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("$FreeBSD$"); 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/dirent.h> 53 #include <sys/event.h> 54 #include <sys/eventhandler.h> 55 #include <sys/extattr.h> 56 #include <sys/file.h> 57 #include <sys/fcntl.h> 58 #include <sys/jail.h> 59 #include <sys/kdb.h> 60 #include <sys/kernel.h> 61 #include <sys/kthread.h> 62 #include <sys/malloc.h> 63 #include <sys/mount.h> 64 #include <sys/namei.h> 65 #include <sys/priv.h> 66 #include <sys/reboot.h> 67 #include <sys/sleepqueue.h> 68 #include <sys/stat.h> 69 #include <sys/sysctl.h> 70 #include <sys/syslog.h> 71 #include <sys/vmmeter.h> 72 #include <sys/vnode.h> 73 74 #include <machine/stdarg.h> 75 76 #include <security/mac/mac_framework.h> 77 78 #include <vm/vm.h> 79 #include <vm/vm_object.h> 80 #include <vm/vm_extern.h> 81 #include <vm/pmap.h> 82 #include <vm/vm_map.h> 83 #include <vm/vm_page.h> 84 #include <vm/vm_kern.h> 85 #include <vm/uma.h> 86 87 #ifdef DDB 88 #include <ddb/ddb.h> 89 #endif 90 91 static MALLOC_DEFINE(M_NETADDR, "subr_export_host", "Export host address structure"); 92 93 static void delmntque(struct vnode *vp); 94 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, 95 int slpflag, int slptimeo); 96 static void syncer_shutdown(void *arg, int howto); 97 static int vtryrecycle(struct vnode *vp); 98 static void vbusy(struct vnode *vp); 99 static void vinactive(struct vnode *, struct thread *); 100 static void v_incr_usecount(struct vnode *); 101 static void v_decr_usecount(struct vnode *); 102 static void v_decr_useonly(struct vnode *); 103 static void v_upgrade_usecount(struct vnode *); 104 static void vfree(struct vnode *); 105 static void vnlru_free(int); 106 static void vdestroy(struct vnode *); 107 static void vgonel(struct vnode *); 108 static void vfs_knllock(void *arg); 109 static void vfs_knlunlock(void *arg); 110 static int vfs_knllocked(void *arg); 111 112 113 /* 114 * Enable Giant pushdown based on whether or not the vm is mpsafe in this 115 * build. Without mpsafevm the buffer cache can not run Giant free. 116 */ 117 int mpsafe_vfs = 1; 118 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs); 119 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0, 120 "MPSAFE VFS"); 121 122 /* 123 * Number of vnodes in existence. Increased whenever getnewvnode() 124 * allocates a new vnode, decreased on vdestroy() called on VI_DOOMed 125 * vnode. 126 */ 127 static unsigned long numvnodes; 128 129 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); 130 131 /* 132 * Conversion tables for conversion from vnode types to inode formats 133 * and back. 134 */ 135 enum vtype iftovt_tab[16] = { 136 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 137 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 138 }; 139 int vttoif_tab[10] = { 140 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 141 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT 142 }; 143 144 /* 145 * List of vnodes that are ready for recycling. 146 */ 147 static TAILQ_HEAD(freelst, vnode) vnode_free_list; 148 149 /* 150 * Free vnode target. Free vnodes may simply be files which have been stat'd 151 * but not read. This is somewhat common, and a small cache of such files 152 * should be kept to avoid recreation costs. 153 */ 154 static u_long wantfreevnodes; 155 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 156 /* Number of vnodes in the free list. */ 157 static u_long freevnodes; 158 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); 159 160 /* 161 * Various variables used for debugging the new implementation of 162 * reassignbuf(). 163 * XXX these are probably of (very) limited utility now. 164 */ 165 static int reassignbufcalls; 166 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); 167 168 /* 169 * Cache for the mount type id assigned to NFS. This is used for 170 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 171 */ 172 int nfs_mount_type = -1; 173 174 /* To keep more than one thread at a time from running vfs_getnewfsid */ 175 static struct mtx mntid_mtx; 176 177 /* 178 * Lock for any access to the following: 179 * vnode_free_list 180 * numvnodes 181 * freevnodes 182 */ 183 static struct mtx vnode_free_list_mtx; 184 185 /* Publicly exported FS */ 186 struct nfs_public nfs_pub; 187 188 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 189 static uma_zone_t vnode_zone; 190 static uma_zone_t vnodepoll_zone; 191 192 /* Set to 1 to print out reclaim of active vnodes */ 193 int prtactive; 194 195 /* 196 * The workitem queue. 197 * 198 * It is useful to delay writes of file data and filesystem metadata 199 * for tens of seconds so that quickly created and deleted files need 200 * not waste disk bandwidth being created and removed. To realize this, 201 * we append vnodes to a "workitem" queue. When running with a soft 202 * updates implementation, most pending metadata dependencies should 203 * not wait for more than a few seconds. Thus, mounted on block devices 204 * are delayed only about a half the time that file data is delayed. 205 * Similarly, directory updates are more critical, so are only delayed 206 * about a third the time that file data is delayed. Thus, there are 207 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 208 * one each second (driven off the filesystem syncer process). The 209 * syncer_delayno variable indicates the next queue that is to be processed. 210 * Items that need to be processed soon are placed in this queue: 211 * 212 * syncer_workitem_pending[syncer_delayno] 213 * 214 * A delay of fifteen seconds is done by placing the request fifteen 215 * entries later in the queue: 216 * 217 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 218 * 219 */ 220 static int syncer_delayno; 221 static long syncer_mask; 222 LIST_HEAD(synclist, bufobj); 223 static struct synclist *syncer_workitem_pending; 224 /* 225 * The sync_mtx protects: 226 * bo->bo_synclist 227 * sync_vnode_count 228 * syncer_delayno 229 * syncer_state 230 * syncer_workitem_pending 231 * syncer_worklist_len 232 * rushjob 233 */ 234 static struct mtx sync_mtx; 235 236 #define SYNCER_MAXDELAY 32 237 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 238 static int syncdelay = 30; /* max time to delay syncing data */ 239 static int filedelay = 30; /* time to delay syncing files */ 240 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); 241 static int dirdelay = 29; /* time to delay syncing directories */ 242 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); 243 static int metadelay = 28; /* time to delay syncing metadata */ 244 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); 245 static int rushjob; /* number of slots to run ASAP */ 246 static int stat_rush_requests; /* number of times I/O speeded up */ 247 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); 248 249 /* 250 * When shutting down the syncer, run it at four times normal speed. 251 */ 252 #define SYNCER_SHUTDOWN_SPEEDUP 4 253 static int sync_vnode_count; 254 static int syncer_worklist_len; 255 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 256 syncer_state; 257 258 /* 259 * Number of vnodes we want to exist at any one time. This is mostly used 260 * to size hash tables in vnode-related code. It is normally not used in 261 * getnewvnode(), as wantfreevnodes is normally nonzero.) 262 * 263 * XXX desiredvnodes is historical cruft and should not exist. 264 */ 265 int desiredvnodes; 266 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 267 &desiredvnodes, 0, "Maximum number of vnodes"); 268 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 269 &wantfreevnodes, 0, "Minimum number of vnodes (legacy)"); 270 static int vnlru_nowhere; 271 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 272 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 273 274 /* 275 * Macros to control when a vnode is freed and recycled. All require 276 * the vnode interlock. 277 */ 278 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt) 279 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt) 280 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt) 281 282 283 /* 284 * Initialize the vnode management data structures. 285 */ 286 #ifndef MAXVNODES_MAX 287 #define MAXVNODES_MAX 100000 288 #endif 289 static void 290 vntblinit(void *dummy __unused) 291 { 292 293 /* 294 * Desiredvnodes is a function of the physical memory size and 295 * the kernel's heap size. Specifically, desiredvnodes scales 296 * in proportion to the physical memory size until two fifths 297 * of the kernel's heap size is consumed by vnodes and vm 298 * objects. 299 */ 300 desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size / 301 (5 * (sizeof(struct vm_object) + sizeof(struct vnode)))); 302 if (desiredvnodes > MAXVNODES_MAX) { 303 if (bootverbose) 304 printf("Reducing kern.maxvnodes %d -> %d\n", 305 desiredvnodes, MAXVNODES_MAX); 306 desiredvnodes = MAXVNODES_MAX; 307 } 308 wantfreevnodes = desiredvnodes / 4; 309 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 310 TAILQ_INIT(&vnode_free_list); 311 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 312 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 313 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 314 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 315 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 316 /* 317 * Initialize the filesystem syncer. 318 */ 319 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 320 &syncer_mask); 321 syncer_maxdelay = syncer_mask + 1; 322 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 323 } 324 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) 325 326 327 /* 328 * Mark a mount point as busy. Used to synchronize access and to delay 329 * unmounting. Interlock is not released on failure. 330 */ 331 int 332 vfs_busy(struct mount *mp, int flags, struct mtx *interlkp, 333 struct thread *td) 334 { 335 int lkflags; 336 337 MNT_ILOCK(mp); 338 MNT_REF(mp); 339 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 340 if (flags & LK_NOWAIT) { 341 MNT_REL(mp); 342 MNT_IUNLOCK(mp); 343 return (ENOENT); 344 } 345 if (interlkp) 346 mtx_unlock(interlkp); 347 mp->mnt_kern_flag |= MNTK_MWAIT; 348 /* 349 * Since all busy locks are shared except the exclusive 350 * lock granted when unmounting, the only place that a 351 * wakeup needs to be done is at the release of the 352 * exclusive lock at the end of dounmount. 353 */ 354 msleep(mp, MNT_MTX(mp), PVFS, "vfs_busy", 0); 355 MNT_REL(mp); 356 MNT_IUNLOCK(mp); 357 if (interlkp) 358 mtx_lock(interlkp); 359 return (ENOENT); 360 } 361 if (interlkp) 362 mtx_unlock(interlkp); 363 lkflags = LK_SHARED | LK_INTERLOCK; 364 if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td)) 365 panic("vfs_busy: unexpected lock failure"); 366 return (0); 367 } 368 369 /* 370 * Free a busy filesystem. 371 */ 372 void 373 vfs_unbusy(struct mount *mp, struct thread *td) 374 { 375 376 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); 377 vfs_rel(mp); 378 } 379 380 /* 381 * Lookup a mount point by filesystem identifier. 382 */ 383 struct mount * 384 vfs_getvfs(fsid_t *fsid) 385 { 386 struct mount *mp; 387 388 mtx_lock(&mountlist_mtx); 389 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 390 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 391 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 392 vfs_ref(mp); 393 mtx_unlock(&mountlist_mtx); 394 return (mp); 395 } 396 } 397 mtx_unlock(&mountlist_mtx); 398 return ((struct mount *) 0); 399 } 400 401 /* 402 * Check if a user can access privileged mount options. 403 */ 404 int 405 vfs_suser(struct mount *mp, struct thread *td) 406 { 407 int error; 408 409 /* 410 * If the thread is jailed, but this is not a jail-friendly file 411 * system, deny immediately. 412 */ 413 if (jailed(td->td_ucred) && !(mp->mnt_vfc->vfc_flags & VFCF_JAIL)) 414 return (EPERM); 415 416 /* 417 * If the file system was mounted outside a jail and a jailed thread 418 * tries to access it, deny immediately. 419 */ 420 if (!jailed(mp->mnt_cred) && jailed(td->td_ucred)) 421 return (EPERM); 422 423 /* 424 * If the file system was mounted inside different jail that the jail of 425 * the calling thread, deny immediately. 426 */ 427 if (jailed(mp->mnt_cred) && jailed(td->td_ucred) && 428 mp->mnt_cred->cr_prison != td->td_ucred->cr_prison) { 429 return (EPERM); 430 } 431 432 if ((mp->mnt_flag & MNT_USER) == 0 || 433 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 434 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) 435 return (error); 436 } 437 return (0); 438 } 439 440 /* 441 * Get a new unique fsid. Try to make its val[0] unique, since this value 442 * will be used to create fake device numbers for stat(). Also try (but 443 * not so hard) make its val[0] unique mod 2^16, since some emulators only 444 * support 16-bit device numbers. We end up with unique val[0]'s for the 445 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 446 * 447 * Keep in mind that several mounts may be running in parallel. Starting 448 * the search one past where the previous search terminated is both a 449 * micro-optimization and a defense against returning the same fsid to 450 * different mounts. 451 */ 452 void 453 vfs_getnewfsid(struct mount *mp) 454 { 455 static u_int16_t mntid_base; 456 struct mount *nmp; 457 fsid_t tfsid; 458 int mtype; 459 460 mtx_lock(&mntid_mtx); 461 mtype = mp->mnt_vfc->vfc_typenum; 462 tfsid.val[1] = mtype; 463 mtype = (mtype & 0xFF) << 24; 464 for (;;) { 465 tfsid.val[0] = makedev(255, 466 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 467 mntid_base++; 468 if ((nmp = vfs_getvfs(&tfsid)) == NULL) 469 break; 470 vfs_rel(nmp); 471 } 472 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 473 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 474 mtx_unlock(&mntid_mtx); 475 } 476 477 /* 478 * Knob to control the precision of file timestamps: 479 * 480 * 0 = seconds only; nanoseconds zeroed. 481 * 1 = seconds and nanoseconds, accurate within 1/HZ. 482 * 2 = seconds and nanoseconds, truncated to microseconds. 483 * >=3 = seconds and nanoseconds, maximum precision. 484 */ 485 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 486 487 static int timestamp_precision = TSP_SEC; 488 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 489 ×tamp_precision, 0, ""); 490 491 /* 492 * Get a current timestamp. 493 */ 494 void 495 vfs_timestamp(struct timespec *tsp) 496 { 497 struct timeval tv; 498 499 switch (timestamp_precision) { 500 case TSP_SEC: 501 tsp->tv_sec = time_second; 502 tsp->tv_nsec = 0; 503 break; 504 case TSP_HZ: 505 getnanotime(tsp); 506 break; 507 case TSP_USEC: 508 microtime(&tv); 509 TIMEVAL_TO_TIMESPEC(&tv, tsp); 510 break; 511 case TSP_NSEC: 512 default: 513 nanotime(tsp); 514 break; 515 } 516 } 517 518 /* 519 * Set vnode attributes to VNOVAL 520 */ 521 void 522 vattr_null(struct vattr *vap) 523 { 524 525 vap->va_type = VNON; 526 vap->va_size = VNOVAL; 527 vap->va_bytes = VNOVAL; 528 vap->va_mode = VNOVAL; 529 vap->va_nlink = VNOVAL; 530 vap->va_uid = VNOVAL; 531 vap->va_gid = VNOVAL; 532 vap->va_fsid = VNOVAL; 533 vap->va_fileid = VNOVAL; 534 vap->va_blocksize = VNOVAL; 535 vap->va_rdev = VNOVAL; 536 vap->va_atime.tv_sec = VNOVAL; 537 vap->va_atime.tv_nsec = VNOVAL; 538 vap->va_mtime.tv_sec = VNOVAL; 539 vap->va_mtime.tv_nsec = VNOVAL; 540 vap->va_ctime.tv_sec = VNOVAL; 541 vap->va_ctime.tv_nsec = VNOVAL; 542 vap->va_birthtime.tv_sec = VNOVAL; 543 vap->va_birthtime.tv_nsec = VNOVAL; 544 vap->va_flags = VNOVAL; 545 vap->va_gen = VNOVAL; 546 vap->va_vaflags = 0; 547 } 548 549 /* 550 * This routine is called when we have too many vnodes. It attempts 551 * to free <count> vnodes and will potentially free vnodes that still 552 * have VM backing store (VM backing store is typically the cause 553 * of a vnode blowout so we want to do this). Therefore, this operation 554 * is not considered cheap. 555 * 556 * A number of conditions may prevent a vnode from being reclaimed. 557 * the buffer cache may have references on the vnode, a directory 558 * vnode may still have references due to the namei cache representing 559 * underlying files, or the vnode may be in active use. It is not 560 * desireable to reuse such vnodes. These conditions may cause the 561 * number of vnodes to reach some minimum value regardless of what 562 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 563 */ 564 static int 565 vlrureclaim(struct mount *mp) 566 { 567 struct thread *td; 568 struct vnode *vp; 569 int done; 570 int trigger; 571 int usevnodes; 572 int count; 573 574 /* 575 * Calculate the trigger point, don't allow user 576 * screwups to blow us up. This prevents us from 577 * recycling vnodes with lots of resident pages. We 578 * aren't trying to free memory, we are trying to 579 * free vnodes. 580 */ 581 usevnodes = desiredvnodes; 582 if (usevnodes <= 0) 583 usevnodes = 1; 584 trigger = cnt.v_page_count * 2 / usevnodes; 585 done = 0; 586 td = curthread; 587 vn_start_write(NULL, &mp, V_WAIT); 588 MNT_ILOCK(mp); 589 count = mp->mnt_nvnodelistsize / 10 + 1; 590 while (count != 0) { 591 vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 592 while (vp != NULL && vp->v_type == VMARKER) 593 vp = TAILQ_NEXT(vp, v_nmntvnodes); 594 if (vp == NULL) 595 break; 596 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 597 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 598 --count; 599 if (!VI_TRYLOCK(vp)) 600 goto next_iter; 601 /* 602 * If it's been deconstructed already, it's still 603 * referenced, or it exceeds the trigger, skip it. 604 */ 605 if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) || 606 (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL && 607 vp->v_object->resident_page_count > trigger)) { 608 VI_UNLOCK(vp); 609 goto next_iter; 610 } 611 MNT_IUNLOCK(mp); 612 vholdl(vp); 613 if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT, td)) { 614 vdrop(vp); 615 goto next_iter_mntunlocked; 616 } 617 VI_LOCK(vp); 618 /* 619 * v_usecount may have been bumped after VOP_LOCK() dropped 620 * the vnode interlock and before it was locked again. 621 * 622 * It is not necessary to recheck VI_DOOMED because it can 623 * only be set by another thread that holds both the vnode 624 * lock and vnode interlock. If another thread has the 625 * vnode lock before we get to VOP_LOCK() and obtains the 626 * vnode interlock after VOP_LOCK() drops the vnode 627 * interlock, the other thread will be unable to drop the 628 * vnode lock before our VOP_LOCK() call fails. 629 */ 630 if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) || 631 (vp->v_object != NULL && 632 vp->v_object->resident_page_count > trigger)) { 633 VOP_UNLOCK(vp, LK_INTERLOCK, td); 634 goto next_iter_mntunlocked; 635 } 636 KASSERT((vp->v_iflag & VI_DOOMED) == 0, 637 ("VI_DOOMED unexpectedly detected in vlrureclaim()")); 638 vgonel(vp); 639 VOP_UNLOCK(vp, 0, td); 640 vdropl(vp); 641 done++; 642 next_iter_mntunlocked: 643 if ((count % 256) != 0) 644 goto relock_mnt; 645 goto yield; 646 next_iter: 647 if ((count % 256) != 0) 648 continue; 649 MNT_IUNLOCK(mp); 650 yield: 651 uio_yield(); 652 relock_mnt: 653 MNT_ILOCK(mp); 654 } 655 MNT_IUNLOCK(mp); 656 vn_finished_write(mp); 657 return done; 658 } 659 660 /* 661 * Attempt to keep the free list at wantfreevnodes length. 662 */ 663 static void 664 vnlru_free(int count) 665 { 666 struct vnode *vp; 667 int vfslocked; 668 669 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 670 for (; count > 0; count--) { 671 vp = TAILQ_FIRST(&vnode_free_list); 672 /* 673 * The list can be modified while the free_list_mtx 674 * has been dropped and vp could be NULL here. 675 */ 676 if (!vp) 677 break; 678 VNASSERT(vp->v_op != NULL, vp, 679 ("vnlru_free: vnode already reclaimed.")); 680 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 681 /* 682 * Don't recycle if we can't get the interlock. 683 */ 684 if (!VI_TRYLOCK(vp)) { 685 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 686 continue; 687 } 688 VNASSERT(VCANRECYCLE(vp), vp, 689 ("vp inconsistent on freelist")); 690 freevnodes--; 691 vp->v_iflag &= ~VI_FREE; 692 vholdl(vp); 693 mtx_unlock(&vnode_free_list_mtx); 694 VI_UNLOCK(vp); 695 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 696 vtryrecycle(vp); 697 VFS_UNLOCK_GIANT(vfslocked); 698 /* 699 * If the recycled succeeded this vdrop will actually free 700 * the vnode. If not it will simply place it back on 701 * the free list. 702 */ 703 vdrop(vp); 704 mtx_lock(&vnode_free_list_mtx); 705 } 706 } 707 /* 708 * Attempt to recycle vnodes in a context that is always safe to block. 709 * Calling vlrurecycle() from the bowels of filesystem code has some 710 * interesting deadlock problems. 711 */ 712 static struct proc *vnlruproc; 713 static int vnlruproc_sig; 714 715 static void 716 vnlru_proc(void) 717 { 718 struct mount *mp, *nmp; 719 int done; 720 struct proc *p = vnlruproc; 721 struct thread *td = curthread; 722 723 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 724 SHUTDOWN_PRI_FIRST); 725 726 mtx_lock(&Giant); 727 728 for (;;) { 729 kproc_suspend_check(p); 730 mtx_lock(&vnode_free_list_mtx); 731 if (freevnodes > wantfreevnodes) 732 vnlru_free(freevnodes - wantfreevnodes); 733 if (numvnodes <= desiredvnodes * 9 / 10) { 734 vnlruproc_sig = 0; 735 wakeup(&vnlruproc_sig); 736 msleep(vnlruproc, &vnode_free_list_mtx, 737 PVFS|PDROP, "vlruwt", hz); 738 continue; 739 } 740 mtx_unlock(&vnode_free_list_mtx); 741 done = 0; 742 mtx_lock(&mountlist_mtx); 743 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 744 int vfsunlocked; 745 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 746 nmp = TAILQ_NEXT(mp, mnt_list); 747 continue; 748 } 749 if (!VFS_NEEDSGIANT(mp)) { 750 mtx_unlock(&Giant); 751 vfsunlocked = 1; 752 } else 753 vfsunlocked = 0; 754 done += vlrureclaim(mp); 755 if (vfsunlocked) 756 mtx_lock(&Giant); 757 mtx_lock(&mountlist_mtx); 758 nmp = TAILQ_NEXT(mp, mnt_list); 759 vfs_unbusy(mp, td); 760 } 761 mtx_unlock(&mountlist_mtx); 762 if (done == 0) { 763 EVENTHANDLER_INVOKE(vfs_lowvnodes, desiredvnodes / 10); 764 #if 0 765 /* These messages are temporary debugging aids */ 766 if (vnlru_nowhere < 5) 767 printf("vnlru process getting nowhere..\n"); 768 else if (vnlru_nowhere == 5) 769 printf("vnlru process messages stopped.\n"); 770 #endif 771 vnlru_nowhere++; 772 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 773 } else 774 uio_yield(); 775 } 776 } 777 778 static struct kproc_desc vnlru_kp = { 779 "vnlru", 780 vnlru_proc, 781 &vnlruproc 782 }; 783 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 784 785 /* 786 * Routines having to do with the management of the vnode table. 787 */ 788 789 static void 790 vdestroy(struct vnode *vp) 791 { 792 struct bufobj *bo; 793 794 CTR1(KTR_VFS, "vdestroy vp %p", vp); 795 mtx_lock(&vnode_free_list_mtx); 796 numvnodes--; 797 mtx_unlock(&vnode_free_list_mtx); 798 bo = &vp->v_bufobj; 799 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 800 ("cleaned vnode still on the free list.")); 801 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 802 VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count")); 803 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 804 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 805 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 806 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 807 VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL")); 808 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 809 VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL")); 810 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); 811 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); 812 VI_UNLOCK(vp); 813 #ifdef MAC 814 mac_vnode_destroy(vp); 815 #endif 816 if (vp->v_pollinfo != NULL) { 817 knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note); 818 mtx_destroy(&vp->v_pollinfo->vpi_lock); 819 uma_zfree(vnodepoll_zone, vp->v_pollinfo); 820 } 821 #ifdef INVARIANTS 822 /* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */ 823 vp->v_op = NULL; 824 #endif 825 lockdestroy(vp->v_vnlock); 826 mtx_destroy(&vp->v_interlock); 827 uma_zfree(vnode_zone, vp); 828 } 829 830 /* 831 * Try to recycle a freed vnode. We abort if anyone picks up a reference 832 * before we actually vgone(). This function must be called with the vnode 833 * held to prevent the vnode from being returned to the free list midway 834 * through vgone(). 835 */ 836 static int 837 vtryrecycle(struct vnode *vp) 838 { 839 struct thread *td = curthread; 840 struct mount *vnmp; 841 842 CTR1(KTR_VFS, "vtryrecycle: trying vp %p", vp); 843 VNASSERT(vp->v_holdcnt, vp, 844 ("vtryrecycle: Recycling vp %p without a reference.", vp)); 845 /* 846 * This vnode may found and locked via some other list, if so we 847 * can't recycle it yet. 848 */ 849 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 850 return (EWOULDBLOCK); 851 /* 852 * Don't recycle if its filesystem is being suspended. 853 */ 854 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 855 VOP_UNLOCK(vp, 0, td); 856 return (EBUSY); 857 } 858 /* 859 * If we got this far, we need to acquire the interlock and see if 860 * anyone picked up this vnode from another list. If not, we will 861 * mark it with DOOMED via vgonel() so that anyone who does find it 862 * will skip over it. 863 */ 864 VI_LOCK(vp); 865 if (vp->v_usecount) { 866 VOP_UNLOCK(vp, LK_INTERLOCK, td); 867 vn_finished_write(vnmp); 868 return (EBUSY); 869 } 870 if ((vp->v_iflag & VI_DOOMED) == 0) 871 vgonel(vp); 872 VOP_UNLOCK(vp, LK_INTERLOCK, td); 873 vn_finished_write(vnmp); 874 CTR1(KTR_VFS, "vtryrecycle: recycled vp %p", vp); 875 return (0); 876 } 877 878 /* 879 * Return the next vnode from the free list. 880 */ 881 int 882 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, 883 struct vnode **vpp) 884 { 885 struct vnode *vp = NULL; 886 struct bufobj *bo; 887 888 mtx_lock(&vnode_free_list_mtx); 889 /* 890 * Lend our context to reclaim vnodes if they've exceeded the max. 891 */ 892 if (freevnodes > wantfreevnodes) 893 vnlru_free(1); 894 /* 895 * Wait for available vnodes. 896 */ 897 if (numvnodes > desiredvnodes) { 898 if (mp != NULL && (mp->mnt_kern_flag & MNTK_SUSPEND)) { 899 /* 900 * File system is beeing suspended, we cannot risk a 901 * deadlock here, so allocate new vnode anyway. 902 */ 903 if (freevnodes > wantfreevnodes) 904 vnlru_free(freevnodes - wantfreevnodes); 905 goto alloc; 906 } 907 if (vnlruproc_sig == 0) { 908 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 909 wakeup(vnlruproc); 910 } 911 msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS, 912 "vlruwk", hz); 913 #if 0 /* XXX Not all VFS_VGET/ffs_vget callers check returns. */ 914 if (numvnodes > desiredvnodes) { 915 mtx_unlock(&vnode_free_list_mtx); 916 return (ENFILE); 917 } 918 #endif 919 } 920 alloc: 921 numvnodes++; 922 mtx_unlock(&vnode_free_list_mtx); 923 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); 924 /* 925 * Setup locks. 926 */ 927 vp->v_vnlock = &vp->v_lock; 928 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 929 /* 930 * By default, don't allow shared locks unless filesystems 931 * opt-in. 932 */ 933 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE); 934 /* 935 * Initialize bufobj. 936 */ 937 bo = &vp->v_bufobj; 938 bo->__bo_vnode = vp; 939 bo->bo_mtx = &vp->v_interlock; 940 bo->bo_ops = &buf_ops_bio; 941 bo->bo_private = vp; 942 TAILQ_INIT(&bo->bo_clean.bv_hd); 943 TAILQ_INIT(&bo->bo_dirty.bv_hd); 944 /* 945 * Initialize namecache. 946 */ 947 LIST_INIT(&vp->v_cache_src); 948 TAILQ_INIT(&vp->v_cache_dst); 949 /* 950 * Finalize various vnode identity bits. 951 */ 952 vp->v_type = VNON; 953 vp->v_tag = tag; 954 vp->v_op = vops; 955 v_incr_usecount(vp); 956 vp->v_data = 0; 957 #ifdef MAC 958 mac_vnode_init(vp); 959 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 960 mac_vnode_associate_singlelabel(mp, vp); 961 else if (mp == NULL) 962 printf("NULL mp in getnewvnode()\n"); 963 #endif 964 if (mp != NULL) { 965 bo->bo_bsize = mp->mnt_stat.f_iosize; 966 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) 967 vp->v_vflag |= VV_NOKNOTE; 968 } 969 970 CTR2(KTR_VFS, "getnewvnode: mp %p vp %p", mp, vp); 971 *vpp = vp; 972 return (0); 973 } 974 975 /* 976 * Delete from old mount point vnode list, if on one. 977 */ 978 static void 979 delmntque(struct vnode *vp) 980 { 981 struct mount *mp; 982 983 mp = vp->v_mount; 984 if (mp == NULL) 985 return; 986 MNT_ILOCK(mp); 987 vp->v_mount = NULL; 988 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 989 ("bad mount point vnode list size")); 990 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 991 mp->mnt_nvnodelistsize--; 992 MNT_REL(mp); 993 MNT_IUNLOCK(mp); 994 } 995 996 static void 997 insmntque_stddtr(struct vnode *vp, void *dtr_arg) 998 { 999 struct thread *td; 1000 1001 td = curthread; /* XXX ? */ 1002 vp->v_data = NULL; 1003 vp->v_op = &dead_vnodeops; 1004 /* XXX non mp-safe fs may still call insmntque with vnode 1005 unlocked */ 1006 if (!VOP_ISLOCKED(vp, td)) 1007 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 1008 vgone(vp); 1009 vput(vp); 1010 } 1011 1012 /* 1013 * Insert into list of vnodes for the new mount point, if available. 1014 */ 1015 int 1016 insmntque1(struct vnode *vp, struct mount *mp, 1017 void (*dtr)(struct vnode *, void *), void *dtr_arg) 1018 { 1019 1020 KASSERT(vp->v_mount == NULL, 1021 ("insmntque: vnode already on per mount vnode list")); 1022 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 1023 MNT_ILOCK(mp); 1024 if ((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 && 1025 mp->mnt_nvnodelistsize == 0) { 1026 MNT_IUNLOCK(mp); 1027 if (dtr != NULL) 1028 dtr(vp, dtr_arg); 1029 return (EBUSY); 1030 } 1031 vp->v_mount = mp; 1032 MNT_REF(mp); 1033 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1034 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 1035 ("neg mount point vnode list size")); 1036 mp->mnt_nvnodelistsize++; 1037 MNT_IUNLOCK(mp); 1038 return (0); 1039 } 1040 1041 int 1042 insmntque(struct vnode *vp, struct mount *mp) 1043 { 1044 1045 return (insmntque1(vp, mp, insmntque_stddtr, NULL)); 1046 } 1047 1048 /* 1049 * Flush out and invalidate all buffers associated with a bufobj 1050 * Called with the underlying object locked. 1051 */ 1052 int 1053 bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag, 1054 int slptimeo) 1055 { 1056 int error; 1057 1058 BO_LOCK(bo); 1059 if (flags & V_SAVE) { 1060 error = bufobj_wwait(bo, slpflag, slptimeo); 1061 if (error) { 1062 BO_UNLOCK(bo); 1063 return (error); 1064 } 1065 if (bo->bo_dirty.bv_cnt > 0) { 1066 BO_UNLOCK(bo); 1067 if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0) 1068 return (error); 1069 /* 1070 * XXX We could save a lock/unlock if this was only 1071 * enabled under INVARIANTS 1072 */ 1073 BO_LOCK(bo); 1074 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 1075 panic("vinvalbuf: dirty bufs"); 1076 } 1077 } 1078 /* 1079 * If you alter this loop please notice that interlock is dropped and 1080 * reacquired in flushbuflist. Special care is needed to ensure that 1081 * no race conditions occur from this. 1082 */ 1083 do { 1084 error = flushbuflist(&bo->bo_clean, 1085 flags, bo, slpflag, slptimeo); 1086 if (error == 0) 1087 error = flushbuflist(&bo->bo_dirty, 1088 flags, bo, slpflag, slptimeo); 1089 if (error != 0 && error != EAGAIN) { 1090 BO_UNLOCK(bo); 1091 return (error); 1092 } 1093 } while (error != 0); 1094 1095 /* 1096 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1097 * have write I/O in-progress but if there is a VM object then the 1098 * VM object can also have read-I/O in-progress. 1099 */ 1100 do { 1101 bufobj_wwait(bo, 0, 0); 1102 BO_UNLOCK(bo); 1103 if (bo->bo_object != NULL) { 1104 VM_OBJECT_LOCK(bo->bo_object); 1105 vm_object_pip_wait(bo->bo_object, "bovlbx"); 1106 VM_OBJECT_UNLOCK(bo->bo_object); 1107 } 1108 BO_LOCK(bo); 1109 } while (bo->bo_numoutput > 0); 1110 BO_UNLOCK(bo); 1111 1112 /* 1113 * Destroy the copy in the VM cache, too. 1114 */ 1115 if (bo->bo_object != NULL) { 1116 VM_OBJECT_LOCK(bo->bo_object); 1117 vm_object_page_remove(bo->bo_object, 0, 0, 1118 (flags & V_SAVE) ? TRUE : FALSE); 1119 VM_OBJECT_UNLOCK(bo->bo_object); 1120 } 1121 1122 #ifdef INVARIANTS 1123 BO_LOCK(bo); 1124 if ((flags & (V_ALT | V_NORMAL)) == 0 && 1125 (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0)) 1126 panic("vinvalbuf: flush failed"); 1127 BO_UNLOCK(bo); 1128 #endif 1129 return (0); 1130 } 1131 1132 /* 1133 * Flush out and invalidate all buffers associated with a vnode. 1134 * Called with the underlying object locked. 1135 */ 1136 int 1137 vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag, 1138 int slptimeo) 1139 { 1140 1141 CTR2(KTR_VFS, "vinvalbuf vp %p flags %d", vp, flags); 1142 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1143 return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo)); 1144 } 1145 1146 /* 1147 * Flush out buffers on the specified list. 1148 * 1149 */ 1150 static int 1151 flushbuflist( struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 1152 int slptimeo) 1153 { 1154 struct buf *bp, *nbp; 1155 int retval, error; 1156 daddr_t lblkno; 1157 b_xflags_t xflags; 1158 1159 ASSERT_BO_LOCKED(bo); 1160 1161 retval = 0; 1162 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 1163 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1164 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1165 continue; 1166 } 1167 lblkno = 0; 1168 xflags = 0; 1169 if (nbp != NULL) { 1170 lblkno = nbp->b_lblkno; 1171 xflags = nbp->b_xflags & 1172 (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN); 1173 } 1174 retval = EAGAIN; 1175 error = BUF_TIMELOCK(bp, 1176 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo), 1177 "flushbuf", slpflag, slptimeo); 1178 if (error) { 1179 BO_LOCK(bo); 1180 return (error != ENOLCK ? error : EAGAIN); 1181 } 1182 KASSERT(bp->b_bufobj == bo, 1183 ("bp %p wrong b_bufobj %p should be %p", 1184 bp, bp->b_bufobj, bo)); 1185 if (bp->b_bufobj != bo) { /* XXX: necessary ? */ 1186 BUF_UNLOCK(bp); 1187 BO_LOCK(bo); 1188 return (EAGAIN); 1189 } 1190 /* 1191 * XXX Since there are no node locks for NFS, I 1192 * believe there is a slight chance that a delayed 1193 * write will occur while sleeping just above, so 1194 * check for it. 1195 */ 1196 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1197 (flags & V_SAVE)) { 1198 bremfree(bp); 1199 bp->b_flags |= B_ASYNC; 1200 bwrite(bp); 1201 BO_LOCK(bo); 1202 return (EAGAIN); /* XXX: why not loop ? */ 1203 } 1204 bremfree(bp); 1205 bp->b_flags |= (B_INVAL | B_RELBUF); 1206 bp->b_flags &= ~B_ASYNC; 1207 brelse(bp); 1208 BO_LOCK(bo); 1209 if (nbp != NULL && 1210 (nbp->b_bufobj != bo || 1211 nbp->b_lblkno != lblkno || 1212 (nbp->b_xflags & 1213 (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN)) != xflags)) 1214 break; /* nbp invalid */ 1215 } 1216 return (retval); 1217 } 1218 1219 /* 1220 * Truncate a file's buffer and pages to a specified length. This 1221 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1222 * sync activity. 1223 */ 1224 int 1225 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, 1226 off_t length, int blksize) 1227 { 1228 struct buf *bp, *nbp; 1229 int anyfreed; 1230 int trunclbn; 1231 struct bufobj *bo; 1232 1233 CTR2(KTR_VFS, "vtruncbuf vp %p length %jd", vp, length); 1234 /* 1235 * Round up to the *next* lbn. 1236 */ 1237 trunclbn = (length + blksize - 1) / blksize; 1238 1239 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1240 restart: 1241 VI_LOCK(vp); 1242 bo = &vp->v_bufobj; 1243 anyfreed = 1; 1244 for (;anyfreed;) { 1245 anyfreed = 0; 1246 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 1247 if (bp->b_lblkno < trunclbn) 1248 continue; 1249 if (BUF_LOCK(bp, 1250 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1251 VI_MTX(vp)) == ENOLCK) 1252 goto restart; 1253 1254 bremfree(bp); 1255 bp->b_flags |= (B_INVAL | B_RELBUF); 1256 bp->b_flags &= ~B_ASYNC; 1257 brelse(bp); 1258 anyfreed = 1; 1259 1260 if (nbp != NULL && 1261 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1262 (nbp->b_vp != vp) || 1263 (nbp->b_flags & B_DELWRI))) { 1264 goto restart; 1265 } 1266 VI_LOCK(vp); 1267 } 1268 1269 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1270 if (bp->b_lblkno < trunclbn) 1271 continue; 1272 if (BUF_LOCK(bp, 1273 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1274 VI_MTX(vp)) == ENOLCK) 1275 goto restart; 1276 bremfree(bp); 1277 bp->b_flags |= (B_INVAL | B_RELBUF); 1278 bp->b_flags &= ~B_ASYNC; 1279 brelse(bp); 1280 anyfreed = 1; 1281 if (nbp != NULL && 1282 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1283 (nbp->b_vp != vp) || 1284 (nbp->b_flags & B_DELWRI) == 0)) { 1285 goto restart; 1286 } 1287 VI_LOCK(vp); 1288 } 1289 } 1290 1291 if (length > 0) { 1292 restartsync: 1293 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1294 if (bp->b_lblkno > 0) 1295 continue; 1296 /* 1297 * Since we hold the vnode lock this should only 1298 * fail if we're racing with the buf daemon. 1299 */ 1300 if (BUF_LOCK(bp, 1301 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1302 VI_MTX(vp)) == ENOLCK) { 1303 goto restart; 1304 } 1305 VNASSERT((bp->b_flags & B_DELWRI), vp, 1306 ("buf(%p) on dirty queue without DELWRI", bp)); 1307 1308 bremfree(bp); 1309 bawrite(bp); 1310 VI_LOCK(vp); 1311 goto restartsync; 1312 } 1313 } 1314 1315 bufobj_wwait(bo, 0, 0); 1316 VI_UNLOCK(vp); 1317 vnode_pager_setsize(vp, length); 1318 1319 return (0); 1320 } 1321 1322 /* 1323 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1324 * a vnode. 1325 * 1326 * NOTE: We have to deal with the special case of a background bitmap 1327 * buffer, a situation where two buffers will have the same logical 1328 * block offset. We want (1) only the foreground buffer to be accessed 1329 * in a lookup and (2) must differentiate between the foreground and 1330 * background buffer in the splay tree algorithm because the splay 1331 * tree cannot normally handle multiple entities with the same 'index'. 1332 * We accomplish this by adding differentiating flags to the splay tree's 1333 * numerical domain. 1334 */ 1335 static 1336 struct buf * 1337 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1338 { 1339 struct buf dummy; 1340 struct buf *lefttreemax, *righttreemin, *y; 1341 1342 if (root == NULL) 1343 return (NULL); 1344 lefttreemax = righttreemin = &dummy; 1345 for (;;) { 1346 if (lblkno < root->b_lblkno || 1347 (lblkno == root->b_lblkno && 1348 (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1349 if ((y = root->b_left) == NULL) 1350 break; 1351 if (lblkno < y->b_lblkno) { 1352 /* Rotate right. */ 1353 root->b_left = y->b_right; 1354 y->b_right = root; 1355 root = y; 1356 if ((y = root->b_left) == NULL) 1357 break; 1358 } 1359 /* Link into the new root's right tree. */ 1360 righttreemin->b_left = root; 1361 righttreemin = root; 1362 } else if (lblkno > root->b_lblkno || 1363 (lblkno == root->b_lblkno && 1364 (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { 1365 if ((y = root->b_right) == NULL) 1366 break; 1367 if (lblkno > y->b_lblkno) { 1368 /* Rotate left. */ 1369 root->b_right = y->b_left; 1370 y->b_left = root; 1371 root = y; 1372 if ((y = root->b_right) == NULL) 1373 break; 1374 } 1375 /* Link into the new root's left tree. */ 1376 lefttreemax->b_right = root; 1377 lefttreemax = root; 1378 } else { 1379 break; 1380 } 1381 root = y; 1382 } 1383 /* Assemble the new root. */ 1384 lefttreemax->b_right = root->b_left; 1385 righttreemin->b_left = root->b_right; 1386 root->b_left = dummy.b_right; 1387 root->b_right = dummy.b_left; 1388 return (root); 1389 } 1390 1391 static void 1392 buf_vlist_remove(struct buf *bp) 1393 { 1394 struct buf *root; 1395 struct bufv *bv; 1396 1397 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1398 ASSERT_BO_LOCKED(bp->b_bufobj); 1399 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) != 1400 (BX_VNDIRTY|BX_VNCLEAN), 1401 ("buf_vlist_remove: Buf %p is on two lists", bp)); 1402 if (bp->b_xflags & BX_VNDIRTY) 1403 bv = &bp->b_bufobj->bo_dirty; 1404 else 1405 bv = &bp->b_bufobj->bo_clean; 1406 if (bp != bv->bv_root) { 1407 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1408 KASSERT(root == bp, ("splay lookup failed in remove")); 1409 } 1410 if (bp->b_left == NULL) { 1411 root = bp->b_right; 1412 } else { 1413 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1414 root->b_right = bp->b_right; 1415 } 1416 bv->bv_root = root; 1417 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 1418 bv->bv_cnt--; 1419 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1420 } 1421 1422 /* 1423 * Add the buffer to the sorted clean or dirty block list using a 1424 * splay tree algorithm. 1425 * 1426 * NOTE: xflags is passed as a constant, optimizing this inline function! 1427 */ 1428 static void 1429 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 1430 { 1431 struct buf *root; 1432 struct bufv *bv; 1433 1434 ASSERT_BO_LOCKED(bo); 1435 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1436 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 1437 bp->b_xflags |= xflags; 1438 if (xflags & BX_VNDIRTY) 1439 bv = &bo->bo_dirty; 1440 else 1441 bv = &bo->bo_clean; 1442 1443 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1444 if (root == NULL) { 1445 bp->b_left = NULL; 1446 bp->b_right = NULL; 1447 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 1448 } else if (bp->b_lblkno < root->b_lblkno || 1449 (bp->b_lblkno == root->b_lblkno && 1450 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1451 bp->b_left = root->b_left; 1452 bp->b_right = root; 1453 root->b_left = NULL; 1454 TAILQ_INSERT_BEFORE(root, bp, b_bobufs); 1455 } else { 1456 bp->b_right = root->b_right; 1457 bp->b_left = root; 1458 root->b_right = NULL; 1459 TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs); 1460 } 1461 bv->bv_cnt++; 1462 bv->bv_root = bp; 1463 } 1464 1465 /* 1466 * Lookup a buffer using the splay tree. Note that we specifically avoid 1467 * shadow buffers used in background bitmap writes. 1468 * 1469 * This code isn't quite efficient as it could be because we are maintaining 1470 * two sorted lists and do not know which list the block resides in. 1471 * 1472 * During a "make buildworld" the desired buffer is found at one of 1473 * the roots more than 60% of the time. Thus, checking both roots 1474 * before performing either splay eliminates unnecessary splays on the 1475 * first tree splayed. 1476 */ 1477 struct buf * 1478 gbincore(struct bufobj *bo, daddr_t lblkno) 1479 { 1480 struct buf *bp; 1481 1482 ASSERT_BO_LOCKED(bo); 1483 if ((bp = bo->bo_clean.bv_root) != NULL && 1484 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1485 return (bp); 1486 if ((bp = bo->bo_dirty.bv_root) != NULL && 1487 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1488 return (bp); 1489 if ((bp = bo->bo_clean.bv_root) != NULL) { 1490 bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp); 1491 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1492 return (bp); 1493 } 1494 if ((bp = bo->bo_dirty.bv_root) != NULL) { 1495 bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp); 1496 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1497 return (bp); 1498 } 1499 return (NULL); 1500 } 1501 1502 /* 1503 * Associate a buffer with a vnode. 1504 */ 1505 void 1506 bgetvp(struct vnode *vp, struct buf *bp) 1507 { 1508 1509 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 1510 1511 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 1512 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 1513 ("bgetvp: bp already attached! %p", bp)); 1514 1515 ASSERT_VI_LOCKED(vp, "bgetvp"); 1516 vholdl(vp); 1517 if (VFS_NEEDSGIANT(vp->v_mount) || 1518 vp->v_bufobj.bo_flag & BO_NEEDSGIANT) 1519 bp->b_flags |= B_NEEDSGIANT; 1520 bp->b_vp = vp; 1521 bp->b_bufobj = &vp->v_bufobj; 1522 /* 1523 * Insert onto list for new vnode. 1524 */ 1525 buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN); 1526 } 1527 1528 /* 1529 * Disassociate a buffer from a vnode. 1530 */ 1531 void 1532 brelvp(struct buf *bp) 1533 { 1534 struct bufobj *bo; 1535 struct vnode *vp; 1536 1537 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1538 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1539 1540 /* 1541 * Delete from old vnode list, if on one. 1542 */ 1543 vp = bp->b_vp; /* XXX */ 1544 bo = bp->b_bufobj; 1545 BO_LOCK(bo); 1546 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1547 buf_vlist_remove(bp); 1548 else 1549 panic("brelvp: Buffer %p not on queue.", bp); 1550 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1551 bo->bo_flag &= ~BO_ONWORKLST; 1552 mtx_lock(&sync_mtx); 1553 LIST_REMOVE(bo, bo_synclist); 1554 syncer_worklist_len--; 1555 mtx_unlock(&sync_mtx); 1556 } 1557 bp->b_flags &= ~B_NEEDSGIANT; 1558 bp->b_vp = NULL; 1559 bp->b_bufobj = NULL; 1560 vdropl(vp); 1561 } 1562 1563 /* 1564 * Add an item to the syncer work queue. 1565 */ 1566 static void 1567 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 1568 { 1569 int slot; 1570 1571 ASSERT_BO_LOCKED(bo); 1572 1573 mtx_lock(&sync_mtx); 1574 if (bo->bo_flag & BO_ONWORKLST) 1575 LIST_REMOVE(bo, bo_synclist); 1576 else { 1577 bo->bo_flag |= BO_ONWORKLST; 1578 syncer_worklist_len++; 1579 } 1580 1581 if (delay > syncer_maxdelay - 2) 1582 delay = syncer_maxdelay - 2; 1583 slot = (syncer_delayno + delay) & syncer_mask; 1584 1585 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 1586 mtx_unlock(&sync_mtx); 1587 } 1588 1589 static int 1590 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 1591 { 1592 int error, len; 1593 1594 mtx_lock(&sync_mtx); 1595 len = syncer_worklist_len - sync_vnode_count; 1596 mtx_unlock(&sync_mtx); 1597 error = SYSCTL_OUT(req, &len, sizeof(len)); 1598 return (error); 1599 } 1600 1601 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0, 1602 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 1603 1604 static struct proc *updateproc; 1605 static void sched_sync(void); 1606 static struct kproc_desc up_kp = { 1607 "syncer", 1608 sched_sync, 1609 &updateproc 1610 }; 1611 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 1612 1613 static int 1614 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) 1615 { 1616 struct vnode *vp; 1617 struct mount *mp; 1618 int vfslocked; 1619 1620 vfslocked = 0; 1621 restart: 1622 *bo = LIST_FIRST(slp); 1623 if (*bo == NULL) { 1624 VFS_UNLOCK_GIANT(vfslocked); 1625 return (0); 1626 } 1627 vp = (*bo)->__bo_vnode; /* XXX */ 1628 if (VFS_NEEDSGIANT(vp->v_mount)) { 1629 if (!vfslocked) { 1630 vfslocked = 1; 1631 if (mtx_trylock(&Giant) == 0) { 1632 mtx_unlock(&sync_mtx); 1633 mtx_lock(&Giant); 1634 mtx_lock(&sync_mtx); 1635 goto restart; 1636 } 1637 } 1638 } else { 1639 VFS_UNLOCK_GIANT(vfslocked); 1640 vfslocked = 0; 1641 } 1642 if (VOP_ISLOCKED(vp, NULL) != 0) { 1643 VFS_UNLOCK_GIANT(vfslocked); 1644 return (1); 1645 } 1646 if (VI_TRYLOCK(vp) == 0) { 1647 VFS_UNLOCK_GIANT(vfslocked); 1648 return (1); 1649 } 1650 /* 1651 * We use vhold in case the vnode does not 1652 * successfully sync. vhold prevents the vnode from 1653 * going away when we unlock the sync_mtx so that 1654 * we can acquire the vnode interlock. 1655 */ 1656 vholdl(vp); 1657 mtx_unlock(&sync_mtx); 1658 VI_UNLOCK(vp); 1659 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1660 vdrop(vp); 1661 VFS_UNLOCK_GIANT(vfslocked); 1662 mtx_lock(&sync_mtx); 1663 return (1); 1664 } 1665 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 1666 (void) VOP_FSYNC(vp, MNT_LAZY, td); 1667 VOP_UNLOCK(vp, 0, td); 1668 vn_finished_write(mp); 1669 VI_LOCK(vp); 1670 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { 1671 /* 1672 * Put us back on the worklist. The worklist 1673 * routine will remove us from our current 1674 * position and then add us back in at a later 1675 * position. 1676 */ 1677 vn_syncer_add_to_worklist(*bo, syncdelay); 1678 } 1679 vdropl(vp); 1680 VFS_UNLOCK_GIANT(vfslocked); 1681 mtx_lock(&sync_mtx); 1682 return (0); 1683 } 1684 1685 /* 1686 * System filesystem synchronizer daemon. 1687 */ 1688 static void 1689 sched_sync(void) 1690 { 1691 struct synclist *next; 1692 struct synclist *slp; 1693 struct bufobj *bo; 1694 long starttime; 1695 struct thread *td = curthread; 1696 static int dummychan; 1697 int last_work_seen; 1698 int net_worklist_len; 1699 int syncer_final_iter; 1700 int first_printf; 1701 int error; 1702 1703 last_work_seen = 0; 1704 syncer_final_iter = 0; 1705 first_printf = 1; 1706 syncer_state = SYNCER_RUNNING; 1707 starttime = time_uptime; 1708 td->td_pflags |= TDP_NORUNNINGBUF; 1709 1710 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 1711 SHUTDOWN_PRI_LAST); 1712 1713 mtx_lock(&sync_mtx); 1714 for (;;) { 1715 if (syncer_state == SYNCER_FINAL_DELAY && 1716 syncer_final_iter == 0) { 1717 mtx_unlock(&sync_mtx); 1718 kproc_suspend_check(td->td_proc); 1719 mtx_lock(&sync_mtx); 1720 } 1721 net_worklist_len = syncer_worklist_len - sync_vnode_count; 1722 if (syncer_state != SYNCER_RUNNING && 1723 starttime != time_uptime) { 1724 if (first_printf) { 1725 printf("\nSyncing disks, vnodes remaining..."); 1726 first_printf = 0; 1727 } 1728 printf("%d ", net_worklist_len); 1729 } 1730 starttime = time_uptime; 1731 1732 /* 1733 * Push files whose dirty time has expired. Be careful 1734 * of interrupt race on slp queue. 1735 * 1736 * Skip over empty worklist slots when shutting down. 1737 */ 1738 do { 1739 slp = &syncer_workitem_pending[syncer_delayno]; 1740 syncer_delayno += 1; 1741 if (syncer_delayno == syncer_maxdelay) 1742 syncer_delayno = 0; 1743 next = &syncer_workitem_pending[syncer_delayno]; 1744 /* 1745 * If the worklist has wrapped since the 1746 * it was emptied of all but syncer vnodes, 1747 * switch to the FINAL_DELAY state and run 1748 * for one more second. 1749 */ 1750 if (syncer_state == SYNCER_SHUTTING_DOWN && 1751 net_worklist_len == 0 && 1752 last_work_seen == syncer_delayno) { 1753 syncer_state = SYNCER_FINAL_DELAY; 1754 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 1755 } 1756 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 1757 syncer_worklist_len > 0); 1758 1759 /* 1760 * Keep track of the last time there was anything 1761 * on the worklist other than syncer vnodes. 1762 * Return to the SHUTTING_DOWN state if any 1763 * new work appears. 1764 */ 1765 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 1766 last_work_seen = syncer_delayno; 1767 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 1768 syncer_state = SYNCER_SHUTTING_DOWN; 1769 while (!LIST_EMPTY(slp)) { 1770 error = sync_vnode(slp, &bo, td); 1771 if (error == 1) { 1772 LIST_REMOVE(bo, bo_synclist); 1773 LIST_INSERT_HEAD(next, bo, bo_synclist); 1774 continue; 1775 } 1776 } 1777 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 1778 syncer_final_iter--; 1779 /* 1780 * The variable rushjob allows the kernel to speed up the 1781 * processing of the filesystem syncer process. A rushjob 1782 * value of N tells the filesystem syncer to process the next 1783 * N seconds worth of work on its queue ASAP. Currently rushjob 1784 * is used by the soft update code to speed up the filesystem 1785 * syncer process when the incore state is getting so far 1786 * ahead of the disk that the kernel memory pool is being 1787 * threatened with exhaustion. 1788 */ 1789 if (rushjob > 0) { 1790 rushjob -= 1; 1791 continue; 1792 } 1793 /* 1794 * Just sleep for a short period of time between 1795 * iterations when shutting down to allow some I/O 1796 * to happen. 1797 * 1798 * If it has taken us less than a second to process the 1799 * current work, then wait. Otherwise start right over 1800 * again. We can still lose time if any single round 1801 * takes more than two seconds, but it does not really 1802 * matter as we are just trying to generally pace the 1803 * filesystem activity. 1804 */ 1805 if (syncer_state != SYNCER_RUNNING) 1806 msleep(&dummychan, &sync_mtx, PPAUSE, "syncfnl", 1807 hz / SYNCER_SHUTDOWN_SPEEDUP); 1808 else if (time_uptime == starttime) 1809 msleep(&lbolt, &sync_mtx, PPAUSE, "syncer", 0); 1810 } 1811 } 1812 1813 /* 1814 * Request the syncer daemon to speed up its work. 1815 * We never push it to speed up more than half of its 1816 * normal turn time, otherwise it could take over the cpu. 1817 */ 1818 int 1819 speedup_syncer(void) 1820 { 1821 struct thread *td; 1822 int ret = 0; 1823 1824 td = FIRST_THREAD_IN_PROC(updateproc); 1825 mtx_lock(&sync_mtx); 1826 if (rushjob < syncdelay / 2) { 1827 rushjob += 1; 1828 stat_rush_requests += 1; 1829 ret = 1; 1830 } 1831 mtx_unlock(&sync_mtx); 1832 sleepq_remove(td, &lbolt); 1833 return (ret); 1834 } 1835 1836 /* 1837 * Tell the syncer to speed up its work and run though its work 1838 * list several times, then tell it to shut down. 1839 */ 1840 static void 1841 syncer_shutdown(void *arg, int howto) 1842 { 1843 struct thread *td; 1844 1845 if (howto & RB_NOSYNC) 1846 return; 1847 td = FIRST_THREAD_IN_PROC(updateproc); 1848 mtx_lock(&sync_mtx); 1849 syncer_state = SYNCER_SHUTTING_DOWN; 1850 rushjob = 0; 1851 mtx_unlock(&sync_mtx); 1852 sleepq_remove(td, &lbolt); 1853 kproc_shutdown(arg, howto); 1854 } 1855 1856 /* 1857 * Reassign a buffer from one vnode to another. 1858 * Used to assign file specific control information 1859 * (indirect blocks) to the vnode to which they belong. 1860 */ 1861 void 1862 reassignbuf(struct buf *bp) 1863 { 1864 struct vnode *vp; 1865 struct bufobj *bo; 1866 int delay; 1867 #ifdef INVARIANTS 1868 struct bufv *bv; 1869 #endif 1870 1871 vp = bp->b_vp; 1872 bo = bp->b_bufobj; 1873 ++reassignbufcalls; 1874 1875 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 1876 bp, bp->b_vp, bp->b_flags); 1877 /* 1878 * B_PAGING flagged buffers cannot be reassigned because their vp 1879 * is not fully linked in. 1880 */ 1881 if (bp->b_flags & B_PAGING) 1882 panic("cannot reassign paging buffer"); 1883 1884 /* 1885 * Delete from old vnode list, if on one. 1886 */ 1887 VI_LOCK(vp); 1888 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1889 buf_vlist_remove(bp); 1890 else 1891 panic("reassignbuf: Buffer %p not on queue.", bp); 1892 /* 1893 * If dirty, put on list of dirty buffers; otherwise insert onto list 1894 * of clean buffers. 1895 */ 1896 if (bp->b_flags & B_DELWRI) { 1897 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 1898 switch (vp->v_type) { 1899 case VDIR: 1900 delay = dirdelay; 1901 break; 1902 case VCHR: 1903 delay = metadelay; 1904 break; 1905 default: 1906 delay = filedelay; 1907 } 1908 vn_syncer_add_to_worklist(bo, delay); 1909 } 1910 buf_vlist_add(bp, bo, BX_VNDIRTY); 1911 } else { 1912 buf_vlist_add(bp, bo, BX_VNCLEAN); 1913 1914 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1915 mtx_lock(&sync_mtx); 1916 LIST_REMOVE(bo, bo_synclist); 1917 syncer_worklist_len--; 1918 mtx_unlock(&sync_mtx); 1919 bo->bo_flag &= ~BO_ONWORKLST; 1920 } 1921 } 1922 #ifdef INVARIANTS 1923 bv = &bo->bo_clean; 1924 bp = TAILQ_FIRST(&bv->bv_hd); 1925 KASSERT(bp == NULL || bp->b_bufobj == bo, 1926 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 1927 bp = TAILQ_LAST(&bv->bv_hd, buflists); 1928 KASSERT(bp == NULL || bp->b_bufobj == bo, 1929 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 1930 bv = &bo->bo_dirty; 1931 bp = TAILQ_FIRST(&bv->bv_hd); 1932 KASSERT(bp == NULL || bp->b_bufobj == bo, 1933 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 1934 bp = TAILQ_LAST(&bv->bv_hd, buflists); 1935 KASSERT(bp == NULL || bp->b_bufobj == bo, 1936 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 1937 #endif 1938 VI_UNLOCK(vp); 1939 } 1940 1941 /* 1942 * Increment the use and hold counts on the vnode, taking care to reference 1943 * the driver's usecount if this is a chardev. The vholdl() will remove 1944 * the vnode from the free list if it is presently free. Requires the 1945 * vnode interlock and returns with it held. 1946 */ 1947 static void 1948 v_incr_usecount(struct vnode *vp) 1949 { 1950 1951 CTR3(KTR_VFS, "v_incr_usecount: vp %p holdcnt %d usecount %d\n", 1952 vp, vp->v_holdcnt, vp->v_usecount); 1953 vp->v_usecount++; 1954 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1955 dev_lock(); 1956 vp->v_rdev->si_usecount++; 1957 dev_unlock(); 1958 } 1959 vholdl(vp); 1960 } 1961 1962 /* 1963 * Turn a holdcnt into a use+holdcnt such that only one call to 1964 * v_decr_usecount is needed. 1965 */ 1966 static void 1967 v_upgrade_usecount(struct vnode *vp) 1968 { 1969 1970 CTR3(KTR_VFS, "v_upgrade_usecount: vp %p holdcnt %d usecount %d\n", 1971 vp, vp->v_holdcnt, vp->v_usecount); 1972 vp->v_usecount++; 1973 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1974 dev_lock(); 1975 vp->v_rdev->si_usecount++; 1976 dev_unlock(); 1977 } 1978 } 1979 1980 /* 1981 * Decrement the vnode use and hold count along with the driver's usecount 1982 * if this is a chardev. The vdropl() below releases the vnode interlock 1983 * as it may free the vnode. 1984 */ 1985 static void 1986 v_decr_usecount(struct vnode *vp) 1987 { 1988 1989 CTR3(KTR_VFS, "v_decr_usecount: vp %p holdcnt %d usecount %d\n", 1990 vp, vp->v_holdcnt, vp->v_usecount); 1991 ASSERT_VI_LOCKED(vp, __FUNCTION__); 1992 VNASSERT(vp->v_usecount > 0, vp, 1993 ("v_decr_usecount: negative usecount")); 1994 vp->v_usecount--; 1995 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1996 dev_lock(); 1997 vp->v_rdev->si_usecount--; 1998 dev_unlock(); 1999 } 2000 vdropl(vp); 2001 } 2002 2003 /* 2004 * Decrement only the use count and driver use count. This is intended to 2005 * be paired with a follow on vdropl() to release the remaining hold count. 2006 * In this way we may vgone() a vnode with a 0 usecount without risk of 2007 * having it end up on a free list because the hold count is kept above 0. 2008 */ 2009 static void 2010 v_decr_useonly(struct vnode *vp) 2011 { 2012 2013 CTR3(KTR_VFS, "v_decr_useonly: vp %p holdcnt %d usecount %d\n", 2014 vp, vp->v_holdcnt, vp->v_usecount); 2015 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2016 VNASSERT(vp->v_usecount > 0, vp, 2017 ("v_decr_useonly: negative usecount")); 2018 vp->v_usecount--; 2019 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2020 dev_lock(); 2021 vp->v_rdev->si_usecount--; 2022 dev_unlock(); 2023 } 2024 } 2025 2026 /* 2027 * Grab a particular vnode from the free list, increment its 2028 * reference count and lock it. The vnode lock bit is set if the 2029 * vnode is being eliminated in vgone. The process is awakened 2030 * when the transition is completed, and an error returned to 2031 * indicate that the vnode is no longer usable (possibly having 2032 * been changed to a new filesystem type). 2033 */ 2034 int 2035 vget(struct vnode *vp, int flags, struct thread *td) 2036 { 2037 int oweinact; 2038 int oldflags; 2039 int error; 2040 2041 error = 0; 2042 oldflags = flags; 2043 oweinact = 0; 2044 VFS_ASSERT_GIANT(vp->v_mount); 2045 if ((flags & LK_INTERLOCK) == 0) 2046 VI_LOCK(vp); 2047 /* 2048 * If the inactive call was deferred because vput() was called 2049 * with a shared lock, we have to do it here before another thread 2050 * gets a reference to data that should be dead. 2051 */ 2052 if (vp->v_iflag & VI_OWEINACT) { 2053 if (flags & LK_NOWAIT) { 2054 VI_UNLOCK(vp); 2055 return (EBUSY); 2056 } 2057 flags &= ~LK_TYPE_MASK; 2058 flags |= LK_EXCLUSIVE; 2059 oweinact = 1; 2060 } 2061 vholdl(vp); 2062 if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { 2063 vdrop(vp); 2064 return (error); 2065 } 2066 VI_LOCK(vp); 2067 /* Upgrade our holdcnt to a usecount. */ 2068 v_upgrade_usecount(vp); 2069 if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0) 2070 panic("vget: vn_lock failed to return ENOENT\n"); 2071 if (oweinact) { 2072 if (vp->v_iflag & VI_OWEINACT) 2073 vinactive(vp, td); 2074 VI_UNLOCK(vp); 2075 if ((oldflags & LK_TYPE_MASK) == 0) 2076 VOP_UNLOCK(vp, 0, td); 2077 } else 2078 VI_UNLOCK(vp); 2079 return (0); 2080 } 2081 2082 /* 2083 * Increase the reference count of a vnode. 2084 */ 2085 void 2086 vref(struct vnode *vp) 2087 { 2088 2089 VI_LOCK(vp); 2090 v_incr_usecount(vp); 2091 VI_UNLOCK(vp); 2092 } 2093 2094 /* 2095 * Return reference count of a vnode. 2096 * 2097 * The results of this call are only guaranteed when some mechanism other 2098 * than the VI lock is used to stop other processes from gaining references 2099 * to the vnode. This may be the case if the caller holds the only reference. 2100 * This is also useful when stale data is acceptable as race conditions may 2101 * be accounted for by some other means. 2102 */ 2103 int 2104 vrefcnt(struct vnode *vp) 2105 { 2106 int usecnt; 2107 2108 VI_LOCK(vp); 2109 usecnt = vp->v_usecount; 2110 VI_UNLOCK(vp); 2111 2112 return (usecnt); 2113 } 2114 2115 2116 /* 2117 * Vnode put/release. 2118 * If count drops to zero, call inactive routine and return to freelist. 2119 */ 2120 void 2121 vrele(struct vnode *vp) 2122 { 2123 struct thread *td = curthread; /* XXX */ 2124 2125 KASSERT(vp != NULL, ("vrele: null vp")); 2126 VFS_ASSERT_GIANT(vp->v_mount); 2127 2128 VI_LOCK(vp); 2129 2130 /* Skip this v_writecount check if we're going to panic below. */ 2131 VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp, 2132 ("vrele: missed vn_close")); 2133 2134 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2135 vp->v_usecount == 1)) { 2136 v_decr_usecount(vp); 2137 return; 2138 } 2139 if (vp->v_usecount != 1) { 2140 #ifdef DIAGNOSTIC 2141 vprint("vrele: negative ref count", vp); 2142 #endif 2143 VI_UNLOCK(vp); 2144 panic("vrele: negative ref cnt"); 2145 } 2146 /* 2147 * We want to hold the vnode until the inactive finishes to 2148 * prevent vgone() races. We drop the use count here and the 2149 * hold count below when we're done. 2150 */ 2151 v_decr_useonly(vp); 2152 /* 2153 * We must call VOP_INACTIVE with the node locked. Mark 2154 * as VI_DOINGINACT to avoid recursion. 2155 */ 2156 vp->v_iflag |= VI_OWEINACT; 2157 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) { 2158 VI_LOCK(vp); 2159 if (vp->v_usecount > 0) 2160 vp->v_iflag &= ~VI_OWEINACT; 2161 if (vp->v_iflag & VI_OWEINACT) 2162 vinactive(vp, td); 2163 VOP_UNLOCK(vp, 0, td); 2164 } else { 2165 VI_LOCK(vp); 2166 if (vp->v_usecount > 0) 2167 vp->v_iflag &= ~VI_OWEINACT; 2168 } 2169 vdropl(vp); 2170 } 2171 2172 /* 2173 * Release an already locked vnode. This give the same effects as 2174 * unlock+vrele(), but takes less time and avoids releasing and 2175 * re-aquiring the lock (as vrele() acquires the lock internally.) 2176 */ 2177 void 2178 vput(struct vnode *vp) 2179 { 2180 struct thread *td = curthread; /* XXX */ 2181 int error; 2182 2183 KASSERT(vp != NULL, ("vput: null vp")); 2184 ASSERT_VOP_LOCKED(vp, "vput"); 2185 VFS_ASSERT_GIANT(vp->v_mount); 2186 VI_LOCK(vp); 2187 /* Skip this v_writecount check if we're going to panic below. */ 2188 VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp, 2189 ("vput: missed vn_close")); 2190 error = 0; 2191 2192 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2193 vp->v_usecount == 1)) { 2194 VOP_UNLOCK(vp, 0, td); 2195 v_decr_usecount(vp); 2196 return; 2197 } 2198 2199 if (vp->v_usecount != 1) { 2200 #ifdef DIAGNOSTIC 2201 vprint("vput: negative ref count", vp); 2202 #endif 2203 panic("vput: negative ref cnt"); 2204 } 2205 /* 2206 * We want to hold the vnode until the inactive finishes to 2207 * prevent vgone() races. We drop the use count here and the 2208 * hold count below when we're done. 2209 */ 2210 v_decr_useonly(vp); 2211 vp->v_iflag |= VI_OWEINACT; 2212 if (VOP_ISLOCKED(vp, NULL) != LK_EXCLUSIVE) { 2213 error = VOP_LOCK(vp, LK_UPGRADE|LK_INTERLOCK|LK_NOWAIT, td); 2214 VI_LOCK(vp); 2215 if (error) { 2216 if (vp->v_usecount > 0) 2217 vp->v_iflag &= ~VI_OWEINACT; 2218 goto done; 2219 } 2220 } 2221 if (vp->v_usecount > 0) 2222 vp->v_iflag &= ~VI_OWEINACT; 2223 if (vp->v_iflag & VI_OWEINACT) 2224 vinactive(vp, td); 2225 VOP_UNLOCK(vp, 0, td); 2226 done: 2227 vdropl(vp); 2228 } 2229 2230 /* 2231 * Somebody doesn't want the vnode recycled. 2232 */ 2233 void 2234 vhold(struct vnode *vp) 2235 { 2236 2237 VI_LOCK(vp); 2238 vholdl(vp); 2239 VI_UNLOCK(vp); 2240 } 2241 2242 void 2243 vholdl(struct vnode *vp) 2244 { 2245 2246 vp->v_holdcnt++; 2247 if (VSHOULDBUSY(vp)) 2248 vbusy(vp); 2249 } 2250 2251 /* 2252 * Note that there is one less who cares about this vnode. vdrop() is the 2253 * opposite of vhold(). 2254 */ 2255 void 2256 vdrop(struct vnode *vp) 2257 { 2258 2259 VI_LOCK(vp); 2260 vdropl(vp); 2261 } 2262 2263 /* 2264 * Drop the hold count of the vnode. If this is the last reference to 2265 * the vnode we will free it if it has been vgone'd otherwise it is 2266 * placed on the free list. 2267 */ 2268 void 2269 vdropl(struct vnode *vp) 2270 { 2271 2272 ASSERT_VI_LOCKED(vp, "vdropl"); 2273 if (vp->v_holdcnt <= 0) 2274 panic("vdrop: holdcnt %d", vp->v_holdcnt); 2275 vp->v_holdcnt--; 2276 if (vp->v_holdcnt == 0) { 2277 if (vp->v_iflag & VI_DOOMED) { 2278 vdestroy(vp); 2279 return; 2280 } else 2281 vfree(vp); 2282 } 2283 VI_UNLOCK(vp); 2284 } 2285 2286 /* 2287 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 2288 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 2289 * OWEINACT tracks whether a vnode missed a call to inactive due to a 2290 * failed lock upgrade. 2291 */ 2292 static void 2293 vinactive(struct vnode *vp, struct thread *td) 2294 { 2295 2296 ASSERT_VOP_LOCKED(vp, "vinactive"); 2297 ASSERT_VI_LOCKED(vp, "vinactive"); 2298 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, 2299 ("vinactive: recursed on VI_DOINGINACT")); 2300 vp->v_iflag |= VI_DOINGINACT; 2301 vp->v_iflag &= ~VI_OWEINACT; 2302 VI_UNLOCK(vp); 2303 VOP_INACTIVE(vp, td); 2304 VI_LOCK(vp); 2305 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, 2306 ("vinactive: lost VI_DOINGINACT")); 2307 vp->v_iflag &= ~VI_DOINGINACT; 2308 } 2309 2310 /* 2311 * Remove any vnodes in the vnode table belonging to mount point mp. 2312 * 2313 * If FORCECLOSE is not specified, there should not be any active ones, 2314 * return error if any are found (nb: this is a user error, not a 2315 * system error). If FORCECLOSE is specified, detach any active vnodes 2316 * that are found. 2317 * 2318 * If WRITECLOSE is set, only flush out regular file vnodes open for 2319 * writing. 2320 * 2321 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2322 * 2323 * `rootrefs' specifies the base reference count for the root vnode 2324 * of this filesystem. The root vnode is considered busy if its 2325 * v_usecount exceeds this value. On a successful return, vflush(, td) 2326 * will call vrele() on the root vnode exactly rootrefs times. 2327 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2328 * be zero. 2329 */ 2330 #ifdef DIAGNOSTIC 2331 static int busyprt = 0; /* print out busy vnodes */ 2332 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 2333 #endif 2334 2335 int 2336 vflush( struct mount *mp, int rootrefs, int flags, struct thread *td) 2337 { 2338 struct vnode *vp, *mvp, *rootvp = NULL; 2339 struct vattr vattr; 2340 int busy = 0, error; 2341 2342 CTR1(KTR_VFS, "vflush: mp %p", mp); 2343 if (rootrefs > 0) { 2344 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2345 ("vflush: bad args")); 2346 /* 2347 * Get the filesystem root vnode. We can vput() it 2348 * immediately, since with rootrefs > 0, it won't go away. 2349 */ 2350 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0) 2351 return (error); 2352 vput(rootvp); 2353 2354 } 2355 MNT_ILOCK(mp); 2356 loop: 2357 MNT_VNODE_FOREACH(vp, mp, mvp) { 2358 2359 VI_LOCK(vp); 2360 vholdl(vp); 2361 MNT_IUNLOCK(mp); 2362 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td); 2363 if (error) { 2364 vdrop(vp); 2365 MNT_ILOCK(mp); 2366 MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp); 2367 goto loop; 2368 } 2369 /* 2370 * Skip over a vnodes marked VV_SYSTEM. 2371 */ 2372 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2373 VOP_UNLOCK(vp, 0, td); 2374 vdrop(vp); 2375 MNT_ILOCK(mp); 2376 continue; 2377 } 2378 /* 2379 * If WRITECLOSE is set, flush out unlinked but still open 2380 * files (even if open only for reading) and regular file 2381 * vnodes open for writing. 2382 */ 2383 if (flags & WRITECLOSE) { 2384 error = VOP_GETATTR(vp, &vattr, td->td_ucred, td); 2385 VI_LOCK(vp); 2386 2387 if ((vp->v_type == VNON || 2388 (error == 0 && vattr.va_nlink > 0)) && 2389 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2390 VOP_UNLOCK(vp, 0, td); 2391 vdropl(vp); 2392 MNT_ILOCK(mp); 2393 continue; 2394 } 2395 } else 2396 VI_LOCK(vp); 2397 /* 2398 * With v_usecount == 0, all we need to do is clear out the 2399 * vnode data structures and we are done. 2400 * 2401 * If FORCECLOSE is set, forcibly close the vnode. 2402 */ 2403 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 2404 VNASSERT(vp->v_usecount == 0 || 2405 (vp->v_type != VCHR && vp->v_type != VBLK), vp, 2406 ("device VNODE %p is FORCECLOSED", vp)); 2407 vgonel(vp); 2408 } else { 2409 busy++; 2410 #ifdef DIAGNOSTIC 2411 if (busyprt) 2412 vprint("vflush: busy vnode", vp); 2413 #endif 2414 } 2415 VOP_UNLOCK(vp, 0, td); 2416 vdropl(vp); 2417 MNT_ILOCK(mp); 2418 } 2419 MNT_IUNLOCK(mp); 2420 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2421 /* 2422 * If just the root vnode is busy, and if its refcount 2423 * is equal to `rootrefs', then go ahead and kill it. 2424 */ 2425 VI_LOCK(rootvp); 2426 KASSERT(busy > 0, ("vflush: not busy")); 2427 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 2428 ("vflush: usecount %d < rootrefs %d", 2429 rootvp->v_usecount, rootrefs)); 2430 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2431 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK, td); 2432 vgone(rootvp); 2433 VOP_UNLOCK(rootvp, 0, td); 2434 busy = 0; 2435 } else 2436 VI_UNLOCK(rootvp); 2437 } 2438 if (busy) 2439 return (EBUSY); 2440 for (; rootrefs > 0; rootrefs--) 2441 vrele(rootvp); 2442 return (0); 2443 } 2444 2445 /* 2446 * Recycle an unused vnode to the front of the free list. 2447 */ 2448 int 2449 vrecycle(struct vnode *vp, struct thread *td) 2450 { 2451 int recycled; 2452 2453 ASSERT_VOP_LOCKED(vp, "vrecycle"); 2454 recycled = 0; 2455 VI_LOCK(vp); 2456 if (vp->v_usecount == 0) { 2457 recycled = 1; 2458 vgonel(vp); 2459 } 2460 VI_UNLOCK(vp); 2461 return (recycled); 2462 } 2463 2464 /* 2465 * Eliminate all activity associated with a vnode 2466 * in preparation for reuse. 2467 */ 2468 void 2469 vgone(struct vnode *vp) 2470 { 2471 VI_LOCK(vp); 2472 vgonel(vp); 2473 VI_UNLOCK(vp); 2474 } 2475 2476 /* 2477 * vgone, with the vp interlock held. 2478 */ 2479 void 2480 vgonel(struct vnode *vp) 2481 { 2482 struct thread *td; 2483 int oweinact; 2484 int active; 2485 struct mount *mp; 2486 2487 CTR1(KTR_VFS, "vgonel: vp %p", vp); 2488 ASSERT_VOP_LOCKED(vp, "vgonel"); 2489 ASSERT_VI_LOCKED(vp, "vgonel"); 2490 VNASSERT(vp->v_holdcnt, vp, 2491 ("vgonel: vp %p has no reference.", vp)); 2492 td = curthread; 2493 2494 /* 2495 * Don't vgonel if we're already doomed. 2496 */ 2497 if (vp->v_iflag & VI_DOOMED) 2498 return; 2499 vp->v_iflag |= VI_DOOMED; 2500 /* 2501 * Check to see if the vnode is in use. If so, we have to call 2502 * VOP_CLOSE() and VOP_INACTIVE(). 2503 */ 2504 active = vp->v_usecount; 2505 oweinact = (vp->v_iflag & VI_OWEINACT); 2506 VI_UNLOCK(vp); 2507 /* 2508 * Clean out any buffers associated with the vnode. 2509 * If the flush fails, just toss the buffers. 2510 */ 2511 mp = NULL; 2512 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 2513 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 2514 if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0) 2515 vinvalbuf(vp, 0, td, 0, 0); 2516 2517 /* 2518 * If purging an active vnode, it must be closed and 2519 * deactivated before being reclaimed. 2520 */ 2521 if (active) 2522 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2523 if (oweinact || active) { 2524 VI_LOCK(vp); 2525 if ((vp->v_iflag & VI_DOINGINACT) == 0) 2526 vinactive(vp, td); 2527 VI_UNLOCK(vp); 2528 } 2529 /* 2530 * Reclaim the vnode. 2531 */ 2532 if (VOP_RECLAIM(vp, td)) 2533 panic("vgone: cannot reclaim"); 2534 if (mp != NULL) 2535 vn_finished_secondary_write(mp); 2536 VNASSERT(vp->v_object == NULL, vp, 2537 ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag)); 2538 /* 2539 * Delete from old mount point vnode list. 2540 */ 2541 delmntque(vp); 2542 cache_purge(vp); 2543 /* 2544 * Done with purge, reset to the standard lock and invalidate 2545 * the vnode. 2546 */ 2547 VI_LOCK(vp); 2548 vp->v_vnlock = &vp->v_lock; 2549 vp->v_op = &dead_vnodeops; 2550 vp->v_tag = "none"; 2551 vp->v_type = VBAD; 2552 } 2553 2554 /* 2555 * Calculate the total number of references to a special device. 2556 */ 2557 int 2558 vcount(struct vnode *vp) 2559 { 2560 int count; 2561 2562 dev_lock(); 2563 count = vp->v_rdev->si_usecount; 2564 dev_unlock(); 2565 return (count); 2566 } 2567 2568 /* 2569 * Same as above, but using the struct cdev *as argument 2570 */ 2571 int 2572 count_dev(struct cdev *dev) 2573 { 2574 int count; 2575 2576 dev_lock(); 2577 count = dev->si_usecount; 2578 dev_unlock(); 2579 return(count); 2580 } 2581 2582 /* 2583 * Print out a description of a vnode. 2584 */ 2585 static char *typename[] = 2586 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", 2587 "VMARKER"}; 2588 2589 void 2590 vn_printf(struct vnode *vp, const char *fmt, ...) 2591 { 2592 va_list ap; 2593 char buf[256], buf2[16]; 2594 u_long flags; 2595 2596 va_start(ap, fmt); 2597 vprintf(fmt, ap); 2598 va_end(ap); 2599 printf("%p: ", (void *)vp); 2600 printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]); 2601 printf(" usecount %d, writecount %d, refcount %d mountedhere %p\n", 2602 vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere); 2603 buf[0] = '\0'; 2604 buf[1] = '\0'; 2605 if (vp->v_vflag & VV_ROOT) 2606 strlcat(buf, "|VV_ROOT", sizeof(buf)); 2607 if (vp->v_vflag & VV_ISTTY) 2608 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 2609 if (vp->v_vflag & VV_NOSYNC) 2610 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 2611 if (vp->v_vflag & VV_CACHEDLABEL) 2612 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 2613 if (vp->v_vflag & VV_TEXT) 2614 strlcat(buf, "|VV_TEXT", sizeof(buf)); 2615 if (vp->v_vflag & VV_COPYONWRITE) 2616 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 2617 if (vp->v_vflag & VV_SYSTEM) 2618 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 2619 if (vp->v_vflag & VV_PROCDEP) 2620 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 2621 if (vp->v_vflag & VV_NOKNOTE) 2622 strlcat(buf, "|VV_NOKNOTE", sizeof(buf)); 2623 if (vp->v_vflag & VV_DELETED) 2624 strlcat(buf, "|VV_DELETED", sizeof(buf)); 2625 if (vp->v_vflag & VV_MD) 2626 strlcat(buf, "|VV_MD", sizeof(buf)); 2627 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | 2628 VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | 2629 VV_NOKNOTE | VV_DELETED | VV_MD); 2630 if (flags != 0) { 2631 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 2632 strlcat(buf, buf2, sizeof(buf)); 2633 } 2634 if (vp->v_iflag & VI_MOUNT) 2635 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 2636 if (vp->v_iflag & VI_AGE) 2637 strlcat(buf, "|VI_AGE", sizeof(buf)); 2638 if (vp->v_iflag & VI_DOOMED) 2639 strlcat(buf, "|VI_DOOMED", sizeof(buf)); 2640 if (vp->v_iflag & VI_FREE) 2641 strlcat(buf, "|VI_FREE", sizeof(buf)); 2642 if (vp->v_iflag & VI_OBJDIRTY) 2643 strlcat(buf, "|VI_OBJDIRTY", sizeof(buf)); 2644 if (vp->v_iflag & VI_DOINGINACT) 2645 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 2646 if (vp->v_iflag & VI_OWEINACT) 2647 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 2648 flags = vp->v_iflag & ~(VI_MOUNT | VI_AGE | VI_DOOMED | VI_FREE | 2649 VI_OBJDIRTY | VI_DOINGINACT | VI_OWEINACT); 2650 if (flags != 0) { 2651 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 2652 strlcat(buf, buf2, sizeof(buf)); 2653 } 2654 printf(" flags (%s)\n", buf + 1); 2655 if (mtx_owned(VI_MTX(vp))) 2656 printf(" VI_LOCKed"); 2657 if (vp->v_object != NULL) 2658 printf(" v_object %p ref %d pages %d\n", 2659 vp->v_object, vp->v_object->ref_count, 2660 vp->v_object->resident_page_count); 2661 printf(" "); 2662 lockmgr_printinfo(vp->v_vnlock); 2663 printf("\n"); 2664 if (vp->v_data != NULL) 2665 VOP_PRINT(vp); 2666 } 2667 2668 #ifdef DDB 2669 /* 2670 * List all of the locked vnodes in the system. 2671 * Called when debugging the kernel. 2672 */ 2673 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 2674 { 2675 struct mount *mp, *nmp; 2676 struct vnode *vp; 2677 2678 /* 2679 * Note: because this is DDB, we can't obey the locking semantics 2680 * for these structures, which means we could catch an inconsistent 2681 * state and dereference a nasty pointer. Not much to be done 2682 * about that. 2683 */ 2684 db_printf("Locked vnodes\n"); 2685 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2686 nmp = TAILQ_NEXT(mp, mnt_list); 2687 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2688 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp, NULL)) 2689 vprint("", vp); 2690 } 2691 nmp = TAILQ_NEXT(mp, mnt_list); 2692 } 2693 } 2694 2695 /* 2696 * Show details about the given vnode. 2697 */ 2698 DB_SHOW_COMMAND(vnode, db_show_vnode) 2699 { 2700 struct vnode *vp; 2701 2702 if (!have_addr) 2703 return; 2704 vp = (struct vnode *)addr; 2705 vn_printf(vp, "vnode "); 2706 } 2707 #endif /* DDB */ 2708 2709 /* 2710 * Fill in a struct xvfsconf based on a struct vfsconf. 2711 */ 2712 static void 2713 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) 2714 { 2715 2716 strcpy(xvfsp->vfc_name, vfsp->vfc_name); 2717 xvfsp->vfc_typenum = vfsp->vfc_typenum; 2718 xvfsp->vfc_refcount = vfsp->vfc_refcount; 2719 xvfsp->vfc_flags = vfsp->vfc_flags; 2720 /* 2721 * These are unused in userland, we keep them 2722 * to not break binary compatibility. 2723 */ 2724 xvfsp->vfc_vfsops = NULL; 2725 xvfsp->vfc_next = NULL; 2726 } 2727 2728 /* 2729 * Top level filesystem related information gathering. 2730 */ 2731 static int 2732 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 2733 { 2734 struct vfsconf *vfsp; 2735 struct xvfsconf xvfsp; 2736 int error; 2737 2738 error = 0; 2739 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2740 bzero(&xvfsp, sizeof(xvfsp)); 2741 vfsconf2x(vfsp, &xvfsp); 2742 error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp); 2743 if (error) 2744 break; 2745 } 2746 return (error); 2747 } 2748 2749 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, 2750 "S,xvfsconf", "List of all configured filesystems"); 2751 2752 #ifndef BURN_BRIDGES 2753 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2754 2755 static int 2756 vfs_sysctl(SYSCTL_HANDLER_ARGS) 2757 { 2758 int *name = (int *)arg1 - 1; /* XXX */ 2759 u_int namelen = arg2 + 1; /* XXX */ 2760 struct vfsconf *vfsp; 2761 struct xvfsconf xvfsp; 2762 2763 printf("WARNING: userland calling deprecated sysctl, " 2764 "please rebuild world\n"); 2765 2766 #if 1 || defined(COMPAT_PRELITE2) 2767 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2768 if (namelen == 1) 2769 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2770 #endif 2771 2772 switch (name[1]) { 2773 case VFS_MAXTYPENUM: 2774 if (namelen != 2) 2775 return (ENOTDIR); 2776 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2777 case VFS_CONF: 2778 if (namelen != 3) 2779 return (ENOTDIR); /* overloaded */ 2780 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) 2781 if (vfsp->vfc_typenum == name[2]) 2782 break; 2783 if (vfsp == NULL) 2784 return (EOPNOTSUPP); 2785 bzero(&xvfsp, sizeof(xvfsp)); 2786 vfsconf2x(vfsp, &xvfsp); 2787 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 2788 } 2789 return (EOPNOTSUPP); 2790 } 2791 2792 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, 2793 vfs_sysctl, "Generic filesystem"); 2794 2795 #if 1 || defined(COMPAT_PRELITE2) 2796 2797 static int 2798 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 2799 { 2800 int error; 2801 struct vfsconf *vfsp; 2802 struct ovfsconf ovfs; 2803 2804 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2805 bzero(&ovfs, sizeof(ovfs)); 2806 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 2807 strcpy(ovfs.vfc_name, vfsp->vfc_name); 2808 ovfs.vfc_index = vfsp->vfc_typenum; 2809 ovfs.vfc_refcount = vfsp->vfc_refcount; 2810 ovfs.vfc_flags = vfsp->vfc_flags; 2811 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 2812 if (error) 2813 return error; 2814 } 2815 return 0; 2816 } 2817 2818 #endif /* 1 || COMPAT_PRELITE2 */ 2819 #endif /* !BURN_BRIDGES */ 2820 2821 #define KINFO_VNODESLOP 10 2822 #ifdef notyet 2823 /* 2824 * Dump vnode list (via sysctl). 2825 */ 2826 /* ARGSUSED */ 2827 static int 2828 sysctl_vnode(SYSCTL_HANDLER_ARGS) 2829 { 2830 struct xvnode *xvn; 2831 struct thread *td = req->td; 2832 struct mount *mp; 2833 struct vnode *vp; 2834 int error, len, n; 2835 2836 /* 2837 * Stale numvnodes access is not fatal here. 2838 */ 2839 req->lock = 0; 2840 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 2841 if (!req->oldptr) 2842 /* Make an estimate */ 2843 return (SYSCTL_OUT(req, 0, len)); 2844 2845 error = sysctl_wire_old_buffer(req, 0); 2846 if (error != 0) 2847 return (error); 2848 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 2849 n = 0; 2850 mtx_lock(&mountlist_mtx); 2851 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 2852 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 2853 continue; 2854 MNT_ILOCK(mp); 2855 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2856 if (n == len) 2857 break; 2858 vref(vp); 2859 xvn[n].xv_size = sizeof *xvn; 2860 xvn[n].xv_vnode = vp; 2861 xvn[n].xv_id = 0; /* XXX compat */ 2862 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 2863 XV_COPY(usecount); 2864 XV_COPY(writecount); 2865 XV_COPY(holdcnt); 2866 XV_COPY(mount); 2867 XV_COPY(numoutput); 2868 XV_COPY(type); 2869 #undef XV_COPY 2870 xvn[n].xv_flag = vp->v_vflag; 2871 2872 switch (vp->v_type) { 2873 case VREG: 2874 case VDIR: 2875 case VLNK: 2876 break; 2877 case VBLK: 2878 case VCHR: 2879 if (vp->v_rdev == NULL) { 2880 vrele(vp); 2881 continue; 2882 } 2883 xvn[n].xv_dev = dev2udev(vp->v_rdev); 2884 break; 2885 case VSOCK: 2886 xvn[n].xv_socket = vp->v_socket; 2887 break; 2888 case VFIFO: 2889 xvn[n].xv_fifo = vp->v_fifoinfo; 2890 break; 2891 case VNON: 2892 case VBAD: 2893 default: 2894 /* shouldn't happen? */ 2895 vrele(vp); 2896 continue; 2897 } 2898 vrele(vp); 2899 ++n; 2900 } 2901 MNT_IUNLOCK(mp); 2902 mtx_lock(&mountlist_mtx); 2903 vfs_unbusy(mp, td); 2904 if (n == len) 2905 break; 2906 } 2907 mtx_unlock(&mountlist_mtx); 2908 2909 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 2910 free(xvn, M_TEMP); 2911 return (error); 2912 } 2913 2914 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 2915 0, 0, sysctl_vnode, "S,xvnode", ""); 2916 #endif 2917 2918 /* 2919 * Unmount all filesystems. The list is traversed in reverse order 2920 * of mounting to avoid dependencies. 2921 */ 2922 void 2923 vfs_unmountall(void) 2924 { 2925 struct mount *mp; 2926 struct thread *td; 2927 int error; 2928 2929 KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread")); 2930 td = curthread; 2931 /* 2932 * Since this only runs when rebooting, it is not interlocked. 2933 */ 2934 while(!TAILQ_EMPTY(&mountlist)) { 2935 mp = TAILQ_LAST(&mountlist, mntlist); 2936 error = dounmount(mp, MNT_FORCE, td); 2937 if (error) { 2938 TAILQ_REMOVE(&mountlist, mp, mnt_list); 2939 /* 2940 * XXX: Due to the way in which we mount the root 2941 * file system off of devfs, devfs will generate a 2942 * "busy" warning when we try to unmount it before 2943 * the root. Don't print a warning as a result in 2944 * order to avoid false positive errors that may 2945 * cause needless upset. 2946 */ 2947 if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) { 2948 printf("unmount of %s failed (", 2949 mp->mnt_stat.f_mntonname); 2950 if (error == EBUSY) 2951 printf("BUSY)\n"); 2952 else 2953 printf("%d)\n", error); 2954 } 2955 } else { 2956 /* The unmount has removed mp from the mountlist */ 2957 } 2958 } 2959 } 2960 2961 /* 2962 * perform msync on all vnodes under a mount point 2963 * the mount point must be locked. 2964 */ 2965 void 2966 vfs_msync(struct mount *mp, int flags) 2967 { 2968 struct vnode *vp, *mvp; 2969 struct vm_object *obj; 2970 2971 MNT_ILOCK(mp); 2972 MNT_VNODE_FOREACH(vp, mp, mvp) { 2973 VI_LOCK(vp); 2974 if ((vp->v_iflag & VI_OBJDIRTY) && 2975 (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { 2976 MNT_IUNLOCK(mp); 2977 if (!vget(vp, 2978 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 2979 curthread)) { 2980 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 2981 vput(vp); 2982 MNT_ILOCK(mp); 2983 continue; 2984 } 2985 2986 obj = vp->v_object; 2987 if (obj != NULL) { 2988 VM_OBJECT_LOCK(obj); 2989 vm_object_page_clean(obj, 0, 0, 2990 flags == MNT_WAIT ? 2991 OBJPC_SYNC : OBJPC_NOSYNC); 2992 VM_OBJECT_UNLOCK(obj); 2993 } 2994 vput(vp); 2995 } 2996 MNT_ILOCK(mp); 2997 } else 2998 VI_UNLOCK(vp); 2999 } 3000 MNT_IUNLOCK(mp); 3001 } 3002 3003 /* 3004 * Mark a vnode as free, putting it up for recycling. 3005 */ 3006 static void 3007 vfree(struct vnode *vp) 3008 { 3009 3010 CTR1(KTR_VFS, "vfree vp %p", vp); 3011 ASSERT_VI_LOCKED(vp, "vfree"); 3012 mtx_lock(&vnode_free_list_mtx); 3013 VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed.")); 3014 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free")); 3015 VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't")); 3016 VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp, 3017 ("vfree: Freeing doomed vnode")); 3018 if (vp->v_iflag & VI_AGE) { 3019 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 3020 } else { 3021 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 3022 } 3023 freevnodes++; 3024 vp->v_iflag &= ~VI_AGE; 3025 vp->v_iflag |= VI_FREE; 3026 mtx_unlock(&vnode_free_list_mtx); 3027 } 3028 3029 /* 3030 * Opposite of vfree() - mark a vnode as in use. 3031 */ 3032 static void 3033 vbusy(struct vnode *vp) 3034 { 3035 CTR1(KTR_VFS, "vbusy vp %p", vp); 3036 ASSERT_VI_LOCKED(vp, "vbusy"); 3037 VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free")); 3038 VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed.")); 3039 3040 mtx_lock(&vnode_free_list_mtx); 3041 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 3042 freevnodes--; 3043 vp->v_iflag &= ~(VI_FREE|VI_AGE); 3044 mtx_unlock(&vnode_free_list_mtx); 3045 } 3046 3047 /* 3048 * Initalize per-vnode helper structure to hold poll-related state. 3049 */ 3050 void 3051 v_addpollinfo(struct vnode *vp) 3052 { 3053 struct vpollinfo *vi; 3054 3055 vi = uma_zalloc(vnodepoll_zone, M_WAITOK); 3056 if (vp->v_pollinfo != NULL) { 3057 uma_zfree(vnodepoll_zone, vi); 3058 return; 3059 } 3060 vp->v_pollinfo = vi; 3061 mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 3062 knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, vp, vfs_knllock, 3063 vfs_knlunlock, vfs_knllocked); 3064 } 3065 3066 /* 3067 * Record a process's interest in events which might happen to 3068 * a vnode. Because poll uses the historic select-style interface 3069 * internally, this routine serves as both the ``check for any 3070 * pending events'' and the ``record my interest in future events'' 3071 * functions. (These are done together, while the lock is held, 3072 * to avoid race conditions.) 3073 */ 3074 int 3075 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 3076 { 3077 3078 if (vp->v_pollinfo == NULL) 3079 v_addpollinfo(vp); 3080 mtx_lock(&vp->v_pollinfo->vpi_lock); 3081 if (vp->v_pollinfo->vpi_revents & events) { 3082 /* 3083 * This leaves events we are not interested 3084 * in available for the other process which 3085 * which presumably had requested them 3086 * (otherwise they would never have been 3087 * recorded). 3088 */ 3089 events &= vp->v_pollinfo->vpi_revents; 3090 vp->v_pollinfo->vpi_revents &= ~events; 3091 3092 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3093 return events; 3094 } 3095 vp->v_pollinfo->vpi_events |= events; 3096 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 3097 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3098 return 0; 3099 } 3100 3101 /* 3102 * Routine to create and manage a filesystem syncer vnode. 3103 */ 3104 #define sync_close ((int (*)(struct vop_close_args *))nullop) 3105 static int sync_fsync(struct vop_fsync_args *); 3106 static int sync_inactive(struct vop_inactive_args *); 3107 static int sync_reclaim(struct vop_reclaim_args *); 3108 3109 static struct vop_vector sync_vnodeops = { 3110 .vop_bypass = VOP_EOPNOTSUPP, 3111 .vop_close = sync_close, /* close */ 3112 .vop_fsync = sync_fsync, /* fsync */ 3113 .vop_inactive = sync_inactive, /* inactive */ 3114 .vop_reclaim = sync_reclaim, /* reclaim */ 3115 .vop_lock1 = vop_stdlock, /* lock */ 3116 .vop_unlock = vop_stdunlock, /* unlock */ 3117 .vop_islocked = vop_stdislocked, /* islocked */ 3118 }; 3119 3120 /* 3121 * Create a new filesystem syncer vnode for the specified mount point. 3122 */ 3123 int 3124 vfs_allocate_syncvnode(struct mount *mp) 3125 { 3126 struct vnode *vp; 3127 static long start, incr, next; 3128 int error; 3129 3130 /* Allocate a new vnode */ 3131 if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) { 3132 mp->mnt_syncer = NULL; 3133 return (error); 3134 } 3135 vp->v_type = VNON; 3136 error = insmntque(vp, mp); 3137 if (error != 0) 3138 panic("vfs_allocate_syncvnode: insmntque failed"); 3139 /* 3140 * Place the vnode onto the syncer worklist. We attempt to 3141 * scatter them about on the list so that they will go off 3142 * at evenly distributed times even if all the filesystems 3143 * are mounted at once. 3144 */ 3145 next += incr; 3146 if (next == 0 || next > syncer_maxdelay) { 3147 start /= 2; 3148 incr /= 2; 3149 if (start == 0) { 3150 start = syncer_maxdelay / 2; 3151 incr = syncer_maxdelay; 3152 } 3153 next = start; 3154 } 3155 VI_LOCK(vp); 3156 vn_syncer_add_to_worklist(&vp->v_bufobj, 3157 syncdelay > 0 ? next % syncdelay : 0); 3158 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 3159 mtx_lock(&sync_mtx); 3160 sync_vnode_count++; 3161 mtx_unlock(&sync_mtx); 3162 VI_UNLOCK(vp); 3163 mp->mnt_syncer = vp; 3164 return (0); 3165 } 3166 3167 /* 3168 * Do a lazy sync of the filesystem. 3169 */ 3170 static int 3171 sync_fsync(struct vop_fsync_args *ap) 3172 { 3173 struct vnode *syncvp = ap->a_vp; 3174 struct mount *mp = syncvp->v_mount; 3175 struct thread *td = ap->a_td; 3176 int error; 3177 struct bufobj *bo; 3178 3179 /* 3180 * We only need to do something if this is a lazy evaluation. 3181 */ 3182 if (ap->a_waitfor != MNT_LAZY) 3183 return (0); 3184 3185 /* 3186 * Move ourselves to the back of the sync list. 3187 */ 3188 bo = &syncvp->v_bufobj; 3189 BO_LOCK(bo); 3190 vn_syncer_add_to_worklist(bo, syncdelay); 3191 BO_UNLOCK(bo); 3192 3193 /* 3194 * Walk the list of vnodes pushing all that are dirty and 3195 * not already on the sync list. 3196 */ 3197 mtx_lock(&mountlist_mtx); 3198 if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { 3199 mtx_unlock(&mountlist_mtx); 3200 return (0); 3201 } 3202 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 3203 vfs_unbusy(mp, td); 3204 return (0); 3205 } 3206 MNT_ILOCK(mp); 3207 mp->mnt_noasync++; 3208 mp->mnt_kern_flag &= ~MNTK_ASYNC; 3209 MNT_IUNLOCK(mp); 3210 vfs_msync(mp, MNT_NOWAIT); 3211 error = VFS_SYNC(mp, MNT_LAZY, td); 3212 MNT_ILOCK(mp); 3213 mp->mnt_noasync--; 3214 if ((mp->mnt_flag & MNT_ASYNC) != 0 && mp->mnt_noasync == 0) 3215 mp->mnt_kern_flag |= MNTK_ASYNC; 3216 MNT_IUNLOCK(mp); 3217 vn_finished_write(mp); 3218 vfs_unbusy(mp, td); 3219 return (error); 3220 } 3221 3222 /* 3223 * The syncer vnode is no referenced. 3224 */ 3225 static int 3226 sync_inactive(struct vop_inactive_args *ap) 3227 { 3228 3229 vgone(ap->a_vp); 3230 return (0); 3231 } 3232 3233 /* 3234 * The syncer vnode is no longer needed and is being decommissioned. 3235 * 3236 * Modifications to the worklist must be protected by sync_mtx. 3237 */ 3238 static int 3239 sync_reclaim(struct vop_reclaim_args *ap) 3240 { 3241 struct vnode *vp = ap->a_vp; 3242 struct bufobj *bo; 3243 3244 VI_LOCK(vp); 3245 bo = &vp->v_bufobj; 3246 vp->v_mount->mnt_syncer = NULL; 3247 if (bo->bo_flag & BO_ONWORKLST) { 3248 mtx_lock(&sync_mtx); 3249 LIST_REMOVE(bo, bo_synclist); 3250 syncer_worklist_len--; 3251 sync_vnode_count--; 3252 mtx_unlock(&sync_mtx); 3253 bo->bo_flag &= ~BO_ONWORKLST; 3254 } 3255 VI_UNLOCK(vp); 3256 3257 return (0); 3258 } 3259 3260 /* 3261 * Check if vnode represents a disk device 3262 */ 3263 int 3264 vn_isdisk(struct vnode *vp, int *errp) 3265 { 3266 int error; 3267 3268 error = 0; 3269 dev_lock(); 3270 if (vp->v_type != VCHR) 3271 error = ENOTBLK; 3272 else if (vp->v_rdev == NULL) 3273 error = ENXIO; 3274 else if (vp->v_rdev->si_devsw == NULL) 3275 error = ENXIO; 3276 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 3277 error = ENOTBLK; 3278 dev_unlock(); 3279 if (errp != NULL) 3280 *errp = error; 3281 return (error == 0); 3282 } 3283 3284 /* 3285 * Common filesystem object access control check routine. Accepts a 3286 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3287 * and optional call-by-reference privused argument allowing vaccess() 3288 * to indicate to the caller whether privilege was used to satisfy the 3289 * request (obsoleted). Returns 0 on success, or an errno on failure. 3290 * 3291 * The ifdef'd CAPABILITIES version is here for reference, but is not 3292 * actually used. 3293 */ 3294 int 3295 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 3296 mode_t acc_mode, struct ucred *cred, int *privused) 3297 { 3298 mode_t dac_granted; 3299 mode_t priv_granted; 3300 3301 /* 3302 * Look for a normal, non-privileged way to access the file/directory 3303 * as requested. If it exists, go with that. 3304 */ 3305 3306 if (privused != NULL) 3307 *privused = 0; 3308 3309 dac_granted = 0; 3310 3311 /* Check the owner. */ 3312 if (cred->cr_uid == file_uid) { 3313 dac_granted |= VADMIN; 3314 if (file_mode & S_IXUSR) 3315 dac_granted |= VEXEC; 3316 if (file_mode & S_IRUSR) 3317 dac_granted |= VREAD; 3318 if (file_mode & S_IWUSR) 3319 dac_granted |= (VWRITE | VAPPEND); 3320 3321 if ((acc_mode & dac_granted) == acc_mode) 3322 return (0); 3323 3324 goto privcheck; 3325 } 3326 3327 /* Otherwise, check the groups (first match) */ 3328 if (groupmember(file_gid, cred)) { 3329 if (file_mode & S_IXGRP) 3330 dac_granted |= VEXEC; 3331 if (file_mode & S_IRGRP) 3332 dac_granted |= VREAD; 3333 if (file_mode & S_IWGRP) 3334 dac_granted |= (VWRITE | VAPPEND); 3335 3336 if ((acc_mode & dac_granted) == acc_mode) 3337 return (0); 3338 3339 goto privcheck; 3340 } 3341 3342 /* Otherwise, check everyone else. */ 3343 if (file_mode & S_IXOTH) 3344 dac_granted |= VEXEC; 3345 if (file_mode & S_IROTH) 3346 dac_granted |= VREAD; 3347 if (file_mode & S_IWOTH) 3348 dac_granted |= (VWRITE | VAPPEND); 3349 if ((acc_mode & dac_granted) == acc_mode) 3350 return (0); 3351 3352 privcheck: 3353 /* 3354 * Build a privilege mask to determine if the set of privileges 3355 * satisfies the requirements when combined with the granted mask 3356 * from above. For each privilege, if the privilege is required, 3357 * bitwise or the request type onto the priv_granted mask. 3358 */ 3359 priv_granted = 0; 3360 3361 if (type == VDIR) { 3362 /* 3363 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 3364 * requests, instead of PRIV_VFS_EXEC. 3365 */ 3366 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3367 !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0)) 3368 priv_granted |= VEXEC; 3369 } else { 3370 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3371 !priv_check_cred(cred, PRIV_VFS_EXEC, 0)) 3372 priv_granted |= VEXEC; 3373 } 3374 3375 if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && 3376 !priv_check_cred(cred, PRIV_VFS_READ, 0)) 3377 priv_granted |= VREAD; 3378 3379 if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3380 !priv_check_cred(cred, PRIV_VFS_WRITE, 0)) 3381 priv_granted |= (VWRITE | VAPPEND); 3382 3383 if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3384 !priv_check_cred(cred, PRIV_VFS_ADMIN, 0)) 3385 priv_granted |= VADMIN; 3386 3387 if ((acc_mode & (priv_granted | dac_granted)) == acc_mode) { 3388 /* XXX audit: privilege used */ 3389 if (privused != NULL) 3390 *privused = 1; 3391 return (0); 3392 } 3393 3394 return ((acc_mode & VADMIN) ? EPERM : EACCES); 3395 } 3396 3397 /* 3398 * Credential check based on process requesting service, and per-attribute 3399 * permissions. 3400 */ 3401 int 3402 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 3403 struct thread *td, int access) 3404 { 3405 3406 /* 3407 * Kernel-invoked always succeeds. 3408 */ 3409 if (cred == NOCRED) 3410 return (0); 3411 3412 /* 3413 * Do not allow privileged processes in jail to directly manipulate 3414 * system attributes. 3415 */ 3416 switch (attrnamespace) { 3417 case EXTATTR_NAMESPACE_SYSTEM: 3418 /* Potentially should be: return (EPERM); */ 3419 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0)); 3420 case EXTATTR_NAMESPACE_USER: 3421 return (VOP_ACCESS(vp, access, cred, td)); 3422 default: 3423 return (EPERM); 3424 } 3425 } 3426 3427 #ifdef DEBUG_VFS_LOCKS 3428 /* 3429 * This only exists to supress warnings from unlocked specfs accesses. It is 3430 * no longer ok to have an unlocked VFS. 3431 */ 3432 #define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD) 3433 3434 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 3435 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, ""); 3436 3437 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 3438 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, ""); 3439 3440 int vfs_badlock_print = 1; /* Print lock violations. */ 3441 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, ""); 3442 3443 #ifdef KDB 3444 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 3445 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, ""); 3446 #endif 3447 3448 static void 3449 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 3450 { 3451 3452 #ifdef KDB 3453 if (vfs_badlock_backtrace) 3454 kdb_backtrace(); 3455 #endif 3456 if (vfs_badlock_print) 3457 printf("%s: %p %s\n", str, (void *)vp, msg); 3458 if (vfs_badlock_ddb) 3459 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 3460 } 3461 3462 void 3463 assert_vi_locked(struct vnode *vp, const char *str) 3464 { 3465 3466 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 3467 vfs_badlock("interlock is not locked but should be", str, vp); 3468 } 3469 3470 void 3471 assert_vi_unlocked(struct vnode *vp, const char *str) 3472 { 3473 3474 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 3475 vfs_badlock("interlock is locked but should not be", str, vp); 3476 } 3477 3478 void 3479 assert_vop_locked(struct vnode *vp, const char *str) 3480 { 3481 3482 if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0) 3483 vfs_badlock("is not locked but should be", str, vp); 3484 } 3485 3486 void 3487 assert_vop_unlocked(struct vnode *vp, const char *str) 3488 { 3489 3490 if (vp && !IGNORE_LOCK(vp) && 3491 VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE) 3492 vfs_badlock("is locked but should not be", str, vp); 3493 } 3494 3495 void 3496 assert_vop_elocked(struct vnode *vp, const char *str) 3497 { 3498 3499 if (vp && !IGNORE_LOCK(vp) && 3500 VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE) 3501 vfs_badlock("is not exclusive locked but should be", str, vp); 3502 } 3503 3504 #if 0 3505 void 3506 assert_vop_elocked_other(struct vnode *vp, const char *str) 3507 { 3508 3509 if (vp && !IGNORE_LOCK(vp) && 3510 VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER) 3511 vfs_badlock("is not exclusive locked by another thread", 3512 str, vp); 3513 } 3514 3515 void 3516 assert_vop_slocked(struct vnode *vp, const char *str) 3517 { 3518 3519 if (vp && !IGNORE_LOCK(vp) && 3520 VOP_ISLOCKED(vp, curthread) != LK_SHARED) 3521 vfs_badlock("is not locked shared but should be", str, vp); 3522 } 3523 #endif /* 0 */ 3524 #endif /* DEBUG_VFS_LOCKS */ 3525 3526 void 3527 vop_rename_pre(void *ap) 3528 { 3529 struct vop_rename_args *a = ap; 3530 3531 #ifdef DEBUG_VFS_LOCKS 3532 if (a->a_tvp) 3533 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 3534 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 3535 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 3536 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 3537 3538 /* Check the source (from). */ 3539 if (a->a_tdvp != a->a_fdvp && a->a_tvp != a->a_fdvp) 3540 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 3541 if (a->a_tvp != a->a_fvp) 3542 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 3543 3544 /* Check the target. */ 3545 if (a->a_tvp) 3546 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 3547 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 3548 #endif 3549 if (a->a_tdvp != a->a_fdvp) 3550 vhold(a->a_fdvp); 3551 if (a->a_tvp != a->a_fvp) 3552 vhold(a->a_fvp); 3553 vhold(a->a_tdvp); 3554 if (a->a_tvp) 3555 vhold(a->a_tvp); 3556 } 3557 3558 void 3559 vop_strategy_pre(void *ap) 3560 { 3561 #ifdef DEBUG_VFS_LOCKS 3562 struct vop_strategy_args *a; 3563 struct buf *bp; 3564 3565 a = ap; 3566 bp = a->a_bp; 3567 3568 /* 3569 * Cluster ops lock their component buffers but not the IO container. 3570 */ 3571 if ((bp->b_flags & B_CLUSTER) != 0) 3572 return; 3573 3574 if (BUF_REFCNT(bp) < 1) { 3575 if (vfs_badlock_print) 3576 printf( 3577 "VOP_STRATEGY: bp is not locked but should be\n"); 3578 if (vfs_badlock_ddb) 3579 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 3580 } 3581 #endif 3582 } 3583 3584 void 3585 vop_lookup_pre(void *ap) 3586 { 3587 #ifdef DEBUG_VFS_LOCKS 3588 struct vop_lookup_args *a; 3589 struct vnode *dvp; 3590 3591 a = ap; 3592 dvp = a->a_dvp; 3593 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3594 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 3595 #endif 3596 } 3597 3598 void 3599 vop_lookup_post(void *ap, int rc) 3600 { 3601 #ifdef DEBUG_VFS_LOCKS 3602 struct vop_lookup_args *a; 3603 struct vnode *dvp; 3604 struct vnode *vp; 3605 3606 a = ap; 3607 dvp = a->a_dvp; 3608 vp = *(a->a_vpp); 3609 3610 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3611 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 3612 3613 if (!rc) 3614 ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)"); 3615 #endif 3616 } 3617 3618 void 3619 vop_lock_pre(void *ap) 3620 { 3621 #ifdef DEBUG_VFS_LOCKS 3622 struct vop_lock1_args *a = ap; 3623 3624 if ((a->a_flags & LK_INTERLOCK) == 0) 3625 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3626 else 3627 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 3628 #endif 3629 } 3630 3631 void 3632 vop_lock_post(void *ap, int rc) 3633 { 3634 #ifdef DEBUG_VFS_LOCKS 3635 struct vop_lock1_args *a = ap; 3636 3637 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3638 if (rc == 0) 3639 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 3640 #endif 3641 } 3642 3643 void 3644 vop_unlock_pre(void *ap) 3645 { 3646 #ifdef DEBUG_VFS_LOCKS 3647 struct vop_unlock_args *a = ap; 3648 3649 if (a->a_flags & LK_INTERLOCK) 3650 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 3651 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 3652 #endif 3653 } 3654 3655 void 3656 vop_unlock_post(void *ap, int rc) 3657 { 3658 #ifdef DEBUG_VFS_LOCKS 3659 struct vop_unlock_args *a = ap; 3660 3661 if (a->a_flags & LK_INTERLOCK) 3662 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 3663 #endif 3664 } 3665 3666 void 3667 vop_create_post(void *ap, int rc) 3668 { 3669 struct vop_create_args *a = ap; 3670 3671 if (!rc) 3672 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 3673 } 3674 3675 void 3676 vop_link_post(void *ap, int rc) 3677 { 3678 struct vop_link_args *a = ap; 3679 3680 if (!rc) { 3681 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK); 3682 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE); 3683 } 3684 } 3685 3686 void 3687 vop_mkdir_post(void *ap, int rc) 3688 { 3689 struct vop_mkdir_args *a = ap; 3690 3691 if (!rc) 3692 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 3693 } 3694 3695 void 3696 vop_mknod_post(void *ap, int rc) 3697 { 3698 struct vop_mknod_args *a = ap; 3699 3700 if (!rc) 3701 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 3702 } 3703 3704 void 3705 vop_remove_post(void *ap, int rc) 3706 { 3707 struct vop_remove_args *a = ap; 3708 3709 if (!rc) { 3710 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 3711 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 3712 } 3713 } 3714 3715 void 3716 vop_rename_post(void *ap, int rc) 3717 { 3718 struct vop_rename_args *a = ap; 3719 3720 if (!rc) { 3721 VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE); 3722 VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE); 3723 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 3724 if (a->a_tvp) 3725 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 3726 } 3727 if (a->a_tdvp != a->a_fdvp) 3728 vdrop(a->a_fdvp); 3729 if (a->a_tvp != a->a_fvp) 3730 vdrop(a->a_fvp); 3731 vdrop(a->a_tdvp); 3732 if (a->a_tvp) 3733 vdrop(a->a_tvp); 3734 } 3735 3736 void 3737 vop_rmdir_post(void *ap, int rc) 3738 { 3739 struct vop_rmdir_args *a = ap; 3740 3741 if (!rc) { 3742 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 3743 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 3744 } 3745 } 3746 3747 void 3748 vop_setattr_post(void *ap, int rc) 3749 { 3750 struct vop_setattr_args *a = ap; 3751 3752 if (!rc) 3753 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 3754 } 3755 3756 void 3757 vop_symlink_post(void *ap, int rc) 3758 { 3759 struct vop_symlink_args *a = ap; 3760 3761 if (!rc) 3762 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 3763 } 3764 3765 static struct knlist fs_knlist; 3766 3767 static void 3768 vfs_event_init(void *arg) 3769 { 3770 knlist_init(&fs_knlist, NULL, NULL, NULL, NULL); 3771 } 3772 /* XXX - correct order? */ 3773 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 3774 3775 void 3776 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused) 3777 { 3778 3779 KNOTE_UNLOCKED(&fs_knlist, event); 3780 } 3781 3782 static int filt_fsattach(struct knote *kn); 3783 static void filt_fsdetach(struct knote *kn); 3784 static int filt_fsevent(struct knote *kn, long hint); 3785 3786 struct filterops fs_filtops = 3787 { 0, filt_fsattach, filt_fsdetach, filt_fsevent }; 3788 3789 static int 3790 filt_fsattach(struct knote *kn) 3791 { 3792 3793 kn->kn_flags |= EV_CLEAR; 3794 knlist_add(&fs_knlist, kn, 0); 3795 return (0); 3796 } 3797 3798 static void 3799 filt_fsdetach(struct knote *kn) 3800 { 3801 3802 knlist_remove(&fs_knlist, kn, 0); 3803 } 3804 3805 static int 3806 filt_fsevent(struct knote *kn, long hint) 3807 { 3808 3809 kn->kn_fflags |= hint; 3810 return (kn->kn_fflags != 0); 3811 } 3812 3813 static int 3814 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 3815 { 3816 struct vfsidctl vc; 3817 int error; 3818 struct mount *mp; 3819 3820 error = SYSCTL_IN(req, &vc, sizeof(vc)); 3821 if (error) 3822 return (error); 3823 if (vc.vc_vers != VFS_CTL_VERS1) 3824 return (EINVAL); 3825 mp = vfs_getvfs(&vc.vc_fsid); 3826 if (mp == NULL) 3827 return (ENOENT); 3828 /* ensure that a specific sysctl goes to the right filesystem. */ 3829 if (strcmp(vc.vc_fstypename, "*") != 0 && 3830 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 3831 vfs_rel(mp); 3832 return (EINVAL); 3833 } 3834 VCTLTOREQ(&vc, req); 3835 error = VFS_SYSCTL(mp, vc.vc_op, req); 3836 vfs_rel(mp); 3837 return (error); 3838 } 3839 3840 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR, NULL, 0, sysctl_vfs_ctl, "", 3841 "Sysctl by fsid"); 3842 3843 /* 3844 * Function to initialize a va_filerev field sensibly. 3845 * XXX: Wouldn't a random number make a lot more sense ?? 3846 */ 3847 u_quad_t 3848 init_va_filerev(void) 3849 { 3850 struct bintime bt; 3851 3852 getbinuptime(&bt); 3853 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 3854 } 3855 3856 static int filt_vfsread(struct knote *kn, long hint); 3857 static int filt_vfswrite(struct knote *kn, long hint); 3858 static int filt_vfsvnode(struct knote *kn, long hint); 3859 static void filt_vfsdetach(struct knote *kn); 3860 static struct filterops vfsread_filtops = 3861 { 1, NULL, filt_vfsdetach, filt_vfsread }; 3862 static struct filterops vfswrite_filtops = 3863 { 1, NULL, filt_vfsdetach, filt_vfswrite }; 3864 static struct filterops vfsvnode_filtops = 3865 { 1, NULL, filt_vfsdetach, filt_vfsvnode }; 3866 3867 static void 3868 vfs_knllock(void *arg) 3869 { 3870 struct vnode *vp = arg; 3871 3872 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread); 3873 } 3874 3875 static void 3876 vfs_knlunlock(void *arg) 3877 { 3878 struct vnode *vp = arg; 3879 3880 VOP_UNLOCK(vp, 0, curthread); 3881 } 3882 3883 static int 3884 vfs_knllocked(void *arg) 3885 { 3886 struct vnode *vp = arg; 3887 3888 return (VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE); 3889 } 3890 3891 int 3892 vfs_kqfilter(struct vop_kqfilter_args *ap) 3893 { 3894 struct vnode *vp = ap->a_vp; 3895 struct knote *kn = ap->a_kn; 3896 struct knlist *knl; 3897 3898 switch (kn->kn_filter) { 3899 case EVFILT_READ: 3900 kn->kn_fop = &vfsread_filtops; 3901 break; 3902 case EVFILT_WRITE: 3903 kn->kn_fop = &vfswrite_filtops; 3904 break; 3905 case EVFILT_VNODE: 3906 kn->kn_fop = &vfsvnode_filtops; 3907 break; 3908 default: 3909 return (EINVAL); 3910 } 3911 3912 kn->kn_hook = (caddr_t)vp; 3913 3914 if (vp->v_pollinfo == NULL) 3915 v_addpollinfo(vp); 3916 if (vp->v_pollinfo == NULL) 3917 return (ENOMEM); 3918 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 3919 knlist_add(knl, kn, 0); 3920 3921 return (0); 3922 } 3923 3924 /* 3925 * Detach knote from vnode 3926 */ 3927 static void 3928 filt_vfsdetach(struct knote *kn) 3929 { 3930 struct vnode *vp = (struct vnode *)kn->kn_hook; 3931 3932 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 3933 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 3934 } 3935 3936 /*ARGSUSED*/ 3937 static int 3938 filt_vfsread(struct knote *kn, long hint) 3939 { 3940 struct vnode *vp = (struct vnode *)kn->kn_hook; 3941 struct vattr va; 3942 3943 /* 3944 * filesystem is gone, so set the EOF flag and schedule 3945 * the knote for deletion. 3946 */ 3947 if (hint == NOTE_REVOKE) { 3948 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 3949 return (1); 3950 } 3951 3952 if (VOP_GETATTR(vp, &va, curthread->td_ucred, curthread)) 3953 return (0); 3954 3955 kn->kn_data = va.va_size - kn->kn_fp->f_offset; 3956 return (kn->kn_data != 0); 3957 } 3958 3959 /*ARGSUSED*/ 3960 static int 3961 filt_vfswrite(struct knote *kn, long hint) 3962 { 3963 /* 3964 * filesystem is gone, so set the EOF flag and schedule 3965 * the knote for deletion. 3966 */ 3967 if (hint == NOTE_REVOKE) 3968 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 3969 3970 kn->kn_data = 0; 3971 return (1); 3972 } 3973 3974 static int 3975 filt_vfsvnode(struct knote *kn, long hint) 3976 { 3977 if (kn->kn_sfflags & hint) 3978 kn->kn_fflags |= hint; 3979 if (hint == NOTE_REVOKE) { 3980 kn->kn_flags |= EV_EOF; 3981 return (1); 3982 } 3983 return (kn->kn_fflags != 0); 3984 } 3985 3986 int 3987 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 3988 { 3989 int error; 3990 3991 if (dp->d_reclen > ap->a_uio->uio_resid) 3992 return (ENAMETOOLONG); 3993 error = uiomove(dp, dp->d_reclen, ap->a_uio); 3994 if (error) { 3995 if (ap->a_ncookies != NULL) { 3996 if (ap->a_cookies != NULL) 3997 free(ap->a_cookies, M_TEMP); 3998 ap->a_cookies = NULL; 3999 *ap->a_ncookies = 0; 4000 } 4001 return (error); 4002 } 4003 if (ap->a_ncookies == NULL) 4004 return (0); 4005 4006 KASSERT(ap->a_cookies, 4007 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 4008 4009 *ap->a_cookies = realloc(*ap->a_cookies, 4010 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); 4011 (*ap->a_cookies)[*ap->a_ncookies] = off; 4012 return (0); 4013 } 4014 4015 /* 4016 * Mark for update the access time of the file if the filesystem 4017 * supports VA_MARK_ATIME. This functionality is used by execve 4018 * and mmap, so we want to avoid the synchronous I/O implied by 4019 * directly setting va_atime for the sake of efficiency. 4020 */ 4021 void 4022 vfs_mark_atime(struct vnode *vp, struct thread *td) 4023 { 4024 struct vattr atimeattr; 4025 4026 if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) { 4027 VATTR_NULL(&atimeattr); 4028 atimeattr.va_vaflags |= VA_MARK_ATIME; 4029 (void)VOP_SETATTR(vp, &atimeattr, td->td_ucred, td); 4030 } 4031 } 4032