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