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 static MALLOC_DEFINE(M_NETADDR, "subr_export_host", "Export host address structure"); 85 86 static void delmntque(struct vnode *vp); 87 static void insmntque(struct vnode *vp, struct mount *mp); 88 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, 89 int slpflag, int slptimeo); 90 static void syncer_shutdown(void *arg, int howto); 91 static int vtryrecycle(struct vnode *vp); 92 static void vbusy(struct vnode *vp); 93 static void vdropl(struct vnode *vp); 94 static void vinactive(struct vnode *, struct thread *); 95 static void v_incr_usecount(struct vnode *); 96 static void v_decr_usecount(struct vnode *); 97 static void v_decr_useonly(struct vnode *); 98 static void v_upgrade_usecount(struct vnode *); 99 static void vfree(struct vnode *); 100 static void vnlru_free(int); 101 static void vdestroy(struct vnode *); 102 static void vgonel(struct vnode *); 103 static void vfs_knllock(void *arg); 104 static void vfs_knlunlock(void *arg); 105 static int vfs_knllocked(void *arg); 106 107 108 /* 109 * Enable Giant pushdown based on whether or not the vm is mpsafe in this 110 * build. Without mpsafevm the buffer cache can not run Giant free. 111 */ 112 #if defined(__alpha__) || defined(__amd64__) || defined(__i386__) || \ 113 defined(__ia64__) || defined(__sparc64__) 114 int mpsafe_vfs = 1; 115 #else 116 int mpsafe_vfs; 117 #endif 118 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs); 119 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0, 120 "MPSAFE VFS"); 121 122 /* 123 * Number of vnodes in existence. Increased whenever getnewvnode() 124 * allocates a new vnode, never decreased. 125 */ 126 static unsigned long numvnodes; 127 128 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); 129 130 /* 131 * Conversion tables for conversion from vnode types to inode formats 132 * and back. 133 */ 134 enum vtype iftovt_tab[16] = { 135 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 136 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 137 }; 138 int vttoif_tab[10] = { 139 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 140 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT 141 }; 142 143 /* 144 * List of vnodes that are ready for recycling. 145 */ 146 static TAILQ_HEAD(freelst, vnode) vnode_free_list; 147 148 /* 149 * Free vnode target. Free vnodes may simply be files which have been stat'd 150 * but not read. This is somewhat common, and a small cache of such files 151 * should be kept to avoid recreation costs. 152 */ 153 static u_long wantfreevnodes; 154 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 155 /* Number of vnodes in the free list. */ 156 static u_long freevnodes; 157 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); 158 159 /* 160 * Various variables used for debugging the new implementation of 161 * reassignbuf(). 162 * XXX these are probably of (very) limited utility now. 163 */ 164 static int reassignbufcalls; 165 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); 166 167 /* 168 * Cache for the mount type id assigned to NFS. This is used for 169 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 170 */ 171 int nfs_mount_type = -1; 172 173 /* To keep more than one thread at a time from running vfs_getnewfsid */ 174 static struct mtx mntid_mtx; 175 176 /* 177 * Lock for any access to the following: 178 * vnode_free_list 179 * numvnodes 180 * freevnodes 181 */ 182 static struct mtx vnode_free_list_mtx; 183 184 /* Publicly exported FS */ 185 struct nfs_public nfs_pub; 186 187 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 188 static uma_zone_t vnode_zone; 189 static uma_zone_t vnodepoll_zone; 190 191 /* Set to 1 to print out reclaim of active vnodes */ 192 int prtactive; 193 194 /* 195 * The workitem queue. 196 * 197 * It is useful to delay writes of file data and filesystem metadata 198 * for tens of seconds so that quickly created and deleted files need 199 * not waste disk bandwidth being created and removed. To realize this, 200 * we append vnodes to a "workitem" queue. When running with a soft 201 * updates implementation, most pending metadata dependencies should 202 * not wait for more than a few seconds. Thus, mounted on block devices 203 * are delayed only about a half the time that file data is delayed. 204 * Similarly, directory updates are more critical, so are only delayed 205 * about a third the time that file data is delayed. Thus, there are 206 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 207 * one each second (driven off the filesystem syncer process). The 208 * syncer_delayno variable indicates the next queue that is to be processed. 209 * Items that need to be processed soon are placed in this queue: 210 * 211 * syncer_workitem_pending[syncer_delayno] 212 * 213 * A delay of fifteen seconds is done by placing the request fifteen 214 * entries later in the queue: 215 * 216 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 217 * 218 */ 219 static int syncer_delayno; 220 static long syncer_mask; 221 LIST_HEAD(synclist, bufobj); 222 static struct synclist *syncer_workitem_pending; 223 /* 224 * The sync_mtx protects: 225 * bo->bo_synclist 226 * sync_vnode_count 227 * syncer_delayno 228 * syncer_state 229 * syncer_workitem_pending 230 * syncer_worklist_len 231 * rushjob 232 */ 233 static struct mtx sync_mtx; 234 235 #define SYNCER_MAXDELAY 32 236 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 237 static int syncdelay = 30; /* max time to delay syncing data */ 238 static int filedelay = 30; /* time to delay syncing files */ 239 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); 240 static int dirdelay = 29; /* time to delay syncing directories */ 241 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); 242 static int metadelay = 28; /* time to delay syncing metadata */ 243 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); 244 static int rushjob; /* number of slots to run ASAP */ 245 static int stat_rush_requests; /* number of times I/O speeded up */ 246 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); 247 248 /* 249 * When shutting down the syncer, run it at four times normal speed. 250 */ 251 #define SYNCER_SHUTDOWN_SPEEDUP 4 252 static int sync_vnode_count; 253 static int syncer_worklist_len; 254 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 255 syncer_state; 256 257 /* 258 * Number of vnodes we want to exist at any one time. This is mostly used 259 * to size hash tables in vnode-related code. It is normally not used in 260 * getnewvnode(), as wantfreevnodes is normally nonzero.) 261 * 262 * XXX desiredvnodes is historical cruft and should not exist. 263 */ 264 int desiredvnodes; 265 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 266 &desiredvnodes, 0, "Maximum number of vnodes"); 267 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 268 &wantfreevnodes, 0, "Minimum number of vnodes (legacy)"); 269 static int vnlru_nowhere; 270 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 271 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 272 273 /* 274 * Macros to control when a vnode is freed and recycled. All require 275 * the vnode interlock. 276 */ 277 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt) 278 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt) 279 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt) 280 281 282 /* 283 * Initialize the vnode management data structures. 284 */ 285 #ifndef MAXVNODES_MAX 286 #define MAXVNODES_MAX 100000 287 #endif 288 static void 289 vntblinit(void *dummy __unused) 290 { 291 292 /* 293 * Desiredvnodes is a function of the physical memory size and 294 * the kernel's heap size. Specifically, desiredvnodes scales 295 * in proportion to the physical memory size until two fifths 296 * of the kernel's heap size is consumed by vnodes and vm 297 * objects. 298 */ 299 desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size / 300 (5 * (sizeof(struct vm_object) + sizeof(struct vnode)))); 301 if (desiredvnodes > MAXVNODES_MAX) { 302 if (bootverbose) 303 printf("Reducing kern.maxvnodes %d -> %d\n", 304 desiredvnodes, MAXVNODES_MAX); 305 desiredvnodes = MAXVNODES_MAX; 306 } 307 wantfreevnodes = desiredvnodes / 4; 308 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 309 TAILQ_INIT(&vnode_free_list); 310 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 311 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 312 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 313 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 314 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 315 /* 316 * Initialize the filesystem syncer. 317 */ 318 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 319 &syncer_mask); 320 syncer_maxdelay = syncer_mask + 1; 321 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 322 } 323 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) 324 325 326 /* 327 * Mark a mount point as busy. Used to synchronize access and to delay 328 * unmounting. Interlock is not released on failure. 329 */ 330 int 331 vfs_busy(struct mount *mp, int flags, struct mtx *interlkp, 332 struct thread *td) 333 { 334 int lkflags; 335 336 MNT_ILOCK(mp); 337 MNT_REF(mp); 338 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 339 if (flags & LK_NOWAIT) { 340 MNT_REL(mp); 341 MNT_IUNLOCK(mp); 342 return (ENOENT); 343 } 344 if (interlkp) 345 mtx_unlock(interlkp); 346 mp->mnt_kern_flag |= MNTK_MWAIT; 347 /* 348 * Since all busy locks are shared except the exclusive 349 * lock granted when unmounting, the only place that a 350 * wakeup needs to be done is at the release of the 351 * exclusive lock at the end of dounmount. 352 */ 353 msleep(mp, MNT_MTX(mp), PVFS, "vfs_busy", 0); 354 MNT_REL(mp); 355 MNT_IUNLOCK(mp); 356 if (interlkp) 357 mtx_lock(interlkp); 358 return (ENOENT); 359 } 360 if (interlkp) 361 mtx_unlock(interlkp); 362 lkflags = LK_SHARED | LK_INTERLOCK; 363 if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td)) 364 panic("vfs_busy: unexpected lock failure"); 365 return (0); 366 } 367 368 /* 369 * Free a busy filesystem. 370 */ 371 void 372 vfs_unbusy(struct mount *mp, struct thread *td) 373 { 374 375 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); 376 vfs_rel(mp); 377 } 378 379 /* 380 * Lookup a mount point by filesystem identifier. 381 */ 382 struct mount * 383 vfs_getvfs(fsid_t *fsid) 384 { 385 struct mount *mp; 386 387 mtx_lock(&mountlist_mtx); 388 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 389 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 390 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 391 vfs_ref(mp); 392 mtx_unlock(&mountlist_mtx); 393 return (mp); 394 } 395 } 396 mtx_unlock(&mountlist_mtx); 397 return ((struct mount *) 0); 398 } 399 400 /* 401 * Check if a user can access priveledged mount options. 402 */ 403 int 404 vfs_suser(struct mount *mp, struct thread *td) 405 { 406 int error; 407 408 if ((mp->mnt_flag & MNT_USER) == 0 || 409 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 410 if ((error = suser(td)) != 0) 411 return (error); 412 } 413 return (0); 414 } 415 416 /* 417 * Get a new unique fsid. Try to make its val[0] unique, since this value 418 * will be used to create fake device numbers for stat(). Also try (but 419 * not so hard) make its val[0] unique mod 2^16, since some emulators only 420 * support 16-bit device numbers. We end up with unique val[0]'s for the 421 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 422 * 423 * Keep in mind that several mounts may be running in parallel. Starting 424 * the search one past where the previous search terminated is both a 425 * micro-optimization and a defense against returning the same fsid to 426 * different mounts. 427 */ 428 void 429 vfs_getnewfsid(struct mount *mp) 430 { 431 static u_int16_t mntid_base; 432 struct mount *nmp; 433 fsid_t tfsid; 434 int mtype; 435 436 mtx_lock(&mntid_mtx); 437 mtype = mp->mnt_vfc->vfc_typenum; 438 tfsid.val[1] = mtype; 439 mtype = (mtype & 0xFF) << 24; 440 for (;;) { 441 tfsid.val[0] = makedev(255, 442 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 443 mntid_base++; 444 if ((nmp = vfs_getvfs(&tfsid)) == NULL) 445 break; 446 vfs_rel(nmp); 447 } 448 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 449 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 450 mtx_unlock(&mntid_mtx); 451 } 452 453 /* 454 * Knob to control the precision of file timestamps: 455 * 456 * 0 = seconds only; nanoseconds zeroed. 457 * 1 = seconds and nanoseconds, accurate within 1/HZ. 458 * 2 = seconds and nanoseconds, truncated to microseconds. 459 * >=3 = seconds and nanoseconds, maximum precision. 460 */ 461 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 462 463 static int timestamp_precision = TSP_SEC; 464 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 465 ×tamp_precision, 0, ""); 466 467 /* 468 * Get a current timestamp. 469 */ 470 void 471 vfs_timestamp(struct timespec *tsp) 472 { 473 struct timeval tv; 474 475 switch (timestamp_precision) { 476 case TSP_SEC: 477 tsp->tv_sec = time_second; 478 tsp->tv_nsec = 0; 479 break; 480 case TSP_HZ: 481 getnanotime(tsp); 482 break; 483 case TSP_USEC: 484 microtime(&tv); 485 TIMEVAL_TO_TIMESPEC(&tv, tsp); 486 break; 487 case TSP_NSEC: 488 default: 489 nanotime(tsp); 490 break; 491 } 492 } 493 494 /* 495 * Set vnode attributes to VNOVAL 496 */ 497 void 498 vattr_null(struct vattr *vap) 499 { 500 501 vap->va_type = VNON; 502 vap->va_size = VNOVAL; 503 vap->va_bytes = VNOVAL; 504 vap->va_mode = VNOVAL; 505 vap->va_nlink = VNOVAL; 506 vap->va_uid = VNOVAL; 507 vap->va_gid = VNOVAL; 508 vap->va_fsid = VNOVAL; 509 vap->va_fileid = VNOVAL; 510 vap->va_blocksize = VNOVAL; 511 vap->va_rdev = VNOVAL; 512 vap->va_atime.tv_sec = VNOVAL; 513 vap->va_atime.tv_nsec = VNOVAL; 514 vap->va_mtime.tv_sec = VNOVAL; 515 vap->va_mtime.tv_nsec = VNOVAL; 516 vap->va_ctime.tv_sec = VNOVAL; 517 vap->va_ctime.tv_nsec = VNOVAL; 518 vap->va_birthtime.tv_sec = VNOVAL; 519 vap->va_birthtime.tv_nsec = VNOVAL; 520 vap->va_flags = VNOVAL; 521 vap->va_gen = VNOVAL; 522 vap->va_vaflags = 0; 523 } 524 525 /* 526 * This routine is called when we have too many vnodes. It attempts 527 * to free <count> vnodes and will potentially free vnodes that still 528 * have VM backing store (VM backing store is typically the cause 529 * of a vnode blowout so we want to do this). Therefore, this operation 530 * is not considered cheap. 531 * 532 * A number of conditions may prevent a vnode from being reclaimed. 533 * the buffer cache may have references on the vnode, a directory 534 * vnode may still have references due to the namei cache representing 535 * underlying files, or the vnode may be in active use. It is not 536 * desireable to reuse such vnodes. These conditions may cause the 537 * number of vnodes to reach some minimum value regardless of what 538 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 539 */ 540 static int 541 vlrureclaim(struct mount *mp) 542 { 543 struct thread *td; 544 struct vnode *vp; 545 int done; 546 int trigger; 547 int usevnodes; 548 int count; 549 550 /* 551 * Calculate the trigger point, don't allow user 552 * screwups to blow us up. This prevents us from 553 * recycling vnodes with lots of resident pages. We 554 * aren't trying to free memory, we are trying to 555 * free vnodes. 556 */ 557 usevnodes = desiredvnodes; 558 if (usevnodes <= 0) 559 usevnodes = 1; 560 trigger = cnt.v_page_count * 2 / usevnodes; 561 done = 0; 562 td = curthread; 563 vn_start_write(NULL, &mp, V_WAIT); 564 MNT_ILOCK(mp); 565 count = mp->mnt_nvnodelistsize / 10 + 1; 566 while (count != 0) { 567 vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 568 while (vp != NULL && vp->v_type == VMARKER) 569 vp = TAILQ_NEXT(vp, v_nmntvnodes); 570 if (vp == NULL) 571 break; 572 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 573 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 574 --count; 575 if (!VI_TRYLOCK(vp)) 576 goto next_iter; 577 /* 578 * If it's been deconstructed already, it's still 579 * referenced, or it exceeds the trigger, skip it. 580 */ 581 if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) || 582 (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL && 583 vp->v_object->resident_page_count > trigger)) { 584 VI_UNLOCK(vp); 585 goto next_iter; 586 } 587 MNT_IUNLOCK(mp); 588 vholdl(vp); 589 if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT, td)) { 590 vdrop(vp); 591 goto next_iter_mntunlocked; 592 } 593 VI_LOCK(vp); 594 /* 595 * v_usecount may have been bumped after VOP_LOCK() dropped 596 * the vnode interlock and before it was locked again. 597 * 598 * It is not necessary to recheck VI_DOOMED because it can 599 * only be set by another thread that holds both the vnode 600 * lock and vnode interlock. If another thread has the 601 * vnode lock before we get to VOP_LOCK() and obtains the 602 * vnode interlock after VOP_LOCK() drops the vnode 603 * interlock, the other thread will be unable to drop the 604 * vnode lock before our VOP_LOCK() call fails. 605 */ 606 if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) || 607 (vp->v_object != NULL && 608 vp->v_object->resident_page_count > trigger)) { 609 VOP_UNLOCK(vp, LK_INTERLOCK, td); 610 goto next_iter_mntunlocked; 611 } 612 KASSERT((vp->v_iflag & VI_DOOMED) == 0, 613 ("VI_DOOMED unexpectedly detected in vlrureclaim()")); 614 vgonel(vp); 615 VOP_UNLOCK(vp, 0, td); 616 vdropl(vp); 617 done++; 618 next_iter_mntunlocked: 619 if ((count % 256) != 0) 620 goto relock_mnt; 621 goto yield; 622 next_iter: 623 if ((count % 256) != 0) 624 continue; 625 MNT_IUNLOCK(mp); 626 yield: 627 uio_yield(); 628 relock_mnt: 629 MNT_ILOCK(mp); 630 } 631 MNT_IUNLOCK(mp); 632 vn_finished_write(mp); 633 return done; 634 } 635 636 /* 637 * Attempt to keep the free list at wantfreevnodes length. 638 */ 639 static void 640 vnlru_free(int count) 641 { 642 struct vnode *vp; 643 int vfslocked; 644 645 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 646 for (; count > 0; count--) { 647 vp = TAILQ_FIRST(&vnode_free_list); 648 /* 649 * The list can be modified while the free_list_mtx 650 * has been dropped and vp could be NULL here. 651 */ 652 if (!vp) 653 break; 654 VNASSERT(vp->v_op != NULL, vp, 655 ("vnlru_free: vnode already reclaimed.")); 656 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 657 /* 658 * Don't recycle if we can't get the interlock. 659 */ 660 if (!VI_TRYLOCK(vp)) { 661 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 662 continue; 663 } 664 VNASSERT(VCANRECYCLE(vp), vp, 665 ("vp inconsistent on freelist")); 666 freevnodes--; 667 vp->v_iflag &= ~VI_FREE; 668 vholdl(vp); 669 mtx_unlock(&vnode_free_list_mtx); 670 VI_UNLOCK(vp); 671 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 672 vtryrecycle(vp); 673 VFS_UNLOCK_GIANT(vfslocked); 674 /* 675 * If the recycled succeeded this vdrop will actually free 676 * the vnode. If not it will simply place it back on 677 * the free list. 678 */ 679 vdrop(vp); 680 mtx_lock(&vnode_free_list_mtx); 681 } 682 } 683 /* 684 * Attempt to recycle vnodes in a context that is always safe to block. 685 * Calling vlrurecycle() from the bowels of filesystem code has some 686 * interesting deadlock problems. 687 */ 688 static struct proc *vnlruproc; 689 static int vnlruproc_sig; 690 691 static void 692 vnlru_proc(void) 693 { 694 struct mount *mp, *nmp; 695 int done; 696 struct proc *p = vnlruproc; 697 struct thread *td = FIRST_THREAD_IN_PROC(p); 698 699 mtx_lock(&Giant); 700 701 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 702 SHUTDOWN_PRI_FIRST); 703 704 for (;;) { 705 kthread_suspend_check(p); 706 mtx_lock(&vnode_free_list_mtx); 707 if (freevnodes > wantfreevnodes) 708 vnlru_free(freevnodes - wantfreevnodes); 709 if (numvnodes <= desiredvnodes * 9 / 10) { 710 vnlruproc_sig = 0; 711 wakeup(&vnlruproc_sig); 712 msleep(vnlruproc, &vnode_free_list_mtx, 713 PVFS|PDROP, "vlruwt", hz); 714 continue; 715 } 716 mtx_unlock(&vnode_free_list_mtx); 717 done = 0; 718 mtx_lock(&mountlist_mtx); 719 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 720 int vfsunlocked; 721 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 722 nmp = TAILQ_NEXT(mp, mnt_list); 723 continue; 724 } 725 if (!VFS_NEEDSGIANT(mp)) { 726 mtx_unlock(&Giant); 727 vfsunlocked = 1; 728 } else 729 vfsunlocked = 0; 730 done += vlrureclaim(mp); 731 if (vfsunlocked) 732 mtx_lock(&Giant); 733 mtx_lock(&mountlist_mtx); 734 nmp = TAILQ_NEXT(mp, mnt_list); 735 vfs_unbusy(mp, td); 736 } 737 mtx_unlock(&mountlist_mtx); 738 if (done == 0) { 739 #if 0 740 /* These messages are temporary debugging aids */ 741 if (vnlru_nowhere < 5) 742 printf("vnlru process getting nowhere..\n"); 743 else if (vnlru_nowhere == 5) 744 printf("vnlru process messages stopped.\n"); 745 #endif 746 vnlru_nowhere++; 747 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 748 } else 749 uio_yield(); 750 } 751 } 752 753 static struct kproc_desc vnlru_kp = { 754 "vnlru", 755 vnlru_proc, 756 &vnlruproc 757 }; 758 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 759 760 /* 761 * Routines having to do with the management of the vnode table. 762 */ 763 764 static void 765 vdestroy(struct vnode *vp) 766 { 767 struct bufobj *bo; 768 769 CTR1(KTR_VFS, "vdestroy vp %p", vp); 770 mtx_lock(&vnode_free_list_mtx); 771 numvnodes--; 772 mtx_unlock(&vnode_free_list_mtx); 773 bo = &vp->v_bufobj; 774 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 775 ("cleaned vnode still on the free list.")); 776 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 777 VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count")); 778 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 779 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 780 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 781 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 782 VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL")); 783 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 784 VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL")); 785 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); 786 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); 787 VI_UNLOCK(vp); 788 #ifdef MAC 789 mac_destroy_vnode(vp); 790 #endif 791 if (vp->v_pollinfo != NULL) { 792 knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note); 793 mtx_destroy(&vp->v_pollinfo->vpi_lock); 794 uma_zfree(vnodepoll_zone, vp->v_pollinfo); 795 } 796 #ifdef INVARIANTS 797 /* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */ 798 vp->v_op = NULL; 799 #endif 800 lockdestroy(vp->v_vnlock); 801 mtx_destroy(&vp->v_interlock); 802 uma_zfree(vnode_zone, vp); 803 } 804 805 /* 806 * Try to recycle a freed vnode. We abort if anyone picks up a reference 807 * before we actually vgone(). This function must be called with the vnode 808 * held to prevent the vnode from being returned to the free list midway 809 * through vgone(). 810 */ 811 static int 812 vtryrecycle(struct vnode *vp) 813 { 814 struct thread *td = curthread; 815 struct mount *vnmp; 816 817 CTR1(KTR_VFS, "vtryrecycle: trying vp %p", vp); 818 VNASSERT(vp->v_holdcnt, vp, 819 ("vtryrecycle: Recycling vp %p without a reference.", vp)); 820 /* 821 * This vnode may found and locked via some other list, if so we 822 * can't recycle it yet. 823 */ 824 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 825 return (EWOULDBLOCK); 826 /* 827 * Don't recycle if its filesystem is being suspended. 828 */ 829 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 830 VOP_UNLOCK(vp, 0, td); 831 return (EBUSY); 832 } 833 /* 834 * If we got this far, we need to acquire the interlock and see if 835 * anyone picked up this vnode from another list. If not, we will 836 * mark it with DOOMED via vgonel() so that anyone who does find it 837 * will skip over it. 838 */ 839 VI_LOCK(vp); 840 if (vp->v_usecount) { 841 VOP_UNLOCK(vp, LK_INTERLOCK, td); 842 vn_finished_write(vnmp); 843 return (EBUSY); 844 } 845 if ((vp->v_iflag & VI_DOOMED) == 0) 846 vgonel(vp); 847 VOP_UNLOCK(vp, LK_INTERLOCK, td); 848 vn_finished_write(vnmp); 849 CTR1(KTR_VFS, "vtryrecycle: recycled vp %p", vp); 850 return (0); 851 } 852 853 /* 854 * Return the next vnode from the free list. 855 */ 856 int 857 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, 858 struct vnode **vpp) 859 { 860 struct vnode *vp = NULL; 861 struct bufobj *bo; 862 863 mtx_lock(&vnode_free_list_mtx); 864 /* 865 * Lend our context to reclaim vnodes if they've exceeded the max. 866 */ 867 if (freevnodes > wantfreevnodes) 868 vnlru_free(1); 869 /* 870 * Wait for available vnodes. 871 */ 872 if (numvnodes > desiredvnodes) { 873 if (vnlruproc_sig == 0) { 874 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 875 wakeup(vnlruproc); 876 } 877 msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS, 878 "vlruwk", hz); 879 #if 0 /* XXX Not all VFS_VGET/ffs_vget callers check returns. */ 880 if (numvnodes > desiredvnodes) { 881 mtx_unlock(&vnode_free_list_mtx); 882 return (ENFILE); 883 } 884 #endif 885 } 886 numvnodes++; 887 mtx_unlock(&vnode_free_list_mtx); 888 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); 889 /* 890 * Setup locks. 891 */ 892 vp->v_vnlock = &vp->v_lock; 893 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 894 /* 895 * By default, don't allow shared locks unless filesystems 896 * opt-in. 897 */ 898 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE); 899 /* 900 * Initialize bufobj. 901 */ 902 bo = &vp->v_bufobj; 903 bo->__bo_vnode = vp; 904 bo->bo_mtx = &vp->v_interlock; 905 bo->bo_ops = &buf_ops_bio; 906 bo->bo_private = vp; 907 TAILQ_INIT(&bo->bo_clean.bv_hd); 908 TAILQ_INIT(&bo->bo_dirty.bv_hd); 909 /* 910 * Initialize namecache. 911 */ 912 LIST_INIT(&vp->v_cache_src); 913 TAILQ_INIT(&vp->v_cache_dst); 914 /* 915 * Finalize various vnode identity bits. 916 */ 917 vp->v_type = VNON; 918 vp->v_tag = tag; 919 vp->v_op = vops; 920 v_incr_usecount(vp); 921 vp->v_data = 0; 922 #ifdef MAC 923 mac_init_vnode(vp); 924 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 925 mac_associate_vnode_singlelabel(mp, vp); 926 else if (mp == NULL) 927 printf("NULL mp in getnewvnode()\n"); 928 #endif 929 if (mp != NULL) { 930 insmntque(vp, mp); 931 bo->bo_bsize = mp->mnt_stat.f_iosize; 932 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) 933 vp->v_vflag |= VV_NOKNOTE; 934 } 935 936 CTR2(KTR_VFS, "getnewvnode: mp %p vp %p", mp, vp); 937 *vpp = vp; 938 return (0); 939 } 940 941 /* 942 * Delete from old mount point vnode list, if on one. 943 */ 944 static void 945 delmntque(struct vnode *vp) 946 { 947 struct mount *mp; 948 949 mp = vp->v_mount; 950 if (mp == NULL) 951 return; 952 MNT_ILOCK(mp); 953 vp->v_mount = NULL; 954 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 955 ("bad mount point vnode list size")); 956 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 957 mp->mnt_nvnodelistsize--; 958 MNT_REL(mp); 959 MNT_IUNLOCK(mp); 960 } 961 962 /* 963 * Insert into list of vnodes for the new mount point, if available. 964 */ 965 static void 966 insmntque(struct vnode *vp, struct mount *mp) 967 { 968 969 vp->v_mount = mp; 970 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 971 MNT_ILOCK(mp); 972 MNT_REF(mp); 973 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 974 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 975 ("neg mount point vnode list size")); 976 mp->mnt_nvnodelistsize++; 977 MNT_IUNLOCK(mp); 978 } 979 980 /* 981 * Flush out and invalidate all buffers associated with a bufobj 982 * Called with the underlying object locked. 983 */ 984 int 985 bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag, 986 int slptimeo) 987 { 988 int error; 989 990 BO_LOCK(bo); 991 if (flags & V_SAVE) { 992 error = bufobj_wwait(bo, slpflag, slptimeo); 993 if (error) { 994 BO_UNLOCK(bo); 995 return (error); 996 } 997 if (bo->bo_dirty.bv_cnt > 0) { 998 BO_UNLOCK(bo); 999 if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0) 1000 return (error); 1001 /* 1002 * XXX We could save a lock/unlock if this was only 1003 * enabled under INVARIANTS 1004 */ 1005 BO_LOCK(bo); 1006 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 1007 panic("vinvalbuf: dirty bufs"); 1008 } 1009 } 1010 /* 1011 * If you alter this loop please notice that interlock is dropped and 1012 * reacquired in flushbuflist. Special care is needed to ensure that 1013 * no race conditions occur from this. 1014 */ 1015 do { 1016 error = flushbuflist(&bo->bo_clean, 1017 flags, bo, slpflag, slptimeo); 1018 if (error == 0) 1019 error = flushbuflist(&bo->bo_dirty, 1020 flags, bo, slpflag, slptimeo); 1021 if (error != 0 && error != EAGAIN) { 1022 BO_UNLOCK(bo); 1023 return (error); 1024 } 1025 } while (error != 0); 1026 1027 /* 1028 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1029 * have write I/O in-progress but if there is a VM object then the 1030 * VM object can also have read-I/O in-progress. 1031 */ 1032 do { 1033 bufobj_wwait(bo, 0, 0); 1034 BO_UNLOCK(bo); 1035 if (bo->bo_object != NULL) { 1036 VM_OBJECT_LOCK(bo->bo_object); 1037 vm_object_pip_wait(bo->bo_object, "bovlbx"); 1038 VM_OBJECT_UNLOCK(bo->bo_object); 1039 } 1040 BO_LOCK(bo); 1041 } while (bo->bo_numoutput > 0); 1042 BO_UNLOCK(bo); 1043 1044 /* 1045 * Destroy the copy in the VM cache, too. 1046 */ 1047 if (bo->bo_object != NULL) { 1048 VM_OBJECT_LOCK(bo->bo_object); 1049 vm_object_page_remove(bo->bo_object, 0, 0, 1050 (flags & V_SAVE) ? TRUE : FALSE); 1051 VM_OBJECT_UNLOCK(bo->bo_object); 1052 } 1053 1054 #ifdef INVARIANTS 1055 BO_LOCK(bo); 1056 if ((flags & (V_ALT | V_NORMAL)) == 0 && 1057 (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0)) 1058 panic("vinvalbuf: flush failed"); 1059 BO_UNLOCK(bo); 1060 #endif 1061 return (0); 1062 } 1063 1064 /* 1065 * Flush out and invalidate all buffers associated with a vnode. 1066 * Called with the underlying object locked. 1067 */ 1068 int 1069 vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag, 1070 int slptimeo) 1071 { 1072 1073 CTR2(KTR_VFS, "vinvalbuf vp %p flags %d", vp, flags); 1074 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1075 return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo)); 1076 } 1077 1078 /* 1079 * Flush out buffers on the specified list. 1080 * 1081 */ 1082 static int 1083 flushbuflist( struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 1084 int slptimeo) 1085 { 1086 struct buf *bp, *nbp; 1087 int retval, error; 1088 daddr_t lblkno; 1089 b_xflags_t xflags; 1090 1091 ASSERT_BO_LOCKED(bo); 1092 1093 retval = 0; 1094 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 1095 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1096 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1097 continue; 1098 } 1099 lblkno = 0; 1100 xflags = 0; 1101 if (nbp != NULL) { 1102 lblkno = nbp->b_lblkno; 1103 xflags = nbp->b_xflags & 1104 (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN); 1105 } 1106 retval = EAGAIN; 1107 error = BUF_TIMELOCK(bp, 1108 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo), 1109 "flushbuf", slpflag, slptimeo); 1110 if (error) { 1111 BO_LOCK(bo); 1112 return (error != ENOLCK ? error : EAGAIN); 1113 } 1114 KASSERT(bp->b_bufobj == bo, 1115 ("bp %p wrong b_bufobj %p should be %p", 1116 bp, bp->b_bufobj, bo)); 1117 if (bp->b_bufobj != bo) { /* XXX: necessary ? */ 1118 BUF_UNLOCK(bp); 1119 BO_LOCK(bo); 1120 return (EAGAIN); 1121 } 1122 /* 1123 * XXX Since there are no node locks for NFS, I 1124 * believe there is a slight chance that a delayed 1125 * write will occur while sleeping just above, so 1126 * check for it. 1127 */ 1128 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1129 (flags & V_SAVE)) { 1130 bremfree(bp); 1131 bp->b_flags |= B_ASYNC; 1132 bwrite(bp); 1133 BO_LOCK(bo); 1134 return (EAGAIN); /* XXX: why not loop ? */ 1135 } 1136 bremfree(bp); 1137 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 1138 bp->b_flags &= ~B_ASYNC; 1139 brelse(bp); 1140 BO_LOCK(bo); 1141 if (nbp != NULL && 1142 (nbp->b_bufobj != bo || 1143 nbp->b_lblkno != lblkno || 1144 (nbp->b_xflags & 1145 (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN)) != xflags)) 1146 break; /* nbp invalid */ 1147 } 1148 return (retval); 1149 } 1150 1151 /* 1152 * Truncate a file's buffer and pages to a specified length. This 1153 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1154 * sync activity. 1155 */ 1156 int 1157 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, 1158 off_t length, int blksize) 1159 { 1160 struct buf *bp, *nbp; 1161 int anyfreed; 1162 int trunclbn; 1163 struct bufobj *bo; 1164 1165 CTR2(KTR_VFS, "vtruncbuf vp %p length %jd", vp, length); 1166 /* 1167 * Round up to the *next* lbn. 1168 */ 1169 trunclbn = (length + blksize - 1) / blksize; 1170 1171 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1172 restart: 1173 VI_LOCK(vp); 1174 bo = &vp->v_bufobj; 1175 anyfreed = 1; 1176 for (;anyfreed;) { 1177 anyfreed = 0; 1178 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 1179 if (bp->b_lblkno < trunclbn) 1180 continue; 1181 if (BUF_LOCK(bp, 1182 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1183 VI_MTX(vp)) == ENOLCK) 1184 goto restart; 1185 1186 bremfree(bp); 1187 bp->b_flags |= (B_INVAL | B_RELBUF); 1188 bp->b_flags &= ~B_ASYNC; 1189 brelse(bp); 1190 anyfreed = 1; 1191 1192 if (nbp != NULL && 1193 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1194 (nbp->b_vp != vp) || 1195 (nbp->b_flags & B_DELWRI))) { 1196 goto restart; 1197 } 1198 VI_LOCK(vp); 1199 } 1200 1201 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1202 if (bp->b_lblkno < trunclbn) 1203 continue; 1204 if (BUF_LOCK(bp, 1205 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1206 VI_MTX(vp)) == ENOLCK) 1207 goto restart; 1208 bremfree(bp); 1209 bp->b_flags |= (B_INVAL | B_RELBUF); 1210 bp->b_flags &= ~B_ASYNC; 1211 brelse(bp); 1212 anyfreed = 1; 1213 if (nbp != NULL && 1214 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1215 (nbp->b_vp != vp) || 1216 (nbp->b_flags & B_DELWRI) == 0)) { 1217 goto restart; 1218 } 1219 VI_LOCK(vp); 1220 } 1221 } 1222 1223 if (length > 0) { 1224 restartsync: 1225 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1226 if (bp->b_lblkno > 0) 1227 continue; 1228 /* 1229 * Since we hold the vnode lock this should only 1230 * fail if we're racing with the buf daemon. 1231 */ 1232 if (BUF_LOCK(bp, 1233 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1234 VI_MTX(vp)) == ENOLCK) { 1235 goto restart; 1236 } 1237 VNASSERT((bp->b_flags & B_DELWRI), vp, 1238 ("buf(%p) on dirty queue without DELWRI", bp)); 1239 1240 bremfree(bp); 1241 bawrite(bp); 1242 VI_LOCK(vp); 1243 goto restartsync; 1244 } 1245 } 1246 1247 bufobj_wwait(bo, 0, 0); 1248 VI_UNLOCK(vp); 1249 vnode_pager_setsize(vp, length); 1250 1251 return (0); 1252 } 1253 1254 /* 1255 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1256 * a vnode. 1257 * 1258 * NOTE: We have to deal with the special case of a background bitmap 1259 * buffer, a situation where two buffers will have the same logical 1260 * block offset. We want (1) only the foreground buffer to be accessed 1261 * in a lookup and (2) must differentiate between the foreground and 1262 * background buffer in the splay tree algorithm because the splay 1263 * tree cannot normally handle multiple entities with the same 'index'. 1264 * We accomplish this by adding differentiating flags to the splay tree's 1265 * numerical domain. 1266 */ 1267 static 1268 struct buf * 1269 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1270 { 1271 struct buf dummy; 1272 struct buf *lefttreemax, *righttreemin, *y; 1273 1274 if (root == NULL) 1275 return (NULL); 1276 lefttreemax = righttreemin = &dummy; 1277 for (;;) { 1278 if (lblkno < root->b_lblkno || 1279 (lblkno == root->b_lblkno && 1280 (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1281 if ((y = root->b_left) == NULL) 1282 break; 1283 if (lblkno < y->b_lblkno) { 1284 /* Rotate right. */ 1285 root->b_left = y->b_right; 1286 y->b_right = root; 1287 root = y; 1288 if ((y = root->b_left) == NULL) 1289 break; 1290 } 1291 /* Link into the new root's right tree. */ 1292 righttreemin->b_left = root; 1293 righttreemin = root; 1294 } else if (lblkno > root->b_lblkno || 1295 (lblkno == root->b_lblkno && 1296 (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { 1297 if ((y = root->b_right) == NULL) 1298 break; 1299 if (lblkno > y->b_lblkno) { 1300 /* Rotate left. */ 1301 root->b_right = y->b_left; 1302 y->b_left = root; 1303 root = y; 1304 if ((y = root->b_right) == NULL) 1305 break; 1306 } 1307 /* Link into the new root's left tree. */ 1308 lefttreemax->b_right = root; 1309 lefttreemax = root; 1310 } else { 1311 break; 1312 } 1313 root = y; 1314 } 1315 /* Assemble the new root. */ 1316 lefttreemax->b_right = root->b_left; 1317 righttreemin->b_left = root->b_right; 1318 root->b_left = dummy.b_right; 1319 root->b_right = dummy.b_left; 1320 return (root); 1321 } 1322 1323 static void 1324 buf_vlist_remove(struct buf *bp) 1325 { 1326 struct buf *root; 1327 struct bufv *bv; 1328 1329 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1330 ASSERT_BO_LOCKED(bp->b_bufobj); 1331 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) != 1332 (BX_VNDIRTY|BX_VNCLEAN), 1333 ("buf_vlist_remove: Buf %p is on two lists", bp)); 1334 if (bp->b_xflags & BX_VNDIRTY) 1335 bv = &bp->b_bufobj->bo_dirty; 1336 else 1337 bv = &bp->b_bufobj->bo_clean; 1338 if (bp != bv->bv_root) { 1339 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1340 KASSERT(root == bp, ("splay lookup failed in remove")); 1341 } 1342 if (bp->b_left == NULL) { 1343 root = bp->b_right; 1344 } else { 1345 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1346 root->b_right = bp->b_right; 1347 } 1348 bv->bv_root = root; 1349 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 1350 bv->bv_cnt--; 1351 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1352 } 1353 1354 /* 1355 * Add the buffer to the sorted clean or dirty block list using a 1356 * splay tree algorithm. 1357 * 1358 * NOTE: xflags is passed as a constant, optimizing this inline function! 1359 */ 1360 static void 1361 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 1362 { 1363 struct buf *root; 1364 struct bufv *bv; 1365 1366 ASSERT_BO_LOCKED(bo); 1367 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1368 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 1369 bp->b_xflags |= xflags; 1370 if (xflags & BX_VNDIRTY) 1371 bv = &bo->bo_dirty; 1372 else 1373 bv = &bo->bo_clean; 1374 1375 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1376 if (root == NULL) { 1377 bp->b_left = NULL; 1378 bp->b_right = NULL; 1379 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 1380 } else if (bp->b_lblkno < root->b_lblkno || 1381 (bp->b_lblkno == root->b_lblkno && 1382 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1383 bp->b_left = root->b_left; 1384 bp->b_right = root; 1385 root->b_left = NULL; 1386 TAILQ_INSERT_BEFORE(root, bp, b_bobufs); 1387 } else { 1388 bp->b_right = root->b_right; 1389 bp->b_left = root; 1390 root->b_right = NULL; 1391 TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs); 1392 } 1393 bv->bv_cnt++; 1394 bv->bv_root = bp; 1395 } 1396 1397 /* 1398 * Lookup a buffer using the splay tree. Note that we specifically avoid 1399 * shadow buffers used in background bitmap writes. 1400 * 1401 * This code isn't quite efficient as it could be because we are maintaining 1402 * two sorted lists and do not know which list the block resides in. 1403 * 1404 * During a "make buildworld" the desired buffer is found at one of 1405 * the roots more than 60% of the time. Thus, checking both roots 1406 * before performing either splay eliminates unnecessary splays on the 1407 * first tree splayed. 1408 */ 1409 struct buf * 1410 gbincore(struct bufobj *bo, daddr_t lblkno) 1411 { 1412 struct buf *bp; 1413 1414 ASSERT_BO_LOCKED(bo); 1415 if ((bp = bo->bo_clean.bv_root) != NULL && 1416 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1417 return (bp); 1418 if ((bp = bo->bo_dirty.bv_root) != NULL && 1419 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1420 return (bp); 1421 if ((bp = bo->bo_clean.bv_root) != NULL) { 1422 bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp); 1423 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1424 return (bp); 1425 } 1426 if ((bp = bo->bo_dirty.bv_root) != NULL) { 1427 bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp); 1428 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1429 return (bp); 1430 } 1431 return (NULL); 1432 } 1433 1434 /* 1435 * Associate a buffer with a vnode. 1436 */ 1437 void 1438 bgetvp(struct vnode *vp, struct buf *bp) 1439 { 1440 1441 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 1442 1443 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 1444 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 1445 ("bgetvp: bp already attached! %p", bp)); 1446 1447 ASSERT_VI_LOCKED(vp, "bgetvp"); 1448 vholdl(vp); 1449 if (VFS_NEEDSGIANT(vp->v_mount) || 1450 vp->v_bufobj.bo_flag & BO_NEEDSGIANT) 1451 bp->b_flags |= B_NEEDSGIANT; 1452 bp->b_vp = vp; 1453 bp->b_bufobj = &vp->v_bufobj; 1454 /* 1455 * Insert onto list for new vnode. 1456 */ 1457 buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN); 1458 } 1459 1460 /* 1461 * Disassociate a buffer from a vnode. 1462 */ 1463 void 1464 brelvp(struct buf *bp) 1465 { 1466 struct bufobj *bo; 1467 struct vnode *vp; 1468 1469 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1470 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1471 1472 /* 1473 * Delete from old vnode list, if on one. 1474 */ 1475 vp = bp->b_vp; /* XXX */ 1476 bo = bp->b_bufobj; 1477 BO_LOCK(bo); 1478 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1479 buf_vlist_remove(bp); 1480 else 1481 panic("brelvp: Buffer %p not on queue.", bp); 1482 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1483 bo->bo_flag &= ~BO_ONWORKLST; 1484 mtx_lock(&sync_mtx); 1485 LIST_REMOVE(bo, bo_synclist); 1486 syncer_worklist_len--; 1487 mtx_unlock(&sync_mtx); 1488 } 1489 bp->b_flags &= ~B_NEEDSGIANT; 1490 bp->b_vp = NULL; 1491 bp->b_bufobj = NULL; 1492 vdropl(vp); 1493 } 1494 1495 /* 1496 * Add an item to the syncer work queue. 1497 */ 1498 static void 1499 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 1500 { 1501 int slot; 1502 1503 ASSERT_BO_LOCKED(bo); 1504 1505 mtx_lock(&sync_mtx); 1506 if (bo->bo_flag & BO_ONWORKLST) 1507 LIST_REMOVE(bo, bo_synclist); 1508 else { 1509 bo->bo_flag |= BO_ONWORKLST; 1510 syncer_worklist_len++; 1511 } 1512 1513 if (delay > syncer_maxdelay - 2) 1514 delay = syncer_maxdelay - 2; 1515 slot = (syncer_delayno + delay) & syncer_mask; 1516 1517 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 1518 mtx_unlock(&sync_mtx); 1519 } 1520 1521 static int 1522 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 1523 { 1524 int error, len; 1525 1526 mtx_lock(&sync_mtx); 1527 len = syncer_worklist_len - sync_vnode_count; 1528 mtx_unlock(&sync_mtx); 1529 error = SYSCTL_OUT(req, &len, sizeof(len)); 1530 return (error); 1531 } 1532 1533 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0, 1534 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 1535 1536 static struct proc *updateproc; 1537 static void sched_sync(void); 1538 static struct kproc_desc up_kp = { 1539 "syncer", 1540 sched_sync, 1541 &updateproc 1542 }; 1543 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 1544 1545 static int 1546 sync_vnode(struct bufobj *bo, struct thread *td) 1547 { 1548 struct vnode *vp; 1549 struct mount *mp; 1550 1551 vp = bo->__bo_vnode; /* XXX */ 1552 if (VOP_ISLOCKED(vp, NULL) != 0) 1553 return (1); 1554 if (VI_TRYLOCK(vp) == 0) 1555 return (1); 1556 /* 1557 * We use vhold in case the vnode does not 1558 * successfully sync. vhold prevents the vnode from 1559 * going away when we unlock the sync_mtx so that 1560 * we can acquire the vnode interlock. 1561 */ 1562 vholdl(vp); 1563 mtx_unlock(&sync_mtx); 1564 VI_UNLOCK(vp); 1565 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1566 vdrop(vp); 1567 mtx_lock(&sync_mtx); 1568 return (1); 1569 } 1570 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 1571 (void) VOP_FSYNC(vp, MNT_LAZY, td); 1572 VOP_UNLOCK(vp, 0, td); 1573 vn_finished_write(mp); 1574 VI_LOCK(vp); 1575 if ((bo->bo_flag & BO_ONWORKLST) != 0) { 1576 /* 1577 * Put us back on the worklist. The worklist 1578 * routine will remove us from our current 1579 * position and then add us back in at a later 1580 * position. 1581 */ 1582 vn_syncer_add_to_worklist(bo, syncdelay); 1583 } 1584 vdropl(vp); 1585 mtx_lock(&sync_mtx); 1586 return (0); 1587 } 1588 1589 /* 1590 * System filesystem synchronizer daemon. 1591 */ 1592 static void 1593 sched_sync(void) 1594 { 1595 struct synclist *next; 1596 struct synclist *slp; 1597 struct bufobj *bo; 1598 long starttime; 1599 struct thread *td = FIRST_THREAD_IN_PROC(updateproc); 1600 static int dummychan; 1601 int last_work_seen; 1602 int net_worklist_len; 1603 int syncer_final_iter; 1604 int first_printf; 1605 int error; 1606 1607 mtx_lock(&Giant); 1608 last_work_seen = 0; 1609 syncer_final_iter = 0; 1610 first_printf = 1; 1611 syncer_state = SYNCER_RUNNING; 1612 starttime = time_uptime; 1613 td->td_pflags |= TDP_NORUNNINGBUF; 1614 1615 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 1616 SHUTDOWN_PRI_LAST); 1617 1618 for (;;) { 1619 mtx_lock(&sync_mtx); 1620 if (syncer_state == SYNCER_FINAL_DELAY && 1621 syncer_final_iter == 0) { 1622 mtx_unlock(&sync_mtx); 1623 kthread_suspend_check(td->td_proc); 1624 mtx_lock(&sync_mtx); 1625 } 1626 net_worklist_len = syncer_worklist_len - sync_vnode_count; 1627 if (syncer_state != SYNCER_RUNNING && 1628 starttime != time_uptime) { 1629 if (first_printf) { 1630 printf("\nSyncing disks, vnodes remaining..."); 1631 first_printf = 0; 1632 } 1633 printf("%d ", net_worklist_len); 1634 } 1635 starttime = time_uptime; 1636 1637 /* 1638 * Push files whose dirty time has expired. Be careful 1639 * of interrupt race on slp queue. 1640 * 1641 * Skip over empty worklist slots when shutting down. 1642 */ 1643 do { 1644 slp = &syncer_workitem_pending[syncer_delayno]; 1645 syncer_delayno += 1; 1646 if (syncer_delayno == syncer_maxdelay) 1647 syncer_delayno = 0; 1648 next = &syncer_workitem_pending[syncer_delayno]; 1649 /* 1650 * If the worklist has wrapped since the 1651 * it was emptied of all but syncer vnodes, 1652 * switch to the FINAL_DELAY state and run 1653 * for one more second. 1654 */ 1655 if (syncer_state == SYNCER_SHUTTING_DOWN && 1656 net_worklist_len == 0 && 1657 last_work_seen == syncer_delayno) { 1658 syncer_state = SYNCER_FINAL_DELAY; 1659 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 1660 } 1661 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 1662 syncer_worklist_len > 0); 1663 1664 /* 1665 * Keep track of the last time there was anything 1666 * on the worklist other than syncer vnodes. 1667 * Return to the SHUTTING_DOWN state if any 1668 * new work appears. 1669 */ 1670 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 1671 last_work_seen = syncer_delayno; 1672 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 1673 syncer_state = SYNCER_SHUTTING_DOWN; 1674 while ((bo = LIST_FIRST(slp)) != NULL) { 1675 error = sync_vnode(bo, td); 1676 if (error == 1) { 1677 LIST_REMOVE(bo, bo_synclist); 1678 LIST_INSERT_HEAD(next, bo, bo_synclist); 1679 continue; 1680 } 1681 } 1682 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 1683 syncer_final_iter--; 1684 mtx_unlock(&sync_mtx); 1685 /* 1686 * The variable rushjob allows the kernel to speed up the 1687 * processing of the filesystem syncer process. A rushjob 1688 * value of N tells the filesystem syncer to process the next 1689 * N seconds worth of work on its queue ASAP. Currently rushjob 1690 * is used by the soft update code to speed up the filesystem 1691 * syncer process when the incore state is getting so far 1692 * ahead of the disk that the kernel memory pool is being 1693 * threatened with exhaustion. 1694 */ 1695 mtx_lock(&sync_mtx); 1696 if (rushjob > 0) { 1697 rushjob -= 1; 1698 mtx_unlock(&sync_mtx); 1699 continue; 1700 } 1701 mtx_unlock(&sync_mtx); 1702 /* 1703 * Just sleep for a short period if time between 1704 * iterations when shutting down to allow some I/O 1705 * to happen. 1706 * 1707 * If it has taken us less than a second to process the 1708 * current work, then wait. Otherwise start right over 1709 * again. We can still lose time if any single round 1710 * takes more than two seconds, but it does not really 1711 * matter as we are just trying to generally pace the 1712 * filesystem activity. 1713 */ 1714 if (syncer_state != SYNCER_RUNNING) 1715 tsleep(&dummychan, PPAUSE, "syncfnl", 1716 hz / SYNCER_SHUTDOWN_SPEEDUP); 1717 else if (time_uptime == starttime) 1718 tsleep(&lbolt, PPAUSE, "syncer", 0); 1719 } 1720 } 1721 1722 /* 1723 * Request the syncer daemon to speed up its work. 1724 * We never push it to speed up more than half of its 1725 * normal turn time, otherwise it could take over the cpu. 1726 */ 1727 int 1728 speedup_syncer() 1729 { 1730 struct thread *td; 1731 int ret = 0; 1732 1733 td = FIRST_THREAD_IN_PROC(updateproc); 1734 sleepq_remove(td, &lbolt); 1735 mtx_lock(&sync_mtx); 1736 if (rushjob < syncdelay / 2) { 1737 rushjob += 1; 1738 stat_rush_requests += 1; 1739 ret = 1; 1740 } 1741 mtx_unlock(&sync_mtx); 1742 return (ret); 1743 } 1744 1745 /* 1746 * Tell the syncer to speed up its work and run though its work 1747 * list several times, then tell it to shut down. 1748 */ 1749 static void 1750 syncer_shutdown(void *arg, int howto) 1751 { 1752 struct thread *td; 1753 1754 if (howto & RB_NOSYNC) 1755 return; 1756 td = FIRST_THREAD_IN_PROC(updateproc); 1757 sleepq_remove(td, &lbolt); 1758 mtx_lock(&sync_mtx); 1759 syncer_state = SYNCER_SHUTTING_DOWN; 1760 rushjob = 0; 1761 mtx_unlock(&sync_mtx); 1762 kproc_shutdown(arg, howto); 1763 } 1764 1765 /* 1766 * Reassign a buffer from one vnode to another. 1767 * Used to assign file specific control information 1768 * (indirect blocks) to the vnode to which they belong. 1769 */ 1770 void 1771 reassignbuf(struct buf *bp) 1772 { 1773 struct vnode *vp; 1774 struct bufobj *bo; 1775 int delay; 1776 #ifdef INVARIANTS 1777 struct bufv *bv; 1778 #endif 1779 1780 vp = bp->b_vp; 1781 bo = bp->b_bufobj; 1782 ++reassignbufcalls; 1783 1784 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 1785 bp, bp->b_vp, bp->b_flags); 1786 /* 1787 * B_PAGING flagged buffers cannot be reassigned because their vp 1788 * is not fully linked in. 1789 */ 1790 if (bp->b_flags & B_PAGING) 1791 panic("cannot reassign paging buffer"); 1792 1793 /* 1794 * Delete from old vnode list, if on one. 1795 */ 1796 VI_LOCK(vp); 1797 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1798 buf_vlist_remove(bp); 1799 else 1800 panic("reassignbuf: Buffer %p not on queue.", bp); 1801 /* 1802 * If dirty, put on list of dirty buffers; otherwise insert onto list 1803 * of clean buffers. 1804 */ 1805 if (bp->b_flags & B_DELWRI) { 1806 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 1807 switch (vp->v_type) { 1808 case VDIR: 1809 delay = dirdelay; 1810 break; 1811 case VCHR: 1812 delay = metadelay; 1813 break; 1814 default: 1815 delay = filedelay; 1816 } 1817 vn_syncer_add_to_worklist(bo, delay); 1818 } 1819 buf_vlist_add(bp, bo, BX_VNDIRTY); 1820 } else { 1821 buf_vlist_add(bp, bo, BX_VNCLEAN); 1822 1823 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1824 mtx_lock(&sync_mtx); 1825 LIST_REMOVE(bo, bo_synclist); 1826 syncer_worklist_len--; 1827 mtx_unlock(&sync_mtx); 1828 bo->bo_flag &= ~BO_ONWORKLST; 1829 } 1830 } 1831 #ifdef INVARIANTS 1832 bv = &bo->bo_clean; 1833 bp = TAILQ_FIRST(&bv->bv_hd); 1834 KASSERT(bp == NULL || bp->b_bufobj == bo, 1835 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 1836 bp = TAILQ_LAST(&bv->bv_hd, buflists); 1837 KASSERT(bp == NULL || bp->b_bufobj == bo, 1838 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 1839 bv = &bo->bo_dirty; 1840 bp = TAILQ_FIRST(&bv->bv_hd); 1841 KASSERT(bp == NULL || bp->b_bufobj == bo, 1842 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 1843 bp = TAILQ_LAST(&bv->bv_hd, buflists); 1844 KASSERT(bp == NULL || bp->b_bufobj == bo, 1845 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 1846 #endif 1847 VI_UNLOCK(vp); 1848 } 1849 1850 /* 1851 * Increment the use and hold counts on the vnode, taking care to reference 1852 * the driver's usecount if this is a chardev. The vholdl() will remove 1853 * the vnode from the free list if it is presently free. Requires the 1854 * vnode interlock and returns with it held. 1855 */ 1856 static void 1857 v_incr_usecount(struct vnode *vp) 1858 { 1859 1860 CTR3(KTR_VFS, "v_incr_usecount: vp %p holdcnt %d usecount %d\n", 1861 vp, vp->v_holdcnt, vp->v_usecount); 1862 vp->v_usecount++; 1863 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1864 dev_lock(); 1865 vp->v_rdev->si_usecount++; 1866 dev_unlock(); 1867 } 1868 vholdl(vp); 1869 } 1870 1871 /* 1872 * Turn a holdcnt into a use+holdcnt such that only one call to 1873 * v_decr_usecount is needed. 1874 */ 1875 static void 1876 v_upgrade_usecount(struct vnode *vp) 1877 { 1878 1879 CTR3(KTR_VFS, "v_upgrade_usecount: vp %p holdcnt %d usecount %d\n", 1880 vp, vp->v_holdcnt, vp->v_usecount); 1881 vp->v_usecount++; 1882 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1883 dev_lock(); 1884 vp->v_rdev->si_usecount++; 1885 dev_unlock(); 1886 } 1887 } 1888 1889 /* 1890 * Decrement the vnode use and hold count along with the driver's usecount 1891 * if this is a chardev. The vdropl() below releases the vnode interlock 1892 * as it may free the vnode. 1893 */ 1894 static void 1895 v_decr_usecount(struct vnode *vp) 1896 { 1897 1898 CTR3(KTR_VFS, "v_decr_usecount: vp %p holdcnt %d usecount %d\n", 1899 vp, vp->v_holdcnt, vp->v_usecount); 1900 ASSERT_VI_LOCKED(vp, __FUNCTION__); 1901 VNASSERT(vp->v_usecount > 0, vp, 1902 ("v_decr_usecount: negative usecount")); 1903 vp->v_usecount--; 1904 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1905 dev_lock(); 1906 vp->v_rdev->si_usecount--; 1907 dev_unlock(); 1908 } 1909 vdropl(vp); 1910 } 1911 1912 /* 1913 * Decrement only the use count and driver use count. This is intended to 1914 * be paired with a follow on vdropl() to release the remaining hold count. 1915 * In this way we may vgone() a vnode with a 0 usecount without risk of 1916 * having it end up on a free list because the hold count is kept above 0. 1917 */ 1918 static void 1919 v_decr_useonly(struct vnode *vp) 1920 { 1921 1922 CTR3(KTR_VFS, "v_decr_useonly: vp %p holdcnt %d usecount %d\n", 1923 vp, vp->v_holdcnt, vp->v_usecount); 1924 ASSERT_VI_LOCKED(vp, __FUNCTION__); 1925 VNASSERT(vp->v_usecount > 0, vp, 1926 ("v_decr_useonly: negative usecount")); 1927 vp->v_usecount--; 1928 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1929 dev_lock(); 1930 vp->v_rdev->si_usecount--; 1931 dev_unlock(); 1932 } 1933 } 1934 1935 /* 1936 * Grab a particular vnode from the free list, increment its 1937 * reference count and lock it. The vnode lock bit is set if the 1938 * vnode is being eliminated in vgone. The process is awakened 1939 * when the transition is completed, and an error returned to 1940 * indicate that the vnode is no longer usable (possibly having 1941 * been changed to a new filesystem type). 1942 */ 1943 int 1944 vget(struct vnode *vp, int flags, struct thread *td) 1945 { 1946 int oweinact; 1947 int oldflags; 1948 int error; 1949 1950 error = 0; 1951 oldflags = flags; 1952 oweinact = 0; 1953 VFS_ASSERT_GIANT(vp->v_mount); 1954 if ((flags & LK_INTERLOCK) == 0) 1955 VI_LOCK(vp); 1956 /* 1957 * If the inactive call was deferred because vput() was called 1958 * with a shared lock, we have to do it here before another thread 1959 * gets a reference to data that should be dead. 1960 */ 1961 if (vp->v_iflag & VI_OWEINACT) { 1962 if (flags & LK_NOWAIT) { 1963 VI_UNLOCK(vp); 1964 return (EBUSY); 1965 } 1966 flags &= ~LK_TYPE_MASK; 1967 flags |= LK_EXCLUSIVE; 1968 oweinact = 1; 1969 } 1970 vholdl(vp); 1971 if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { 1972 vdrop(vp); 1973 return (error); 1974 } 1975 VI_LOCK(vp); 1976 /* Upgrade our holdcnt to a usecount. */ 1977 v_upgrade_usecount(vp); 1978 if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0) 1979 panic("vget: vn_lock failed to return ENOENT\n"); 1980 if (oweinact) { 1981 if (vp->v_iflag & VI_OWEINACT) 1982 vinactive(vp, td); 1983 VI_UNLOCK(vp); 1984 if ((oldflags & LK_TYPE_MASK) == 0) 1985 VOP_UNLOCK(vp, 0, td); 1986 } else 1987 VI_UNLOCK(vp); 1988 return (0); 1989 } 1990 1991 /* 1992 * Increase the reference count of a vnode. 1993 */ 1994 void 1995 vref(struct vnode *vp) 1996 { 1997 1998 VI_LOCK(vp); 1999 v_incr_usecount(vp); 2000 VI_UNLOCK(vp); 2001 } 2002 2003 /* 2004 * Return reference count of a vnode. 2005 * 2006 * The results of this call are only guaranteed when some mechanism other 2007 * than the VI lock is used to stop other processes from gaining references 2008 * to the vnode. This may be the case if the caller holds the only reference. 2009 * This is also useful when stale data is acceptable as race conditions may 2010 * be accounted for by some other means. 2011 */ 2012 int 2013 vrefcnt(struct vnode *vp) 2014 { 2015 int usecnt; 2016 2017 VI_LOCK(vp); 2018 usecnt = vp->v_usecount; 2019 VI_UNLOCK(vp); 2020 2021 return (usecnt); 2022 } 2023 2024 2025 /* 2026 * Vnode put/release. 2027 * If count drops to zero, call inactive routine and return to freelist. 2028 */ 2029 void 2030 vrele(struct vnode *vp) 2031 { 2032 struct thread *td = curthread; /* XXX */ 2033 2034 KASSERT(vp != NULL, ("vrele: null vp")); 2035 VFS_ASSERT_GIANT(vp->v_mount); 2036 2037 VI_LOCK(vp); 2038 2039 /* Skip this v_writecount check if we're going to panic below. */ 2040 VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp, 2041 ("vrele: missed vn_close")); 2042 2043 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2044 vp->v_usecount == 1)) { 2045 v_decr_usecount(vp); 2046 return; 2047 } 2048 if (vp->v_usecount != 1) { 2049 #ifdef DIAGNOSTIC 2050 vprint("vrele: negative ref count", vp); 2051 #endif 2052 VI_UNLOCK(vp); 2053 panic("vrele: negative ref cnt"); 2054 } 2055 /* 2056 * We want to hold the vnode until the inactive finishes to 2057 * prevent vgone() races. We drop the use count here and the 2058 * hold count below when we're done. 2059 */ 2060 v_decr_useonly(vp); 2061 /* 2062 * We must call VOP_INACTIVE with the node locked. Mark 2063 * as VI_DOINGINACT to avoid recursion. 2064 */ 2065 vp->v_iflag |= VI_OWEINACT; 2066 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) { 2067 VI_LOCK(vp); 2068 if (vp->v_usecount > 0) 2069 vp->v_iflag &= ~VI_OWEINACT; 2070 if (vp->v_iflag & VI_OWEINACT) 2071 vinactive(vp, td); 2072 VOP_UNLOCK(vp, 0, td); 2073 } else { 2074 VI_LOCK(vp); 2075 if (vp->v_usecount > 0) 2076 vp->v_iflag &= ~VI_OWEINACT; 2077 } 2078 vdropl(vp); 2079 } 2080 2081 /* 2082 * Release an already locked vnode. This give the same effects as 2083 * unlock+vrele(), but takes less time and avoids releasing and 2084 * re-aquiring the lock (as vrele() aquires the lock internally.) 2085 */ 2086 void 2087 vput(struct vnode *vp) 2088 { 2089 struct thread *td = curthread; /* XXX */ 2090 int error; 2091 2092 KASSERT(vp != NULL, ("vput: null vp")); 2093 ASSERT_VOP_LOCKED(vp, "vput"); 2094 VFS_ASSERT_GIANT(vp->v_mount); 2095 VI_LOCK(vp); 2096 /* Skip this v_writecount check if we're going to panic below. */ 2097 VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp, 2098 ("vput: missed vn_close")); 2099 error = 0; 2100 2101 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2102 vp->v_usecount == 1)) { 2103 VOP_UNLOCK(vp, 0, td); 2104 v_decr_usecount(vp); 2105 return; 2106 } 2107 2108 if (vp->v_usecount != 1) { 2109 #ifdef DIAGNOSTIC 2110 vprint("vput: negative ref count", vp); 2111 #endif 2112 panic("vput: negative ref cnt"); 2113 } 2114 /* 2115 * We want to hold the vnode until the inactive finishes to 2116 * prevent vgone() races. We drop the use count here and the 2117 * hold count below when we're done. 2118 */ 2119 v_decr_useonly(vp); 2120 vp->v_iflag |= VI_OWEINACT; 2121 if (VOP_ISLOCKED(vp, NULL) != LK_EXCLUSIVE) { 2122 error = VOP_LOCK(vp, LK_EXCLUPGRADE|LK_INTERLOCK|LK_NOWAIT, td); 2123 VI_LOCK(vp); 2124 if (error) { 2125 if (vp->v_usecount > 0) 2126 vp->v_iflag &= ~VI_OWEINACT; 2127 goto done; 2128 } 2129 } 2130 if (vp->v_usecount > 0) 2131 vp->v_iflag &= ~VI_OWEINACT; 2132 if (vp->v_iflag & VI_OWEINACT) 2133 vinactive(vp, td); 2134 VOP_UNLOCK(vp, 0, td); 2135 done: 2136 vdropl(vp); 2137 } 2138 2139 /* 2140 * Somebody doesn't want the vnode recycled. 2141 */ 2142 void 2143 vhold(struct vnode *vp) 2144 { 2145 2146 VI_LOCK(vp); 2147 vholdl(vp); 2148 VI_UNLOCK(vp); 2149 } 2150 2151 void 2152 vholdl(struct vnode *vp) 2153 { 2154 2155 vp->v_holdcnt++; 2156 if (VSHOULDBUSY(vp)) 2157 vbusy(vp); 2158 } 2159 2160 /* 2161 * Note that there is one less who cares about this vnode. vdrop() is the 2162 * opposite of vhold(). 2163 */ 2164 void 2165 vdrop(struct vnode *vp) 2166 { 2167 2168 VI_LOCK(vp); 2169 vdropl(vp); 2170 } 2171 2172 /* 2173 * Drop the hold count of the vnode. If this is the last reference to 2174 * the vnode we will free it if it has been vgone'd otherwise it is 2175 * placed on the free list. 2176 */ 2177 static void 2178 vdropl(struct vnode *vp) 2179 { 2180 2181 if (vp->v_holdcnt <= 0) 2182 panic("vdrop: holdcnt %d", vp->v_holdcnt); 2183 vp->v_holdcnt--; 2184 if (vp->v_holdcnt == 0) { 2185 if (vp->v_iflag & VI_DOOMED) { 2186 vdestroy(vp); 2187 return; 2188 } else 2189 vfree(vp); 2190 } 2191 VI_UNLOCK(vp); 2192 } 2193 2194 /* 2195 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 2196 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 2197 * OWEINACT tracks whether a vnode missed a call to inactive due to a 2198 * failed lock upgrade. 2199 */ 2200 static void 2201 vinactive(struct vnode *vp, struct thread *td) 2202 { 2203 2204 ASSERT_VOP_LOCKED(vp, "vinactive"); 2205 ASSERT_VI_LOCKED(vp, "vinactive"); 2206 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, 2207 ("vinactive: recursed on VI_DOINGINACT")); 2208 vp->v_iflag |= VI_DOINGINACT; 2209 vp->v_iflag &= ~VI_OWEINACT; 2210 VI_UNLOCK(vp); 2211 VOP_INACTIVE(vp, td); 2212 VI_LOCK(vp); 2213 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, 2214 ("vinactive: lost VI_DOINGINACT")); 2215 vp->v_iflag &= ~VI_DOINGINACT; 2216 } 2217 2218 /* 2219 * Remove any vnodes in the vnode table belonging to mount point mp. 2220 * 2221 * If FORCECLOSE is not specified, there should not be any active ones, 2222 * return error if any are found (nb: this is a user error, not a 2223 * system error). If FORCECLOSE is specified, detach any active vnodes 2224 * that are found. 2225 * 2226 * If WRITECLOSE is set, only flush out regular file vnodes open for 2227 * writing. 2228 * 2229 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2230 * 2231 * `rootrefs' specifies the base reference count for the root vnode 2232 * of this filesystem. The root vnode is considered busy if its 2233 * v_usecount exceeds this value. On a successful return, vflush(, td) 2234 * will call vrele() on the root vnode exactly rootrefs times. 2235 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2236 * be zero. 2237 */ 2238 #ifdef DIAGNOSTIC 2239 static int busyprt = 0; /* print out busy vnodes */ 2240 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 2241 #endif 2242 2243 int 2244 vflush( struct mount *mp, int rootrefs, int flags, struct thread *td) 2245 { 2246 struct vnode *vp, *mvp, *rootvp = NULL; 2247 struct vattr vattr; 2248 int busy = 0, error; 2249 2250 CTR1(KTR_VFS, "vflush: mp %p", mp); 2251 if (rootrefs > 0) { 2252 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2253 ("vflush: bad args")); 2254 /* 2255 * Get the filesystem root vnode. We can vput() it 2256 * immediately, since with rootrefs > 0, it won't go away. 2257 */ 2258 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0) 2259 return (error); 2260 vput(rootvp); 2261 2262 } 2263 MNT_ILOCK(mp); 2264 loop: 2265 MNT_VNODE_FOREACH(vp, mp, mvp) { 2266 2267 VI_LOCK(vp); 2268 vholdl(vp); 2269 MNT_IUNLOCK(mp); 2270 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td); 2271 if (error) { 2272 vdrop(vp); 2273 MNT_ILOCK(mp); 2274 MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp); 2275 goto loop; 2276 } 2277 /* 2278 * Skip over a vnodes marked VV_SYSTEM. 2279 */ 2280 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2281 VOP_UNLOCK(vp, 0, td); 2282 vdrop(vp); 2283 MNT_ILOCK(mp); 2284 continue; 2285 } 2286 /* 2287 * If WRITECLOSE is set, flush out unlinked but still open 2288 * files (even if open only for reading) and regular file 2289 * vnodes open for writing. 2290 */ 2291 if (flags & WRITECLOSE) { 2292 error = VOP_GETATTR(vp, &vattr, td->td_ucred, td); 2293 VI_LOCK(vp); 2294 2295 if ((vp->v_type == VNON || 2296 (error == 0 && vattr.va_nlink > 0)) && 2297 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2298 VOP_UNLOCK(vp, 0, td); 2299 vdropl(vp); 2300 MNT_ILOCK(mp); 2301 continue; 2302 } 2303 } else 2304 VI_LOCK(vp); 2305 /* 2306 * With v_usecount == 0, all we need to do is clear out the 2307 * vnode data structures and we are done. 2308 * 2309 * If FORCECLOSE is set, forcibly close the vnode. 2310 */ 2311 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 2312 VNASSERT(vp->v_usecount == 0 || 2313 (vp->v_type != VCHR && vp->v_type != VBLK), vp, 2314 ("device VNODE %p is FORCECLOSED", vp)); 2315 vgonel(vp); 2316 } else { 2317 busy++; 2318 #ifdef DIAGNOSTIC 2319 if (busyprt) 2320 vprint("vflush: busy vnode", vp); 2321 #endif 2322 } 2323 VOP_UNLOCK(vp, 0, td); 2324 vdropl(vp); 2325 MNT_ILOCK(mp); 2326 } 2327 MNT_IUNLOCK(mp); 2328 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2329 /* 2330 * If just the root vnode is busy, and if its refcount 2331 * is equal to `rootrefs', then go ahead and kill it. 2332 */ 2333 VI_LOCK(rootvp); 2334 KASSERT(busy > 0, ("vflush: not busy")); 2335 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 2336 ("vflush: usecount %d < rootrefs %d", 2337 rootvp->v_usecount, rootrefs)); 2338 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2339 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK, td); 2340 vgone(rootvp); 2341 VOP_UNLOCK(rootvp, 0, td); 2342 busy = 0; 2343 } else 2344 VI_UNLOCK(rootvp); 2345 } 2346 if (busy) 2347 return (EBUSY); 2348 for (; rootrefs > 0; rootrefs--) 2349 vrele(rootvp); 2350 return (0); 2351 } 2352 2353 /* 2354 * Recycle an unused vnode to the front of the free list. 2355 */ 2356 int 2357 vrecycle(struct vnode *vp, struct thread *td) 2358 { 2359 int recycled; 2360 2361 ASSERT_VOP_LOCKED(vp, "vrecycle"); 2362 recycled = 0; 2363 VI_LOCK(vp); 2364 if (vp->v_usecount == 0) { 2365 recycled = 1; 2366 vgonel(vp); 2367 } 2368 VI_UNLOCK(vp); 2369 return (recycled); 2370 } 2371 2372 /* 2373 * Eliminate all activity associated with a vnode 2374 * in preparation for reuse. 2375 */ 2376 void 2377 vgone(struct vnode *vp) 2378 { 2379 VI_LOCK(vp); 2380 vgonel(vp); 2381 VI_UNLOCK(vp); 2382 } 2383 2384 /* 2385 * vgone, with the vp interlock held. 2386 */ 2387 void 2388 vgonel(struct vnode *vp) 2389 { 2390 struct thread *td; 2391 int oweinact; 2392 int active; 2393 struct mount *mp; 2394 2395 CTR1(KTR_VFS, "vgonel: vp %p", vp); 2396 ASSERT_VOP_LOCKED(vp, "vgonel"); 2397 ASSERT_VI_LOCKED(vp, "vgonel"); 2398 VNASSERT(vp->v_holdcnt, vp, 2399 ("vgonel: vp %p has no reference.", vp)); 2400 td = curthread; 2401 2402 /* 2403 * Don't vgonel if we're already doomed. 2404 */ 2405 if (vp->v_iflag & VI_DOOMED) 2406 return; 2407 vp->v_iflag |= VI_DOOMED; 2408 /* 2409 * Check to see if the vnode is in use. If so, we have to call 2410 * VOP_CLOSE() and VOP_INACTIVE(). 2411 */ 2412 active = vp->v_usecount; 2413 oweinact = (vp->v_iflag & VI_OWEINACT); 2414 VI_UNLOCK(vp); 2415 /* 2416 * Clean out any buffers associated with the vnode. 2417 * If the flush fails, just toss the buffers. 2418 */ 2419 mp = NULL; 2420 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 2421 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 2422 if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0) 2423 vinvalbuf(vp, 0, td, 0, 0); 2424 2425 /* 2426 * If purging an active vnode, it must be closed and 2427 * deactivated before being reclaimed. 2428 */ 2429 if (active) 2430 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2431 if (oweinact || active) { 2432 VI_LOCK(vp); 2433 if ((vp->v_iflag & VI_DOINGINACT) == 0) 2434 vinactive(vp, td); 2435 VI_UNLOCK(vp); 2436 } 2437 /* 2438 * Reclaim the vnode. 2439 */ 2440 if (VOP_RECLAIM(vp, td)) 2441 panic("vgone: cannot reclaim"); 2442 if (mp != NULL) 2443 vn_finished_secondary_write(mp); 2444 VNASSERT(vp->v_object == NULL, vp, 2445 ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag)); 2446 /* 2447 * Delete from old mount point vnode list. 2448 */ 2449 delmntque(vp); 2450 cache_purge(vp); 2451 /* 2452 * Done with purge, reset to the standard lock and invalidate 2453 * the vnode. 2454 */ 2455 VI_LOCK(vp); 2456 vp->v_vnlock = &vp->v_lock; 2457 vp->v_op = &dead_vnodeops; 2458 vp->v_tag = "none"; 2459 vp->v_type = VBAD; 2460 } 2461 2462 /* 2463 * Calculate the total number of references to a special device. 2464 */ 2465 int 2466 vcount(struct vnode *vp) 2467 { 2468 int count; 2469 2470 dev_lock(); 2471 count = vp->v_rdev->si_usecount; 2472 dev_unlock(); 2473 return (count); 2474 } 2475 2476 /* 2477 * Same as above, but using the struct cdev *as argument 2478 */ 2479 int 2480 count_dev(struct cdev *dev) 2481 { 2482 int count; 2483 2484 dev_lock(); 2485 count = dev->si_usecount; 2486 dev_unlock(); 2487 return(count); 2488 } 2489 2490 /* 2491 * Print out a description of a vnode. 2492 */ 2493 static char *typename[] = 2494 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", 2495 "VMARKER"}; 2496 2497 void 2498 vn_printf(struct vnode *vp, const char *fmt, ...) 2499 { 2500 va_list ap; 2501 char buf[96]; 2502 2503 va_start(ap, fmt); 2504 vprintf(fmt, ap); 2505 va_end(ap); 2506 printf("%p: ", (void *)vp); 2507 printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]); 2508 printf(" usecount %d, writecount %d, refcount %d mountedhere %p\n", 2509 vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere); 2510 buf[0] = '\0'; 2511 buf[1] = '\0'; 2512 if (vp->v_vflag & VV_ROOT) 2513 strcat(buf, "|VV_ROOT"); 2514 if (vp->v_vflag & VV_TEXT) 2515 strcat(buf, "|VV_TEXT"); 2516 if (vp->v_vflag & VV_SYSTEM) 2517 strcat(buf, "|VV_SYSTEM"); 2518 if (vp->v_iflag & VI_DOOMED) 2519 strcat(buf, "|VI_DOOMED"); 2520 if (vp->v_iflag & VI_FREE) 2521 strcat(buf, "|VI_FREE"); 2522 printf(" flags (%s)\n", buf + 1); 2523 if (mtx_owned(VI_MTX(vp))) 2524 printf(" VI_LOCKed"); 2525 if (vp->v_object != NULL) 2526 printf(" v_object %p ref %d pages %d\n", 2527 vp->v_object, vp->v_object->ref_count, 2528 vp->v_object->resident_page_count); 2529 printf(" "); 2530 lockmgr_printinfo(vp->v_vnlock); 2531 printf("\n"); 2532 if (vp->v_data != NULL) 2533 VOP_PRINT(vp); 2534 } 2535 2536 #ifdef DDB 2537 #include <ddb/ddb.h> 2538 /* 2539 * List all of the locked vnodes in the system. 2540 * Called when debugging the kernel. 2541 */ 2542 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 2543 { 2544 struct mount *mp, *nmp; 2545 struct vnode *vp; 2546 2547 /* 2548 * Note: because this is DDB, we can't obey the locking semantics 2549 * for these structures, which means we could catch an inconsistent 2550 * state and dereference a nasty pointer. Not much to be done 2551 * about that. 2552 */ 2553 printf("Locked vnodes\n"); 2554 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2555 nmp = TAILQ_NEXT(mp, mnt_list); 2556 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2557 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp, NULL)) 2558 vprint("", vp); 2559 } 2560 nmp = TAILQ_NEXT(mp, mnt_list); 2561 } 2562 } 2563 #endif 2564 2565 /* 2566 * Fill in a struct xvfsconf based on a struct vfsconf. 2567 */ 2568 static void 2569 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) 2570 { 2571 2572 strcpy(xvfsp->vfc_name, vfsp->vfc_name); 2573 xvfsp->vfc_typenum = vfsp->vfc_typenum; 2574 xvfsp->vfc_refcount = vfsp->vfc_refcount; 2575 xvfsp->vfc_flags = vfsp->vfc_flags; 2576 /* 2577 * These are unused in userland, we keep them 2578 * to not break binary compatibility. 2579 */ 2580 xvfsp->vfc_vfsops = NULL; 2581 xvfsp->vfc_next = NULL; 2582 } 2583 2584 /* 2585 * Top level filesystem related information gathering. 2586 */ 2587 static int 2588 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 2589 { 2590 struct vfsconf *vfsp; 2591 struct xvfsconf xvfsp; 2592 int error; 2593 2594 error = 0; 2595 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2596 bzero(&xvfsp, sizeof(xvfsp)); 2597 vfsconf2x(vfsp, &xvfsp); 2598 error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp); 2599 if (error) 2600 break; 2601 } 2602 return (error); 2603 } 2604 2605 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, 2606 "S,xvfsconf", "List of all configured filesystems"); 2607 2608 #ifndef BURN_BRIDGES 2609 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2610 2611 static int 2612 vfs_sysctl(SYSCTL_HANDLER_ARGS) 2613 { 2614 int *name = (int *)arg1 - 1; /* XXX */ 2615 u_int namelen = arg2 + 1; /* XXX */ 2616 struct vfsconf *vfsp; 2617 struct xvfsconf xvfsp; 2618 2619 printf("WARNING: userland calling deprecated sysctl, " 2620 "please rebuild world\n"); 2621 2622 #if 1 || defined(COMPAT_PRELITE2) 2623 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2624 if (namelen == 1) 2625 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2626 #endif 2627 2628 switch (name[1]) { 2629 case VFS_MAXTYPENUM: 2630 if (namelen != 2) 2631 return (ENOTDIR); 2632 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2633 case VFS_CONF: 2634 if (namelen != 3) 2635 return (ENOTDIR); /* overloaded */ 2636 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) 2637 if (vfsp->vfc_typenum == name[2]) 2638 break; 2639 if (vfsp == NULL) 2640 return (EOPNOTSUPP); 2641 bzero(&xvfsp, sizeof(xvfsp)); 2642 vfsconf2x(vfsp, &xvfsp); 2643 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 2644 } 2645 return (EOPNOTSUPP); 2646 } 2647 2648 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, 2649 vfs_sysctl, "Generic filesystem"); 2650 2651 #if 1 || defined(COMPAT_PRELITE2) 2652 2653 static int 2654 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 2655 { 2656 int error; 2657 struct vfsconf *vfsp; 2658 struct ovfsconf ovfs; 2659 2660 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2661 bzero(&ovfs, sizeof(ovfs)); 2662 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 2663 strcpy(ovfs.vfc_name, vfsp->vfc_name); 2664 ovfs.vfc_index = vfsp->vfc_typenum; 2665 ovfs.vfc_refcount = vfsp->vfc_refcount; 2666 ovfs.vfc_flags = vfsp->vfc_flags; 2667 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 2668 if (error) 2669 return error; 2670 } 2671 return 0; 2672 } 2673 2674 #endif /* 1 || COMPAT_PRELITE2 */ 2675 #endif /* !BURN_BRIDGES */ 2676 2677 #define KINFO_VNODESLOP 10 2678 #ifdef notyet 2679 /* 2680 * Dump vnode list (via sysctl). 2681 */ 2682 /* ARGSUSED */ 2683 static int 2684 sysctl_vnode(SYSCTL_HANDLER_ARGS) 2685 { 2686 struct xvnode *xvn; 2687 struct thread *td = req->td; 2688 struct mount *mp; 2689 struct vnode *vp; 2690 int error, len, n; 2691 2692 /* 2693 * Stale numvnodes access is not fatal here. 2694 */ 2695 req->lock = 0; 2696 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 2697 if (!req->oldptr) 2698 /* Make an estimate */ 2699 return (SYSCTL_OUT(req, 0, len)); 2700 2701 error = sysctl_wire_old_buffer(req, 0); 2702 if (error != 0) 2703 return (error); 2704 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 2705 n = 0; 2706 mtx_lock(&mountlist_mtx); 2707 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 2708 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 2709 continue; 2710 MNT_ILOCK(mp); 2711 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2712 if (n == len) 2713 break; 2714 vref(vp); 2715 xvn[n].xv_size = sizeof *xvn; 2716 xvn[n].xv_vnode = vp; 2717 xvn[n].xv_id = 0; /* XXX compat */ 2718 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 2719 XV_COPY(usecount); 2720 XV_COPY(writecount); 2721 XV_COPY(holdcnt); 2722 XV_COPY(mount); 2723 XV_COPY(numoutput); 2724 XV_COPY(type); 2725 #undef XV_COPY 2726 xvn[n].xv_flag = vp->v_vflag; 2727 2728 switch (vp->v_type) { 2729 case VREG: 2730 case VDIR: 2731 case VLNK: 2732 break; 2733 case VBLK: 2734 case VCHR: 2735 if (vp->v_rdev == NULL) { 2736 vrele(vp); 2737 continue; 2738 } 2739 xvn[n].xv_dev = dev2udev(vp->v_rdev); 2740 break; 2741 case VSOCK: 2742 xvn[n].xv_socket = vp->v_socket; 2743 break; 2744 case VFIFO: 2745 xvn[n].xv_fifo = vp->v_fifoinfo; 2746 break; 2747 case VNON: 2748 case VBAD: 2749 default: 2750 /* shouldn't happen? */ 2751 vrele(vp); 2752 continue; 2753 } 2754 vrele(vp); 2755 ++n; 2756 } 2757 MNT_IUNLOCK(mp); 2758 mtx_lock(&mountlist_mtx); 2759 vfs_unbusy(mp, td); 2760 if (n == len) 2761 break; 2762 } 2763 mtx_unlock(&mountlist_mtx); 2764 2765 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 2766 free(xvn, M_TEMP); 2767 return (error); 2768 } 2769 2770 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 2771 0, 0, sysctl_vnode, "S,xvnode", ""); 2772 #endif 2773 2774 /* 2775 * Unmount all filesystems. The list is traversed in reverse order 2776 * of mounting to avoid dependencies. 2777 */ 2778 void 2779 vfs_unmountall(void) 2780 { 2781 struct mount *mp; 2782 struct thread *td; 2783 int error; 2784 2785 KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread")); 2786 td = curthread; 2787 /* 2788 * Since this only runs when rebooting, it is not interlocked. 2789 */ 2790 while(!TAILQ_EMPTY(&mountlist)) { 2791 mp = TAILQ_LAST(&mountlist, mntlist); 2792 error = dounmount(mp, MNT_FORCE, td); 2793 if (error) { 2794 TAILQ_REMOVE(&mountlist, mp, mnt_list); 2795 /* 2796 * XXX: Due to the way in which we mount the root 2797 * file system off of devfs, devfs will generate a 2798 * "busy" warning when we try to unmount it before 2799 * the root. Don't print a warning as a result in 2800 * order to avoid false positive errors that may 2801 * cause needless upset. 2802 */ 2803 if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) { 2804 printf("unmount of %s failed (", 2805 mp->mnt_stat.f_mntonname); 2806 if (error == EBUSY) 2807 printf("BUSY)\n"); 2808 else 2809 printf("%d)\n", error); 2810 } 2811 } else { 2812 /* The unmount has removed mp from the mountlist */ 2813 } 2814 } 2815 } 2816 2817 /* 2818 * perform msync on all vnodes under a mount point 2819 * the mount point must be locked. 2820 */ 2821 void 2822 vfs_msync(struct mount *mp, int flags) 2823 { 2824 struct vnode *vp, *mvp; 2825 struct vm_object *obj; 2826 2827 MNT_ILOCK(mp); 2828 MNT_VNODE_FOREACH(vp, mp, mvp) { 2829 VI_LOCK(vp); 2830 if ((vp->v_iflag & VI_OBJDIRTY) && 2831 (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { 2832 MNT_IUNLOCK(mp); 2833 if (!vget(vp, 2834 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 2835 curthread)) { 2836 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 2837 vput(vp); 2838 MNT_ILOCK(mp); 2839 continue; 2840 } 2841 2842 obj = vp->v_object; 2843 if (obj != NULL) { 2844 VM_OBJECT_LOCK(obj); 2845 vm_object_page_clean(obj, 0, 0, 2846 flags == MNT_WAIT ? 2847 OBJPC_SYNC : OBJPC_NOSYNC); 2848 VM_OBJECT_UNLOCK(obj); 2849 } 2850 vput(vp); 2851 } 2852 MNT_ILOCK(mp); 2853 } else 2854 VI_UNLOCK(vp); 2855 } 2856 MNT_IUNLOCK(mp); 2857 } 2858 2859 /* 2860 * Mark a vnode as free, putting it up for recycling. 2861 */ 2862 static void 2863 vfree(struct vnode *vp) 2864 { 2865 2866 CTR1(KTR_VFS, "vfree vp %p", vp); 2867 ASSERT_VI_LOCKED(vp, "vfree"); 2868 mtx_lock(&vnode_free_list_mtx); 2869 VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed.")); 2870 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free")); 2871 VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't")); 2872 VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp, 2873 ("vfree: Freeing doomed vnode")); 2874 if (vp->v_iflag & VI_AGE) { 2875 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2876 } else { 2877 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 2878 } 2879 freevnodes++; 2880 vp->v_iflag &= ~VI_AGE; 2881 vp->v_iflag |= VI_FREE; 2882 mtx_unlock(&vnode_free_list_mtx); 2883 } 2884 2885 /* 2886 * Opposite of vfree() - mark a vnode as in use. 2887 */ 2888 static void 2889 vbusy(struct vnode *vp) 2890 { 2891 CTR1(KTR_VFS, "vbusy vp %p", vp); 2892 ASSERT_VI_LOCKED(vp, "vbusy"); 2893 VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free")); 2894 VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed.")); 2895 2896 mtx_lock(&vnode_free_list_mtx); 2897 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2898 freevnodes--; 2899 vp->v_iflag &= ~(VI_FREE|VI_AGE); 2900 mtx_unlock(&vnode_free_list_mtx); 2901 } 2902 2903 /* 2904 * Initalize per-vnode helper structure to hold poll-related state. 2905 */ 2906 void 2907 v_addpollinfo(struct vnode *vp) 2908 { 2909 struct vpollinfo *vi; 2910 2911 vi = uma_zalloc(vnodepoll_zone, M_WAITOK); 2912 if (vp->v_pollinfo != NULL) { 2913 uma_zfree(vnodepoll_zone, vi); 2914 return; 2915 } 2916 vp->v_pollinfo = vi; 2917 mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 2918 knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, vp, vfs_knllock, 2919 vfs_knlunlock, vfs_knllocked); 2920 } 2921 2922 /* 2923 * Record a process's interest in events which might happen to 2924 * a vnode. Because poll uses the historic select-style interface 2925 * internally, this routine serves as both the ``check for any 2926 * pending events'' and the ``record my interest in future events'' 2927 * functions. (These are done together, while the lock is held, 2928 * to avoid race conditions.) 2929 */ 2930 int 2931 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 2932 { 2933 2934 if (vp->v_pollinfo == NULL) 2935 v_addpollinfo(vp); 2936 mtx_lock(&vp->v_pollinfo->vpi_lock); 2937 if (vp->v_pollinfo->vpi_revents & events) { 2938 /* 2939 * This leaves events we are not interested 2940 * in available for the other process which 2941 * which presumably had requested them 2942 * (otherwise they would never have been 2943 * recorded). 2944 */ 2945 events &= vp->v_pollinfo->vpi_revents; 2946 vp->v_pollinfo->vpi_revents &= ~events; 2947 2948 mtx_unlock(&vp->v_pollinfo->vpi_lock); 2949 return events; 2950 } 2951 vp->v_pollinfo->vpi_events |= events; 2952 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 2953 mtx_unlock(&vp->v_pollinfo->vpi_lock); 2954 return 0; 2955 } 2956 2957 /* 2958 * Routine to create and manage a filesystem syncer vnode. 2959 */ 2960 #define sync_close ((int (*)(struct vop_close_args *))nullop) 2961 static int sync_fsync(struct vop_fsync_args *); 2962 static int sync_inactive(struct vop_inactive_args *); 2963 static int sync_reclaim(struct vop_reclaim_args *); 2964 2965 static struct vop_vector sync_vnodeops = { 2966 .vop_bypass = VOP_EOPNOTSUPP, 2967 .vop_close = sync_close, /* close */ 2968 .vop_fsync = sync_fsync, /* fsync */ 2969 .vop_inactive = sync_inactive, /* inactive */ 2970 .vop_reclaim = sync_reclaim, /* reclaim */ 2971 .vop_lock = vop_stdlock, /* lock */ 2972 .vop_unlock = vop_stdunlock, /* unlock */ 2973 .vop_islocked = vop_stdislocked, /* islocked */ 2974 }; 2975 2976 /* 2977 * Create a new filesystem syncer vnode for the specified mount point. 2978 */ 2979 int 2980 vfs_allocate_syncvnode(struct mount *mp) 2981 { 2982 struct vnode *vp; 2983 static long start, incr, next; 2984 int error; 2985 2986 /* Allocate a new vnode */ 2987 if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) { 2988 mp->mnt_syncer = NULL; 2989 return (error); 2990 } 2991 vp->v_type = VNON; 2992 /* 2993 * Place the vnode onto the syncer worklist. We attempt to 2994 * scatter them about on the list so that they will go off 2995 * at evenly distributed times even if all the filesystems 2996 * are mounted at once. 2997 */ 2998 next += incr; 2999 if (next == 0 || next > syncer_maxdelay) { 3000 start /= 2; 3001 incr /= 2; 3002 if (start == 0) { 3003 start = syncer_maxdelay / 2; 3004 incr = syncer_maxdelay; 3005 } 3006 next = start; 3007 } 3008 VI_LOCK(vp); 3009 vn_syncer_add_to_worklist(&vp->v_bufobj, 3010 syncdelay > 0 ? next % syncdelay : 0); 3011 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 3012 mtx_lock(&sync_mtx); 3013 sync_vnode_count++; 3014 mtx_unlock(&sync_mtx); 3015 VI_UNLOCK(vp); 3016 mp->mnt_syncer = vp; 3017 return (0); 3018 } 3019 3020 /* 3021 * Do a lazy sync of the filesystem. 3022 */ 3023 static int 3024 sync_fsync(struct vop_fsync_args *ap) 3025 { 3026 struct vnode *syncvp = ap->a_vp; 3027 struct mount *mp = syncvp->v_mount; 3028 struct thread *td = ap->a_td; 3029 int error, asyncflag; 3030 struct bufobj *bo; 3031 3032 /* 3033 * We only need to do something if this is a lazy evaluation. 3034 */ 3035 if (ap->a_waitfor != MNT_LAZY) 3036 return (0); 3037 3038 /* 3039 * Move ourselves to the back of the sync list. 3040 */ 3041 bo = &syncvp->v_bufobj; 3042 BO_LOCK(bo); 3043 vn_syncer_add_to_worklist(bo, syncdelay); 3044 BO_UNLOCK(bo); 3045 3046 /* 3047 * Walk the list of vnodes pushing all that are dirty and 3048 * not already on the sync list. 3049 */ 3050 mtx_lock(&mountlist_mtx); 3051 if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { 3052 mtx_unlock(&mountlist_mtx); 3053 return (0); 3054 } 3055 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 3056 vfs_unbusy(mp, td); 3057 return (0); 3058 } 3059 asyncflag = mp->mnt_flag & MNT_ASYNC; 3060 mp->mnt_flag &= ~MNT_ASYNC; 3061 vfs_msync(mp, MNT_NOWAIT); 3062 error = VFS_SYNC(mp, MNT_LAZY, td); 3063 if (asyncflag) 3064 mp->mnt_flag |= MNT_ASYNC; 3065 vn_finished_write(mp); 3066 vfs_unbusy(mp, td); 3067 return (error); 3068 } 3069 3070 /* 3071 * The syncer vnode is no referenced. 3072 */ 3073 static int 3074 sync_inactive(struct vop_inactive_args *ap) 3075 { 3076 3077 vgone(ap->a_vp); 3078 return (0); 3079 } 3080 3081 /* 3082 * The syncer vnode is no longer needed and is being decommissioned. 3083 * 3084 * Modifications to the worklist must be protected by sync_mtx. 3085 */ 3086 static int 3087 sync_reclaim(struct vop_reclaim_args *ap) 3088 { 3089 struct vnode *vp = ap->a_vp; 3090 struct bufobj *bo; 3091 3092 VI_LOCK(vp); 3093 bo = &vp->v_bufobj; 3094 vp->v_mount->mnt_syncer = NULL; 3095 if (bo->bo_flag & BO_ONWORKLST) { 3096 mtx_lock(&sync_mtx); 3097 LIST_REMOVE(bo, bo_synclist); 3098 syncer_worklist_len--; 3099 sync_vnode_count--; 3100 mtx_unlock(&sync_mtx); 3101 bo->bo_flag &= ~BO_ONWORKLST; 3102 } 3103 VI_UNLOCK(vp); 3104 3105 return (0); 3106 } 3107 3108 /* 3109 * Check if vnode represents a disk device 3110 */ 3111 int 3112 vn_isdisk(struct vnode *vp, int *errp) 3113 { 3114 int error; 3115 3116 error = 0; 3117 dev_lock(); 3118 if (vp->v_type != VCHR) 3119 error = ENOTBLK; 3120 else if (vp->v_rdev == NULL) 3121 error = ENXIO; 3122 else if (vp->v_rdev->si_devsw == NULL) 3123 error = ENXIO; 3124 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 3125 error = ENOTBLK; 3126 dev_unlock(); 3127 if (errp != NULL) 3128 *errp = error; 3129 return (error == 0); 3130 } 3131 3132 /* 3133 * Common filesystem object access control check routine. Accepts a 3134 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3135 * and optional call-by-reference privused argument allowing vaccess() 3136 * to indicate to the caller whether privilege was used to satisfy the 3137 * request (obsoleted). Returns 0 on success, or an errno on failure. 3138 */ 3139 int 3140 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 3141 mode_t acc_mode, struct ucred *cred, int *privused) 3142 { 3143 mode_t dac_granted; 3144 #ifdef CAPABILITIES 3145 mode_t cap_granted; 3146 #endif 3147 3148 /* 3149 * Look for a normal, non-privileged way to access the file/directory 3150 * as requested. If it exists, go with that. 3151 */ 3152 3153 if (privused != NULL) 3154 *privused = 0; 3155 3156 dac_granted = 0; 3157 3158 /* Check the owner. */ 3159 if (cred->cr_uid == file_uid) { 3160 dac_granted |= VADMIN; 3161 if (file_mode & S_IXUSR) 3162 dac_granted |= VEXEC; 3163 if (file_mode & S_IRUSR) 3164 dac_granted |= VREAD; 3165 if (file_mode & S_IWUSR) 3166 dac_granted |= (VWRITE | VAPPEND); 3167 3168 if ((acc_mode & dac_granted) == acc_mode) 3169 return (0); 3170 3171 goto privcheck; 3172 } 3173 3174 /* Otherwise, check the groups (first match) */ 3175 if (groupmember(file_gid, cred)) { 3176 if (file_mode & S_IXGRP) 3177 dac_granted |= VEXEC; 3178 if (file_mode & S_IRGRP) 3179 dac_granted |= VREAD; 3180 if (file_mode & S_IWGRP) 3181 dac_granted |= (VWRITE | VAPPEND); 3182 3183 if ((acc_mode & dac_granted) == acc_mode) 3184 return (0); 3185 3186 goto privcheck; 3187 } 3188 3189 /* Otherwise, check everyone else. */ 3190 if (file_mode & S_IXOTH) 3191 dac_granted |= VEXEC; 3192 if (file_mode & S_IROTH) 3193 dac_granted |= VREAD; 3194 if (file_mode & S_IWOTH) 3195 dac_granted |= (VWRITE | VAPPEND); 3196 if ((acc_mode & dac_granted) == acc_mode) 3197 return (0); 3198 3199 privcheck: 3200 if (!suser_cred(cred, SUSER_ALLOWJAIL)) { 3201 /* XXX audit: privilege used */ 3202 if (privused != NULL) 3203 *privused = 1; 3204 return (0); 3205 } 3206 3207 #ifdef CAPABILITIES 3208 /* 3209 * Build a capability mask to determine if the set of capabilities 3210 * satisfies the requirements when combined with the granted mask 3211 * from above. 3212 * For each capability, if the capability is required, bitwise 3213 * or the request type onto the cap_granted mask. 3214 */ 3215 cap_granted = 0; 3216 3217 if (type == VDIR) { 3218 /* 3219 * For directories, use CAP_DAC_READ_SEARCH to satisfy 3220 * VEXEC requests, instead of CAP_DAC_EXECUTE. 3221 */ 3222 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3223 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL)) 3224 cap_granted |= VEXEC; 3225 } else { 3226 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3227 !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL)) 3228 cap_granted |= VEXEC; 3229 } 3230 3231 if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && 3232 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL)) 3233 cap_granted |= VREAD; 3234 3235 if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3236 !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL)) 3237 cap_granted |= (VWRITE | VAPPEND); 3238 3239 if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3240 !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL)) 3241 cap_granted |= VADMIN; 3242 3243 if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { 3244 /* XXX audit: privilege used */ 3245 if (privused != NULL) 3246 *privused = 1; 3247 return (0); 3248 } 3249 #endif 3250 3251 return ((acc_mode & VADMIN) ? EPERM : EACCES); 3252 } 3253 3254 /* 3255 * Credential check based on process requesting service, and per-attribute 3256 * permissions. 3257 */ 3258 int 3259 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 3260 struct thread *td, int access) 3261 { 3262 3263 /* 3264 * Kernel-invoked always succeeds. 3265 */ 3266 if (cred == NOCRED) 3267 return (0); 3268 3269 /* 3270 * Do not allow privileged processes in jail to directly 3271 * manipulate system attributes. 3272 * 3273 * XXX What capability should apply here? 3274 * Probably CAP_SYS_SETFFLAG. 3275 */ 3276 switch (attrnamespace) { 3277 case EXTATTR_NAMESPACE_SYSTEM: 3278 /* Potentially should be: return (EPERM); */ 3279 return (suser_cred(cred, 0)); 3280 case EXTATTR_NAMESPACE_USER: 3281 return (VOP_ACCESS(vp, access, cred, td)); 3282 default: 3283 return (EPERM); 3284 } 3285 } 3286 3287 #ifdef DEBUG_VFS_LOCKS 3288 /* 3289 * This only exists to supress warnings from unlocked specfs accesses. It is 3290 * no longer ok to have an unlocked VFS. 3291 */ 3292 #define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD) 3293 3294 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 3295 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, ""); 3296 3297 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 3298 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, ""); 3299 3300 int vfs_badlock_print = 1; /* Print lock violations. */ 3301 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, ""); 3302 3303 #ifdef KDB 3304 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 3305 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, ""); 3306 #endif 3307 3308 static void 3309 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 3310 { 3311 3312 #ifdef KDB 3313 if (vfs_badlock_backtrace) 3314 kdb_backtrace(); 3315 #endif 3316 if (vfs_badlock_print) 3317 printf("%s: %p %s\n", str, (void *)vp, msg); 3318 if (vfs_badlock_ddb) 3319 kdb_enter("lock violation"); 3320 } 3321 3322 void 3323 assert_vi_locked(struct vnode *vp, const char *str) 3324 { 3325 3326 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 3327 vfs_badlock("interlock is not locked but should be", str, vp); 3328 } 3329 3330 void 3331 assert_vi_unlocked(struct vnode *vp, const char *str) 3332 { 3333 3334 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 3335 vfs_badlock("interlock is locked but should not be", str, vp); 3336 } 3337 3338 void 3339 assert_vop_locked(struct vnode *vp, const char *str) 3340 { 3341 3342 if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0) 3343 vfs_badlock("is not locked but should be", str, vp); 3344 } 3345 3346 void 3347 assert_vop_unlocked(struct vnode *vp, const char *str) 3348 { 3349 3350 if (vp && !IGNORE_LOCK(vp) && 3351 VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE) 3352 vfs_badlock("is locked but should not be", str, vp); 3353 } 3354 3355 void 3356 assert_vop_elocked(struct vnode *vp, const char *str) 3357 { 3358 3359 if (vp && !IGNORE_LOCK(vp) && 3360 VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE) 3361 vfs_badlock("is not exclusive locked but should be", str, vp); 3362 } 3363 3364 #if 0 3365 void 3366 assert_vop_elocked_other(struct vnode *vp, const char *str) 3367 { 3368 3369 if (vp && !IGNORE_LOCK(vp) && 3370 VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER) 3371 vfs_badlock("is not exclusive locked by another thread", 3372 str, vp); 3373 } 3374 3375 void 3376 assert_vop_slocked(struct vnode *vp, const char *str) 3377 { 3378 3379 if (vp && !IGNORE_LOCK(vp) && 3380 VOP_ISLOCKED(vp, curthread) != LK_SHARED) 3381 vfs_badlock("is not locked shared but should be", str, vp); 3382 } 3383 #endif /* 0 */ 3384 #endif /* DEBUG_VFS_LOCKS */ 3385 3386 void 3387 vop_rename_pre(void *ap) 3388 { 3389 struct vop_rename_args *a = ap; 3390 3391 #ifdef DEBUG_VFS_LOCKS 3392 if (a->a_tvp) 3393 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 3394 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 3395 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 3396 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 3397 3398 /* Check the source (from). */ 3399 if (a->a_tdvp != a->a_fdvp && a->a_tvp != a->a_fdvp) 3400 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 3401 if (a->a_tvp != a->a_fvp) 3402 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked"); 3403 3404 /* Check the target. */ 3405 if (a->a_tvp) 3406 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 3407 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 3408 #endif 3409 if (a->a_tdvp != a->a_fdvp) 3410 vhold(a->a_fdvp); 3411 if (a->a_tvp != a->a_fvp) 3412 vhold(a->a_fvp); 3413 vhold(a->a_tdvp); 3414 if (a->a_tvp) 3415 vhold(a->a_tvp); 3416 } 3417 3418 void 3419 vop_strategy_pre(void *ap) 3420 { 3421 #ifdef DEBUG_VFS_LOCKS 3422 struct vop_strategy_args *a; 3423 struct buf *bp; 3424 3425 a = ap; 3426 bp = a->a_bp; 3427 3428 /* 3429 * Cluster ops lock their component buffers but not the IO container. 3430 */ 3431 if ((bp->b_flags & B_CLUSTER) != 0) 3432 return; 3433 3434 if (BUF_REFCNT(bp) < 1) { 3435 if (vfs_badlock_print) 3436 printf( 3437 "VOP_STRATEGY: bp is not locked but should be\n"); 3438 if (vfs_badlock_ddb) 3439 kdb_enter("lock violation"); 3440 } 3441 #endif 3442 } 3443 3444 void 3445 vop_lookup_pre(void *ap) 3446 { 3447 #ifdef DEBUG_VFS_LOCKS 3448 struct vop_lookup_args *a; 3449 struct vnode *dvp; 3450 3451 a = ap; 3452 dvp = a->a_dvp; 3453 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3454 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 3455 #endif 3456 } 3457 3458 void 3459 vop_lookup_post(void *ap, int rc) 3460 { 3461 #ifdef DEBUG_VFS_LOCKS 3462 struct vop_lookup_args *a; 3463 struct vnode *dvp; 3464 struct vnode *vp; 3465 3466 a = ap; 3467 dvp = a->a_dvp; 3468 vp = *(a->a_vpp); 3469 3470 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3471 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 3472 3473 if (!rc) 3474 ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)"); 3475 #endif 3476 } 3477 3478 void 3479 vop_lock_pre(void *ap) 3480 { 3481 #ifdef DEBUG_VFS_LOCKS 3482 struct vop_lock_args *a = ap; 3483 3484 if ((a->a_flags & LK_INTERLOCK) == 0) 3485 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3486 else 3487 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 3488 #endif 3489 } 3490 3491 void 3492 vop_lock_post(void *ap, int rc) 3493 { 3494 #ifdef DEBUG_VFS_LOCKS 3495 struct vop_lock_args *a = ap; 3496 3497 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3498 if (rc == 0) 3499 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 3500 #endif 3501 } 3502 3503 void 3504 vop_unlock_pre(void *ap) 3505 { 3506 #ifdef DEBUG_VFS_LOCKS 3507 struct vop_unlock_args *a = ap; 3508 3509 if (a->a_flags & LK_INTERLOCK) 3510 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 3511 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 3512 #endif 3513 } 3514 3515 void 3516 vop_unlock_post(void *ap, int rc) 3517 { 3518 #ifdef DEBUG_VFS_LOCKS 3519 struct vop_unlock_args *a = ap; 3520 3521 if (a->a_flags & LK_INTERLOCK) 3522 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 3523 #endif 3524 } 3525 3526 void 3527 vop_create_post(void *ap, int rc) 3528 { 3529 struct vop_create_args *a = ap; 3530 3531 if (!rc) 3532 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 3533 } 3534 3535 void 3536 vop_link_post(void *ap, int rc) 3537 { 3538 struct vop_link_args *a = ap; 3539 3540 if (!rc) { 3541 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK); 3542 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE); 3543 } 3544 } 3545 3546 void 3547 vop_mkdir_post(void *ap, int rc) 3548 { 3549 struct vop_mkdir_args *a = ap; 3550 3551 if (!rc) 3552 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 3553 } 3554 3555 void 3556 vop_mknod_post(void *ap, int rc) 3557 { 3558 struct vop_mknod_args *a = ap; 3559 3560 if (!rc) 3561 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 3562 } 3563 3564 void 3565 vop_remove_post(void *ap, int rc) 3566 { 3567 struct vop_remove_args *a = ap; 3568 3569 if (!rc) { 3570 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 3571 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 3572 } 3573 } 3574 3575 void 3576 vop_rename_post(void *ap, int rc) 3577 { 3578 struct vop_rename_args *a = ap; 3579 3580 if (!rc) { 3581 VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE); 3582 VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE); 3583 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 3584 if (a->a_tvp) 3585 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 3586 } 3587 if (a->a_tdvp != a->a_fdvp) 3588 vdrop(a->a_fdvp); 3589 if (a->a_tvp != a->a_fvp) 3590 vdrop(a->a_fvp); 3591 vdrop(a->a_tdvp); 3592 if (a->a_tvp) 3593 vdrop(a->a_tvp); 3594 } 3595 3596 void 3597 vop_rmdir_post(void *ap, int rc) 3598 { 3599 struct vop_rmdir_args *a = ap; 3600 3601 if (!rc) { 3602 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 3603 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 3604 } 3605 } 3606 3607 void 3608 vop_setattr_post(void *ap, int rc) 3609 { 3610 struct vop_setattr_args *a = ap; 3611 3612 if (!rc) 3613 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 3614 } 3615 3616 void 3617 vop_symlink_post(void *ap, int rc) 3618 { 3619 struct vop_symlink_args *a = ap; 3620 3621 if (!rc) 3622 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 3623 } 3624 3625 static struct knlist fs_knlist; 3626 3627 static void 3628 vfs_event_init(void *arg) 3629 { 3630 knlist_init(&fs_knlist, NULL, NULL, NULL, NULL); 3631 } 3632 /* XXX - correct order? */ 3633 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 3634 3635 void 3636 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused) 3637 { 3638 3639 KNOTE_UNLOCKED(&fs_knlist, event); 3640 } 3641 3642 static int filt_fsattach(struct knote *kn); 3643 static void filt_fsdetach(struct knote *kn); 3644 static int filt_fsevent(struct knote *kn, long hint); 3645 3646 struct filterops fs_filtops = 3647 { 0, filt_fsattach, filt_fsdetach, filt_fsevent }; 3648 3649 static int 3650 filt_fsattach(struct knote *kn) 3651 { 3652 3653 kn->kn_flags |= EV_CLEAR; 3654 knlist_add(&fs_knlist, kn, 0); 3655 return (0); 3656 } 3657 3658 static void 3659 filt_fsdetach(struct knote *kn) 3660 { 3661 3662 knlist_remove(&fs_knlist, kn, 0); 3663 } 3664 3665 static int 3666 filt_fsevent(struct knote *kn, long hint) 3667 { 3668 3669 kn->kn_fflags |= hint; 3670 return (kn->kn_fflags != 0); 3671 } 3672 3673 static int 3674 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 3675 { 3676 struct vfsidctl vc; 3677 int error; 3678 struct mount *mp; 3679 3680 error = SYSCTL_IN(req, &vc, sizeof(vc)); 3681 if (error) 3682 return (error); 3683 if (vc.vc_vers != VFS_CTL_VERS1) 3684 return (EINVAL); 3685 mp = vfs_getvfs(&vc.vc_fsid); 3686 if (mp == NULL) 3687 return (ENOENT); 3688 /* ensure that a specific sysctl goes to the right filesystem. */ 3689 if (strcmp(vc.vc_fstypename, "*") != 0 && 3690 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 3691 vfs_rel(mp); 3692 return (EINVAL); 3693 } 3694 VCTLTOREQ(&vc, req); 3695 error = VFS_SYSCTL(mp, vc.vc_op, req); 3696 vfs_rel(mp); 3697 return (error); 3698 } 3699 3700 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR, 3701 NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid"); 3702 3703 /* 3704 * Function to initialize a va_filerev field sensibly. 3705 * XXX: Wouldn't a random number make a lot more sense ?? 3706 */ 3707 u_quad_t 3708 init_va_filerev(void) 3709 { 3710 struct bintime bt; 3711 3712 getbinuptime(&bt); 3713 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 3714 } 3715 3716 static int filt_vfsread(struct knote *kn, long hint); 3717 static int filt_vfswrite(struct knote *kn, long hint); 3718 static int filt_vfsvnode(struct knote *kn, long hint); 3719 static void filt_vfsdetach(struct knote *kn); 3720 static struct filterops vfsread_filtops = 3721 { 1, NULL, filt_vfsdetach, filt_vfsread }; 3722 static struct filterops vfswrite_filtops = 3723 { 1, NULL, filt_vfsdetach, filt_vfswrite }; 3724 static struct filterops vfsvnode_filtops = 3725 { 1, NULL, filt_vfsdetach, filt_vfsvnode }; 3726 3727 static void 3728 vfs_knllock(void *arg) 3729 { 3730 struct vnode *vp = arg; 3731 3732 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread); 3733 } 3734 3735 static void 3736 vfs_knlunlock(void *arg) 3737 { 3738 struct vnode *vp = arg; 3739 3740 VOP_UNLOCK(vp, 0, curthread); 3741 } 3742 3743 static int 3744 vfs_knllocked(void *arg) 3745 { 3746 struct vnode *vp = arg; 3747 3748 return (VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE); 3749 } 3750 3751 int 3752 vfs_kqfilter(struct vop_kqfilter_args *ap) 3753 { 3754 struct vnode *vp = ap->a_vp; 3755 struct knote *kn = ap->a_kn; 3756 struct knlist *knl; 3757 3758 switch (kn->kn_filter) { 3759 case EVFILT_READ: 3760 kn->kn_fop = &vfsread_filtops; 3761 break; 3762 case EVFILT_WRITE: 3763 kn->kn_fop = &vfswrite_filtops; 3764 break; 3765 case EVFILT_VNODE: 3766 kn->kn_fop = &vfsvnode_filtops; 3767 break; 3768 default: 3769 return (EINVAL); 3770 } 3771 3772 kn->kn_hook = (caddr_t)vp; 3773 3774 if (vp->v_pollinfo == NULL) 3775 v_addpollinfo(vp); 3776 if (vp->v_pollinfo == NULL) 3777 return (ENOMEM); 3778 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 3779 knlist_add(knl, kn, 0); 3780 3781 return (0); 3782 } 3783 3784 /* 3785 * Detach knote from vnode 3786 */ 3787 static void 3788 filt_vfsdetach(struct knote *kn) 3789 { 3790 struct vnode *vp = (struct vnode *)kn->kn_hook; 3791 3792 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 3793 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 3794 } 3795 3796 /*ARGSUSED*/ 3797 static int 3798 filt_vfsread(struct knote *kn, long hint) 3799 { 3800 struct vnode *vp = (struct vnode *)kn->kn_hook; 3801 struct vattr va; 3802 3803 /* 3804 * filesystem is gone, so set the EOF flag and schedule 3805 * the knote for deletion. 3806 */ 3807 if (hint == NOTE_REVOKE) { 3808 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 3809 return (1); 3810 } 3811 3812 if (VOP_GETATTR(vp, &va, curthread->td_ucred, curthread)) 3813 return (0); 3814 3815 kn->kn_data = va.va_size - kn->kn_fp->f_offset; 3816 return (kn->kn_data != 0); 3817 } 3818 3819 /*ARGSUSED*/ 3820 static int 3821 filt_vfswrite(struct knote *kn, long hint) 3822 { 3823 /* 3824 * filesystem is gone, so set the EOF flag and schedule 3825 * the knote for deletion. 3826 */ 3827 if (hint == NOTE_REVOKE) 3828 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 3829 3830 kn->kn_data = 0; 3831 return (1); 3832 } 3833 3834 static int 3835 filt_vfsvnode(struct knote *kn, long hint) 3836 { 3837 if (kn->kn_sfflags & hint) 3838 kn->kn_fflags |= hint; 3839 if (hint == NOTE_REVOKE) { 3840 kn->kn_flags |= EV_EOF; 3841 return (1); 3842 } 3843 return (kn->kn_fflags != 0); 3844 } 3845 3846 int 3847 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 3848 { 3849 int error; 3850 3851 if (dp->d_reclen > ap->a_uio->uio_resid) 3852 return (ENAMETOOLONG); 3853 error = uiomove(dp, dp->d_reclen, ap->a_uio); 3854 if (error) { 3855 if (ap->a_ncookies != NULL) { 3856 if (ap->a_cookies != NULL) 3857 free(ap->a_cookies, M_TEMP); 3858 ap->a_cookies = NULL; 3859 *ap->a_ncookies = 0; 3860 } 3861 return (error); 3862 } 3863 if (ap->a_ncookies == NULL) 3864 return (0); 3865 3866 KASSERT(ap->a_cookies, 3867 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 3868 3869 *ap->a_cookies = realloc(*ap->a_cookies, 3870 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); 3871 (*ap->a_cookies)[*ap->a_ncookies] = off; 3872 return (0); 3873 } 3874 3875 /* 3876 * Mark for update the access time of the file if the filesystem 3877 * supports VA_MARK_ATIME. This functionality is used by execve 3878 * and mmap, so we want to avoid the synchronous I/O implied by 3879 * directly setting va_atime for the sake of efficiency. 3880 */ 3881 void 3882 vfs_mark_atime(struct vnode *vp, struct thread *td) 3883 { 3884 struct vattr atimeattr; 3885 3886 if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) { 3887 VATTR_NULL(&atimeattr); 3888 atimeattr.va_vaflags |= VA_MARK_ATIME; 3889 (void)VOP_SETATTR(vp, &atimeattr, td->td_ucred, td); 3890 } 3891 } 3892