 1185 * numerical domain. 1186 */ 1187static 1188struct buf * 1189buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1190{ 1191 struct buf dummy; 1192 struct buf *lefttreemax, *righttreemin, *y; 1193 1194 if (root == NULL) 1195 return (NULL); 1196 lefttreemax = righttreemin = &dummy; 1197 for (;;) { 1198 if (lblkno < root->b_lblkno || 1199 (lblkno == root->b_lblkno && 1200 (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1201 if ((y = root->b_left) == NULL) 1202 break; 1203 if (lblkno < y->b_lblkno) { 1204 /* Rotate right. */ 1205 root->b_left = y->b_right; 1206 y->b_right = root; 1207 root = y; 1208 if ((y = root->b_left) == NULL) 1209 break; 1210 } 1211 /* Link into the new root's right tree. */ 1212 righttreemin->b_left = root; 1213 righttreemin = root; 1214 } else if (lblkno > root->b_lblkno || 1215 (lblkno == root->b_lblkno && 1216 (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { 1217 if ((y = root->b_right) == NULL) 1218 break; 1219 if (lblkno > y->b_lblkno) { 1220 /* Rotate left. */ 1221 root->b_right = y->b_left; 1222 y->b_left = root; 1223 root = y; 1224 if ((y = root->b_right) == NULL) 1225 break; 1226 } 1227 /* Link into the new root's left tree. */ 1228 lefttreemax->b_right = root; 1229 lefttreemax = root; 1230 } else { 1231 break; 1232 } 1233 root = y; 1234 } 1235 /* Assemble the new root. */ 1236 lefttreemax->b_right = root->b_left; 1237 righttreemin->b_left = root->b_right; 1238 root->b_left = dummy.b_right; 1239 root->b_right = dummy.b_left; 1240 return (root); 1241} 1242 1243static void 1244buf_vlist_remove(struct buf *bp) 1245{ 1246 struct buf *root; 1247 struct bufv *bv; 1248 1249 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1250 ASSERT_BO_LOCKED(bp->b_bufobj); 1251 if (bp->b_xflags & BX_VNDIRTY) 1252 bv = &bp->b_bufobj->bo_dirty; 1253 else 1254 bv = &bp->b_bufobj->bo_clean; 1255 if (bp != bv->bv_root) { 1256 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1257 KASSERT(root == bp, ("splay lookup failed in remove")); 1258 } 1259 if (bp->b_left == NULL) { 1260 root = bp->b_right; 1261 } else { 1262 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1263 root->b_right = bp->b_right; 1264 } 1265 bv->bv_root = root; 1266 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 1267 bv->bv_cnt--; 1268 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1269} 1270 1271/* 1272 * Add the buffer to the sorted clean or dirty block list using a 1273 * splay tree algorithm. 1274 * 1275 * NOTE: xflags is passed as a constant, optimizing this inline function! 1276 */ 1277static void 1278buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 1279{ 1280 struct buf *root; 1281 struct bufv *bv; 1282 1283 ASSERT_BO_LOCKED(bo); 1284 bp->b_xflags |= xflags; 1285 if (xflags & BX_VNDIRTY) 1286 bv = &bo->bo_dirty; 1287 else 1288 bv = &bo->bo_clean; 1289 1290 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1291 if (root == NULL) { 1292 bp->b_left = NULL; 1293 bp->b_right = NULL; 1294 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 1295 } else if (bp->b_lblkno < root->b_lblkno || 1296 (bp->b_lblkno == root->b_lblkno && 1297 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1298 bp->b_left = root->b_left; 1299 bp->b_right = root; 1300 root->b_left = NULL; 1301 TAILQ_INSERT_BEFORE(root, bp, b_bobufs); 1302 } else { 1303 bp->b_right = root->b_right; 1304 bp->b_left = root; 1305 root->b_right = NULL; 1306 TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs); 1307 } 1308 bv->bv_cnt++; 1309 bv->bv_root = bp; 1310} 1311 1312/* 1313 * Lookup a buffer using the splay tree. Note that we specifically avoid 1314 * shadow buffers used in background bitmap writes. 1315 * 1316 * This code isn't quite efficient as it could be because we are maintaining 1317 * two sorted lists and do not know which list the block resides in. 1318 * 1319 * During a "make buildworld" the desired buffer is found at one of 1320 * the roots more than 60% of the time. Thus, checking both roots 1321 * before performing either splay eliminates unnecessary splays on the 1322 * first tree splayed. 1323 */ 1324struct buf * 1325gbincore(struct bufobj *bo, daddr_t lblkno) 1326{ 1327 struct buf *bp; 1328 1329 ASSERT_BO_LOCKED(bo); 1330 if ((bp = bo->bo_clean.bv_root) != NULL && 1331 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1332 return (bp); 1333 if ((bp = bo->bo_dirty.bv_root) != NULL && 1334 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1335 return (bp); 1336 if ((bp = bo->bo_clean.bv_root) != NULL) { 1337 bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp); 1338 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1339 return (bp); 1340 } 1341 if ((bp = bo->bo_dirty.bv_root) != NULL) { 1342 bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp); 1343 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1344 return (bp); 1345 } 1346 return (NULL); 1347} 1348 1349/* 1350 * Associate a buffer with a vnode. 1351 */ 1352void 1353bgetvp(struct vnode *vp, struct buf *bp) 1354{ 1355 1356 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 1357 1358 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 1359 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 1360 ("bgetvp: bp already attached! %p", bp)); 1361 1362 ASSERT_VI_LOCKED(vp, "bgetvp"); 1363 vholdl(vp); 1364 bp->b_vp = vp; 1365 bp->b_bufobj = &vp->v_bufobj; 1366 /* 1367 * Insert onto list for new vnode. 1368 */ 1369 buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN); 1370} 1371 1372/* 1373 * Disassociate a buffer from a vnode. 1374 */ 1375void 1376brelvp(struct buf *bp) 1377{ 1378 struct bufobj *bo; 1379 struct vnode *vp; 1380 1381 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1382 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1383 1384 /* 1385 * Delete from old vnode list, if on one. 1386 */ 1387 vp = bp->b_vp; /* XXX */ 1388 bo = bp->b_bufobj; 1389 BO_LOCK(bo); 1390 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1391 buf_vlist_remove(bp); 1392 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1393 bo->bo_flag &= ~BO_ONWORKLST; 1394 mtx_lock(&sync_mtx); 1395 LIST_REMOVE(bo, bo_synclist); 1396 syncer_worklist_len--; 1397 mtx_unlock(&sync_mtx); 1398 } 1399 vdropl(vp); 1400 bp->b_vp = NULL; 1401 bp->b_bufobj = NULL; 1402 BO_UNLOCK(bo); 1403} 1404 1405/* 1406 * Add an item to the syncer work queue. 1407 */ 1408static void 1409vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 1410{ 1411 int slot; 1412 1413 ASSERT_BO_LOCKED(bo); 1414 1415 mtx_lock(&sync_mtx); 1416 if (bo->bo_flag & BO_ONWORKLST) 1417 LIST_REMOVE(bo, bo_synclist); 1418 else { 1419 bo->bo_flag |= BO_ONWORKLST; 1420 syncer_worklist_len++; 1421 } 1422 1423 if (delay > syncer_maxdelay - 2) 1424 delay = syncer_maxdelay - 2; 1425 slot = (syncer_delayno + delay) & syncer_mask; 1426 1427 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 1428 mtx_unlock(&sync_mtx); 1429} 1430 1431static int 1432sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 1433{ 1434 int error, len; 1435 1436 mtx_lock(&sync_mtx); 1437 len = syncer_worklist_len - sync_vnode_count; 1438 mtx_unlock(&sync_mtx); 1439 error = SYSCTL_OUT(req, &len, sizeof(len)); 1440 return (error); 1441} 1442 1443SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0, 1444 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 1445 1446struct proc *updateproc; 1447static void sched_sync(void); 1448static struct kproc_desc up_kp = { 1449 "syncer", 1450 sched_sync, 1451 &updateproc 1452}; 1453SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 1454 1455static int 1456sync_vnode(struct bufobj *bo, struct thread *td) 1457{ 1458 struct vnode *vp; 1459 struct mount *mp; 1460 1461 vp = bo->__bo_vnode; /* XXX */ 1462 if (VOP_ISLOCKED(vp, NULL) != 0) 1463 return (1); 1464 if (VI_TRYLOCK(vp) == 0) 1465 return (1); 1466 /* 1467 * We use vhold in case the vnode does not 1468 * successfully sync. vhold prevents the vnode from 1469 * going away when we unlock the sync_mtx so that 1470 * we can acquire the vnode interlock. 1471 */ 1472 vholdl(vp); 1473 mtx_unlock(&sync_mtx); 1474 VI_UNLOCK(vp); 1475 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1476 vdrop(vp); 1477 mtx_lock(&sync_mtx); 1478 return (1); 1479 } 1480 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 1481 (void) VOP_FSYNC(vp, MNT_LAZY, td); 1482 VOP_UNLOCK(vp, 0, td); 1483 vn_finished_write(mp); 1484 VI_LOCK(vp); 1485 if ((bo->bo_flag & BO_ONWORKLST) != 0) { 1486 /* 1487 * Put us back on the worklist. The worklist 1488 * routine will remove us from our current 1489 * position and then add us back in at a later 1490 * position. 1491 */ 1492 vn_syncer_add_to_worklist(bo, syncdelay); 1493 } 1494 vdropl(vp); 1495 VI_UNLOCK(vp); 1496 mtx_lock(&sync_mtx); 1497 return (0); 1498} 1499 1500/* 1501 * System filesystem synchronizer daemon. 1502 */ 1503static void 1504sched_sync(void) 1505{ 1506 struct synclist *next; 1507 struct synclist *slp; 1508 struct bufobj *bo; 1509 long starttime; 1510 struct thread *td = FIRST_THREAD_IN_PROC(updateproc); 1511 static int dummychan; 1512 int last_work_seen; 1513 int net_worklist_len; 1514 int syncer_final_iter; 1515 int first_printf; 1516 int error; 1517 1518 mtx_lock(&Giant); 1519 last_work_seen = 0; 1520 syncer_final_iter = 0; 1521 first_printf = 1; 1522 syncer_state = SYNCER_RUNNING; 1523 starttime = time_second; 1524 1525 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 1526 SHUTDOWN_PRI_LAST); 1527 1528 for (;;) { 1529 mtx_lock(&sync_mtx); 1530 if (syncer_state == SYNCER_FINAL_DELAY && 1531 syncer_final_iter == 0) { 1532 mtx_unlock(&sync_mtx); 1533 kthread_suspend_check(td->td_proc); 1534 mtx_lock(&sync_mtx); 1535 } 1536 net_worklist_len = syncer_worklist_len - sync_vnode_count; 1537 if (syncer_state != SYNCER_RUNNING && 1538 starttime != time_second) { 1539 if (first_printf) { 1540 printf("\nSyncing disks, vnodes remaining..."); 1541 first_printf = 0; 1542 } 1543 printf("%d ", net_worklist_len); 1544 } 1545 starttime = time_second; 1546 1547 /* 1548 * Push files whose dirty time has expired. Be careful 1549 * of interrupt race on slp queue. 1550 * 1551 * Skip over empty worklist slots when shutting down. 1552 */ 1553 do { 1554 slp = &syncer_workitem_pending[syncer_delayno]; 1555 syncer_delayno += 1; 1556 if (syncer_delayno == syncer_maxdelay) 1557 syncer_delayno = 0; 1558 next = &syncer_workitem_pending[syncer_delayno]; 1559 /* 1560 * If the worklist has wrapped since the 1561 * it was emptied of all but syncer vnodes, 1562 * switch to the FINAL_DELAY state and run 1563 * for one more second. 1564 */ 1565 if (syncer_state == SYNCER_SHUTTING_DOWN && 1566 net_worklist_len == 0 && 1567 last_work_seen == syncer_delayno) { 1568 syncer_state = SYNCER_FINAL_DELAY; 1569 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 1570 } 1571 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 1572 syncer_worklist_len > 0); 1573 1574 /* 1575 * Keep track of the last time there was anything 1576 * on the worklist other than syncer vnodes. 1577 * Return to the SHUTTING_DOWN state if any 1578 * new work appears. 1579 */ 1580 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 1581 last_work_seen = syncer_delayno; 1582 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 1583 syncer_state = SYNCER_SHUTTING_DOWN; 1584 while ((bo = LIST_FIRST(slp)) != NULL) { 1585 error = sync_vnode(bo, td); 1586 if (error == 1) { 1587 LIST_REMOVE(bo, bo_synclist); 1588 LIST_INSERT_HEAD(next, bo, bo_synclist); 1589 continue; 1590 } 1591 } 1592 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 1593 syncer_final_iter--; 1594 mtx_unlock(&sync_mtx); 1595 1596 /* 1597 * Do soft update processing. 1598 */ 1599 if (softdep_process_worklist_hook != NULL) 1600 (*softdep_process_worklist_hook)(NULL); 1601 1602 /* 1603 * The variable rushjob allows the kernel to speed up the 1604 * processing of the filesystem syncer process. A rushjob 1605 * value of N tells the filesystem syncer to process the next 1606 * N seconds worth of work on its queue ASAP. Currently rushjob 1607 * is used by the soft update code to speed up the filesystem 1608 * syncer process when the incore state is getting so far 1609 * ahead of the disk that the kernel memory pool is being 1610 * threatened with exhaustion. 1611 */ 1612 mtx_lock(&sync_mtx); 1613 if (rushjob > 0) { 1614 rushjob -= 1; 1615 mtx_unlock(&sync_mtx); 1616 continue; 1617 } 1618 mtx_unlock(&sync_mtx); 1619 /* 1620 * Just sleep for a short period if time between 1621 * iterations when shutting down to allow some I/O 1622 * to happen. 1623 * 1624 * If it has taken us less than a second to process the 1625 * current work, then wait. Otherwise start right over 1626 * again. We can still lose time if any single round 1627 * takes more than two seconds, but it does not really 1628 * matter as we are just trying to generally pace the 1629 * filesystem activity. 1630 */ 1631 if (syncer_state != SYNCER_RUNNING) 1632 tsleep(&dummychan, PPAUSE, "syncfnl", 1633 hz / SYNCER_SHUTDOWN_SPEEDUP); 1634 else if (time_second == starttime) 1635 tsleep(&lbolt, PPAUSE, "syncer", 0); 1636 } 1637} 1638 1639/* 1640 * Request the syncer daemon to speed up its work. 1641 * We never push it to speed up more than half of its 1642 * normal turn time, otherwise it could take over the cpu. 1643 */ 1644int 1645speedup_syncer() 1646{ 1647 struct thread *td; 1648 int ret = 0; 1649 1650 td = FIRST_THREAD_IN_PROC(updateproc); 1651 sleepq_remove(td, &lbolt); 1652 mtx_lock(&sync_mtx); 1653 if (rushjob < syncdelay / 2) { 1654 rushjob += 1; 1655 stat_rush_requests += 1; 1656 ret = 1; 1657 } 1658 mtx_unlock(&sync_mtx); 1659 return (ret); 1660} 1661 1662/* 1663 * Tell the syncer to speed up its work and run though its work 1664 * list several times, then tell it to shut down. 1665 */ 1666static void 1667syncer_shutdown(void *arg, int howto) 1668{ 1669 struct thread *td; 1670 1671 if (howto & RB_NOSYNC) 1672 return; 1673 td = FIRST_THREAD_IN_PROC(updateproc); 1674 sleepq_remove(td, &lbolt); 1675 mtx_lock(&sync_mtx); 1676 syncer_state = SYNCER_SHUTTING_DOWN; 1677 rushjob = 0; 1678 mtx_unlock(&sync_mtx); 1679 kproc_shutdown(arg, howto); 1680} 1681 1682/* 1683 * Reassign a buffer from one vnode to another. 1684 * Used to assign file specific control information 1685 * (indirect blocks) to the vnode to which they belong. 1686 */ 1687void 1688reassignbuf(struct buf *bp) 1689{ 1690 struct vnode *vp; 1691 struct bufobj *bo; 1692 int delay; 1693 1694 vp = bp->b_vp; 1695 bo = bp->b_bufobj; 1696 ++reassignbufcalls; 1697 1698 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 1699 bp, bp->b_vp, bp->b_flags); 1700 /* 1701 * B_PAGING flagged buffers cannot be reassigned because their vp 1702 * is not fully linked in. 1703 */ 1704 if (bp->b_flags & B_PAGING) 1705 panic("cannot reassign paging buffer"); 1706 1707 /* 1708 * Delete from old vnode list, if on one. 1709 */ 1710 VI_LOCK(vp); 1711 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1712 buf_vlist_remove(bp); 1713 /* 1714 * If dirty, put on list of dirty buffers; otherwise insert onto list 1715 * of clean buffers. 1716 */ 1717 if (bp->b_flags & B_DELWRI) { 1718 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 1719 switch (vp->v_type) { 1720 case VDIR: 1721 delay = dirdelay; 1722 break; 1723 case VCHR: 1724 delay = metadelay; 1725 break; 1726 default: 1727 delay = filedelay; 1728 } 1729 vn_syncer_add_to_worklist(bo, delay); 1730 } 1731 buf_vlist_add(bp, bo, BX_VNDIRTY); 1732 } else { 1733 buf_vlist_add(bp, bo, BX_VNCLEAN); 1734 1735 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1736 mtx_lock(&sync_mtx); 1737 LIST_REMOVE(bo, bo_synclist); 1738 syncer_worklist_len--; 1739 mtx_unlock(&sync_mtx); 1740 bo->bo_flag &= ~BO_ONWORKLST; 1741 } 1742 } 1743 VI_UNLOCK(vp); 1744} 1745 1746static void 1747v_incr_usecount(struct vnode *vp, int delta) 1748{ 1749 1750 vp->v_usecount += delta; 1751 vp->v_holdcnt += delta; 1752 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1753 dev_lock(); 1754 vp->v_rdev->si_usecount += delta; 1755 dev_unlock(); 1756 } 1757} 1758 1759/* 1760 * Grab a particular vnode from the free list, increment its 1761 * reference count and lock it. The vnode lock bit is set if the 1762 * vnode is being eliminated in vgone. The process is awakened 1763 * when the transition is completed, and an error returned to 1764 * indicate that the vnode is no longer usable (possibly having 1765 * been changed to a new filesystem type). 1766 */ 1767int 1768vget(vp, flags, td) 1769 struct vnode *vp; 1770 int flags; 1771 struct thread *td; 1772{ 1773 int oweinact; 1774 int oldflags; 1775 int error; 1776 1777 error = 0; 1778 oldflags = flags; 1779 oweinact = 0; 1780 if ((flags & LK_INTERLOCK) == 0) 1781 VI_LOCK(vp); 1782 /* 1783 * If the inactive call was deferred because vput() was called 1784 * with a shared lock, we have to do it here before another thread 1785 * gets a reference to data that should be dead. 1786 */ 1787 if (vp->v_iflag & VI_OWEINACT) { 1788 if (flags & LK_NOWAIT) { 1789 VI_UNLOCK(vp); 1790 return (EBUSY); 1791 } 1792 flags &= ~LK_TYPE_MASK; 1793 flags |= LK_EXCLUSIVE; 1794 oweinact = 1; 1795 } 1796 v_incr_usecount(vp, 1); 1797 if (VSHOULDBUSY(vp)) 1798 vbusy(vp); 1799 if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { 1800 VI_LOCK(vp); 1801 /* 1802 * must expand vrele here because we do not want 1803 * to call VOP_INACTIVE if the reference count 1804 * drops back to zero since it was never really 1805 * active. 1806 */ 1807 v_incr_usecount(vp, -1); 1808 if (VSHOULDFREE(vp)) 1809 vfree(vp); 1810 else 1811 vlruvp(vp); 1812 VI_UNLOCK(vp); 1813 return (error); 1814 } 1815 if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0) 1816 panic("vget: vn_lock failed to return ENOENT\n"); 1817 if (oweinact) { 1818 VI_LOCK(vp); 1819 if (vp->v_iflag & VI_OWEINACT) 1820 vinactive(vp, td); 1821 VI_UNLOCK(vp); 1822 if ((oldflags & LK_TYPE_MASK) == 0) 1823 VOP_UNLOCK(vp, 0, td); 1824 } 1825 return (0); 1826} 1827 1828/* 1829 * Increase the reference count of a vnode. 1830 */ 1831void 1832vref(struct vnode *vp) 1833{ 1834 1835 VI_LOCK(vp); 1836 v_incr_usecount(vp, 1); 1837 VI_UNLOCK(vp); 1838} 1839 1840/* 1841 * Return reference count of a vnode. 1842 * 1843 * The results of this call are only guaranteed when some mechanism other 1844 * than the VI lock is used to stop other processes from gaining references 1845 * to the vnode. This may be the case if the caller holds the only reference. 1846 * This is also useful when stale data is acceptable as race conditions may 1847 * be accounted for by some other means. 1848 */ 1849int 1850vrefcnt(struct vnode *vp) 1851{ 1852 int usecnt; 1853 1854 VI_LOCK(vp); 1855 usecnt = vp->v_usecount; 1856 VI_UNLOCK(vp); 1857 1858 return (usecnt); 1859} 1860 1861 1862/* 1863 * Vnode put/release. 1864 * If count drops to zero, call inactive routine and return to freelist. 1865 */ 1866void 1867vrele(vp) 1868 struct vnode *vp; 1869{ 1870 struct thread *td = curthread; /* XXX */ 1871 1872 KASSERT(vp != NULL, ("vrele: null vp")); 1873 1874 VI_LOCK(vp); 1875 1876 /* Skip this v_writecount check if we're going to panic below. */ 1877 VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp, 1878 ("vrele: missed vn_close")); 1879 1880 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 1881 vp->v_usecount == 1)) { 1882 v_incr_usecount(vp, -1); 1883 VI_UNLOCK(vp); 1884 1885 return; 1886 } 1887 if (vp->v_usecount != 1) { 1888#ifdef DIAGNOSTIC 1889 vprint("vrele: negative ref count", vp); 1890#endif 1891 VI_UNLOCK(vp); 1892 panic("vrele: negative ref cnt"); 1893 } 1894 v_incr_usecount(vp, -1); 1895 /* 1896 * We must call VOP_INACTIVE with the node locked. Mark 1897 * as VI_DOINGINACT to avoid recursion. 1898 */ 1899 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) { 1900 VI_LOCK(vp); 1901 vinactive(vp, td); 1902 VOP_UNLOCK(vp, 0, td); 1903 } else 1904 VI_LOCK(vp); 1905 if (VSHOULDFREE(vp)) 1906 vfree(vp); 1907 else 1908 vlruvp(vp); 1909 VI_UNLOCK(vp); 1910} 1911 1912/* 1913 * Release an already locked vnode. This give the same effects as 1914 * unlock+vrele(), but takes less time and avoids releasing and 1915 * re-aquiring the lock (as vrele() aquires the lock internally.) 1916 */ 1917void 1918vput(vp) 1919 struct vnode *vp; 1920{ 1921 struct thread *td = curthread; /* XXX */ 1922 int error; 1923 1924 KASSERT(vp != NULL, ("vput: null vp")); 1925 ASSERT_VOP_LOCKED(vp, "vput"); 1926 VI_LOCK(vp); 1927 /* Skip this v_writecount check if we're going to panic below. */ 1928 VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp, 1929 ("vput: missed vn_close")); 1930 error = 0; 1931 1932 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 1933 vp->v_usecount == 1)) { 1934 v_incr_usecount(vp, -1); 1935 VOP_UNLOCK(vp, LK_INTERLOCK, td); 1936 return; 1937 } 1938 1939 if (vp->v_usecount != 1) { 1940#ifdef DIAGNOSTIC 1941 vprint("vput: negative ref count", vp); 1942#endif 1943 panic("vput: negative ref cnt"); 1944 } 1945 v_incr_usecount(vp, -1); 1946 vp->v_iflag |= VI_OWEINACT; 1947 if (VOP_ISLOCKED(vp, td) != LK_EXCLUSIVE) { 1948 error = VOP_LOCK(vp, LK_EXCLUPGRADE|LK_INTERLOCK|LK_NOWAIT, td); 1949 VI_LOCK(vp); 1950 if (error) 1951 goto done; 1952 } 1953 if (vp->v_iflag & VI_OWEINACT) 1954 vinactive(vp, td); 1955 VOP_UNLOCK(vp, 0, td); 1956done: 1957 if (VSHOULDFREE(vp)) 1958 vfree(vp); 1959 else 1960 vlruvp(vp); 1961 VI_UNLOCK(vp); 1962} 1963 1964/* 1965 * Somebody doesn't want the vnode recycled. 1966 */ 1967void 1968vhold(struct vnode *vp) 1969{ 1970 1971 VI_LOCK(vp); 1972 vholdl(vp); 1973 VI_UNLOCK(vp); 1974} 1975 1976void 1977vholdl(struct vnode *vp) 1978{ 1979 1980 vp->v_holdcnt++; 1981 if (VSHOULDBUSY(vp)) 1982 vbusy(vp); 1983} 1984 1985/* 1986 * Note that there is one less who cares about this vnode. vdrop() is the 1987 * opposite of vhold(). 1988 */ 1989void 1990vdrop(struct vnode *vp) 1991{ 1992 1993 VI_LOCK(vp); 1994 vdropl(vp); 1995 VI_UNLOCK(vp); 1996} 1997 1998static void 1999vdropl(struct vnode *vp) 2000{ 2001 2002 if (vp->v_holdcnt <= 0) 2003 panic("vdrop: holdcnt %d", vp->v_holdcnt); 2004 vp->v_holdcnt--; 2005 if (VSHOULDFREE(vp)) 2006 vfree(vp); 2007 else 2008 vlruvp(vp); 2009} 2010 2011static void 2012vinactive(struct vnode *vp, struct thread *td) 2013{ 2014 ASSERT_VOP_LOCKED(vp, "vinactive"); 2015 ASSERT_VI_LOCKED(vp, "vinactive"); 2016 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, 2017 ("vinactive: recursed on VI_DOINGINACT")); 2018 vp->v_iflag |= VI_DOINGINACT; 2019 VI_UNLOCK(vp); 2020 VOP_INACTIVE(vp, td); 2021 VI_LOCK(vp); 2022 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, 2023 ("vinactive: lost VI_DOINGINACT")); 2024 vp->v_iflag &= ~(VI_DOINGINACT|VI_OWEINACT); 2025} 2026 2027/* 2028 * Remove any vnodes in the vnode table belonging to mount point mp. 2029 * 2030 * If FORCECLOSE is not specified, there should not be any active ones, 2031 * return error if any are found (nb: this is a user error, not a 2032 * system error). If FORCECLOSE is specified, detach any active vnodes 2033 * that are found. 2034 * 2035 * If WRITECLOSE is set, only flush out regular file vnodes open for 2036 * writing. 2037 * 2038 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2039 * 2040 * `rootrefs' specifies the base reference count for the root vnode 2041 * of this filesystem. The root vnode is considered busy if its 2042 * v_usecount exceeds this value. On a successful return, vflush(, td) 2043 * will call vrele() on the root vnode exactly rootrefs times. 2044 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2045 * be zero. 2046 */ 2047#ifdef DIAGNOSTIC 2048static int busyprt = 0; /* print out busy vnodes */ 2049SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 2050#endif 2051 2052int 2053vflush(mp, rootrefs, flags, td) 2054 struct mount *mp; 2055 int rootrefs; 2056 int flags; 2057 struct thread *td; 2058{ 2059 struct vnode *vp, *nvp, *rootvp = NULL; 2060 struct vattr vattr; 2061 int busy = 0, error; 2062 2063 if (rootrefs > 0) { 2064 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2065 ("vflush: bad args")); 2066 /* 2067 * Get the filesystem root vnode. We can vput() it 2068 * immediately, since with rootrefs > 0, it won't go away. 2069 */ 2070 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0) 2071 return (error); 2072 vput(rootvp); 2073 2074 } 2075 MNT_ILOCK(mp); 2076loop: 2077 MNT_VNODE_FOREACH(vp, mp, nvp) { 2078 2079 VI_LOCK(vp); 2080 MNT_IUNLOCK(mp); 2081 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td); 2082 if (error) { 2083 MNT_ILOCK(mp); 2084 goto loop; 2085 } 2086 /* 2087 * Skip over a vnodes marked VV_SYSTEM. 2088 */ 2089 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2090 VOP_UNLOCK(vp, 0, td); 2091 MNT_ILOCK(mp); 2092 continue; 2093 } 2094 /* 2095 * If WRITECLOSE is set, flush out unlinked but still open 2096 * files (even if open only for reading) and regular file 2097 * vnodes open for writing. 2098 */ 2099 if (flags & WRITECLOSE) { 2100 error = VOP_GETATTR(vp, &vattr, td->td_ucred, td); 2101 VI_LOCK(vp); 2102 2103 if ((vp->v_type == VNON || 2104 (error == 0 && vattr.va_nlink > 0)) && 2105 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2106 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2107 MNT_ILOCK(mp); 2108 continue; 2109 } 2110 } else 2111 VI_LOCK(vp); 2112 /* 2113 * With v_usecount == 0, all we need to do is clear out the 2114 * vnode data structures and we are done. 2115 */ 2116 if (vp->v_usecount == 0) { 2117 vgonel(vp, td); 2118 VOP_UNLOCK(vp, 0, td); 2119 MNT_ILOCK(mp); 2120 continue; 2121 } 2122 /* 2123 * If FORCECLOSE is set, forcibly close the vnode. For block 2124 * or character devices, revert to an anonymous device. For 2125 * all other files, just kill them. 2126 */ 2127 if (flags & FORCECLOSE) { 2128 VNASSERT(vp->v_type != VCHR && vp->v_type != VBLK, vp, 2129 ("device VNODE %p is FORCECLOSED", vp)); 2130 vgonel(vp, td); 2131 VOP_UNLOCK(vp, 0, td); 2132 MNT_ILOCK(mp); 2133 continue; 2134 } 2135 VOP_UNLOCK(vp, 0, td); 2136#ifdef DIAGNOSTIC 2137 if (busyprt) 2138 vprint("vflush: busy vnode", vp); 2139#endif 2140 VI_UNLOCK(vp); 2141 MNT_ILOCK(mp); 2142 busy++; 2143 } 2144 MNT_IUNLOCK(mp); 2145 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2146 /* 2147 * If just the root vnode is busy, and if its refcount 2148 * is equal to `rootrefs', then go ahead and kill it. 2149 */ 2150 VI_LOCK(rootvp); 2151 KASSERT(busy > 0, ("vflush: not busy")); 2152 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 2153 ("vflush: usecount %d < rootrefs %d", 2154 rootvp->v_usecount, rootrefs)); 2155 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2156 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK, td); 2157 vgone(rootvp); 2158 VOP_UNLOCK(rootvp, 0, td); 2159 busy = 0; 2160 } else 2161 VI_UNLOCK(rootvp); 2162 } 2163 if (busy) 2164 return (EBUSY); 2165 for (; rootrefs > 0; rootrefs--) 2166 vrele(rootvp); 2167 return (0); 2168} 2169 2170/* 2171 * This moves a now (likely recyclable) vnode to the end of the 2172 * mountlist. XXX However, it is temporarily disabled until we 2173 * can clean up ffs_sync() and friends, which have loop restart 2174 * conditions which this code causes to operate O(N^2). 2175 */ 2176static void 2177vlruvp(struct vnode *vp) 2178{ 2179#if 0 2180 struct mount *mp; 2181 2182 if ((mp = vp->v_mount) != NULL) { 2183 MNT_ILOCK(mp); 2184 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2185 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2186 MNT_IUNLOCK(mp); 2187 } 2188#endif 2189} 2190 2191/* 2192 * Recycle an unused vnode to the front of the free list. 2193 * Release the passed interlock if the vnode will be recycled. 2194 */ 2195int 2196vrecycle(struct vnode *vp, struct thread *td) 2197{ 2198 2199 ASSERT_VOP_LOCKED(vp, "vrecycle"); 2200 VI_LOCK(vp); 2201 if (vp->v_usecount == 0) { 2202 vgonel(vp, td); 2203 return (1); 2204 } 2205 VI_UNLOCK(vp); 2206 return (0); 2207} 2208 2209/* 2210 * Eliminate all activity associated with a vnode 2211 * in preparation for reuse. 2212 */ 2213void 2214vgone(struct vnode *vp) 2215{ 2216 struct thread *td = curthread; /* XXX */ 2217 ASSERT_VOP_LOCKED(vp, "vgone"); 2218 2219 VI_LOCK(vp); 2220 vgonel(vp, td); 2221} 2222 2223/* 2224 * vgone, with the vp interlock held. 2225 */ 2226void 2227vgonel(struct vnode *vp, struct thread *td) 2228{ 2229 int oweinact; 2230 int active; 2231 int doomed; 2232 2233 ASSERT_VOP_LOCKED(vp, "vgonel"); 2234 ASSERT_VI_LOCKED(vp, "vgonel"); 2235 2236 /* 2237 * Check to see if the vnode is in use. If so we have to reference it 2238 * before we clean it out so that its count cannot fall to zero and 2239 * generate a race against ourselves to recycle it. 2240 */ 2241 if ((active = vp->v_usecount)) 2242 v_incr_usecount(vp, 1); 2243 2244 /* 2245 * See if we're already doomed, if so, this is coming from a 2246 * successful vtryrecycle(); 2247 */ 2248 doomed = (vp->v_iflag & VI_DOOMED); 2249 vp->v_iflag |= VI_DOOMED; 2250 vp->v_vxthread = curthread; 2251 oweinact = (vp->v_iflag & VI_OWEINACT); 2252 VI_UNLOCK(vp); 2253 2254 /* 2255 * Clean out any buffers associated with the vnode. 2256 * If the flush fails, just toss the buffers. 2257 */ 2258 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 2259 (void) vn_write_suspend_wait(vp, NULL, V_WAIT); 2260 if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0) 2261 vinvalbuf(vp, 0, td, 0, 0); 2262 2263 /* 2264 * If purging an active vnode, it must be closed and 2265 * deactivated before being reclaimed. 2266 */ 2267 if (active) 2268 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2269 if (oweinact || active) { 2270 VI_LOCK(vp); 2271 if ((vp->v_iflag & VI_DOINGINACT) == 0) 2272 vinactive(vp, td); 2273 VI_UNLOCK(vp); 2274 } 2275 /* 2276 * Reclaim the vnode. 2277 */ 2278 if (VOP_RECLAIM(vp, td)) 2279 panic("vgone: cannot reclaim"); 2280 2281 VNASSERT(vp->v_object == NULL, vp, 2282 ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag)); 2283 2284 /* 2285 * Delete from old mount point vnode list. 2286 */ 2287 delmntque(vp); 2288 cache_purge(vp); 2289 VI_LOCK(vp); 2290 if (active) { 2291 v_incr_usecount(vp, -1); 2292 VNASSERT(vp->v_usecount >= 0, vp, ("vgone: bad ref count")); 2293 } 2294 /* 2295 * Done with purge, reset to the standard lock and 2296 * notify sleepers of the grim news. 2297 */ 2298 vp->v_vnlock = &vp->v_lock; 2299 vp->v_op = &dead_vnodeops; 2300 vp->v_tag = "none"; 2301 vp->v_type = VBAD; 2302 vp->v_vxthread = NULL; 2303 2304 /* 2305 * If it is on the freelist and not already at the head, 2306 * move it to the head of the list. The test of the 2307 * VDOOMED flag and the reference count of zero is because 2308 * it will be removed from the free list by getnewvnode, 2309 * but will not have its reference count incremented until 2310 * after calling vgone. If the reference count were 2311 * incremented first, vgone would (incorrectly) try to 2312 * close the previous instance of the underlying object. 2313 */ 2314 if (vp->v_holdcnt == 0 && !doomed) { 2315 mtx_lock(&vnode_free_list_mtx); 2316 if (vp->v_iflag & VI_FREE) { 2317 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2318 } else { 2319 vp->v_iflag |= VI_FREE; 2320 freevnodes++; 2321 } 2322 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2323 mtx_unlock(&vnode_free_list_mtx); 2324 } 2325 VI_UNLOCK(vp); 2326} 2327 2328/* 2329 * Calculate the total number of references to a special device. 2330 */ 2331int 2332vcount(vp) 2333 struct vnode *vp; 2334{ 2335 int count; 2336 2337 dev_lock(); 2338 count = vp->v_rdev->si_usecount; 2339 dev_unlock(); 2340 return (count); 2341} 2342 2343/* 2344 * Same as above, but using the struct cdev *as argument 2345 */ 2346int 2347count_dev(dev) 2348 struct cdev *dev; 2349{ 2350 int count; 2351 2352 dev_lock(); 2353 count = dev->si_usecount; 2354 dev_unlock(); 2355 return(count); 2356} 2357 2358/* 2359 * Print out a description of a vnode. 2360 */ 2361static char *typename[] = 2362{"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 2363 2364void 2365vn_printf(struct vnode *vp, const char *fmt, ...) 2366{ 2367 va_list ap; 2368 char buf[96]; 2369 2370 va_start(ap, fmt); 2371 vprintf(fmt, ap); 2372 va_end(ap); 2373 printf("%p: ", (void *)vp); 2374 printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]); 2375 printf(" usecount %d, writecount %d, refcount %d mountedhere %p\n", 2376 vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere); 2377 buf[0] = '\0'; 2378 buf[1] = '\0'; 2379 if (vp->v_vflag & VV_ROOT) 2380 strcat(buf, "|VV_ROOT"); 2381 if (vp->v_vflag & VV_TEXT) 2382 strcat(buf, "|VV_TEXT"); 2383 if (vp->v_vflag & VV_SYSTEM) 2384 strcat(buf, "|VV_SYSTEM"); 2385 if (vp->v_iflag & VI_DOOMED) 2386 strcat(buf, "|VI_DOOMED"); 2387 if (vp->v_iflag & VI_FREE) 2388 strcat(buf, "|VI_FREE"); 2389 printf(" flags (%s)\n", buf + 1); 2390 if (mtx_owned(VI_MTX(vp))) 2391 printf(" VI_LOCKed"); 2392 if (vp->v_object != NULL) 2393 printf(" v_object %p ref %d pages %d\n", 2394 vp->v_object, vp->v_object->ref_count, 2395 vp->v_object->resident_page_count); 2396 printf(" "); 2397 lockmgr_printinfo(vp->v_vnlock); 2398 printf("\n"); 2399 if (vp->v_data != NULL) 2400 VOP_PRINT(vp); 2401} 2402 2403#ifdef DDB 2404#include <ddb/ddb.h> 2405/* 2406 * List all of the locked vnodes in the system. 2407 * Called when debugging the kernel. 2408 */ 2409DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 2410{ 2411 struct mount *mp, *nmp; 2412 struct vnode *vp; 2413 2414 /* 2415 * Note: because this is DDB, we can't obey the locking semantics 2416 * for these structures, which means we could catch an inconsistent 2417 * state and dereference a nasty pointer. Not much to be done 2418 * about that. 2419 */ 2420 printf("Locked vnodes\n"); 2421 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2422 nmp = TAILQ_NEXT(mp, mnt_list); 2423 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2424 if (VOP_ISLOCKED(vp, NULL)) 2425 vprint("", vp); 2426 } 2427 nmp = TAILQ_NEXT(mp, mnt_list); 2428 } 2429} 2430#endif 2431 2432/* 2433 * Fill in a struct xvfsconf based on a struct vfsconf. 2434 */ 2435static void 2436vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) 2437{ 2438 2439 strcpy(xvfsp->vfc_name, vfsp->vfc_name); 2440 xvfsp->vfc_typenum = vfsp->vfc_typenum; 2441 xvfsp->vfc_refcount = vfsp->vfc_refcount; 2442 xvfsp->vfc_flags = vfsp->vfc_flags; 2443 /* 2444 * These are unused in userland, we keep them 2445 * to not break binary compatibility. 2446 */ 2447 xvfsp->vfc_vfsops = NULL; 2448 xvfsp->vfc_next = NULL; 2449} 2450 2451/* 2452 * Top level filesystem related information gathering. 2453 */ 2454static int 2455sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 2456{ 2457 struct vfsconf *vfsp; 2458 struct xvfsconf xvfsp; 2459 int error; 2460 2461 error = 0; 2462 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2463 vfsconf2x(vfsp, &xvfsp); 2464 error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp); 2465 if (error) 2466 break; 2467 } 2468 return (error); 2469} 2470 2471SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, 2472 "S,xvfsconf", "List of all configured filesystems"); 2473 2474#ifndef BURN_BRIDGES 2475static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2476 2477static int 2478vfs_sysctl(SYSCTL_HANDLER_ARGS) 2479{ 2480 int *name = (int *)arg1 - 1; /* XXX */ 2481 u_int namelen = arg2 + 1; /* XXX */ 2482 struct vfsconf *vfsp; 2483 struct xvfsconf xvfsp; 2484 2485 printf("WARNING: userland calling deprecated sysctl, " 2486 "please rebuild world\n"); 2487 2488#if 1 || defined(COMPAT_PRELITE2) 2489 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2490 if (namelen == 1) 2491 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2492#endif 2493 2494 switch (name[1]) { 2495 case VFS_MAXTYPENUM: 2496 if (namelen != 2) 2497 return (ENOTDIR); 2498 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2499 case VFS_CONF: 2500 if (namelen != 3) 2501 return (ENOTDIR); /* overloaded */ 2502 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) 2503 if (vfsp->vfc_typenum == name[2]) 2504 break; 2505 if (vfsp == NULL) 2506 return (EOPNOTSUPP); 2507 vfsconf2x(vfsp, &xvfsp); 2508 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 2509 } 2510 return (EOPNOTSUPP); 2511} 2512 2513static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, 2514 vfs_sysctl, "Generic filesystem"); 2515 2516#if 1 || defined(COMPAT_PRELITE2) 2517 2518static int 2519sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 2520{ 2521 int error; 2522 struct vfsconf *vfsp; 2523 struct ovfsconf ovfs; 2524 2525 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2526 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 2527 strcpy(ovfs.vfc_name, vfsp->vfc_name); 2528 ovfs.vfc_index = vfsp->vfc_typenum; 2529 ovfs.vfc_refcount = vfsp->vfc_refcount; 2530 ovfs.vfc_flags = vfsp->vfc_flags; 2531 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 2532 if (error) 2533 return error; 2534 } 2535 return 0; 2536} 2537 2538#endif /* 1 || COMPAT_PRELITE2 */ 2539#endif /* !BURN_BRIDGES */ 2540 2541#define KINFO_VNODESLOP 10 2542#ifdef notyet 2543/* 2544 * Dump vnode list (via sysctl). 2545 */ 2546/* ARGSUSED */ 2547static int 2548sysctl_vnode(SYSCTL_HANDLER_ARGS) 2549{ 2550 struct xvnode *xvn; 2551 struct thread *td = req->td; 2552 struct mount *mp; 2553 struct vnode *vp; 2554 int error, len, n; 2555 2556 /* 2557 * Stale numvnodes access is not fatal here. 2558 */ 2559 req->lock = 0; 2560 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 2561 if (!req->oldptr) 2562 /* Make an estimate */ 2563 return (SYSCTL_OUT(req, 0, len)); 2564 2565 error = sysctl_wire_old_buffer(req, 0); 2566 if (error != 0) 2567 return (error); 2568 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 2569 n = 0; 2570 mtx_lock(&mountlist_mtx); 2571 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 2572 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 2573 continue; 2574 MNT_ILOCK(mp); 2575 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2576 if (n == len) 2577 break; 2578 vref(vp); 2579 xvn[n].xv_size = sizeof *xvn; 2580 xvn[n].xv_vnode = vp; 2581 xvn[n].xv_id = 0; /* XXX compat */ 2582#define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 2583 XV_COPY(usecount); 2584 XV_COPY(writecount); 2585 XV_COPY(holdcnt); 2586 XV_COPY(mount); 2587 XV_COPY(numoutput); 2588 XV_COPY(type); 2589#undef XV_COPY 2590 xvn[n].xv_flag = vp->v_vflag; 2591 2592 switch (vp->v_type) { 2593 case VREG: 2594 case VDIR: 2595 case VLNK: 2596 break; 2597 case VBLK: 2598 case VCHR: 2599 if (vp->v_rdev == NULL) { 2600 vrele(vp); 2601 continue; 2602 } 2603 xvn[n].xv_dev = dev2udev(vp->v_rdev); 2604 break; 2605 case VSOCK: 2606 xvn[n].xv_socket = vp->v_socket; 2607 break; 2608 case VFIFO: 2609 xvn[n].xv_fifo = vp->v_fifoinfo; 2610 break; 2611 case VNON: 2612 case VBAD: 2613 default: 2614 /* shouldn't happen? */ 2615 vrele(vp); 2616 continue; 2617 } 2618 vrele(vp); 2619 ++n; 2620 } 2621 MNT_IUNLOCK(mp); 2622 mtx_lock(&mountlist_mtx); 2623 vfs_unbusy(mp, td); 2624 if (n == len) 2625 break; 2626 } 2627 mtx_unlock(&mountlist_mtx); 2628 2629 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 2630 free(xvn, M_TEMP); 2631 return (error); 2632} 2633 2634SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 2635 0, 0, sysctl_vnode, "S,xvnode", ""); 2636#endif 2637 2638/* 2639 * Unmount all filesystems. The list is traversed in reverse order 2640 * of mounting to avoid dependencies. 2641 */ 2642void 2643vfs_unmountall() 2644{ 2645 struct mount *mp; 2646 struct thread *td; 2647 int error; 2648 2649 if (curthread != NULL) 2650 td = curthread; 2651 else 2652 td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */ 2653 /* 2654 * Since this only runs when rebooting, it is not interlocked. 2655 */ 2656 while(!TAILQ_EMPTY(&mountlist)) { 2657 mp = TAILQ_LAST(&mountlist, mntlist); 2658 error = dounmount(mp, MNT_FORCE, td); 2659 if (error) { 2660 TAILQ_REMOVE(&mountlist, mp, mnt_list); 2661 printf("unmount of %s failed (", 2662 mp->mnt_stat.f_mntonname); 2663 if (error == EBUSY) 2664 printf("BUSY)\n"); 2665 else 2666 printf("%d)\n", error); 2667 } else { 2668 /* The unmount has removed mp from the mountlist */ 2669 } 2670 } 2671} 2672 2673/* 2674 * perform msync on all vnodes under a mount point 2675 * the mount point must be locked. 2676 */ 2677void 2678vfs_msync(struct mount *mp, int flags) 2679{ 2680 struct vnode *vp, *nvp; 2681 struct vm_object *obj; 2682 int tries; 2683 2684 tries = 5; 2685 MNT_ILOCK(mp); 2686loop: 2687 TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) { 2688 if (vp->v_mount != mp) { 2689 if (--tries > 0) 2690 goto loop; 2691 break; 2692 } 2693 2694 VI_LOCK(vp); 2695 if ((vp->v_iflag & VI_OBJDIRTY) && 2696 (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { 2697 MNT_IUNLOCK(mp); 2698 if (!vget(vp, 2699 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 2700 curthread)) { 2701 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 2702 vput(vp); 2703 MNT_ILOCK(mp); 2704 continue; 2705 } 2706 2707 obj = vp->v_object; 2708 if (obj != NULL) { 2709 VM_OBJECT_LOCK(obj); 2710 vm_object_page_clean(obj, 0, 0, 2711 flags == MNT_WAIT ? 2712 OBJPC_SYNC : OBJPC_NOSYNC); 2713 VM_OBJECT_UNLOCK(obj); 2714 } 2715 vput(vp); 2716 } 2717 MNT_ILOCK(mp); 2718 if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) { 2719 if (--tries > 0) 2720 goto loop; 2721 break; 2722 } 2723 } else 2724 VI_UNLOCK(vp); 2725 } 2726 MNT_IUNLOCK(mp); 2727} 2728 2729/* 2730 * Mark a vnode as free, putting it up for recycling. 2731 */ 2732static void 2733vfree(struct vnode *vp) 2734{ 2735 2736 ASSERT_VI_LOCKED(vp, "vfree"); 2737 mtx_lock(&vnode_free_list_mtx); 2738 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free")); 2739 VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't")); 2740 if (vp->v_iflag & (VI_AGE|VI_DOOMED)) { 2741 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2742 } else { 2743 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 2744 } 2745 freevnodes++; 2746 mtx_unlock(&vnode_free_list_mtx); 2747 vp->v_iflag &= ~(VI_AGE|VI_DOOMED); 2748 vp->v_iflag |= VI_FREE; 2749} 2750 2751/* 2752 * Opposite of vfree() - mark a vnode as in use. 2753 */ 2754static void 2755vbusy(struct vnode *vp) 2756{ 2757 2758 ASSERT_VI_LOCKED(vp, "vbusy"); 2759 VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free")); 2760 2761 mtx_lock(&vnode_free_list_mtx); 2762 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2763 freevnodes--; 2764 mtx_unlock(&vnode_free_list_mtx); 2765 2766 vp->v_iflag &= ~(VI_FREE|VI_AGE); 2767} 2768 2769/* 2770 * Initalize per-vnode helper structure to hold poll-related state. 2771 */ 2772void 2773v_addpollinfo(struct vnode *vp) 2774{ 2775 struct vpollinfo *vi; 2776 2777 vi = uma_zalloc(vnodepoll_zone, M_WAITOK); 2778 if (vp->v_pollinfo != NULL) { 2779 uma_zfree(vnodepoll_zone, vi); 2780 return; 2781 } 2782 vp->v_pollinfo = vi; 2783 mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 2784 knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, 2785 &vp->v_pollinfo->vpi_lock); 2786} 2787 2788/* 2789 * Record a process's interest in events which might happen to 2790 * a vnode. Because poll uses the historic select-style interface 2791 * internally, this routine serves as both the ``check for any 2792 * pending events'' and the ``record my interest in future events'' 2793 * functions. (These are done together, while the lock is held, 2794 * to avoid race conditions.) 2795 */ 2796int 2797vn_pollrecord(vp, td, events) 2798 struct vnode *vp; 2799 struct thread *td; 2800 short events; 2801{ 2802 2803 if (vp->v_pollinfo == NULL) 2804 v_addpollinfo(vp); 2805 mtx_lock(&vp->v_pollinfo->vpi_lock); 2806 if (vp->v_pollinfo->vpi_revents & events) { 2807 /* 2808 * This leaves events we are not interested 2809 * in available for the other process which 2810 * which presumably had requested them 2811 * (otherwise they would never have been 2812 * recorded). 2813 */ 2814 events &= vp->v_pollinfo->vpi_revents; 2815 vp->v_pollinfo->vpi_revents &= ~events; 2816 2817 mtx_unlock(&vp->v_pollinfo->vpi_lock); 2818 return events; 2819 } 2820 vp->v_pollinfo->vpi_events |= events; 2821 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 2822 mtx_unlock(&vp->v_pollinfo->vpi_lock); 2823 return 0; 2824} 2825 2826/* 2827 * Routine to create and manage a filesystem syncer vnode. 2828 */ 2829#define sync_close ((int (*)(struct vop_close_args *))nullop) 2830static int sync_fsync(struct vop_fsync_args *); 2831static int sync_inactive(struct vop_inactive_args *); 2832static int sync_reclaim(struct vop_reclaim_args *); 2833 2834static struct vop_vector sync_vnodeops = { 2835 .vop_bypass = VOP_EOPNOTSUPP, 2836 .vop_close = sync_close, /* close */ 2837 .vop_fsync = sync_fsync, /* fsync */ 2838 .vop_inactive = sync_inactive, /* inactive */ 2839 .vop_reclaim = sync_reclaim, /* reclaim */ 2840 .vop_lock = vop_stdlock, /* lock */ 2841 .vop_unlock = vop_stdunlock, /* unlock */ 2842 .vop_islocked = vop_stdislocked, /* islocked */ 2843}; 2844 2845/* 2846 * Create a new filesystem syncer vnode for the specified mount point. 2847 */ 2848int 2849vfs_allocate_syncvnode(mp) 2850 struct mount *mp; 2851{ 2852 struct vnode *vp; 2853 static long start, incr, next; 2854 int error; 2855 2856 /* Allocate a new vnode */ 2857 if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) { 2858 mp->mnt_syncer = NULL; 2859 return (error); 2860 } 2861 vp->v_type = VNON; 2862 /* 2863 * Place the vnode onto the syncer worklist. We attempt to 2864 * scatter them about on the list so that they will go off 2865 * at evenly distributed times even if all the filesystems 2866 * are mounted at once. 2867 */ 2868 next += incr; 2869 if (next == 0 || next > syncer_maxdelay) { 2870 start /= 2; 2871 incr /= 2; 2872 if (start == 0) { 2873 start = syncer_maxdelay / 2; 2874 incr = syncer_maxdelay; 2875 } 2876 next = start; 2877 } 2878 VI_LOCK(vp); 2879 vn_syncer_add_to_worklist(&vp->v_bufobj, 2880 syncdelay > 0 ? next % syncdelay : 0); 2881 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 2882 mtx_lock(&sync_mtx); 2883 sync_vnode_count++; 2884 mtx_unlock(&sync_mtx); 2885 VI_UNLOCK(vp); 2886 mp->mnt_syncer = vp; 2887 return (0); 2888} 2889 2890/* 2891 * Do a lazy sync of the filesystem. 2892 */ 2893static int 2894sync_fsync(ap) 2895 struct vop_fsync_args /* { 2896 struct vnode *a_vp; 2897 struct ucred *a_cred; 2898 int a_waitfor; 2899 struct thread *a_td; 2900 } */ *ap; 2901{ 2902 struct vnode *syncvp = ap->a_vp; 2903 struct mount *mp = syncvp->v_mount; 2904 struct thread *td = ap->a_td; 2905 int error, asyncflag; 2906 struct bufobj *bo; 2907 2908 /* 2909 * We only need to do something if this is a lazy evaluation. 2910 */ 2911 if (ap->a_waitfor != MNT_LAZY) 2912 return (0); 2913 2914 /* 2915 * Move ourselves to the back of the sync list. 2916 */ 2917 bo = &syncvp->v_bufobj; 2918 BO_LOCK(bo); 2919 vn_syncer_add_to_worklist(bo, syncdelay); 2920 BO_UNLOCK(bo); 2921 2922 /* 2923 * Walk the list of vnodes pushing all that are dirty and 2924 * not already on the sync list. 2925 */ 2926 mtx_lock(&mountlist_mtx); 2927 if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { 2928 mtx_unlock(&mountlist_mtx); 2929 return (0); 2930 } 2931 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 2932 vfs_unbusy(mp, td); 2933 return (0); 2934 } 2935 asyncflag = mp->mnt_flag & MNT_ASYNC; 2936 mp->mnt_flag &= ~MNT_ASYNC; 2937 vfs_msync(mp, MNT_NOWAIT); 2938 error = VFS_SYNC(mp, MNT_LAZY, td); 2939 if (asyncflag) 2940 mp->mnt_flag |= MNT_ASYNC; 2941 vn_finished_write(mp); 2942 vfs_unbusy(mp, td); 2943 return (error); 2944} 2945 2946/* 2947 * The syncer vnode is no referenced. 2948 */ 2949static int 2950sync_inactive(ap) 2951 struct vop_inactive_args /* { 2952 struct vnode *a_vp; 2953 struct thread *a_td; 2954 } */ *ap; 2955{ 2956 2957 vgone(ap->a_vp); 2958 return (0); 2959} 2960 2961/* 2962 * The syncer vnode is no longer needed and is being decommissioned. 2963 * 2964 * Modifications to the worklist must be protected by sync_mtx. 2965 */ 2966static int 2967sync_reclaim(ap) 2968 struct vop_reclaim_args /* { 2969 struct vnode *a_vp; 2970 } */ *ap; 2971{ 2972 struct vnode *vp = ap->a_vp; 2973 struct bufobj *bo; 2974 2975 VI_LOCK(vp); 2976 bo = &vp->v_bufobj; 2977 vp->v_mount->mnt_syncer = NULL; 2978 if (bo->bo_flag & BO_ONWORKLST) { 2979 mtx_lock(&sync_mtx); 2980 LIST_REMOVE(bo, bo_synclist); 2981 syncer_worklist_len--; 2982 sync_vnode_count--; 2983 mtx_unlock(&sync_mtx); 2984 bo->bo_flag &= ~BO_ONWORKLST; 2985 } 2986 VI_UNLOCK(vp); 2987 2988 return (0); 2989} 2990 2991/* 2992 * Check if vnode represents a disk device 2993 */ 2994int 2995vn_isdisk(vp, errp) 2996 struct vnode *vp; 2997 int *errp; 2998{ 2999 int error; 3000 3001 error = 0; 3002 dev_lock(); 3003 if (vp->v_type != VCHR) 3004 error = ENOTBLK; 3005 else if (vp->v_rdev == NULL) 3006 error = ENXIO; 3007 else if (vp->v_rdev->si_devsw == NULL) 3008 error = ENXIO; 3009 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 3010 error = ENOTBLK; 3011 dev_unlock(); 3012 if (errp != NULL) 3013 *errp = error; 3014 return (error == 0); 3015} 3016 3017/* 3018 * Common filesystem object access control check routine. Accepts a 3019 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3020 * and optional call-by-reference privused argument allowing vaccess() 3021 * to indicate to the caller whether privilege was used to satisfy the 3022 * request (obsoleted). Returns 0 on success, or an errno on failure. 3023 */ 3024int 3025vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused) 3026 enum vtype type; 3027 mode_t file_mode; 3028 uid_t file_uid; 3029 gid_t file_gid; 3030 mode_t acc_mode; 3031 struct ucred *cred; 3032 int *privused; 3033{ 3034 mode_t dac_granted; 3035#ifdef CAPABILITIES 3036 mode_t cap_granted; 3037#endif 3038 3039 /* 3040 * Look for a normal, non-privileged way to access the file/directory 3041 * as requested. If it exists, go with that. 3042 */ 3043 3044 if (privused != NULL) 3045 *privused = 0; 3046 3047 dac_granted = 0; 3048 3049 /* Check the owner. */ 3050 if (cred->cr_uid == file_uid) { 3051 dac_granted |= VADMIN; 3052 if (file_mode & S_IXUSR) 3053 dac_granted |= VEXEC; 3054 if (file_mode & S_IRUSR) 3055 dac_granted |= VREAD; 3056 if (file_mode & S_IWUSR) 3057 dac_granted |= (VWRITE | VAPPEND); 3058 3059 if ((acc_mode & dac_granted) == acc_mode) 3060 return (0); 3061 3062 goto privcheck; 3063 } 3064 3065 /* Otherwise, check the groups (first match) */ 3066 if (groupmember(file_gid, cred)) { 3067 if (file_mode & S_IXGRP) 3068 dac_granted |= VEXEC; 3069 if (file_mode & S_IRGRP) 3070 dac_granted |= VREAD; 3071 if (file_mode & S_IWGRP) 3072 dac_granted |= (VWRITE | VAPPEND); 3073 3074 if ((acc_mode & dac_granted) == acc_mode) 3075 return (0); 3076 3077 goto privcheck; 3078 } 3079 3080 /* Otherwise, check everyone else. */ 3081 if (file_mode & S_IXOTH) 3082 dac_granted |= VEXEC; 3083 if (file_mode & S_IROTH) 3084 dac_granted |= VREAD; 3085 if (file_mode & S_IWOTH) 3086 dac_granted |= (VWRITE | VAPPEND); 3087 if ((acc_mode & dac_granted) == acc_mode) 3088 return (0); 3089 3090privcheck: 3091 if (!suser_cred(cred, SUSER_ALLOWJAIL)) { 3092 /* XXX audit: privilege used */ 3093 if (privused != NULL) 3094 *privused = 1; 3095 return (0); 3096 } 3097 3098#ifdef CAPABILITIES 3099 /* 3100 * Build a capability mask to determine if the set of capabilities 3101 * satisfies the requirements when combined with the granted mask 3102 * from above. 3103 * For each capability, if the capability is required, bitwise 3104 * or the request type onto the cap_granted mask. 3105 */ 3106 cap_granted = 0; 3107 3108 if (type == VDIR) { 3109 /* 3110 * For directories, use CAP_DAC_READ_SEARCH to satisfy 3111 * VEXEC requests, instead of CAP_DAC_EXECUTE. 3112 */ 3113 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3114 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL)) 3115 cap_granted |= VEXEC; 3116 } else { 3117 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3118 !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL)) 3119 cap_granted |= VEXEC; 3120 } 3121 3122 if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && 3123 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL)) 3124 cap_granted |= VREAD; 3125 3126 if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3127 !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL)) 3128 cap_granted |= (VWRITE | VAPPEND); 3129 3130 if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3131 !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL)) 3132 cap_granted |= VADMIN; 3133 3134 if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { 3135 /* XXX audit: privilege used */ 3136 if (privused != NULL) 3137 *privused = 1; 3138 return (0); 3139 } 3140#endif 3141 3142 return ((acc_mode & VADMIN) ? EPERM : EACCES); 3143} 3144 3145/* 3146 * Credential check based on process requesting service, and per-attribute 3147 * permissions. 3148 */ 3149int 3150extattr_check_cred(struct vnode *vp, int attrnamespace, 3151 struct ucred *cred, struct thread *td, int access) 3152{ 3153 3154 /* 3155 * Kernel-invoked always succeeds. 3156 */ 3157 if (cred == NOCRED) 3158 return (0); 3159 3160 /* 3161 * Do not allow privileged processes in jail to directly 3162 * manipulate system attributes. 3163 * 3164 * XXX What capability should apply here? 3165 * Probably CAP_SYS_SETFFLAG. 3166 */ 3167 switch (attrnamespace) { 3168 case EXTATTR_NAMESPACE_SYSTEM: 3169 /* Potentially should be: return (EPERM); */ 3170 return (suser_cred(cred, 0)); 3171 case EXTATTR_NAMESPACE_USER: 3172 return (VOP_ACCESS(vp, access, cred, td)); 3173 default: 3174 return (EPERM); 3175 } 3176} 3177 3178#ifdef DEBUG_VFS_LOCKS 3179/* 3180 * This only exists to supress warnings from unlocked specfs accesses. It is 3181 * no longer ok to have an unlocked VFS. 3182 */ 3183#define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD) 3184 3185int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 3186SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, ""); 3187 3188int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 3189SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, ""); 3190 3191int vfs_badlock_print = 1; /* Print lock violations. */ 3192SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, ""); 3193 3194#ifdef KDB 3195int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 3196SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, ""); 3197#endif 3198 3199static void 3200vfs_badlock(const char *msg, const char *str, struct vnode *vp) 3201{ 3202 3203#ifdef KDB 3204 if (vfs_badlock_backtrace) 3205 kdb_backtrace(); 3206#endif 3207 if (vfs_badlock_print) 3208 printf("%s: %p %s\n", str, (void *)vp, msg); 3209 if (vfs_badlock_ddb) 3210 kdb_enter("lock violation"); 3211} 3212 3213void 3214assert_vi_locked(struct vnode *vp, const char *str) 3215{ 3216 3217 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 3218 vfs_badlock("interlock is not locked but should be", str, vp); 3219} 3220 3221void 3222assert_vi_unlocked(struct vnode *vp, const char *str) 3223{ 3224 3225 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 3226 vfs_badlock("interlock is locked but should not be", str, vp); 3227} 3228 3229void 3230assert_vop_locked(struct vnode *vp, const char *str) 3231{ 3232 3233 if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0) 3234 vfs_badlock("is not locked but should be", str, vp); 3235} 3236 3237void 3238assert_vop_unlocked(struct vnode *vp, const char *str) 3239{ 3240 3241 if (vp && !IGNORE_LOCK(vp) && 3242 VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE) 3243 vfs_badlock("is locked but should not be", str, vp); 3244} 3245 3246#if 0 3247void 3248assert_vop_elocked(struct vnode *vp, const char *str) 3249{ 3250 3251 if (vp && !IGNORE_LOCK(vp) && 3252 VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE) 3253 vfs_badlock("is not exclusive locked but should be", str, vp); 3254} 3255 3256void 3257assert_vop_elocked_other(struct vnode *vp, const char *str) 3258{ 3259 3260 if (vp && !IGNORE_LOCK(vp) && 3261 VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER) 3262 vfs_badlock("is not exclusive locked by another thread", 3263 str, vp); 3264} 3265 3266void 3267assert_vop_slocked(struct vnode *vp, const char *str) 3268{ 3269 3270 if (vp && !IGNORE_LOCK(vp) && 3271 VOP_ISLOCKED(vp, curthread) != LK_SHARED) 3272 vfs_badlock("is not locked shared but should be", str, vp); 3273} 3274#endif /* 0 */ 3275 3276void 3277vop_rename_pre(void *ap) 3278{ 3279 struct vop_rename_args *a = ap; 3280 3281 if (a->a_tvp) 3282 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 3283 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 3284 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 3285 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 3286 3287 /* Check the source (from). */ 3288 if (a->a_tdvp != a->a_fdvp) 3289 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 3290 if (a->a_tvp != a->a_fvp) 3291 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked"); 3292 3293 /* Check the target. */ 3294 if (a->a_tvp) 3295 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 3296 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 3297} 3298 3299void 3300vop_strategy_pre(void *ap) 3301{ 3302 struct vop_strategy_args *a; 3303 struct buf *bp; 3304 3305 a = ap; 3306 bp = a->a_bp; 3307 3308 /* 3309 * Cluster ops lock their component buffers but not the IO container. 3310 */ 3311 if ((bp->b_flags & B_CLUSTER) != 0) 3312 return; 3313 3314 if (BUF_REFCNT(bp) < 1) { 3315 if (vfs_badlock_print) 3316 printf( 3317 "VOP_STRATEGY: bp is not locked but should be\n"); 3318 if (vfs_badlock_ddb) 3319 kdb_enter("lock violation"); 3320 } 3321} 3322 3323void 3324vop_lookup_pre(void *ap) 3325{ 3326 struct vop_lookup_args *a; 3327 struct vnode *dvp; 3328 3329 a = ap; 3330 dvp = a->a_dvp; 3331 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3332 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 3333} 3334 3335void 3336vop_lookup_post(void *ap, int rc) 3337{ 3338 struct vop_lookup_args *a; 3339 struct componentname *cnp; 3340 struct vnode *dvp; 3341 struct vnode *vp; 3342 int flags; 3343 3344 a = ap; 3345 dvp = a->a_dvp; 3346 cnp = a->a_cnp; 3347 vp = *(a->a_vpp); 3348 flags = cnp->cn_flags; 3349 3350 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3351 3352 if (rc) 3353 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)"); 3354 else if (flags & ISDOTDOT) 3355 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (ISDOTDOT)"); 3356 else 3357 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 3358 3359 if (!rc) 3360 ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)"); 3361} 3362 3363void 3364vop_lock_pre(void *ap) 3365{ 3366 struct vop_lock_args *a = ap; 3367 3368 if ((a->a_flags & LK_INTERLOCK) == 0) 3369 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3370 else 3371 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 3372} 3373 3374void 3375vop_lock_post(void *ap, int rc) 3376{ 3377 struct vop_lock_args *a = ap; 3378 3379 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3380 if (rc == 0) 3381 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 3382} 3383 3384void 3385vop_unlock_pre(void *ap) 3386{ 3387 struct vop_unlock_args *a = ap; 3388 3389 if (a->a_flags & LK_INTERLOCK) 3390 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 3391 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 3392} 3393 3394void 3395vop_unlock_post(void *ap, int rc) 3396{ 3397 struct vop_unlock_args *a = ap; 3398 3399 if (a->a_flags & LK_INTERLOCK) 3400 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 3401} 3402#endif /* DEBUG_VFS_LOCKS */ 3403 3404static struct knlist fs_knlist; 3405 3406static void 3407vfs_event_init(void *arg) 3408{ 3409 knlist_init(&fs_knlist, NULL); 3410} 3411/* XXX - correct order? */ 3412SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 3413 3414void 3415vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused) 3416{ 3417 3418 KNOTE_UNLOCKED(&fs_knlist, event); 3419} 3420 3421static int filt_fsattach(struct knote *kn); 3422static void filt_fsdetach(struct knote *kn); 3423static int filt_fsevent(struct knote *kn, long hint); 3424 3425struct filterops fs_filtops = 3426 { 0, filt_fsattach, filt_fsdetach, filt_fsevent }; 3427 3428static int 3429filt_fsattach(struct knote *kn) 3430{ 3431 3432 kn->kn_flags |= EV_CLEAR; 3433 knlist_add(&fs_knlist, kn, 0); 3434 return (0); 3435} 3436 3437static void 3438filt_fsdetach(struct knote *kn) 3439{ 3440 3441 knlist_remove(&fs_knlist, kn, 0); 3442} 3443 3444static int 3445filt_fsevent(struct knote *kn, long hint) 3446{ 3447 3448 kn->kn_fflags |= hint; 3449 return (kn->kn_fflags != 0); 3450} 3451 3452static int 3453sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 3454{ 3455 struct vfsidctl vc; 3456 int error; 3457 struct mount *mp; 3458 3459 error = SYSCTL_IN(req, &vc, sizeof(vc)); 3460 if (error) 3461 return (error); 3462 if (vc.vc_vers != VFS_CTL_VERS1) 3463 return (EINVAL); 3464 mp = vfs_getvfs(&vc.vc_fsid); 3465 if (mp == NULL) 3466 return (ENOENT); 3467 /* ensure that a specific sysctl goes to the right filesystem. */ 3468 if (strcmp(vc.vc_fstypename, "*") != 0 && 3469 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 3470 return (EINVAL); 3471 } 3472 VCTLTOREQ(&vc, req); 3473 return (VFS_SYSCTL(mp, vc.vc_op, req)); 3474} 3475 3476SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR, 3477 NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid"); 3478 3479/* 3480 * Function to initialize a va_filerev field sensibly. 3481 * XXX: Wouldn't a random number make a lot more sense ?? 3482 */ 3483u_quad_t 3484init_va_filerev(void) 3485{ 3486 struct bintime bt; 3487 3488 getbinuptime(&bt); 3489 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 3490} 3491