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_compat.h" 45 #include "opt_ddb.h" 46 #include "opt_watchdog.h" 47 48 #include <sys/param.h> 49 #include <sys/systm.h> 50 #include <sys/bio.h> 51 #include <sys/buf.h> 52 #include <sys/condvar.h> 53 #include <sys/conf.h> 54 #include <sys/dirent.h> 55 #include <sys/event.h> 56 #include <sys/eventhandler.h> 57 #include <sys/extattr.h> 58 #include <sys/file.h> 59 #include <sys/fcntl.h> 60 #include <sys/jail.h> 61 #include <sys/kdb.h> 62 #include <sys/kernel.h> 63 #include <sys/kthread.h> 64 #include <sys/lockf.h> 65 #include <sys/malloc.h> 66 #include <sys/mount.h> 67 #include <sys/namei.h> 68 #include <sys/pctrie.h> 69 #include <sys/priv.h> 70 #include <sys/reboot.h> 71 #include <sys/refcount.h> 72 #include <sys/rwlock.h> 73 #include <sys/sched.h> 74 #include <sys/sleepqueue.h> 75 #include <sys/smp.h> 76 #include <sys/stat.h> 77 #include <sys/sysctl.h> 78 #include <sys/syslog.h> 79 #include <sys/vmmeter.h> 80 #include <sys/vnode.h> 81 #include <sys/watchdog.h> 82 83 #include <machine/stdarg.h> 84 85 #include <security/mac/mac_framework.h> 86 87 #include <vm/vm.h> 88 #include <vm/vm_object.h> 89 #include <vm/vm_extern.h> 90 #include <vm/pmap.h> 91 #include <vm/vm_map.h> 92 #include <vm/vm_page.h> 93 #include <vm/vm_kern.h> 94 #include <vm/uma.h> 95 96 #ifdef DDB 97 #include <ddb/ddb.h> 98 #endif 99 100 static void delmntque(struct vnode *vp); 101 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, 102 int slpflag, int slptimeo); 103 static void syncer_shutdown(void *arg, int howto); 104 static int vtryrecycle(struct vnode *vp); 105 static void v_init_counters(struct vnode *); 106 static void v_incr_usecount(struct vnode *); 107 static void v_incr_devcount(struct vnode *); 108 static void v_decr_devcount(struct vnode *); 109 static void vnlru_free(int); 110 static void vgonel(struct vnode *); 111 static void vfs_knllock(void *arg); 112 static void vfs_knlunlock(void *arg); 113 static void vfs_knl_assert_locked(void *arg); 114 static void vfs_knl_assert_unlocked(void *arg); 115 static void destroy_vpollinfo(struct vpollinfo *vi); 116 117 /* 118 * Number of vnodes in existence. Increased whenever getnewvnode() 119 * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode. 120 */ 121 static unsigned long numvnodes; 122 123 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, 124 "Number of vnodes in existence"); 125 126 static u_long vnodes_created; 127 SYSCTL_ULONG(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created, 128 0, "Number of vnodes created by getnewvnode"); 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 are rarely completely free, but are 150 * just ones that are cheap to recycle. Usually they are for files which 151 * have been stat'd but not read; these usually have inode and namecache 152 * data attached to them. This target is the preferred minimum size of a 153 * sub-cache consisting mostly of such files. The system balances the size 154 * of this sub-cache with its complement to try to prevent either from 155 * thrashing while the other is relatively inactive. The targets express 156 * a preference for the best balance. 157 * 158 * "Above" this target there are 2 further targets (watermarks) related 159 * to recyling of free vnodes. In the best-operating case, the cache is 160 * exactly full, the free list has size between vlowat and vhiwat above the 161 * free target, and recycling from it and normal use maintains this state. 162 * Sometimes the free list is below vlowat or even empty, but this state 163 * is even better for immediate use provided the cache is not full. 164 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free 165 * ones) to reach one of these states. The watermarks are currently hard- 166 * coded as 4% and 9% of the available space higher. These and the default 167 * of 25% for wantfreevnodes are too large if the memory size is large. 168 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim 169 * whenever vnlru_proc() becomes active. 170 */ 171 static u_long wantfreevnodes; 172 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, 173 &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes"); 174 static u_long freevnodes; 175 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, 176 &freevnodes, 0, "Number of \"free\" vnodes"); 177 178 static u_long recycles_count; 179 SYSCTL_ULONG(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count, 0, 180 "Number of vnodes recycled to meet vnode cache targets"); 181 182 /* 183 * Various variables used for debugging the new implementation of 184 * reassignbuf(). 185 * XXX these are probably of (very) limited utility now. 186 */ 187 static int reassignbufcalls; 188 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, 189 "Number of calls to reassignbuf"); 190 191 static u_long free_owe_inact; 192 SYSCTL_ULONG(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact, 0, 193 "Number of times free vnodes kept on active list due to VFS " 194 "owing inactivation"); 195 196 /* To keep more than one thread at a time from running vfs_getnewfsid */ 197 static struct mtx mntid_mtx; 198 199 /* 200 * Lock for any access to the following: 201 * vnode_free_list 202 * numvnodes 203 * freevnodes 204 */ 205 static struct mtx vnode_free_list_mtx; 206 207 /* Publicly exported FS */ 208 struct nfs_public nfs_pub; 209 210 static uma_zone_t buf_trie_zone; 211 212 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 213 static uma_zone_t vnode_zone; 214 static uma_zone_t vnodepoll_zone; 215 216 /* 217 * The workitem queue. 218 * 219 * It is useful to delay writes of file data and filesystem metadata 220 * for tens of seconds so that quickly created and deleted files need 221 * not waste disk bandwidth being created and removed. To realize this, 222 * we append vnodes to a "workitem" queue. When running with a soft 223 * updates implementation, most pending metadata dependencies should 224 * not wait for more than a few seconds. Thus, mounted on block devices 225 * are delayed only about a half the time that file data is delayed. 226 * Similarly, directory updates are more critical, so are only delayed 227 * about a third the time that file data is delayed. Thus, there are 228 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 229 * one each second (driven off the filesystem syncer process). The 230 * syncer_delayno variable indicates the next queue that is to be processed. 231 * Items that need to be processed soon are placed in this queue: 232 * 233 * syncer_workitem_pending[syncer_delayno] 234 * 235 * A delay of fifteen seconds is done by placing the request fifteen 236 * entries later in the queue: 237 * 238 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 239 * 240 */ 241 static int syncer_delayno; 242 static long syncer_mask; 243 LIST_HEAD(synclist, bufobj); 244 static struct synclist *syncer_workitem_pending; 245 /* 246 * The sync_mtx protects: 247 * bo->bo_synclist 248 * sync_vnode_count 249 * syncer_delayno 250 * syncer_state 251 * syncer_workitem_pending 252 * syncer_worklist_len 253 * rushjob 254 */ 255 static struct mtx sync_mtx; 256 static struct cv sync_wakeup; 257 258 #define SYNCER_MAXDELAY 32 259 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 260 static int syncdelay = 30; /* max time to delay syncing data */ 261 static int filedelay = 30; /* time to delay syncing files */ 262 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, 263 "Time to delay syncing files (in seconds)"); 264 static int dirdelay = 29; /* time to delay syncing directories */ 265 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, 266 "Time to delay syncing directories (in seconds)"); 267 static int metadelay = 28; /* time to delay syncing metadata */ 268 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, 269 "Time to delay syncing metadata (in seconds)"); 270 static int rushjob; /* number of slots to run ASAP */ 271 static int stat_rush_requests; /* number of times I/O speeded up */ 272 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, 273 "Number of times I/O speeded up (rush requests)"); 274 275 /* 276 * When shutting down the syncer, run it at four times normal speed. 277 */ 278 #define SYNCER_SHUTDOWN_SPEEDUP 4 279 static int sync_vnode_count; 280 static int syncer_worklist_len; 281 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 282 syncer_state; 283 284 /* Target for maximum number of vnodes. */ 285 int desiredvnodes; 286 static int gapvnodes; /* gap between wanted and desired */ 287 static int vhiwat; /* enough extras after expansion */ 288 static int vlowat; /* minimal extras before expansion */ 289 static int vstir; /* nonzero to stir non-free vnodes */ 290 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */ 291 292 static int 293 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS) 294 { 295 int error, old_desiredvnodes; 296 297 old_desiredvnodes = desiredvnodes; 298 if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0) 299 return (error); 300 if (old_desiredvnodes != desiredvnodes) { 301 wantfreevnodes = desiredvnodes / 4; 302 /* XXX locking seems to be incomplete. */ 303 vfs_hash_changesize(desiredvnodes); 304 cache_changesize(desiredvnodes); 305 } 306 return (0); 307 } 308 309 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes, 310 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0, 311 sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes"); 312 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 313 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)"); 314 static int vnlru_nowhere; 315 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 316 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 317 318 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */ 319 static int vnsz2log; 320 321 /* 322 * Support for the bufobj clean & dirty pctrie. 323 */ 324 static void * 325 buf_trie_alloc(struct pctrie *ptree) 326 { 327 328 return uma_zalloc(buf_trie_zone, M_NOWAIT); 329 } 330 331 static void 332 buf_trie_free(struct pctrie *ptree, void *node) 333 { 334 335 uma_zfree(buf_trie_zone, node); 336 } 337 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free); 338 339 /* 340 * Initialize the vnode management data structures. 341 * 342 * Reevaluate the following cap on the number of vnodes after the physical 343 * memory size exceeds 512GB. In the limit, as the physical memory size 344 * grows, the ratio of the memory size in KB to to vnodes approaches 64:1. 345 */ 346 #ifndef MAXVNODES_MAX 347 #define MAXVNODES_MAX (512 * 1024 * 1024 / 64) /* 8M */ 348 #endif 349 350 /* 351 * Initialize a vnode as it first enters the zone. 352 */ 353 static int 354 vnode_init(void *mem, int size, int flags) 355 { 356 struct vnode *vp; 357 struct bufobj *bo; 358 359 vp = mem; 360 bzero(vp, size); 361 /* 362 * Setup locks. 363 */ 364 vp->v_vnlock = &vp->v_lock; 365 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 366 /* 367 * By default, don't allow shared locks unless filesystems opt-in. 368 */ 369 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT, 370 LK_NOSHARE | LK_IS_VNODE); 371 /* 372 * Initialize bufobj. 373 */ 374 bo = &vp->v_bufobj; 375 bo->__bo_vnode = vp; 376 rw_init(BO_LOCKPTR(bo), "bufobj interlock"); 377 bo->bo_private = vp; 378 TAILQ_INIT(&bo->bo_clean.bv_hd); 379 TAILQ_INIT(&bo->bo_dirty.bv_hd); 380 /* 381 * Initialize namecache. 382 */ 383 LIST_INIT(&vp->v_cache_src); 384 TAILQ_INIT(&vp->v_cache_dst); 385 /* 386 * Initialize rangelocks. 387 */ 388 rangelock_init(&vp->v_rl); 389 return (0); 390 } 391 392 /* 393 * Free a vnode when it is cleared from the zone. 394 */ 395 static void 396 vnode_fini(void *mem, int size) 397 { 398 struct vnode *vp; 399 struct bufobj *bo; 400 401 vp = mem; 402 rangelock_destroy(&vp->v_rl); 403 lockdestroy(vp->v_vnlock); 404 mtx_destroy(&vp->v_interlock); 405 bo = &vp->v_bufobj; 406 rw_destroy(BO_LOCKPTR(bo)); 407 } 408 409 static void 410 vntblinit(void *dummy __unused) 411 { 412 u_int i; 413 int physvnodes, virtvnodes; 414 415 /* 416 * Desiredvnodes is a function of the physical memory size and the 417 * kernel's heap size. Generally speaking, it scales with the 418 * physical memory size. The ratio of desiredvnodes to the physical 419 * memory size is 1:16 until desiredvnodes exceeds 98,304. 420 * Thereafter, the 421 * marginal ratio of desiredvnodes to the physical memory size is 422 * 1:64. However, desiredvnodes is limited by the kernel's heap 423 * size. The memory required by desiredvnodes vnodes and vm objects 424 * must not exceed 1/7th of the kernel's heap size. 425 */ 426 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 + 427 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64; 428 virtvnodes = vm_kmem_size / (7 * (sizeof(struct vm_object) + 429 sizeof(struct vnode))); 430 desiredvnodes = min(physvnodes, virtvnodes); 431 if (desiredvnodes > MAXVNODES_MAX) { 432 if (bootverbose) 433 printf("Reducing kern.maxvnodes %d -> %d\n", 434 desiredvnodes, MAXVNODES_MAX); 435 desiredvnodes = MAXVNODES_MAX; 436 } 437 wantfreevnodes = desiredvnodes / 4; 438 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 439 TAILQ_INIT(&vnode_free_list); 440 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 441 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 442 vnode_init, vnode_fini, UMA_ALIGN_PTR, 0); 443 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 444 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 445 /* 446 * Preallocate enough nodes to support one-per buf so that 447 * we can not fail an insert. reassignbuf() callers can not 448 * tolerate the insertion failure. 449 */ 450 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(), 451 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, 452 UMA_ZONE_NOFREE | UMA_ZONE_VM); 453 uma_prealloc(buf_trie_zone, nbuf); 454 /* 455 * Initialize the filesystem syncer. 456 */ 457 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 458 &syncer_mask); 459 syncer_maxdelay = syncer_mask + 1; 460 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 461 cv_init(&sync_wakeup, "syncer"); 462 for (i = 1; i <= sizeof(struct vnode); i <<= 1) 463 vnsz2log++; 464 vnsz2log--; 465 } 466 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL); 467 468 469 /* 470 * Mark a mount point as busy. Used to synchronize access and to delay 471 * unmounting. Eventually, mountlist_mtx is not released on failure. 472 * 473 * vfs_busy() is a custom lock, it can block the caller. 474 * vfs_busy() only sleeps if the unmount is active on the mount point. 475 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any 476 * vnode belonging to mp. 477 * 478 * Lookup uses vfs_busy() to traverse mount points. 479 * root fs var fs 480 * / vnode lock A / vnode lock (/var) D 481 * /var vnode lock B /log vnode lock(/var/log) E 482 * vfs_busy lock C vfs_busy lock F 483 * 484 * Within each file system, the lock order is C->A->B and F->D->E. 485 * 486 * When traversing across mounts, the system follows that lock order: 487 * 488 * C->A->B 489 * | 490 * +->F->D->E 491 * 492 * The lookup() process for namei("/var") illustrates the process: 493 * VOP_LOOKUP() obtains B while A is held 494 * vfs_busy() obtains a shared lock on F while A and B are held 495 * vput() releases lock on B 496 * vput() releases lock on A 497 * VFS_ROOT() obtains lock on D while shared lock on F is held 498 * vfs_unbusy() releases shared lock on F 499 * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A. 500 * Attempt to lock A (instead of vp_crossmp) while D is held would 501 * violate the global order, causing deadlocks. 502 * 503 * dounmount() locks B while F is drained. 504 */ 505 int 506 vfs_busy(struct mount *mp, int flags) 507 { 508 509 MPASS((flags & ~MBF_MASK) == 0); 510 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags); 511 512 MNT_ILOCK(mp); 513 MNT_REF(mp); 514 /* 515 * If mount point is currenly being unmounted, sleep until the 516 * mount point fate is decided. If thread doing the unmounting fails, 517 * it will clear MNTK_UNMOUNT flag before waking us up, indicating 518 * that this mount point has survived the unmount attempt and vfs_busy 519 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE 520 * flag in addition to MNTK_UNMOUNT, indicating that mount point is 521 * about to be really destroyed. vfs_busy needs to release its 522 * reference on the mount point in this case and return with ENOENT, 523 * telling the caller that mount mount it tried to busy is no longer 524 * valid. 525 */ 526 while (mp->mnt_kern_flag & MNTK_UNMOUNT) { 527 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) { 528 MNT_REL(mp); 529 MNT_IUNLOCK(mp); 530 CTR1(KTR_VFS, "%s: failed busying before sleeping", 531 __func__); 532 return (ENOENT); 533 } 534 if (flags & MBF_MNTLSTLOCK) 535 mtx_unlock(&mountlist_mtx); 536 mp->mnt_kern_flag |= MNTK_MWAIT; 537 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0); 538 if (flags & MBF_MNTLSTLOCK) 539 mtx_lock(&mountlist_mtx); 540 MNT_ILOCK(mp); 541 } 542 if (flags & MBF_MNTLSTLOCK) 543 mtx_unlock(&mountlist_mtx); 544 mp->mnt_lockref++; 545 MNT_IUNLOCK(mp); 546 return (0); 547 } 548 549 /* 550 * Free a busy filesystem. 551 */ 552 void 553 vfs_unbusy(struct mount *mp) 554 { 555 556 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 557 MNT_ILOCK(mp); 558 MNT_REL(mp); 559 KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref")); 560 mp->mnt_lockref--; 561 if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) { 562 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT); 563 CTR1(KTR_VFS, "%s: waking up waiters", __func__); 564 mp->mnt_kern_flag &= ~MNTK_DRAINING; 565 wakeup(&mp->mnt_lockref); 566 } 567 MNT_IUNLOCK(mp); 568 } 569 570 /* 571 * Lookup a mount point by filesystem identifier. 572 */ 573 struct mount * 574 vfs_getvfs(fsid_t *fsid) 575 { 576 struct mount *mp; 577 578 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 579 mtx_lock(&mountlist_mtx); 580 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 581 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 582 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 583 vfs_ref(mp); 584 mtx_unlock(&mountlist_mtx); 585 return (mp); 586 } 587 } 588 mtx_unlock(&mountlist_mtx); 589 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 590 return ((struct mount *) 0); 591 } 592 593 /* 594 * Lookup a mount point by filesystem identifier, busying it before 595 * returning. 596 * 597 * To avoid congestion on mountlist_mtx, implement simple direct-mapped 598 * cache for popular filesystem identifiers. The cache is lockess, using 599 * the fact that struct mount's are never freed. In worst case we may 600 * get pointer to unmounted or even different filesystem, so we have to 601 * check what we got, and go slow way if so. 602 */ 603 struct mount * 604 vfs_busyfs(fsid_t *fsid) 605 { 606 #define FSID_CACHE_SIZE 256 607 typedef struct mount * volatile vmp_t; 608 static vmp_t cache[FSID_CACHE_SIZE]; 609 struct mount *mp; 610 int error; 611 uint32_t hash; 612 613 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 614 hash = fsid->val[0] ^ fsid->val[1]; 615 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1); 616 mp = cache[hash]; 617 if (mp == NULL || 618 mp->mnt_stat.f_fsid.val[0] != fsid->val[0] || 619 mp->mnt_stat.f_fsid.val[1] != fsid->val[1]) 620 goto slow; 621 if (vfs_busy(mp, 0) != 0) { 622 cache[hash] = NULL; 623 goto slow; 624 } 625 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 626 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) 627 return (mp); 628 else 629 vfs_unbusy(mp); 630 631 slow: 632 mtx_lock(&mountlist_mtx); 633 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 634 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 635 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 636 error = vfs_busy(mp, MBF_MNTLSTLOCK); 637 if (error) { 638 cache[hash] = NULL; 639 mtx_unlock(&mountlist_mtx); 640 return (NULL); 641 } 642 cache[hash] = mp; 643 return (mp); 644 } 645 } 646 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 647 mtx_unlock(&mountlist_mtx); 648 return ((struct mount *) 0); 649 } 650 651 /* 652 * Check if a user can access privileged mount options. 653 */ 654 int 655 vfs_suser(struct mount *mp, struct thread *td) 656 { 657 int error; 658 659 /* 660 * If the thread is jailed, but this is not a jail-friendly file 661 * system, deny immediately. 662 */ 663 if (!(mp->mnt_vfc->vfc_flags & VFCF_JAIL) && jailed(td->td_ucred)) 664 return (EPERM); 665 666 /* 667 * If the file system was mounted outside the jail of the calling 668 * thread, deny immediately. 669 */ 670 if (prison_check(td->td_ucred, mp->mnt_cred) != 0) 671 return (EPERM); 672 673 /* 674 * If file system supports delegated administration, we don't check 675 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified 676 * by the file system itself. 677 * If this is not the user that did original mount, we check for 678 * the PRIV_VFS_MOUNT_OWNER privilege. 679 */ 680 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) && 681 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 682 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) 683 return (error); 684 } 685 return (0); 686 } 687 688 /* 689 * Get a new unique fsid. Try to make its val[0] unique, since this value 690 * will be used to create fake device numbers for stat(). Also try (but 691 * not so hard) make its val[0] unique mod 2^16, since some emulators only 692 * support 16-bit device numbers. We end up with unique val[0]'s for the 693 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 694 * 695 * Keep in mind that several mounts may be running in parallel. Starting 696 * the search one past where the previous search terminated is both a 697 * micro-optimization and a defense against returning the same fsid to 698 * different mounts. 699 */ 700 void 701 vfs_getnewfsid(struct mount *mp) 702 { 703 static uint16_t mntid_base; 704 struct mount *nmp; 705 fsid_t tfsid; 706 int mtype; 707 708 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 709 mtx_lock(&mntid_mtx); 710 mtype = mp->mnt_vfc->vfc_typenum; 711 tfsid.val[1] = mtype; 712 mtype = (mtype & 0xFF) << 24; 713 for (;;) { 714 tfsid.val[0] = makedev(255, 715 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 716 mntid_base++; 717 if ((nmp = vfs_getvfs(&tfsid)) == NULL) 718 break; 719 vfs_rel(nmp); 720 } 721 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 722 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 723 mtx_unlock(&mntid_mtx); 724 } 725 726 /* 727 * Knob to control the precision of file timestamps: 728 * 729 * 0 = seconds only; nanoseconds zeroed. 730 * 1 = seconds and nanoseconds, accurate within 1/HZ. 731 * 2 = seconds and nanoseconds, truncated to microseconds. 732 * >=3 = seconds and nanoseconds, maximum precision. 733 */ 734 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 735 736 static int timestamp_precision = TSP_USEC; 737 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 738 ×tamp_precision, 0, "File timestamp precision (0: seconds, " 739 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to ms, " 740 "3+: sec + ns (max. precision))"); 741 742 /* 743 * Get a current timestamp. 744 */ 745 void 746 vfs_timestamp(struct timespec *tsp) 747 { 748 struct timeval tv; 749 750 switch (timestamp_precision) { 751 case TSP_SEC: 752 tsp->tv_sec = time_second; 753 tsp->tv_nsec = 0; 754 break; 755 case TSP_HZ: 756 getnanotime(tsp); 757 break; 758 case TSP_USEC: 759 microtime(&tv); 760 TIMEVAL_TO_TIMESPEC(&tv, tsp); 761 break; 762 case TSP_NSEC: 763 default: 764 nanotime(tsp); 765 break; 766 } 767 } 768 769 /* 770 * Set vnode attributes to VNOVAL 771 */ 772 void 773 vattr_null(struct vattr *vap) 774 { 775 776 vap->va_type = VNON; 777 vap->va_size = VNOVAL; 778 vap->va_bytes = VNOVAL; 779 vap->va_mode = VNOVAL; 780 vap->va_nlink = VNOVAL; 781 vap->va_uid = VNOVAL; 782 vap->va_gid = VNOVAL; 783 vap->va_fsid = VNOVAL; 784 vap->va_fileid = VNOVAL; 785 vap->va_blocksize = VNOVAL; 786 vap->va_rdev = VNOVAL; 787 vap->va_atime.tv_sec = VNOVAL; 788 vap->va_atime.tv_nsec = VNOVAL; 789 vap->va_mtime.tv_sec = VNOVAL; 790 vap->va_mtime.tv_nsec = VNOVAL; 791 vap->va_ctime.tv_sec = VNOVAL; 792 vap->va_ctime.tv_nsec = VNOVAL; 793 vap->va_birthtime.tv_sec = VNOVAL; 794 vap->va_birthtime.tv_nsec = VNOVAL; 795 vap->va_flags = VNOVAL; 796 vap->va_gen = VNOVAL; 797 vap->va_vaflags = 0; 798 } 799 800 /* 801 * This routine is called when we have too many vnodes. It attempts 802 * to free <count> vnodes and will potentially free vnodes that still 803 * have VM backing store (VM backing store is typically the cause 804 * of a vnode blowout so we want to do this). Therefore, this operation 805 * is not considered cheap. 806 * 807 * A number of conditions may prevent a vnode from being reclaimed. 808 * the buffer cache may have references on the vnode, a directory 809 * vnode may still have references due to the namei cache representing 810 * underlying files, or the vnode may be in active use. It is not 811 * desireable to reuse such vnodes. These conditions may cause the 812 * number of vnodes to reach some minimum value regardless of what 813 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 814 */ 815 static int 816 vlrureclaim(struct mount *mp, int reclaim_nc_src, int trigger) 817 { 818 struct vnode *vp; 819 int count, done, target; 820 821 done = 0; 822 vn_start_write(NULL, &mp, V_WAIT); 823 MNT_ILOCK(mp); 824 count = mp->mnt_nvnodelistsize; 825 target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1); 826 target = target / 10 + 1; 827 while (count != 0 && done < target) { 828 vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 829 while (vp != NULL && vp->v_type == VMARKER) 830 vp = TAILQ_NEXT(vp, v_nmntvnodes); 831 if (vp == NULL) 832 break; 833 /* 834 * XXX LRU is completely broken for non-free vnodes. First 835 * by calling here in mountpoint order, then by moving 836 * unselected vnodes to the end here, and most grossly by 837 * removing the vlruvp() function that was supposed to 838 * maintain the order. (This function was born broken 839 * since syncer problems prevented it doing anything.) The 840 * order is closer to LRC (C = Created). 841 * 842 * LRU reclaiming of vnodes seems to have last worked in 843 * FreeBSD-3 where LRU wasn't mentioned under any spelling. 844 * Then there was no hold count, and inactive vnodes were 845 * simply put on the free list in LRU order. The separate 846 * lists also break LRU. We prefer to reclaim from the 847 * free list for technical reasons. This tends to thrash 848 * the free list to keep very unrecently used held vnodes. 849 * The problem is mitigated by keeping the free list large. 850 */ 851 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 852 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 853 --count; 854 if (!VI_TRYLOCK(vp)) 855 goto next_iter; 856 /* 857 * If it's been deconstructed already, it's still 858 * referenced, or it exceeds the trigger, skip it. 859 * Also skip free vnodes. We are trying to make space 860 * to expand the free list, not reduce it. 861 */ 862 if (vp->v_usecount || 863 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 864 ((vp->v_iflag & VI_FREE) != 0) || 865 (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL && 866 vp->v_object->resident_page_count > trigger)) { 867 VI_UNLOCK(vp); 868 goto next_iter; 869 } 870 MNT_IUNLOCK(mp); 871 vholdl(vp); 872 if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) { 873 vdrop(vp); 874 goto next_iter_mntunlocked; 875 } 876 VI_LOCK(vp); 877 /* 878 * v_usecount may have been bumped after VOP_LOCK() dropped 879 * the vnode interlock and before it was locked again. 880 * 881 * It is not necessary to recheck VI_DOOMED because it can 882 * only be set by another thread that holds both the vnode 883 * lock and vnode interlock. If another thread has the 884 * vnode lock before we get to VOP_LOCK() and obtains the 885 * vnode interlock after VOP_LOCK() drops the vnode 886 * interlock, the other thread will be unable to drop the 887 * vnode lock before our VOP_LOCK() call fails. 888 */ 889 if (vp->v_usecount || 890 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 891 (vp->v_iflag & VI_FREE) != 0 || 892 (vp->v_object != NULL && 893 vp->v_object->resident_page_count > trigger)) { 894 VOP_UNLOCK(vp, LK_INTERLOCK); 895 vdrop(vp); 896 goto next_iter_mntunlocked; 897 } 898 KASSERT((vp->v_iflag & VI_DOOMED) == 0, 899 ("VI_DOOMED unexpectedly detected in vlrureclaim()")); 900 atomic_add_long(&recycles_count, 1); 901 vgonel(vp); 902 VOP_UNLOCK(vp, 0); 903 vdropl(vp); 904 done++; 905 next_iter_mntunlocked: 906 if (!should_yield()) 907 goto relock_mnt; 908 goto yield; 909 next_iter: 910 if (!should_yield()) 911 continue; 912 MNT_IUNLOCK(mp); 913 yield: 914 kern_yield(PRI_USER); 915 relock_mnt: 916 MNT_ILOCK(mp); 917 } 918 MNT_IUNLOCK(mp); 919 vn_finished_write(mp); 920 return done; 921 } 922 923 /* 924 * Attempt to reduce the free list by the requested amount. 925 */ 926 static void 927 vnlru_free(int count) 928 { 929 struct vnode *vp; 930 931 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 932 for (; count > 0; count--) { 933 vp = TAILQ_FIRST(&vnode_free_list); 934 /* 935 * The list can be modified while the free_list_mtx 936 * has been dropped and vp could be NULL here. 937 */ 938 if (!vp) 939 break; 940 VNASSERT(vp->v_op != NULL, vp, 941 ("vnlru_free: vnode already reclaimed.")); 942 KASSERT((vp->v_iflag & VI_FREE) != 0, 943 ("Removing vnode not on freelist")); 944 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 945 ("Mangling active vnode")); 946 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist); 947 /* 948 * Don't recycle if we can't get the interlock. 949 */ 950 if (!VI_TRYLOCK(vp)) { 951 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist); 952 continue; 953 } 954 VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0, 955 vp, ("vp inconsistent on freelist")); 956 957 /* 958 * The clear of VI_FREE prevents activation of the 959 * vnode. There is no sense in putting the vnode on 960 * the mount point active list, only to remove it 961 * later during recycling. Inline the relevant part 962 * of vholdl(), to avoid triggering assertions or 963 * activating. 964 */ 965 freevnodes--; 966 vp->v_iflag &= ~VI_FREE; 967 refcount_acquire(&vp->v_holdcnt); 968 969 mtx_unlock(&vnode_free_list_mtx); 970 VI_UNLOCK(vp); 971 vtryrecycle(vp); 972 /* 973 * If the recycled succeeded this vdrop will actually free 974 * the vnode. If not it will simply place it back on 975 * the free list. 976 */ 977 vdrop(vp); 978 mtx_lock(&vnode_free_list_mtx); 979 } 980 } 981 982 /* XXX some names and initialization are bad for limits and watermarks. */ 983 static int 984 vspace(void) 985 { 986 int space; 987 988 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100); 989 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */ 990 vlowat = vhiwat / 2; 991 if (numvnodes > desiredvnodes) 992 return (0); 993 space = desiredvnodes - numvnodes; 994 if (freevnodes > wantfreevnodes) 995 space += freevnodes - wantfreevnodes; 996 return (space); 997 } 998 999 /* 1000 * Attempt to recycle vnodes in a context that is always safe to block. 1001 * Calling vlrurecycle() from the bowels of filesystem code has some 1002 * interesting deadlock problems. 1003 */ 1004 static struct proc *vnlruproc; 1005 static int vnlruproc_sig; 1006 1007 static void 1008 vnlru_proc(void) 1009 { 1010 struct mount *mp, *nmp; 1011 unsigned long ofreevnodes, onumvnodes; 1012 int done, force, reclaim_nc_src, trigger, usevnodes; 1013 1014 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc, 1015 SHUTDOWN_PRI_FIRST); 1016 1017 force = 0; 1018 for (;;) { 1019 kproc_suspend_check(vnlruproc); 1020 mtx_lock(&vnode_free_list_mtx); 1021 /* 1022 * If numvnodes is too large (due to desiredvnodes being 1023 * adjusted using its sysctl, or emergency growth), first 1024 * try to reduce it by discarding from the free list. 1025 */ 1026 if (numvnodes > desiredvnodes && freevnodes > 0) 1027 vnlru_free(ulmin(numvnodes - desiredvnodes, 1028 freevnodes)); 1029 /* 1030 * Sleep if the vnode cache is in a good state. This is 1031 * when it is not over-full and has space for about a 4% 1032 * or 9% expansion (by growing its size or inexcessively 1033 * reducing its free list). Otherwise, try to reclaim 1034 * space for a 10% expansion. 1035 */ 1036 if (vstir && force == 0) { 1037 force = 1; 1038 vstir = 0; 1039 } 1040 if (vspace() >= vlowat && force == 0) { 1041 vnlruproc_sig = 0; 1042 wakeup(&vnlruproc_sig); 1043 msleep(vnlruproc, &vnode_free_list_mtx, 1044 PVFS|PDROP, "vlruwt", hz); 1045 continue; 1046 } 1047 mtx_unlock(&vnode_free_list_mtx); 1048 done = 0; 1049 ofreevnodes = freevnodes; 1050 onumvnodes = numvnodes; 1051 /* 1052 * Calculate parameters for recycling. These are the same 1053 * throughout the loop to give some semblance of fairness. 1054 * The trigger point is to avoid recycling vnodes with lots 1055 * of resident pages. We aren't trying to free memory; we 1056 * are trying to recycle or at least free vnodes. 1057 */ 1058 if (numvnodes <= desiredvnodes) 1059 usevnodes = numvnodes - freevnodes; 1060 else 1061 usevnodes = numvnodes; 1062 if (usevnodes <= 0) 1063 usevnodes = 1; 1064 /* 1065 * The trigger value is is chosen to give a conservatively 1066 * large value to ensure that it alone doesn't prevent 1067 * making progress. The value can easily be so large that 1068 * it is effectively infinite in some congested and 1069 * misconfigured cases, and this is necessary. Normally 1070 * it is about 8 to 100 (pages), which is quite large. 1071 */ 1072 trigger = vm_cnt.v_page_count * 2 / usevnodes; 1073 if (force < 2) 1074 trigger = vsmalltrigger; 1075 reclaim_nc_src = force >= 3; 1076 mtx_lock(&mountlist_mtx); 1077 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 1078 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) { 1079 nmp = TAILQ_NEXT(mp, mnt_list); 1080 continue; 1081 } 1082 done += vlrureclaim(mp, reclaim_nc_src, trigger); 1083 mtx_lock(&mountlist_mtx); 1084 nmp = TAILQ_NEXT(mp, mnt_list); 1085 vfs_unbusy(mp); 1086 } 1087 mtx_unlock(&mountlist_mtx); 1088 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes) 1089 uma_reclaim(); 1090 if (done == 0) { 1091 if (force == 0 || force == 1) { 1092 force = 2; 1093 continue; 1094 } 1095 if (force == 2) { 1096 force = 3; 1097 continue; 1098 } 1099 force = 0; 1100 vnlru_nowhere++; 1101 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 1102 } else 1103 kern_yield(PRI_USER); 1104 /* 1105 * After becoming active to expand above low water, keep 1106 * active until above high water. 1107 */ 1108 force = vspace() < vhiwat; 1109 } 1110 } 1111 1112 static struct kproc_desc vnlru_kp = { 1113 "vnlru", 1114 vnlru_proc, 1115 &vnlruproc 1116 }; 1117 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, 1118 &vnlru_kp); 1119 1120 /* 1121 * Routines having to do with the management of the vnode table. 1122 */ 1123 1124 /* 1125 * Try to recycle a freed vnode. We abort if anyone picks up a reference 1126 * before we actually vgone(). This function must be called with the vnode 1127 * held to prevent the vnode from being returned to the free list midway 1128 * through vgone(). 1129 */ 1130 static int 1131 vtryrecycle(struct vnode *vp) 1132 { 1133 struct mount *vnmp; 1134 1135 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 1136 VNASSERT(vp->v_holdcnt, vp, 1137 ("vtryrecycle: Recycling vp %p without a reference.", vp)); 1138 /* 1139 * This vnode may found and locked via some other list, if so we 1140 * can't recycle it yet. 1141 */ 1142 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 1143 CTR2(KTR_VFS, 1144 "%s: impossible to recycle, vp %p lock is already held", 1145 __func__, vp); 1146 return (EWOULDBLOCK); 1147 } 1148 /* 1149 * Don't recycle if its filesystem is being suspended. 1150 */ 1151 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 1152 VOP_UNLOCK(vp, 0); 1153 CTR2(KTR_VFS, 1154 "%s: impossible to recycle, cannot start the write for %p", 1155 __func__, vp); 1156 return (EBUSY); 1157 } 1158 /* 1159 * If we got this far, we need to acquire the interlock and see if 1160 * anyone picked up this vnode from another list. If not, we will 1161 * mark it with DOOMED via vgonel() so that anyone who does find it 1162 * will skip over it. 1163 */ 1164 VI_LOCK(vp); 1165 if (vp->v_usecount) { 1166 VOP_UNLOCK(vp, LK_INTERLOCK); 1167 vn_finished_write(vnmp); 1168 CTR2(KTR_VFS, 1169 "%s: impossible to recycle, %p is already referenced", 1170 __func__, vp); 1171 return (EBUSY); 1172 } 1173 if ((vp->v_iflag & VI_DOOMED) == 0) { 1174 atomic_add_long(&recycles_count, 1); 1175 vgonel(vp); 1176 } 1177 VOP_UNLOCK(vp, LK_INTERLOCK); 1178 vn_finished_write(vnmp); 1179 return (0); 1180 } 1181 1182 static void 1183 vcheckspace(void) 1184 { 1185 1186 if (vspace() < vlowat && vnlruproc_sig == 0) { 1187 vnlruproc_sig = 1; 1188 wakeup(vnlruproc); 1189 } 1190 } 1191 1192 /* 1193 * Wait if necessary for space for a new vnode. 1194 */ 1195 static int 1196 getnewvnode_wait(int suspended) 1197 { 1198 1199 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 1200 if (numvnodes >= desiredvnodes) { 1201 if (suspended) { 1202 /* 1203 * The file system is being suspended. We cannot 1204 * risk a deadlock here, so allow allocation of 1205 * another vnode even if this would give too many. 1206 */ 1207 return (0); 1208 } 1209 if (vnlruproc_sig == 0) { 1210 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 1211 wakeup(vnlruproc); 1212 } 1213 msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS, 1214 "vlruwk", hz); 1215 } 1216 /* Post-adjust like the pre-adjust in getnewvnode(). */ 1217 if (numvnodes + 1 > desiredvnodes && freevnodes > 1) 1218 vnlru_free(1); 1219 return (numvnodes >= desiredvnodes ? ENFILE : 0); 1220 } 1221 1222 /* 1223 * This hack is fragile, and probably not needed any more now that the 1224 * watermark handling works. 1225 */ 1226 void 1227 getnewvnode_reserve(u_int count) 1228 { 1229 struct thread *td; 1230 1231 /* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */ 1232 /* XXX no longer so quick, but this part is not racy. */ 1233 mtx_lock(&vnode_free_list_mtx); 1234 if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes) 1235 vnlru_free(ulmin(numvnodes + count - desiredvnodes, 1236 freevnodes - wantfreevnodes)); 1237 mtx_unlock(&vnode_free_list_mtx); 1238 1239 td = curthread; 1240 /* First try to be quick and racy. */ 1241 if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) { 1242 td->td_vp_reserv += count; 1243 vcheckspace(); /* XXX no longer so quick, but more racy */ 1244 return; 1245 } else 1246 atomic_subtract_long(&numvnodes, count); 1247 1248 mtx_lock(&vnode_free_list_mtx); 1249 while (count > 0) { 1250 if (getnewvnode_wait(0) == 0) { 1251 count--; 1252 td->td_vp_reserv++; 1253 atomic_add_long(&numvnodes, 1); 1254 } 1255 } 1256 vcheckspace(); 1257 mtx_unlock(&vnode_free_list_mtx); 1258 } 1259 1260 /* 1261 * This hack is fragile, especially if desiredvnodes or wantvnodes are 1262 * misconfgured or changed significantly. Reducing desiredvnodes below 1263 * the reserved amount should cause bizarre behaviour like reducing it 1264 * below the number of active vnodes -- the system will try to reduce 1265 * numvnodes to match, but should fail, so the subtraction below should 1266 * not overflow. 1267 */ 1268 void 1269 getnewvnode_drop_reserve(void) 1270 { 1271 struct thread *td; 1272 1273 td = curthread; 1274 atomic_subtract_long(&numvnodes, td->td_vp_reserv); 1275 td->td_vp_reserv = 0; 1276 } 1277 1278 /* 1279 * Return the next vnode from the free list. 1280 */ 1281 int 1282 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, 1283 struct vnode **vpp) 1284 { 1285 struct vnode *vp; 1286 struct thread *td; 1287 struct lock_object *lo; 1288 static int cyclecount; 1289 int error; 1290 1291 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag); 1292 vp = NULL; 1293 td = curthread; 1294 if (td->td_vp_reserv > 0) { 1295 td->td_vp_reserv -= 1; 1296 goto alloc; 1297 } 1298 mtx_lock(&vnode_free_list_mtx); 1299 if (numvnodes < desiredvnodes) 1300 cyclecount = 0; 1301 else if (cyclecount++ >= freevnodes) { 1302 cyclecount = 0; 1303 vstir = 1; 1304 } 1305 /* 1306 * Grow the vnode cache if it will not be above its target max 1307 * after growing. Otherwise, if the free list is nonempty, try 1308 * to reclaim 1 item from it before growing the cache (possibly 1309 * above its target max if the reclamation failed or is delayed). 1310 * Otherwise, wait for some space. In all cases, schedule 1311 * vnlru_proc() if we are getting short of space. The watermarks 1312 * should be chosen so that we never wait or even reclaim from 1313 * the free list to below its target minimum. 1314 */ 1315 if (numvnodes + 1 <= desiredvnodes) 1316 ; 1317 else if (freevnodes > 0) 1318 vnlru_free(1); 1319 else { 1320 error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag & 1321 MNTK_SUSPEND)); 1322 #if 0 /* XXX Not all VFS_VGET/ffs_vget callers check returns. */ 1323 if (error != 0) { 1324 mtx_unlock(&vnode_free_list_mtx); 1325 return (error); 1326 } 1327 #endif 1328 } 1329 vcheckspace(); 1330 atomic_add_long(&numvnodes, 1); 1331 mtx_unlock(&vnode_free_list_mtx); 1332 alloc: 1333 atomic_add_long(&vnodes_created, 1); 1334 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK); 1335 /* 1336 * Locks are given the generic name "vnode" when created. 1337 * Follow the historic practice of using the filesystem 1338 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc. 1339 * 1340 * Locks live in a witness group keyed on their name. Thus, 1341 * when a lock is renamed, it must also move from the witness 1342 * group of its old name to the witness group of its new name. 1343 * 1344 * The change only needs to be made when the vnode moves 1345 * from one filesystem type to another. We ensure that each 1346 * filesystem use a single static name pointer for its tag so 1347 * that we can compare pointers rather than doing a strcmp(). 1348 */ 1349 lo = &vp->v_vnlock->lock_object; 1350 if (lo->lo_name != tag) { 1351 lo->lo_name = tag; 1352 WITNESS_DESTROY(lo); 1353 WITNESS_INIT(lo, tag); 1354 } 1355 /* 1356 * By default, don't allow shared locks unless filesystems opt-in. 1357 */ 1358 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE; 1359 /* 1360 * Finalize various vnode identity bits. 1361 */ 1362 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp)); 1363 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp)); 1364 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp)); 1365 vp->v_type = VNON; 1366 vp->v_tag = tag; 1367 vp->v_op = vops; 1368 v_init_counters(vp); 1369 vp->v_bufobj.bo_ops = &buf_ops_bio; 1370 #ifdef MAC 1371 mac_vnode_init(vp); 1372 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 1373 mac_vnode_associate_singlelabel(mp, vp); 1374 else if (mp == NULL && vops != &dead_vnodeops) 1375 printf("NULL mp in getnewvnode()\n"); 1376 #endif 1377 if (mp != NULL) { 1378 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize; 1379 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) 1380 vp->v_vflag |= VV_NOKNOTE; 1381 } 1382 1383 /* 1384 * For the filesystems which do not use vfs_hash_insert(), 1385 * still initialize v_hash to have vfs_hash_index() useful. 1386 * E.g., nullfs uses vfs_hash_index() on the lower vnode for 1387 * its own hashing. 1388 */ 1389 vp->v_hash = (uintptr_t)vp >> vnsz2log; 1390 1391 *vpp = vp; 1392 return (0); 1393 } 1394 1395 /* 1396 * Delete from old mount point vnode list, if on one. 1397 */ 1398 static void 1399 delmntque(struct vnode *vp) 1400 { 1401 struct mount *mp; 1402 int active; 1403 1404 mp = vp->v_mount; 1405 if (mp == NULL) 1406 return; 1407 MNT_ILOCK(mp); 1408 VI_LOCK(vp); 1409 KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize, 1410 ("Active vnode list size %d > Vnode list size %d", 1411 mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize)); 1412 active = vp->v_iflag & VI_ACTIVE; 1413 vp->v_iflag &= ~VI_ACTIVE; 1414 if (active) { 1415 mtx_lock(&vnode_free_list_mtx); 1416 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist); 1417 mp->mnt_activevnodelistsize--; 1418 mtx_unlock(&vnode_free_list_mtx); 1419 } 1420 vp->v_mount = NULL; 1421 VI_UNLOCK(vp); 1422 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 1423 ("bad mount point vnode list size")); 1424 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1425 mp->mnt_nvnodelistsize--; 1426 MNT_REL(mp); 1427 MNT_IUNLOCK(mp); 1428 } 1429 1430 static void 1431 insmntque_stddtr(struct vnode *vp, void *dtr_arg) 1432 { 1433 1434 vp->v_data = NULL; 1435 vp->v_op = &dead_vnodeops; 1436 vgone(vp); 1437 vput(vp); 1438 } 1439 1440 /* 1441 * Insert into list of vnodes for the new mount point, if available. 1442 */ 1443 int 1444 insmntque1(struct vnode *vp, struct mount *mp, 1445 void (*dtr)(struct vnode *, void *), void *dtr_arg) 1446 { 1447 1448 KASSERT(vp->v_mount == NULL, 1449 ("insmntque: vnode already on per mount vnode list")); 1450 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 1451 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); 1452 1453 /* 1454 * We acquire the vnode interlock early to ensure that the 1455 * vnode cannot be recycled by another process releasing a 1456 * holdcnt on it before we get it on both the vnode list 1457 * and the active vnode list. The mount mutex protects only 1458 * manipulation of the vnode list and the vnode freelist 1459 * mutex protects only manipulation of the active vnode list. 1460 * Hence the need to hold the vnode interlock throughout. 1461 */ 1462 MNT_ILOCK(mp); 1463 VI_LOCK(vp); 1464 if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 && 1465 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || 1466 mp->mnt_nvnodelistsize == 0)) && 1467 (vp->v_vflag & VV_FORCEINSMQ) == 0) { 1468 VI_UNLOCK(vp); 1469 MNT_IUNLOCK(mp); 1470 if (dtr != NULL) 1471 dtr(vp, dtr_arg); 1472 return (EBUSY); 1473 } 1474 vp->v_mount = mp; 1475 MNT_REF(mp); 1476 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1477 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 1478 ("neg mount point vnode list size")); 1479 mp->mnt_nvnodelistsize++; 1480 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 1481 ("Activating already active vnode")); 1482 vp->v_iflag |= VI_ACTIVE; 1483 mtx_lock(&vnode_free_list_mtx); 1484 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist); 1485 mp->mnt_activevnodelistsize++; 1486 mtx_unlock(&vnode_free_list_mtx); 1487 VI_UNLOCK(vp); 1488 MNT_IUNLOCK(mp); 1489 return (0); 1490 } 1491 1492 int 1493 insmntque(struct vnode *vp, struct mount *mp) 1494 { 1495 1496 return (insmntque1(vp, mp, insmntque_stddtr, NULL)); 1497 } 1498 1499 /* 1500 * Flush out and invalidate all buffers associated with a bufobj 1501 * Called with the underlying object locked. 1502 */ 1503 int 1504 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) 1505 { 1506 int error; 1507 1508 BO_LOCK(bo); 1509 if (flags & V_SAVE) { 1510 error = bufobj_wwait(bo, slpflag, slptimeo); 1511 if (error) { 1512 BO_UNLOCK(bo); 1513 return (error); 1514 } 1515 if (bo->bo_dirty.bv_cnt > 0) { 1516 BO_UNLOCK(bo); 1517 if ((error = BO_SYNC(bo, MNT_WAIT)) != 0) 1518 return (error); 1519 /* 1520 * XXX We could save a lock/unlock if this was only 1521 * enabled under INVARIANTS 1522 */ 1523 BO_LOCK(bo); 1524 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 1525 panic("vinvalbuf: dirty bufs"); 1526 } 1527 } 1528 /* 1529 * If you alter this loop please notice that interlock is dropped and 1530 * reacquired in flushbuflist. Special care is needed to ensure that 1531 * no race conditions occur from this. 1532 */ 1533 do { 1534 error = flushbuflist(&bo->bo_clean, 1535 flags, bo, slpflag, slptimeo); 1536 if (error == 0 && !(flags & V_CLEANONLY)) 1537 error = flushbuflist(&bo->bo_dirty, 1538 flags, bo, slpflag, slptimeo); 1539 if (error != 0 && error != EAGAIN) { 1540 BO_UNLOCK(bo); 1541 return (error); 1542 } 1543 } while (error != 0); 1544 1545 /* 1546 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1547 * have write I/O in-progress but if there is a VM object then the 1548 * VM object can also have read-I/O in-progress. 1549 */ 1550 do { 1551 bufobj_wwait(bo, 0, 0); 1552 BO_UNLOCK(bo); 1553 if (bo->bo_object != NULL) { 1554 VM_OBJECT_WLOCK(bo->bo_object); 1555 vm_object_pip_wait(bo->bo_object, "bovlbx"); 1556 VM_OBJECT_WUNLOCK(bo->bo_object); 1557 } 1558 BO_LOCK(bo); 1559 } while (bo->bo_numoutput > 0); 1560 BO_UNLOCK(bo); 1561 1562 /* 1563 * Destroy the copy in the VM cache, too. 1564 */ 1565 if (bo->bo_object != NULL && 1566 (flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0) { 1567 VM_OBJECT_WLOCK(bo->bo_object); 1568 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? 1569 OBJPR_CLEANONLY : 0); 1570 VM_OBJECT_WUNLOCK(bo->bo_object); 1571 } 1572 1573 #ifdef INVARIANTS 1574 BO_LOCK(bo); 1575 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0 && 1576 (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0)) 1577 panic("vinvalbuf: flush failed"); 1578 BO_UNLOCK(bo); 1579 #endif 1580 return (0); 1581 } 1582 1583 /* 1584 * Flush out and invalidate all buffers associated with a vnode. 1585 * Called with the underlying object locked. 1586 */ 1587 int 1588 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 1589 { 1590 1591 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 1592 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1593 if (vp->v_object != NULL && vp->v_object->handle != vp) 1594 return (0); 1595 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); 1596 } 1597 1598 /* 1599 * Flush out buffers on the specified list. 1600 * 1601 */ 1602 static int 1603 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 1604 int slptimeo) 1605 { 1606 struct buf *bp, *nbp; 1607 int retval, error; 1608 daddr_t lblkno; 1609 b_xflags_t xflags; 1610 1611 ASSERT_BO_WLOCKED(bo); 1612 1613 retval = 0; 1614 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 1615 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1616 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1617 continue; 1618 } 1619 lblkno = 0; 1620 xflags = 0; 1621 if (nbp != NULL) { 1622 lblkno = nbp->b_lblkno; 1623 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN); 1624 } 1625 retval = EAGAIN; 1626 error = BUF_TIMELOCK(bp, 1627 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), 1628 "flushbuf", slpflag, slptimeo); 1629 if (error) { 1630 BO_LOCK(bo); 1631 return (error != ENOLCK ? error : EAGAIN); 1632 } 1633 KASSERT(bp->b_bufobj == bo, 1634 ("bp %p wrong b_bufobj %p should be %p", 1635 bp, bp->b_bufobj, bo)); 1636 /* 1637 * XXX Since there are no node locks for NFS, I 1638 * believe there is a slight chance that a delayed 1639 * write will occur while sleeping just above, so 1640 * check for it. 1641 */ 1642 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1643 (flags & V_SAVE)) { 1644 bremfree(bp); 1645 bp->b_flags |= B_ASYNC; 1646 bwrite(bp); 1647 BO_LOCK(bo); 1648 return (EAGAIN); /* XXX: why not loop ? */ 1649 } 1650 bremfree(bp); 1651 bp->b_flags |= (B_INVAL | B_RELBUF); 1652 bp->b_flags &= ~B_ASYNC; 1653 brelse(bp); 1654 BO_LOCK(bo); 1655 nbp = gbincore(bo, lblkno); 1656 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1657 != xflags) 1658 break; /* nbp invalid */ 1659 } 1660 return (retval); 1661 } 1662 1663 int 1664 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn) 1665 { 1666 struct buf *bp; 1667 int error; 1668 daddr_t lblkno; 1669 1670 ASSERT_BO_LOCKED(bo); 1671 1672 for (lblkno = startn;;) { 1673 again: 1674 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno); 1675 if (bp == NULL || bp->b_lblkno >= endn) 1676 break; 1677 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 1678 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0); 1679 if (error != 0) { 1680 BO_RLOCK(bo); 1681 if (error == ENOLCK) 1682 goto again; 1683 return (error); 1684 } 1685 KASSERT(bp->b_bufobj == bo, 1686 ("bp %p wrong b_bufobj %p should be %p", 1687 bp, bp->b_bufobj, bo)); 1688 lblkno = bp->b_lblkno + 1; 1689 if ((bp->b_flags & B_MANAGED) == 0) 1690 bremfree(bp); 1691 bp->b_flags |= B_RELBUF; 1692 /* 1693 * In the VMIO case, use the B_NOREUSE flag to hint that the 1694 * pages backing each buffer in the range are unlikely to be 1695 * reused. Dirty buffers will have the hint applied once 1696 * they've been written. 1697 */ 1698 if (bp->b_vp->v_object != NULL) 1699 bp->b_flags |= B_NOREUSE; 1700 brelse(bp); 1701 BO_RLOCK(bo); 1702 } 1703 return (0); 1704 } 1705 1706 /* 1707 * Truncate a file's buffer and pages to a specified length. This 1708 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1709 * sync activity. 1710 */ 1711 int 1712 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize) 1713 { 1714 struct buf *bp, *nbp; 1715 int anyfreed; 1716 int trunclbn; 1717 struct bufobj *bo; 1718 1719 CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__, 1720 vp, cred, blksize, (uintmax_t)length); 1721 1722 /* 1723 * Round up to the *next* lbn. 1724 */ 1725 trunclbn = (length + blksize - 1) / blksize; 1726 1727 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1728 restart: 1729 bo = &vp->v_bufobj; 1730 BO_LOCK(bo); 1731 anyfreed = 1; 1732 for (;anyfreed;) { 1733 anyfreed = 0; 1734 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 1735 if (bp->b_lblkno < trunclbn) 1736 continue; 1737 if (BUF_LOCK(bp, 1738 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1739 BO_LOCKPTR(bo)) == ENOLCK) 1740 goto restart; 1741 1742 bremfree(bp); 1743 bp->b_flags |= (B_INVAL | B_RELBUF); 1744 bp->b_flags &= ~B_ASYNC; 1745 brelse(bp); 1746 anyfreed = 1; 1747 1748 BO_LOCK(bo); 1749 if (nbp != NULL && 1750 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1751 (nbp->b_vp != vp) || 1752 (nbp->b_flags & B_DELWRI))) { 1753 BO_UNLOCK(bo); 1754 goto restart; 1755 } 1756 } 1757 1758 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1759 if (bp->b_lblkno < trunclbn) 1760 continue; 1761 if (BUF_LOCK(bp, 1762 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1763 BO_LOCKPTR(bo)) == ENOLCK) 1764 goto restart; 1765 bremfree(bp); 1766 bp->b_flags |= (B_INVAL | B_RELBUF); 1767 bp->b_flags &= ~B_ASYNC; 1768 brelse(bp); 1769 anyfreed = 1; 1770 1771 BO_LOCK(bo); 1772 if (nbp != NULL && 1773 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1774 (nbp->b_vp != vp) || 1775 (nbp->b_flags & B_DELWRI) == 0)) { 1776 BO_UNLOCK(bo); 1777 goto restart; 1778 } 1779 } 1780 } 1781 1782 if (length > 0) { 1783 restartsync: 1784 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1785 if (bp->b_lblkno > 0) 1786 continue; 1787 /* 1788 * Since we hold the vnode lock this should only 1789 * fail if we're racing with the buf daemon. 1790 */ 1791 if (BUF_LOCK(bp, 1792 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1793 BO_LOCKPTR(bo)) == ENOLCK) { 1794 goto restart; 1795 } 1796 VNASSERT((bp->b_flags & B_DELWRI), vp, 1797 ("buf(%p) on dirty queue without DELWRI", bp)); 1798 1799 bremfree(bp); 1800 bawrite(bp); 1801 BO_LOCK(bo); 1802 goto restartsync; 1803 } 1804 } 1805 1806 bufobj_wwait(bo, 0, 0); 1807 BO_UNLOCK(bo); 1808 vnode_pager_setsize(vp, length); 1809 1810 return (0); 1811 } 1812 1813 static void 1814 buf_vlist_remove(struct buf *bp) 1815 { 1816 struct bufv *bv; 1817 1818 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1819 ASSERT_BO_WLOCKED(bp->b_bufobj); 1820 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) != 1821 (BX_VNDIRTY|BX_VNCLEAN), 1822 ("buf_vlist_remove: Buf %p is on two lists", bp)); 1823 if (bp->b_xflags & BX_VNDIRTY) 1824 bv = &bp->b_bufobj->bo_dirty; 1825 else 1826 bv = &bp->b_bufobj->bo_clean; 1827 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno); 1828 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 1829 bv->bv_cnt--; 1830 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1831 } 1832 1833 /* 1834 * Add the buffer to the sorted clean or dirty block list. 1835 * 1836 * NOTE: xflags is passed as a constant, optimizing this inline function! 1837 */ 1838 static void 1839 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 1840 { 1841 struct bufv *bv; 1842 struct buf *n; 1843 int error; 1844 1845 ASSERT_BO_WLOCKED(bo); 1846 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0, 1847 ("dead bo %p", bo)); 1848 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1849 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 1850 bp->b_xflags |= xflags; 1851 if (xflags & BX_VNDIRTY) 1852 bv = &bo->bo_dirty; 1853 else 1854 bv = &bo->bo_clean; 1855 1856 /* 1857 * Keep the list ordered. Optimize empty list insertion. Assume 1858 * we tend to grow at the tail so lookup_le should usually be cheaper 1859 * than _ge. 1860 */ 1861 if (bv->bv_cnt == 0 || 1862 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno) 1863 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 1864 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL) 1865 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs); 1866 else 1867 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs); 1868 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp); 1869 if (error) 1870 panic("buf_vlist_add: Preallocated nodes insufficient."); 1871 bv->bv_cnt++; 1872 } 1873 1874 /* 1875 * Look up a buffer using the buffer tries. 1876 */ 1877 struct buf * 1878 gbincore(struct bufobj *bo, daddr_t lblkno) 1879 { 1880 struct buf *bp; 1881 1882 ASSERT_BO_LOCKED(bo); 1883 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno); 1884 if (bp != NULL) 1885 return (bp); 1886 return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno); 1887 } 1888 1889 /* 1890 * Associate a buffer with a vnode. 1891 */ 1892 void 1893 bgetvp(struct vnode *vp, struct buf *bp) 1894 { 1895 struct bufobj *bo; 1896 1897 bo = &vp->v_bufobj; 1898 ASSERT_BO_WLOCKED(bo); 1899 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 1900 1901 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 1902 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 1903 ("bgetvp: bp already attached! %p", bp)); 1904 1905 vhold(vp); 1906 bp->b_vp = vp; 1907 bp->b_bufobj = bo; 1908 /* 1909 * Insert onto list for new vnode. 1910 */ 1911 buf_vlist_add(bp, bo, BX_VNCLEAN); 1912 } 1913 1914 /* 1915 * Disassociate a buffer from a vnode. 1916 */ 1917 void 1918 brelvp(struct buf *bp) 1919 { 1920 struct bufobj *bo; 1921 struct vnode *vp; 1922 1923 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1924 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1925 1926 /* 1927 * Delete from old vnode list, if on one. 1928 */ 1929 vp = bp->b_vp; /* XXX */ 1930 bo = bp->b_bufobj; 1931 BO_LOCK(bo); 1932 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1933 buf_vlist_remove(bp); 1934 else 1935 panic("brelvp: Buffer %p not on queue.", bp); 1936 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1937 bo->bo_flag &= ~BO_ONWORKLST; 1938 mtx_lock(&sync_mtx); 1939 LIST_REMOVE(bo, bo_synclist); 1940 syncer_worklist_len--; 1941 mtx_unlock(&sync_mtx); 1942 } 1943 bp->b_vp = NULL; 1944 bp->b_bufobj = NULL; 1945 BO_UNLOCK(bo); 1946 vdrop(vp); 1947 } 1948 1949 /* 1950 * Add an item to the syncer work queue. 1951 */ 1952 static void 1953 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 1954 { 1955 int slot; 1956 1957 ASSERT_BO_WLOCKED(bo); 1958 1959 mtx_lock(&sync_mtx); 1960 if (bo->bo_flag & BO_ONWORKLST) 1961 LIST_REMOVE(bo, bo_synclist); 1962 else { 1963 bo->bo_flag |= BO_ONWORKLST; 1964 syncer_worklist_len++; 1965 } 1966 1967 if (delay > syncer_maxdelay - 2) 1968 delay = syncer_maxdelay - 2; 1969 slot = (syncer_delayno + delay) & syncer_mask; 1970 1971 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 1972 mtx_unlock(&sync_mtx); 1973 } 1974 1975 static int 1976 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 1977 { 1978 int error, len; 1979 1980 mtx_lock(&sync_mtx); 1981 len = syncer_worklist_len - sync_vnode_count; 1982 mtx_unlock(&sync_mtx); 1983 error = SYSCTL_OUT(req, &len, sizeof(len)); 1984 return (error); 1985 } 1986 1987 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0, 1988 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 1989 1990 static struct proc *updateproc; 1991 static void sched_sync(void); 1992 static struct kproc_desc up_kp = { 1993 "syncer", 1994 sched_sync, 1995 &updateproc 1996 }; 1997 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); 1998 1999 static int 2000 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) 2001 { 2002 struct vnode *vp; 2003 struct mount *mp; 2004 2005 *bo = LIST_FIRST(slp); 2006 if (*bo == NULL) 2007 return (0); 2008 vp = (*bo)->__bo_vnode; /* XXX */ 2009 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) 2010 return (1); 2011 /* 2012 * We use vhold in case the vnode does not 2013 * successfully sync. vhold prevents the vnode from 2014 * going away when we unlock the sync_mtx so that 2015 * we can acquire the vnode interlock. 2016 */ 2017 vholdl(vp); 2018 mtx_unlock(&sync_mtx); 2019 VI_UNLOCK(vp); 2020 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2021 vdrop(vp); 2022 mtx_lock(&sync_mtx); 2023 return (*bo == LIST_FIRST(slp)); 2024 } 2025 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2026 (void) VOP_FSYNC(vp, MNT_LAZY, td); 2027 VOP_UNLOCK(vp, 0); 2028 vn_finished_write(mp); 2029 BO_LOCK(*bo); 2030 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { 2031 /* 2032 * Put us back on the worklist. The worklist 2033 * routine will remove us from our current 2034 * position and then add us back in at a later 2035 * position. 2036 */ 2037 vn_syncer_add_to_worklist(*bo, syncdelay); 2038 } 2039 BO_UNLOCK(*bo); 2040 vdrop(vp); 2041 mtx_lock(&sync_mtx); 2042 return (0); 2043 } 2044 2045 static int first_printf = 1; 2046 2047 /* 2048 * System filesystem synchronizer daemon. 2049 */ 2050 static void 2051 sched_sync(void) 2052 { 2053 struct synclist *next, *slp; 2054 struct bufobj *bo; 2055 long starttime; 2056 struct thread *td = curthread; 2057 int last_work_seen; 2058 int net_worklist_len; 2059 int syncer_final_iter; 2060 int error; 2061 2062 last_work_seen = 0; 2063 syncer_final_iter = 0; 2064 syncer_state = SYNCER_RUNNING; 2065 starttime = time_uptime; 2066 td->td_pflags |= TDP_NORUNNINGBUF; 2067 2068 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 2069 SHUTDOWN_PRI_LAST); 2070 2071 mtx_lock(&sync_mtx); 2072 for (;;) { 2073 if (syncer_state == SYNCER_FINAL_DELAY && 2074 syncer_final_iter == 0) { 2075 mtx_unlock(&sync_mtx); 2076 kproc_suspend_check(td->td_proc); 2077 mtx_lock(&sync_mtx); 2078 } 2079 net_worklist_len = syncer_worklist_len - sync_vnode_count; 2080 if (syncer_state != SYNCER_RUNNING && 2081 starttime != time_uptime) { 2082 if (first_printf) { 2083 printf("\nSyncing disks, vnodes remaining..."); 2084 first_printf = 0; 2085 } 2086 printf("%d ", net_worklist_len); 2087 } 2088 starttime = time_uptime; 2089 2090 /* 2091 * Push files whose dirty time has expired. Be careful 2092 * of interrupt race on slp queue. 2093 * 2094 * Skip over empty worklist slots when shutting down. 2095 */ 2096 do { 2097 slp = &syncer_workitem_pending[syncer_delayno]; 2098 syncer_delayno += 1; 2099 if (syncer_delayno == syncer_maxdelay) 2100 syncer_delayno = 0; 2101 next = &syncer_workitem_pending[syncer_delayno]; 2102 /* 2103 * If the worklist has wrapped since the 2104 * it was emptied of all but syncer vnodes, 2105 * switch to the FINAL_DELAY state and run 2106 * for one more second. 2107 */ 2108 if (syncer_state == SYNCER_SHUTTING_DOWN && 2109 net_worklist_len == 0 && 2110 last_work_seen == syncer_delayno) { 2111 syncer_state = SYNCER_FINAL_DELAY; 2112 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 2113 } 2114 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 2115 syncer_worklist_len > 0); 2116 2117 /* 2118 * Keep track of the last time there was anything 2119 * on the worklist other than syncer vnodes. 2120 * Return to the SHUTTING_DOWN state if any 2121 * new work appears. 2122 */ 2123 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 2124 last_work_seen = syncer_delayno; 2125 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 2126 syncer_state = SYNCER_SHUTTING_DOWN; 2127 while (!LIST_EMPTY(slp)) { 2128 error = sync_vnode(slp, &bo, td); 2129 if (error == 1) { 2130 LIST_REMOVE(bo, bo_synclist); 2131 LIST_INSERT_HEAD(next, bo, bo_synclist); 2132 continue; 2133 } 2134 2135 if (first_printf == 0) { 2136 /* 2137 * Drop the sync mutex, because some watchdog 2138 * drivers need to sleep while patting 2139 */ 2140 mtx_unlock(&sync_mtx); 2141 wdog_kern_pat(WD_LASTVAL); 2142 mtx_lock(&sync_mtx); 2143 } 2144 2145 } 2146 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 2147 syncer_final_iter--; 2148 /* 2149 * The variable rushjob allows the kernel to speed up the 2150 * processing of the filesystem syncer process. A rushjob 2151 * value of N tells the filesystem syncer to process the next 2152 * N seconds worth of work on its queue ASAP. Currently rushjob 2153 * is used by the soft update code to speed up the filesystem 2154 * syncer process when the incore state is getting so far 2155 * ahead of the disk that the kernel memory pool is being 2156 * threatened with exhaustion. 2157 */ 2158 if (rushjob > 0) { 2159 rushjob -= 1; 2160 continue; 2161 } 2162 /* 2163 * Just sleep for a short period of time between 2164 * iterations when shutting down to allow some I/O 2165 * to happen. 2166 * 2167 * If it has taken us less than a second to process the 2168 * current work, then wait. Otherwise start right over 2169 * again. We can still lose time if any single round 2170 * takes more than two seconds, but it does not really 2171 * matter as we are just trying to generally pace the 2172 * filesystem activity. 2173 */ 2174 if (syncer_state != SYNCER_RUNNING || 2175 time_uptime == starttime) { 2176 thread_lock(td); 2177 sched_prio(td, PPAUSE); 2178 thread_unlock(td); 2179 } 2180 if (syncer_state != SYNCER_RUNNING) 2181 cv_timedwait(&sync_wakeup, &sync_mtx, 2182 hz / SYNCER_SHUTDOWN_SPEEDUP); 2183 else if (time_uptime == starttime) 2184 cv_timedwait(&sync_wakeup, &sync_mtx, hz); 2185 } 2186 } 2187 2188 /* 2189 * Request the syncer daemon to speed up its work. 2190 * We never push it to speed up more than half of its 2191 * normal turn time, otherwise it could take over the cpu. 2192 */ 2193 int 2194 speedup_syncer(void) 2195 { 2196 int ret = 0; 2197 2198 mtx_lock(&sync_mtx); 2199 if (rushjob < syncdelay / 2) { 2200 rushjob += 1; 2201 stat_rush_requests += 1; 2202 ret = 1; 2203 } 2204 mtx_unlock(&sync_mtx); 2205 cv_broadcast(&sync_wakeup); 2206 return (ret); 2207 } 2208 2209 /* 2210 * Tell the syncer to speed up its work and run though its work 2211 * list several times, then tell it to shut down. 2212 */ 2213 static void 2214 syncer_shutdown(void *arg, int howto) 2215 { 2216 2217 if (howto & RB_NOSYNC) 2218 return; 2219 mtx_lock(&sync_mtx); 2220 syncer_state = SYNCER_SHUTTING_DOWN; 2221 rushjob = 0; 2222 mtx_unlock(&sync_mtx); 2223 cv_broadcast(&sync_wakeup); 2224 kproc_shutdown(arg, howto); 2225 } 2226 2227 void 2228 syncer_suspend(void) 2229 { 2230 2231 syncer_shutdown(updateproc, 0); 2232 } 2233 2234 void 2235 syncer_resume(void) 2236 { 2237 2238 mtx_lock(&sync_mtx); 2239 first_printf = 1; 2240 syncer_state = SYNCER_RUNNING; 2241 mtx_unlock(&sync_mtx); 2242 cv_broadcast(&sync_wakeup); 2243 kproc_resume(updateproc); 2244 } 2245 2246 /* 2247 * Reassign a buffer from one vnode to another. 2248 * Used to assign file specific control information 2249 * (indirect blocks) to the vnode to which they belong. 2250 */ 2251 void 2252 reassignbuf(struct buf *bp) 2253 { 2254 struct vnode *vp; 2255 struct bufobj *bo; 2256 int delay; 2257 #ifdef INVARIANTS 2258 struct bufv *bv; 2259 #endif 2260 2261 vp = bp->b_vp; 2262 bo = bp->b_bufobj; 2263 ++reassignbufcalls; 2264 2265 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 2266 bp, bp->b_vp, bp->b_flags); 2267 /* 2268 * B_PAGING flagged buffers cannot be reassigned because their vp 2269 * is not fully linked in. 2270 */ 2271 if (bp->b_flags & B_PAGING) 2272 panic("cannot reassign paging buffer"); 2273 2274 /* 2275 * Delete from old vnode list, if on one. 2276 */ 2277 BO_LOCK(bo); 2278 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2279 buf_vlist_remove(bp); 2280 else 2281 panic("reassignbuf: Buffer %p not on queue.", bp); 2282 /* 2283 * If dirty, put on list of dirty buffers; otherwise insert onto list 2284 * of clean buffers. 2285 */ 2286 if (bp->b_flags & B_DELWRI) { 2287 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 2288 switch (vp->v_type) { 2289 case VDIR: 2290 delay = dirdelay; 2291 break; 2292 case VCHR: 2293 delay = metadelay; 2294 break; 2295 default: 2296 delay = filedelay; 2297 } 2298 vn_syncer_add_to_worklist(bo, delay); 2299 } 2300 buf_vlist_add(bp, bo, BX_VNDIRTY); 2301 } else { 2302 buf_vlist_add(bp, bo, BX_VNCLEAN); 2303 2304 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2305 mtx_lock(&sync_mtx); 2306 LIST_REMOVE(bo, bo_synclist); 2307 syncer_worklist_len--; 2308 mtx_unlock(&sync_mtx); 2309 bo->bo_flag &= ~BO_ONWORKLST; 2310 } 2311 } 2312 #ifdef INVARIANTS 2313 bv = &bo->bo_clean; 2314 bp = TAILQ_FIRST(&bv->bv_hd); 2315 KASSERT(bp == NULL || bp->b_bufobj == bo, 2316 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2317 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2318 KASSERT(bp == NULL || bp->b_bufobj == bo, 2319 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2320 bv = &bo->bo_dirty; 2321 bp = TAILQ_FIRST(&bv->bv_hd); 2322 KASSERT(bp == NULL || bp->b_bufobj == bo, 2323 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2324 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2325 KASSERT(bp == NULL || bp->b_bufobj == bo, 2326 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2327 #endif 2328 BO_UNLOCK(bo); 2329 } 2330 2331 /* 2332 * A temporary hack until refcount_* APIs are sorted out. 2333 */ 2334 static __inline int 2335 vfs_refcount_acquire_if_not_zero(volatile u_int *count) 2336 { 2337 u_int old; 2338 2339 for (;;) { 2340 old = *count; 2341 if (old == 0) 2342 return (0); 2343 if (atomic_cmpset_int(count, old, old + 1)) 2344 return (1); 2345 } 2346 } 2347 2348 static __inline int 2349 vfs_refcount_release_if_not_last(volatile u_int *count) 2350 { 2351 u_int old; 2352 2353 for (;;) { 2354 old = *count; 2355 if (old == 1) 2356 return (0); 2357 if (atomic_cmpset_int(count, old, old - 1)) 2358 return (1); 2359 } 2360 } 2361 2362 static void 2363 v_init_counters(struct vnode *vp) 2364 { 2365 2366 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0, 2367 vp, ("%s called for an initialized vnode", __FUNCTION__)); 2368 ASSERT_VI_UNLOCKED(vp, __FUNCTION__); 2369 2370 refcount_init(&vp->v_holdcnt, 1); 2371 refcount_init(&vp->v_usecount, 1); 2372 } 2373 2374 /* 2375 * Increment the use and hold counts on the vnode, taking care to reference 2376 * the driver's usecount if this is a chardev. The _vhold() will remove 2377 * the vnode from the free list if it is presently free. 2378 */ 2379 static void 2380 v_incr_usecount(struct vnode *vp) 2381 { 2382 2383 ASSERT_VI_UNLOCKED(vp, __func__); 2384 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2385 2386 if (vp->v_type == VCHR) { 2387 VI_LOCK(vp); 2388 _vhold(vp, true); 2389 if (vp->v_iflag & VI_OWEINACT) { 2390 VNASSERT(vp->v_usecount == 0, vp, 2391 ("vnode with usecount and VI_OWEINACT set")); 2392 vp->v_iflag &= ~VI_OWEINACT; 2393 } 2394 refcount_acquire(&vp->v_usecount); 2395 v_incr_devcount(vp); 2396 VI_UNLOCK(vp); 2397 return; 2398 } 2399 2400 _vhold(vp, false); 2401 if (vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) { 2402 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2403 ("vnode with usecount and VI_OWEINACT set")); 2404 } else { 2405 VI_LOCK(vp); 2406 if (vp->v_iflag & VI_OWEINACT) 2407 vp->v_iflag &= ~VI_OWEINACT; 2408 refcount_acquire(&vp->v_usecount); 2409 VI_UNLOCK(vp); 2410 } 2411 } 2412 2413 /* 2414 * Increment si_usecount of the associated device, if any. 2415 */ 2416 static void 2417 v_incr_devcount(struct vnode *vp) 2418 { 2419 2420 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2421 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2422 dev_lock(); 2423 vp->v_rdev->si_usecount++; 2424 dev_unlock(); 2425 } 2426 } 2427 2428 /* 2429 * Decrement si_usecount of the associated device, if any. 2430 */ 2431 static void 2432 v_decr_devcount(struct vnode *vp) 2433 { 2434 2435 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2436 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2437 dev_lock(); 2438 vp->v_rdev->si_usecount--; 2439 dev_unlock(); 2440 } 2441 } 2442 2443 /* 2444 * Grab a particular vnode from the free list, increment its 2445 * reference count and lock it. VI_DOOMED is set if the vnode 2446 * is being destroyed. Only callers who specify LK_RETRY will 2447 * see doomed vnodes. If inactive processing was delayed in 2448 * vput try to do it here. 2449 * 2450 * Notes on lockless counter manipulation: 2451 * _vhold, vputx and other routines make various decisions based 2452 * on either holdcnt or usecount being 0. As long as either contuner 2453 * is not transitioning 0->1 nor 1->0, the manipulation can be done 2454 * with atomic operations. Otherwise the interlock is taken. 2455 */ 2456 int 2457 vget(struct vnode *vp, int flags, struct thread *td) 2458 { 2459 int error, oweinact; 2460 2461 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 2462 ("vget: invalid lock operation")); 2463 2464 if ((flags & LK_INTERLOCK) != 0) 2465 ASSERT_VI_LOCKED(vp, __func__); 2466 else 2467 ASSERT_VI_UNLOCKED(vp, __func__); 2468 if ((flags & LK_VNHELD) != 0) 2469 VNASSERT((vp->v_holdcnt > 0), vp, 2470 ("vget: LK_VNHELD passed but vnode not held")); 2471 2472 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 2473 2474 if ((flags & LK_VNHELD) == 0) 2475 _vhold(vp, (flags & LK_INTERLOCK) != 0); 2476 2477 if ((error = vn_lock(vp, flags)) != 0) { 2478 vdrop(vp); 2479 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, 2480 vp); 2481 return (error); 2482 } 2483 if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0) 2484 panic("vget: vn_lock failed to return ENOENT\n"); 2485 /* 2486 * We don't guarantee that any particular close will 2487 * trigger inactive processing so just make a best effort 2488 * here at preventing a reference to a removed file. If 2489 * we don't succeed no harm is done. 2490 * 2491 * Upgrade our holdcnt to a usecount. 2492 */ 2493 if (vp->v_type != VCHR && 2494 vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) { 2495 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2496 ("vnode with usecount and VI_OWEINACT set")); 2497 } else { 2498 VI_LOCK(vp); 2499 if ((vp->v_iflag & VI_OWEINACT) == 0) { 2500 oweinact = 0; 2501 } else { 2502 oweinact = 1; 2503 vp->v_iflag &= ~VI_OWEINACT; 2504 } 2505 refcount_acquire(&vp->v_usecount); 2506 v_incr_devcount(vp); 2507 if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE && 2508 (flags & LK_NOWAIT) == 0) 2509 vinactive(vp, td); 2510 VI_UNLOCK(vp); 2511 } 2512 return (0); 2513 } 2514 2515 /* 2516 * Increase the reference count of a vnode. 2517 */ 2518 void 2519 vref(struct vnode *vp) 2520 { 2521 2522 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2523 v_incr_usecount(vp); 2524 } 2525 2526 /* 2527 * Return reference count of a vnode. 2528 * 2529 * The results of this call are only guaranteed when some mechanism is used to 2530 * stop other processes from gaining references to the vnode. This may be the 2531 * case if the caller holds the only reference. This is also useful when stale 2532 * data is acceptable as race conditions may be accounted for by some other 2533 * means. 2534 */ 2535 int 2536 vrefcnt(struct vnode *vp) 2537 { 2538 2539 return (vp->v_usecount); 2540 } 2541 2542 #define VPUTX_VRELE 1 2543 #define VPUTX_VPUT 2 2544 #define VPUTX_VUNREF 3 2545 2546 /* 2547 * Decrement the use and hold counts for a vnode. 2548 * 2549 * See an explanation near vget() as to why atomic operation is safe. 2550 */ 2551 static void 2552 vputx(struct vnode *vp, int func) 2553 { 2554 int error; 2555 2556 KASSERT(vp != NULL, ("vputx: null vp")); 2557 if (func == VPUTX_VUNREF) 2558 ASSERT_VOP_LOCKED(vp, "vunref"); 2559 else if (func == VPUTX_VPUT) 2560 ASSERT_VOP_LOCKED(vp, "vput"); 2561 else 2562 KASSERT(func == VPUTX_VRELE, ("vputx: wrong func")); 2563 ASSERT_VI_UNLOCKED(vp, __func__); 2564 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2565 2566 if (vp->v_type != VCHR && 2567 vfs_refcount_release_if_not_last(&vp->v_usecount)) { 2568 if (func == VPUTX_VPUT) 2569 VOP_UNLOCK(vp, 0); 2570 vdrop(vp); 2571 return; 2572 } 2573 2574 VI_LOCK(vp); 2575 2576 /* 2577 * We want to hold the vnode until the inactive finishes to 2578 * prevent vgone() races. We drop the use count here and the 2579 * hold count below when we're done. 2580 */ 2581 if (!refcount_release(&vp->v_usecount) || 2582 (vp->v_iflag & VI_DOINGINACT)) { 2583 if (func == VPUTX_VPUT) 2584 VOP_UNLOCK(vp, 0); 2585 v_decr_devcount(vp); 2586 vdropl(vp); 2587 return; 2588 } 2589 2590 v_decr_devcount(vp); 2591 2592 error = 0; 2593 2594 if (vp->v_usecount != 0) { 2595 vprint("vputx: usecount not zero", vp); 2596 panic("vputx: usecount not zero"); 2597 } 2598 2599 CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp); 2600 2601 /* 2602 * We must call VOP_INACTIVE with the node locked. Mark 2603 * as VI_DOINGINACT to avoid recursion. 2604 */ 2605 vp->v_iflag |= VI_OWEINACT; 2606 switch (func) { 2607 case VPUTX_VRELE: 2608 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 2609 VI_LOCK(vp); 2610 break; 2611 case VPUTX_VPUT: 2612 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 2613 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | 2614 LK_NOWAIT); 2615 VI_LOCK(vp); 2616 } 2617 break; 2618 case VPUTX_VUNREF: 2619 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 2620 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); 2621 VI_LOCK(vp); 2622 } 2623 break; 2624 } 2625 VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp, 2626 ("vnode with usecount and VI_OWEINACT set")); 2627 if (error == 0) { 2628 if (vp->v_iflag & VI_OWEINACT) 2629 vinactive(vp, curthread); 2630 if (func != VPUTX_VUNREF) 2631 VOP_UNLOCK(vp, 0); 2632 } 2633 vdropl(vp); 2634 } 2635 2636 /* 2637 * Vnode put/release. 2638 * If count drops to zero, call inactive routine and return to freelist. 2639 */ 2640 void 2641 vrele(struct vnode *vp) 2642 { 2643 2644 vputx(vp, VPUTX_VRELE); 2645 } 2646 2647 /* 2648 * Release an already locked vnode. This give the same effects as 2649 * unlock+vrele(), but takes less time and avoids releasing and 2650 * re-aquiring the lock (as vrele() acquires the lock internally.) 2651 */ 2652 void 2653 vput(struct vnode *vp) 2654 { 2655 2656 vputx(vp, VPUTX_VPUT); 2657 } 2658 2659 /* 2660 * Release an exclusively locked vnode. Do not unlock the vnode lock. 2661 */ 2662 void 2663 vunref(struct vnode *vp) 2664 { 2665 2666 vputx(vp, VPUTX_VUNREF); 2667 } 2668 2669 /* 2670 * Increase the hold count and activate if this is the first reference. 2671 */ 2672 void 2673 _vhold(struct vnode *vp, bool locked) 2674 { 2675 struct mount *mp; 2676 2677 if (locked) 2678 ASSERT_VI_LOCKED(vp, __func__); 2679 else 2680 ASSERT_VI_UNLOCKED(vp, __func__); 2681 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2682 if (!locked && vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt)) { 2683 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 2684 ("_vhold: vnode with holdcnt is free")); 2685 return; 2686 } 2687 2688 if (!locked) 2689 VI_LOCK(vp); 2690 if ((vp->v_iflag & VI_FREE) == 0) { 2691 refcount_acquire(&vp->v_holdcnt); 2692 if (!locked) 2693 VI_UNLOCK(vp); 2694 return; 2695 } 2696 VNASSERT(vp->v_holdcnt == 0, vp, 2697 ("%s: wrong hold count", __func__)); 2698 VNASSERT(vp->v_op != NULL, vp, 2699 ("%s: vnode already reclaimed.", __func__)); 2700 /* 2701 * Remove a vnode from the free list, mark it as in use, 2702 * and put it on the active list. 2703 */ 2704 mtx_lock(&vnode_free_list_mtx); 2705 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist); 2706 freevnodes--; 2707 vp->v_iflag &= ~VI_FREE; 2708 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 2709 ("Activating already active vnode")); 2710 vp->v_iflag |= VI_ACTIVE; 2711 mp = vp->v_mount; 2712 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist); 2713 mp->mnt_activevnodelistsize++; 2714 mtx_unlock(&vnode_free_list_mtx); 2715 refcount_acquire(&vp->v_holdcnt); 2716 if (!locked) 2717 VI_UNLOCK(vp); 2718 } 2719 2720 /* 2721 * Drop the hold count of the vnode. If this is the last reference to 2722 * the vnode we place it on the free list unless it has been vgone'd 2723 * (marked VI_DOOMED) in which case we will free it. 2724 * 2725 * Because the vnode vm object keeps a hold reference on the vnode if 2726 * there is at least one resident non-cached page, the vnode cannot 2727 * leave the active list without the page cleanup done. 2728 */ 2729 void 2730 _vdrop(struct vnode *vp, bool locked) 2731 { 2732 struct bufobj *bo; 2733 struct mount *mp; 2734 int active; 2735 2736 if (locked) 2737 ASSERT_VI_LOCKED(vp, __func__); 2738 else 2739 ASSERT_VI_UNLOCKED(vp, __func__); 2740 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2741 if ((int)vp->v_holdcnt <= 0) 2742 panic("vdrop: holdcnt %d", vp->v_holdcnt); 2743 if (vfs_refcount_release_if_not_last(&vp->v_holdcnt)) { 2744 if (locked) 2745 VI_UNLOCK(vp); 2746 return; 2747 } 2748 2749 if (!locked) 2750 VI_LOCK(vp); 2751 if (refcount_release(&vp->v_holdcnt) == 0) { 2752 VI_UNLOCK(vp); 2753 return; 2754 } 2755 if ((vp->v_iflag & VI_DOOMED) == 0) { 2756 /* 2757 * Mark a vnode as free: remove it from its active list 2758 * and put it up for recycling on the freelist. 2759 */ 2760 VNASSERT(vp->v_op != NULL, vp, 2761 ("vdropl: vnode already reclaimed.")); 2762 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 2763 ("vnode already free")); 2764 VNASSERT(vp->v_holdcnt == 0, vp, 2765 ("vdropl: freeing when we shouldn't")); 2766 active = vp->v_iflag & VI_ACTIVE; 2767 if ((vp->v_iflag & VI_OWEINACT) == 0) { 2768 vp->v_iflag &= ~VI_ACTIVE; 2769 mp = vp->v_mount; 2770 mtx_lock(&vnode_free_list_mtx); 2771 if (active) { 2772 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, 2773 v_actfreelist); 2774 mp->mnt_activevnodelistsize--; 2775 } 2776 TAILQ_INSERT_TAIL(&vnode_free_list, vp, 2777 v_actfreelist); 2778 freevnodes++; 2779 vp->v_iflag |= VI_FREE; 2780 mtx_unlock(&vnode_free_list_mtx); 2781 } else { 2782 atomic_add_long(&free_owe_inact, 1); 2783 } 2784 VI_UNLOCK(vp); 2785 return; 2786 } 2787 /* 2788 * The vnode has been marked for destruction, so free it. 2789 * 2790 * The vnode will be returned to the zone where it will 2791 * normally remain until it is needed for another vnode. We 2792 * need to cleanup (or verify that the cleanup has already 2793 * been done) any residual data left from its current use 2794 * so as not to contaminate the freshly allocated vnode. 2795 */ 2796 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp); 2797 atomic_subtract_long(&numvnodes, 1); 2798 bo = &vp->v_bufobj; 2799 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 2800 ("cleaned vnode still on the free list.")); 2801 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 2802 VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count")); 2803 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 2804 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 2805 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 2806 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 2807 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp, 2808 ("clean blk trie not empty")); 2809 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 2810 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp, 2811 ("dirty blk trie not empty")); 2812 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); 2813 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); 2814 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for ..")); 2815 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp, 2816 ("Dangling rangelock waiters")); 2817 VI_UNLOCK(vp); 2818 #ifdef MAC 2819 mac_vnode_destroy(vp); 2820 #endif 2821 if (vp->v_pollinfo != NULL) { 2822 destroy_vpollinfo(vp->v_pollinfo); 2823 vp->v_pollinfo = NULL; 2824 } 2825 #ifdef INVARIANTS 2826 /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */ 2827 vp->v_op = NULL; 2828 #endif 2829 bzero(&vp->v_un, sizeof(vp->v_un)); 2830 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 2831 vp->v_iflag = 0; 2832 vp->v_vflag = 0; 2833 bo->bo_flag = 0; 2834 uma_zfree(vnode_zone, vp); 2835 } 2836 2837 /* 2838 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 2839 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 2840 * OWEINACT tracks whether a vnode missed a call to inactive due to a 2841 * failed lock upgrade. 2842 */ 2843 void 2844 vinactive(struct vnode *vp, struct thread *td) 2845 { 2846 struct vm_object *obj; 2847 2848 ASSERT_VOP_ELOCKED(vp, "vinactive"); 2849 ASSERT_VI_LOCKED(vp, "vinactive"); 2850 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, 2851 ("vinactive: recursed on VI_DOINGINACT")); 2852 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2853 vp->v_iflag |= VI_DOINGINACT; 2854 vp->v_iflag &= ~VI_OWEINACT; 2855 VI_UNLOCK(vp); 2856 /* 2857 * Before moving off the active list, we must be sure that any 2858 * modified pages are converted into the vnode's dirty 2859 * buffers, since these will no longer be checked once the 2860 * vnode is on the inactive list. 2861 * 2862 * The write-out of the dirty pages is asynchronous. At the 2863 * point that VOP_INACTIVE() is called, there could still be 2864 * pending I/O and dirty pages in the object. 2865 */ 2866 obj = vp->v_object; 2867 if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0) { 2868 VM_OBJECT_WLOCK(obj); 2869 vm_object_page_clean(obj, 0, 0, OBJPC_NOSYNC); 2870 VM_OBJECT_WUNLOCK(obj); 2871 } 2872 VOP_INACTIVE(vp, td); 2873 VI_LOCK(vp); 2874 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, 2875 ("vinactive: lost VI_DOINGINACT")); 2876 vp->v_iflag &= ~VI_DOINGINACT; 2877 } 2878 2879 /* 2880 * Remove any vnodes in the vnode table belonging to mount point mp. 2881 * 2882 * If FORCECLOSE is not specified, there should not be any active ones, 2883 * return error if any are found (nb: this is a user error, not a 2884 * system error). If FORCECLOSE is specified, detach any active vnodes 2885 * that are found. 2886 * 2887 * If WRITECLOSE is set, only flush out regular file vnodes open for 2888 * writing. 2889 * 2890 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2891 * 2892 * `rootrefs' specifies the base reference count for the root vnode 2893 * of this filesystem. The root vnode is considered busy if its 2894 * v_usecount exceeds this value. On a successful return, vflush(, td) 2895 * will call vrele() on the root vnode exactly rootrefs times. 2896 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2897 * be zero. 2898 */ 2899 #ifdef DIAGNOSTIC 2900 static int busyprt = 0; /* print out busy vnodes */ 2901 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 2902 #endif 2903 2904 int 2905 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 2906 { 2907 struct vnode *vp, *mvp, *rootvp = NULL; 2908 struct vattr vattr; 2909 int busy = 0, error; 2910 2911 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 2912 rootrefs, flags); 2913 if (rootrefs > 0) { 2914 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2915 ("vflush: bad args")); 2916 /* 2917 * Get the filesystem root vnode. We can vput() it 2918 * immediately, since with rootrefs > 0, it won't go away. 2919 */ 2920 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 2921 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 2922 __func__, error); 2923 return (error); 2924 } 2925 vput(rootvp); 2926 } 2927 loop: 2928 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 2929 vholdl(vp); 2930 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 2931 if (error) { 2932 vdrop(vp); 2933 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 2934 goto loop; 2935 } 2936 /* 2937 * Skip over a vnodes marked VV_SYSTEM. 2938 */ 2939 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2940 VOP_UNLOCK(vp, 0); 2941 vdrop(vp); 2942 continue; 2943 } 2944 /* 2945 * If WRITECLOSE is set, flush out unlinked but still open 2946 * files (even if open only for reading) and regular file 2947 * vnodes open for writing. 2948 */ 2949 if (flags & WRITECLOSE) { 2950 if (vp->v_object != NULL) { 2951 VM_OBJECT_WLOCK(vp->v_object); 2952 vm_object_page_clean(vp->v_object, 0, 0, 0); 2953 VM_OBJECT_WUNLOCK(vp->v_object); 2954 } 2955 error = VOP_FSYNC(vp, MNT_WAIT, td); 2956 if (error != 0) { 2957 VOP_UNLOCK(vp, 0); 2958 vdrop(vp); 2959 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 2960 return (error); 2961 } 2962 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 2963 VI_LOCK(vp); 2964 2965 if ((vp->v_type == VNON || 2966 (error == 0 && vattr.va_nlink > 0)) && 2967 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2968 VOP_UNLOCK(vp, 0); 2969 vdropl(vp); 2970 continue; 2971 } 2972 } else 2973 VI_LOCK(vp); 2974 /* 2975 * With v_usecount == 0, all we need to do is clear out the 2976 * vnode data structures and we are done. 2977 * 2978 * If FORCECLOSE is set, forcibly close the vnode. 2979 */ 2980 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 2981 vgonel(vp); 2982 } else { 2983 busy++; 2984 #ifdef DIAGNOSTIC 2985 if (busyprt) 2986 vprint("vflush: busy vnode", vp); 2987 #endif 2988 } 2989 VOP_UNLOCK(vp, 0); 2990 vdropl(vp); 2991 } 2992 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2993 /* 2994 * If just the root vnode is busy, and if its refcount 2995 * is equal to `rootrefs', then go ahead and kill it. 2996 */ 2997 VI_LOCK(rootvp); 2998 KASSERT(busy > 0, ("vflush: not busy")); 2999 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 3000 ("vflush: usecount %d < rootrefs %d", 3001 rootvp->v_usecount, rootrefs)); 3002 if (busy == 1 && rootvp->v_usecount == rootrefs) { 3003 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 3004 vgone(rootvp); 3005 VOP_UNLOCK(rootvp, 0); 3006 busy = 0; 3007 } else 3008 VI_UNLOCK(rootvp); 3009 } 3010 if (busy) { 3011 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 3012 busy); 3013 return (EBUSY); 3014 } 3015 for (; rootrefs > 0; rootrefs--) 3016 vrele(rootvp); 3017 return (0); 3018 } 3019 3020 /* 3021 * Recycle an unused vnode to the front of the free list. 3022 */ 3023 int 3024 vrecycle(struct vnode *vp) 3025 { 3026 int recycled; 3027 3028 ASSERT_VOP_ELOCKED(vp, "vrecycle"); 3029 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3030 recycled = 0; 3031 VI_LOCK(vp); 3032 if (vp->v_usecount == 0) { 3033 recycled = 1; 3034 vgonel(vp); 3035 } 3036 VI_UNLOCK(vp); 3037 return (recycled); 3038 } 3039 3040 /* 3041 * Eliminate all activity associated with a vnode 3042 * in preparation for reuse. 3043 */ 3044 void 3045 vgone(struct vnode *vp) 3046 { 3047 VI_LOCK(vp); 3048 vgonel(vp); 3049 VI_UNLOCK(vp); 3050 } 3051 3052 static void 3053 notify_lowervp_vfs_dummy(struct mount *mp __unused, 3054 struct vnode *lowervp __unused) 3055 { 3056 } 3057 3058 /* 3059 * Notify upper mounts about reclaimed or unlinked vnode. 3060 */ 3061 void 3062 vfs_notify_upper(struct vnode *vp, int event) 3063 { 3064 static struct vfsops vgonel_vfsops = { 3065 .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy, 3066 .vfs_unlink_lowervp = notify_lowervp_vfs_dummy, 3067 }; 3068 struct mount *mp, *ump, *mmp; 3069 3070 mp = vp->v_mount; 3071 if (mp == NULL) 3072 return; 3073 3074 MNT_ILOCK(mp); 3075 if (TAILQ_EMPTY(&mp->mnt_uppers)) 3076 goto unlock; 3077 MNT_IUNLOCK(mp); 3078 mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO); 3079 mmp->mnt_op = &vgonel_vfsops; 3080 mmp->mnt_kern_flag |= MNTK_MARKER; 3081 MNT_ILOCK(mp); 3082 mp->mnt_kern_flag |= MNTK_VGONE_UPPER; 3083 for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) { 3084 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) { 3085 ump = TAILQ_NEXT(ump, mnt_upper_link); 3086 continue; 3087 } 3088 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link); 3089 MNT_IUNLOCK(mp); 3090 switch (event) { 3091 case VFS_NOTIFY_UPPER_RECLAIM: 3092 VFS_RECLAIM_LOWERVP(ump, vp); 3093 break; 3094 case VFS_NOTIFY_UPPER_UNLINK: 3095 VFS_UNLINK_LOWERVP(ump, vp); 3096 break; 3097 default: 3098 KASSERT(0, ("invalid event %d", event)); 3099 break; 3100 } 3101 MNT_ILOCK(mp); 3102 ump = TAILQ_NEXT(mmp, mnt_upper_link); 3103 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link); 3104 } 3105 free(mmp, M_TEMP); 3106 mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER; 3107 if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) { 3108 mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER; 3109 wakeup(&mp->mnt_uppers); 3110 } 3111 unlock: 3112 MNT_IUNLOCK(mp); 3113 } 3114 3115 /* 3116 * vgone, with the vp interlock held. 3117 */ 3118 static void 3119 vgonel(struct vnode *vp) 3120 { 3121 struct thread *td; 3122 int oweinact; 3123 int active; 3124 struct mount *mp; 3125 3126 ASSERT_VOP_ELOCKED(vp, "vgonel"); 3127 ASSERT_VI_LOCKED(vp, "vgonel"); 3128 VNASSERT(vp->v_holdcnt, vp, 3129 ("vgonel: vp %p has no reference.", vp)); 3130 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3131 td = curthread; 3132 3133 /* 3134 * Don't vgonel if we're already doomed. 3135 */ 3136 if (vp->v_iflag & VI_DOOMED) 3137 return; 3138 vp->v_iflag |= VI_DOOMED; 3139 3140 /* 3141 * Check to see if the vnode is in use. If so, we have to call 3142 * VOP_CLOSE() and VOP_INACTIVE(). 3143 */ 3144 active = vp->v_usecount; 3145 oweinact = (vp->v_iflag & VI_OWEINACT); 3146 VI_UNLOCK(vp); 3147 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); 3148 3149 /* 3150 * If purging an active vnode, it must be closed and 3151 * deactivated before being reclaimed. 3152 */ 3153 if (active) 3154 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 3155 if (oweinact || active) { 3156 VI_LOCK(vp); 3157 if ((vp->v_iflag & VI_DOINGINACT) == 0) 3158 vinactive(vp, td); 3159 VI_UNLOCK(vp); 3160 } 3161 if (vp->v_type == VSOCK) 3162 vfs_unp_reclaim(vp); 3163 3164 /* 3165 * Clean out any buffers associated with the vnode. 3166 * If the flush fails, just toss the buffers. 3167 */ 3168 mp = NULL; 3169 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 3170 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 3171 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { 3172 while (vinvalbuf(vp, 0, 0, 0) != 0) 3173 ; 3174 } 3175 3176 BO_LOCK(&vp->v_bufobj); 3177 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && 3178 vp->v_bufobj.bo_dirty.bv_cnt == 0 && 3179 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && 3180 vp->v_bufobj.bo_clean.bv_cnt == 0, 3181 ("vp %p bufobj not invalidated", vp)); 3182 vp->v_bufobj.bo_flag |= BO_DEAD; 3183 BO_UNLOCK(&vp->v_bufobj); 3184 3185 /* 3186 * Reclaim the vnode. 3187 */ 3188 if (VOP_RECLAIM(vp, td)) 3189 panic("vgone: cannot reclaim"); 3190 if (mp != NULL) 3191 vn_finished_secondary_write(mp); 3192 VNASSERT(vp->v_object == NULL, vp, 3193 ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag)); 3194 /* 3195 * Clear the advisory locks and wake up waiting threads. 3196 */ 3197 (void)VOP_ADVLOCKPURGE(vp); 3198 vp->v_lockf = NULL; 3199 /* 3200 * Delete from old mount point vnode list. 3201 */ 3202 delmntque(vp); 3203 cache_purge(vp); 3204 /* 3205 * Done with purge, reset to the standard lock and invalidate 3206 * the vnode. 3207 */ 3208 VI_LOCK(vp); 3209 vp->v_vnlock = &vp->v_lock; 3210 vp->v_op = &dead_vnodeops; 3211 vp->v_tag = "none"; 3212 vp->v_type = VBAD; 3213 } 3214 3215 /* 3216 * Calculate the total number of references to a special device. 3217 */ 3218 int 3219 vcount(struct vnode *vp) 3220 { 3221 int count; 3222 3223 dev_lock(); 3224 count = vp->v_rdev->si_usecount; 3225 dev_unlock(); 3226 return (count); 3227 } 3228 3229 /* 3230 * Same as above, but using the struct cdev *as argument 3231 */ 3232 int 3233 count_dev(struct cdev *dev) 3234 { 3235 int count; 3236 3237 dev_lock(); 3238 count = dev->si_usecount; 3239 dev_unlock(); 3240 return(count); 3241 } 3242 3243 /* 3244 * Print out a description of a vnode. 3245 */ 3246 static char *typename[] = 3247 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", 3248 "VMARKER"}; 3249 3250 void 3251 vn_printf(struct vnode *vp, const char *fmt, ...) 3252 { 3253 va_list ap; 3254 char buf[256], buf2[16]; 3255 u_long flags; 3256 3257 va_start(ap, fmt); 3258 vprintf(fmt, ap); 3259 va_end(ap); 3260 printf("%p: ", (void *)vp); 3261 printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]); 3262 printf(" usecount %d, writecount %d, refcount %d mountedhere %p\n", 3263 vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere); 3264 buf[0] = '\0'; 3265 buf[1] = '\0'; 3266 if (vp->v_vflag & VV_ROOT) 3267 strlcat(buf, "|VV_ROOT", sizeof(buf)); 3268 if (vp->v_vflag & VV_ISTTY) 3269 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 3270 if (vp->v_vflag & VV_NOSYNC) 3271 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 3272 if (vp->v_vflag & VV_ETERNALDEV) 3273 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 3274 if (vp->v_vflag & VV_CACHEDLABEL) 3275 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 3276 if (vp->v_vflag & VV_TEXT) 3277 strlcat(buf, "|VV_TEXT", sizeof(buf)); 3278 if (vp->v_vflag & VV_COPYONWRITE) 3279 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 3280 if (vp->v_vflag & VV_SYSTEM) 3281 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 3282 if (vp->v_vflag & VV_PROCDEP) 3283 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 3284 if (vp->v_vflag & VV_NOKNOTE) 3285 strlcat(buf, "|VV_NOKNOTE", sizeof(buf)); 3286 if (vp->v_vflag & VV_DELETED) 3287 strlcat(buf, "|VV_DELETED", sizeof(buf)); 3288 if (vp->v_vflag & VV_MD) 3289 strlcat(buf, "|VV_MD", sizeof(buf)); 3290 if (vp->v_vflag & VV_FORCEINSMQ) 3291 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 3292 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 3293 VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | 3294 VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ); 3295 if (flags != 0) { 3296 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 3297 strlcat(buf, buf2, sizeof(buf)); 3298 } 3299 if (vp->v_iflag & VI_MOUNT) 3300 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 3301 if (vp->v_iflag & VI_DOOMED) 3302 strlcat(buf, "|VI_DOOMED", sizeof(buf)); 3303 if (vp->v_iflag & VI_FREE) 3304 strlcat(buf, "|VI_FREE", sizeof(buf)); 3305 if (vp->v_iflag & VI_ACTIVE) 3306 strlcat(buf, "|VI_ACTIVE", sizeof(buf)); 3307 if (vp->v_iflag & VI_DOINGINACT) 3308 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 3309 if (vp->v_iflag & VI_OWEINACT) 3310 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 3311 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE | 3312 VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT); 3313 if (flags != 0) { 3314 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 3315 strlcat(buf, buf2, sizeof(buf)); 3316 } 3317 printf(" flags (%s)\n", buf + 1); 3318 if (mtx_owned(VI_MTX(vp))) 3319 printf(" VI_LOCKed"); 3320 if (vp->v_object != NULL) 3321 printf(" v_object %p ref %d pages %d " 3322 "cleanbuf %d dirtybuf %d\n", 3323 vp->v_object, vp->v_object->ref_count, 3324 vp->v_object->resident_page_count, 3325 vp->v_bufobj.bo_clean.bv_cnt, 3326 vp->v_bufobj.bo_dirty.bv_cnt); 3327 printf(" "); 3328 lockmgr_printinfo(vp->v_vnlock); 3329 if (vp->v_data != NULL) 3330 VOP_PRINT(vp); 3331 } 3332 3333 #ifdef DDB 3334 /* 3335 * List all of the locked vnodes in the system. 3336 * Called when debugging the kernel. 3337 */ 3338 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 3339 { 3340 struct mount *mp; 3341 struct vnode *vp; 3342 3343 /* 3344 * Note: because this is DDB, we can't obey the locking semantics 3345 * for these structures, which means we could catch an inconsistent 3346 * state and dereference a nasty pointer. Not much to be done 3347 * about that. 3348 */ 3349 db_printf("Locked vnodes\n"); 3350 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3351 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3352 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) 3353 vprint("", vp); 3354 } 3355 } 3356 } 3357 3358 /* 3359 * Show details about the given vnode. 3360 */ 3361 DB_SHOW_COMMAND(vnode, db_show_vnode) 3362 { 3363 struct vnode *vp; 3364 3365 if (!have_addr) 3366 return; 3367 vp = (struct vnode *)addr; 3368 vn_printf(vp, "vnode "); 3369 } 3370 3371 /* 3372 * Show details about the given mount point. 3373 */ 3374 DB_SHOW_COMMAND(mount, db_show_mount) 3375 { 3376 struct mount *mp; 3377 struct vfsopt *opt; 3378 struct statfs *sp; 3379 struct vnode *vp; 3380 char buf[512]; 3381 uint64_t mflags; 3382 u_int flags; 3383 3384 if (!have_addr) { 3385 /* No address given, print short info about all mount points. */ 3386 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3387 db_printf("%p %s on %s (%s)\n", mp, 3388 mp->mnt_stat.f_mntfromname, 3389 mp->mnt_stat.f_mntonname, 3390 mp->mnt_stat.f_fstypename); 3391 if (db_pager_quit) 3392 break; 3393 } 3394 db_printf("\nMore info: show mount <addr>\n"); 3395 return; 3396 } 3397 3398 mp = (struct mount *)addr; 3399 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 3400 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 3401 3402 buf[0] = '\0'; 3403 mflags = mp->mnt_flag; 3404 #define MNT_FLAG(flag) do { \ 3405 if (mflags & (flag)) { \ 3406 if (buf[0] != '\0') \ 3407 strlcat(buf, ", ", sizeof(buf)); \ 3408 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 3409 mflags &= ~(flag); \ 3410 } \ 3411 } while (0) 3412 MNT_FLAG(MNT_RDONLY); 3413 MNT_FLAG(MNT_SYNCHRONOUS); 3414 MNT_FLAG(MNT_NOEXEC); 3415 MNT_FLAG(MNT_NOSUID); 3416 MNT_FLAG(MNT_NFS4ACLS); 3417 MNT_FLAG(MNT_UNION); 3418 MNT_FLAG(MNT_ASYNC); 3419 MNT_FLAG(MNT_SUIDDIR); 3420 MNT_FLAG(MNT_SOFTDEP); 3421 MNT_FLAG(MNT_NOSYMFOLLOW); 3422 MNT_FLAG(MNT_GJOURNAL); 3423 MNT_FLAG(MNT_MULTILABEL); 3424 MNT_FLAG(MNT_ACLS); 3425 MNT_FLAG(MNT_NOATIME); 3426 MNT_FLAG(MNT_NOCLUSTERR); 3427 MNT_FLAG(MNT_NOCLUSTERW); 3428 MNT_FLAG(MNT_SUJ); 3429 MNT_FLAG(MNT_EXRDONLY); 3430 MNT_FLAG(MNT_EXPORTED); 3431 MNT_FLAG(MNT_DEFEXPORTED); 3432 MNT_FLAG(MNT_EXPORTANON); 3433 MNT_FLAG(MNT_EXKERB); 3434 MNT_FLAG(MNT_EXPUBLIC); 3435 MNT_FLAG(MNT_LOCAL); 3436 MNT_FLAG(MNT_QUOTA); 3437 MNT_FLAG(MNT_ROOTFS); 3438 MNT_FLAG(MNT_USER); 3439 MNT_FLAG(MNT_IGNORE); 3440 MNT_FLAG(MNT_UPDATE); 3441 MNT_FLAG(MNT_DELEXPORT); 3442 MNT_FLAG(MNT_RELOAD); 3443 MNT_FLAG(MNT_FORCE); 3444 MNT_FLAG(MNT_SNAPSHOT); 3445 MNT_FLAG(MNT_BYFSID); 3446 #undef MNT_FLAG 3447 if (mflags != 0) { 3448 if (buf[0] != '\0') 3449 strlcat(buf, ", ", sizeof(buf)); 3450 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 3451 "0x%016jx", mflags); 3452 } 3453 db_printf(" mnt_flag = %s\n", buf); 3454 3455 buf[0] = '\0'; 3456 flags = mp->mnt_kern_flag; 3457 #define MNT_KERN_FLAG(flag) do { \ 3458 if (flags & (flag)) { \ 3459 if (buf[0] != '\0') \ 3460 strlcat(buf, ", ", sizeof(buf)); \ 3461 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 3462 flags &= ~(flag); \ 3463 } \ 3464 } while (0) 3465 MNT_KERN_FLAG(MNTK_UNMOUNTF); 3466 MNT_KERN_FLAG(MNTK_ASYNC); 3467 MNT_KERN_FLAG(MNTK_SOFTDEP); 3468 MNT_KERN_FLAG(MNTK_NOINSMNTQ); 3469 MNT_KERN_FLAG(MNTK_DRAINING); 3470 MNT_KERN_FLAG(MNTK_REFEXPIRE); 3471 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 3472 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 3473 MNT_KERN_FLAG(MNTK_NO_IOPF); 3474 MNT_KERN_FLAG(MNTK_VGONE_UPPER); 3475 MNT_KERN_FLAG(MNTK_VGONE_WAITER); 3476 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT); 3477 MNT_KERN_FLAG(MNTK_MARKER); 3478 MNT_KERN_FLAG(MNTK_USES_BCACHE); 3479 MNT_KERN_FLAG(MNTK_NOASYNC); 3480 MNT_KERN_FLAG(MNTK_UNMOUNT); 3481 MNT_KERN_FLAG(MNTK_MWAIT); 3482 MNT_KERN_FLAG(MNTK_SUSPEND); 3483 MNT_KERN_FLAG(MNTK_SUSPEND2); 3484 MNT_KERN_FLAG(MNTK_SUSPENDED); 3485 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 3486 MNT_KERN_FLAG(MNTK_NOKNOTE); 3487 #undef MNT_KERN_FLAG 3488 if (flags != 0) { 3489 if (buf[0] != '\0') 3490 strlcat(buf, ", ", sizeof(buf)); 3491 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 3492 "0x%08x", flags); 3493 } 3494 db_printf(" mnt_kern_flag = %s\n", buf); 3495 3496 db_printf(" mnt_opt = "); 3497 opt = TAILQ_FIRST(mp->mnt_opt); 3498 if (opt != NULL) { 3499 db_printf("%s", opt->name); 3500 opt = TAILQ_NEXT(opt, link); 3501 while (opt != NULL) { 3502 db_printf(", %s", opt->name); 3503 opt = TAILQ_NEXT(opt, link); 3504 } 3505 } 3506 db_printf("\n"); 3507 3508 sp = &mp->mnt_stat; 3509 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 3510 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 3511 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 3512 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 3513 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 3514 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 3515 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 3516 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 3517 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 3518 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 3519 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 3520 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 3521 3522 db_printf(" mnt_cred = { uid=%u ruid=%u", 3523 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 3524 if (jailed(mp->mnt_cred)) 3525 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 3526 db_printf(" }\n"); 3527 db_printf(" mnt_ref = %d\n", mp->mnt_ref); 3528 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 3529 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 3530 db_printf(" mnt_activevnodelistsize = %d\n", 3531 mp->mnt_activevnodelistsize); 3532 db_printf(" mnt_writeopcount = %d\n", mp->mnt_writeopcount); 3533 db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen); 3534 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 3535 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 3536 db_printf(" mnt_lockref = %d\n", mp->mnt_lockref); 3537 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 3538 db_printf(" mnt_secondary_accwrites = %d\n", 3539 mp->mnt_secondary_accwrites); 3540 db_printf(" mnt_gjprovider = %s\n", 3541 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 3542 3543 db_printf("\n\nList of active vnodes\n"); 3544 TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) { 3545 if (vp->v_type != VMARKER) { 3546 vn_printf(vp, "vnode "); 3547 if (db_pager_quit) 3548 break; 3549 } 3550 } 3551 db_printf("\n\nList of inactive vnodes\n"); 3552 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3553 if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) { 3554 vn_printf(vp, "vnode "); 3555 if (db_pager_quit) 3556 break; 3557 } 3558 } 3559 } 3560 #endif /* DDB */ 3561 3562 /* 3563 * Fill in a struct xvfsconf based on a struct vfsconf. 3564 */ 3565 static int 3566 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 3567 { 3568 struct xvfsconf xvfsp; 3569 3570 bzero(&xvfsp, sizeof(xvfsp)); 3571 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 3572 xvfsp.vfc_typenum = vfsp->vfc_typenum; 3573 xvfsp.vfc_refcount = vfsp->vfc_refcount; 3574 xvfsp.vfc_flags = vfsp->vfc_flags; 3575 /* 3576 * These are unused in userland, we keep them 3577 * to not break binary compatibility. 3578 */ 3579 xvfsp.vfc_vfsops = NULL; 3580 xvfsp.vfc_next = NULL; 3581 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 3582 } 3583 3584 #ifdef COMPAT_FREEBSD32 3585 struct xvfsconf32 { 3586 uint32_t vfc_vfsops; 3587 char vfc_name[MFSNAMELEN]; 3588 int32_t vfc_typenum; 3589 int32_t vfc_refcount; 3590 int32_t vfc_flags; 3591 uint32_t vfc_next; 3592 }; 3593 3594 static int 3595 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 3596 { 3597 struct xvfsconf32 xvfsp; 3598 3599 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 3600 xvfsp.vfc_typenum = vfsp->vfc_typenum; 3601 xvfsp.vfc_refcount = vfsp->vfc_refcount; 3602 xvfsp.vfc_flags = vfsp->vfc_flags; 3603 xvfsp.vfc_vfsops = 0; 3604 xvfsp.vfc_next = 0; 3605 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 3606 } 3607 #endif 3608 3609 /* 3610 * Top level filesystem related information gathering. 3611 */ 3612 static int 3613 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 3614 { 3615 struct vfsconf *vfsp; 3616 int error; 3617 3618 error = 0; 3619 vfsconf_slock(); 3620 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 3621 #ifdef COMPAT_FREEBSD32 3622 if (req->flags & SCTL_MASK32) 3623 error = vfsconf2x32(req, vfsp); 3624 else 3625 #endif 3626 error = vfsconf2x(req, vfsp); 3627 if (error) 3628 break; 3629 } 3630 vfsconf_sunlock(); 3631 return (error); 3632 } 3633 3634 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | 3635 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, 3636 "S,xvfsconf", "List of all configured filesystems"); 3637 3638 #ifndef BURN_BRIDGES 3639 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 3640 3641 static int 3642 vfs_sysctl(SYSCTL_HANDLER_ARGS) 3643 { 3644 int *name = (int *)arg1 - 1; /* XXX */ 3645 u_int namelen = arg2 + 1; /* XXX */ 3646 struct vfsconf *vfsp; 3647 3648 log(LOG_WARNING, "userland calling deprecated sysctl, " 3649 "please rebuild world\n"); 3650 3651 #if 1 || defined(COMPAT_PRELITE2) 3652 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 3653 if (namelen == 1) 3654 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 3655 #endif 3656 3657 switch (name[1]) { 3658 case VFS_MAXTYPENUM: 3659 if (namelen != 2) 3660 return (ENOTDIR); 3661 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 3662 case VFS_CONF: 3663 if (namelen != 3) 3664 return (ENOTDIR); /* overloaded */ 3665 vfsconf_slock(); 3666 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 3667 if (vfsp->vfc_typenum == name[2]) 3668 break; 3669 } 3670 vfsconf_sunlock(); 3671 if (vfsp == NULL) 3672 return (EOPNOTSUPP); 3673 #ifdef COMPAT_FREEBSD32 3674 if (req->flags & SCTL_MASK32) 3675 return (vfsconf2x32(req, vfsp)); 3676 else 3677 #endif 3678 return (vfsconf2x(req, vfsp)); 3679 } 3680 return (EOPNOTSUPP); 3681 } 3682 3683 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | 3684 CTLFLAG_MPSAFE, vfs_sysctl, 3685 "Generic filesystem"); 3686 3687 #if 1 || defined(COMPAT_PRELITE2) 3688 3689 static int 3690 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 3691 { 3692 int error; 3693 struct vfsconf *vfsp; 3694 struct ovfsconf ovfs; 3695 3696 vfsconf_slock(); 3697 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 3698 bzero(&ovfs, sizeof(ovfs)); 3699 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 3700 strcpy(ovfs.vfc_name, vfsp->vfc_name); 3701 ovfs.vfc_index = vfsp->vfc_typenum; 3702 ovfs.vfc_refcount = vfsp->vfc_refcount; 3703 ovfs.vfc_flags = vfsp->vfc_flags; 3704 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 3705 if (error != 0) { 3706 vfsconf_sunlock(); 3707 return (error); 3708 } 3709 } 3710 vfsconf_sunlock(); 3711 return (0); 3712 } 3713 3714 #endif /* 1 || COMPAT_PRELITE2 */ 3715 #endif /* !BURN_BRIDGES */ 3716 3717 #define KINFO_VNODESLOP 10 3718 #ifdef notyet 3719 /* 3720 * Dump vnode list (via sysctl). 3721 */ 3722 /* ARGSUSED */ 3723 static int 3724 sysctl_vnode(SYSCTL_HANDLER_ARGS) 3725 { 3726 struct xvnode *xvn; 3727 struct mount *mp; 3728 struct vnode *vp; 3729 int error, len, n; 3730 3731 /* 3732 * Stale numvnodes access is not fatal here. 3733 */ 3734 req->lock = 0; 3735 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 3736 if (!req->oldptr) 3737 /* Make an estimate */ 3738 return (SYSCTL_OUT(req, 0, len)); 3739 3740 error = sysctl_wire_old_buffer(req, 0); 3741 if (error != 0) 3742 return (error); 3743 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 3744 n = 0; 3745 mtx_lock(&mountlist_mtx); 3746 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3747 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) 3748 continue; 3749 MNT_ILOCK(mp); 3750 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3751 if (n == len) 3752 break; 3753 vref(vp); 3754 xvn[n].xv_size = sizeof *xvn; 3755 xvn[n].xv_vnode = vp; 3756 xvn[n].xv_id = 0; /* XXX compat */ 3757 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 3758 XV_COPY(usecount); 3759 XV_COPY(writecount); 3760 XV_COPY(holdcnt); 3761 XV_COPY(mount); 3762 XV_COPY(numoutput); 3763 XV_COPY(type); 3764 #undef XV_COPY 3765 xvn[n].xv_flag = vp->v_vflag; 3766 3767 switch (vp->v_type) { 3768 case VREG: 3769 case VDIR: 3770 case VLNK: 3771 break; 3772 case VBLK: 3773 case VCHR: 3774 if (vp->v_rdev == NULL) { 3775 vrele(vp); 3776 continue; 3777 } 3778 xvn[n].xv_dev = dev2udev(vp->v_rdev); 3779 break; 3780 case VSOCK: 3781 xvn[n].xv_socket = vp->v_socket; 3782 break; 3783 case VFIFO: 3784 xvn[n].xv_fifo = vp->v_fifoinfo; 3785 break; 3786 case VNON: 3787 case VBAD: 3788 default: 3789 /* shouldn't happen? */ 3790 vrele(vp); 3791 continue; 3792 } 3793 vrele(vp); 3794 ++n; 3795 } 3796 MNT_IUNLOCK(mp); 3797 mtx_lock(&mountlist_mtx); 3798 vfs_unbusy(mp); 3799 if (n == len) 3800 break; 3801 } 3802 mtx_unlock(&mountlist_mtx); 3803 3804 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 3805 free(xvn, M_TEMP); 3806 return (error); 3807 } 3808 3809 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD | 3810 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode", 3811 ""); 3812 #endif 3813 3814 static void 3815 unmount_or_warn(struct mount *mp) 3816 { 3817 int error; 3818 3819 error = dounmount(mp, MNT_FORCE, curthread); 3820 if (error != 0) { 3821 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); 3822 if (error == EBUSY) 3823 printf("BUSY)\n"); 3824 else 3825 printf("%d)\n", error); 3826 } 3827 } 3828 3829 /* 3830 * Unmount all filesystems. The list is traversed in reverse order 3831 * of mounting to avoid dependencies. 3832 */ 3833 void 3834 vfs_unmountall(void) 3835 { 3836 struct mount *mp, *tmp; 3837 3838 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 3839 3840 /* 3841 * Since this only runs when rebooting, it is not interlocked. 3842 */ 3843 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { 3844 vfs_ref(mp); 3845 3846 /* 3847 * Forcibly unmounting "/dev" before "/" would prevent clean 3848 * unmount of the latter. 3849 */ 3850 if (mp == rootdevmp) 3851 continue; 3852 3853 unmount_or_warn(mp); 3854 } 3855 3856 if (rootdevmp != NULL) 3857 unmount_or_warn(rootdevmp); 3858 } 3859 3860 /* 3861 * perform msync on all vnodes under a mount point 3862 * the mount point must be locked. 3863 */ 3864 void 3865 vfs_msync(struct mount *mp, int flags) 3866 { 3867 struct vnode *vp, *mvp; 3868 struct vm_object *obj; 3869 3870 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 3871 MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) { 3872 obj = vp->v_object; 3873 if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 && 3874 (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) { 3875 if (!vget(vp, 3876 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 3877 curthread)) { 3878 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 3879 vput(vp); 3880 continue; 3881 } 3882 3883 obj = vp->v_object; 3884 if (obj != NULL) { 3885 VM_OBJECT_WLOCK(obj); 3886 vm_object_page_clean(obj, 0, 0, 3887 flags == MNT_WAIT ? 3888 OBJPC_SYNC : OBJPC_NOSYNC); 3889 VM_OBJECT_WUNLOCK(obj); 3890 } 3891 vput(vp); 3892 } 3893 } else 3894 VI_UNLOCK(vp); 3895 } 3896 } 3897 3898 static void 3899 destroy_vpollinfo_free(struct vpollinfo *vi) 3900 { 3901 3902 knlist_destroy(&vi->vpi_selinfo.si_note); 3903 mtx_destroy(&vi->vpi_lock); 3904 uma_zfree(vnodepoll_zone, vi); 3905 } 3906 3907 static void 3908 destroy_vpollinfo(struct vpollinfo *vi) 3909 { 3910 3911 knlist_clear(&vi->vpi_selinfo.si_note, 1); 3912 seldrain(&vi->vpi_selinfo); 3913 destroy_vpollinfo_free(vi); 3914 } 3915 3916 /* 3917 * Initalize per-vnode helper structure to hold poll-related state. 3918 */ 3919 void 3920 v_addpollinfo(struct vnode *vp) 3921 { 3922 struct vpollinfo *vi; 3923 3924 if (vp->v_pollinfo != NULL) 3925 return; 3926 vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO); 3927 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 3928 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 3929 vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked); 3930 VI_LOCK(vp); 3931 if (vp->v_pollinfo != NULL) { 3932 VI_UNLOCK(vp); 3933 destroy_vpollinfo_free(vi); 3934 return; 3935 } 3936 vp->v_pollinfo = vi; 3937 VI_UNLOCK(vp); 3938 } 3939 3940 /* 3941 * Record a process's interest in events which might happen to 3942 * a vnode. Because poll uses the historic select-style interface 3943 * internally, this routine serves as both the ``check for any 3944 * pending events'' and the ``record my interest in future events'' 3945 * functions. (These are done together, while the lock is held, 3946 * to avoid race conditions.) 3947 */ 3948 int 3949 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 3950 { 3951 3952 v_addpollinfo(vp); 3953 mtx_lock(&vp->v_pollinfo->vpi_lock); 3954 if (vp->v_pollinfo->vpi_revents & events) { 3955 /* 3956 * This leaves events we are not interested 3957 * in available for the other process which 3958 * which presumably had requested them 3959 * (otherwise they would never have been 3960 * recorded). 3961 */ 3962 events &= vp->v_pollinfo->vpi_revents; 3963 vp->v_pollinfo->vpi_revents &= ~events; 3964 3965 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3966 return (events); 3967 } 3968 vp->v_pollinfo->vpi_events |= events; 3969 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 3970 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3971 return (0); 3972 } 3973 3974 /* 3975 * Routine to create and manage a filesystem syncer vnode. 3976 */ 3977 #define sync_close ((int (*)(struct vop_close_args *))nullop) 3978 static int sync_fsync(struct vop_fsync_args *); 3979 static int sync_inactive(struct vop_inactive_args *); 3980 static int sync_reclaim(struct vop_reclaim_args *); 3981 3982 static struct vop_vector sync_vnodeops = { 3983 .vop_bypass = VOP_EOPNOTSUPP, 3984 .vop_close = sync_close, /* close */ 3985 .vop_fsync = sync_fsync, /* fsync */ 3986 .vop_inactive = sync_inactive, /* inactive */ 3987 .vop_reclaim = sync_reclaim, /* reclaim */ 3988 .vop_lock1 = vop_stdlock, /* lock */ 3989 .vop_unlock = vop_stdunlock, /* unlock */ 3990 .vop_islocked = vop_stdislocked, /* islocked */ 3991 }; 3992 3993 /* 3994 * Create a new filesystem syncer vnode for the specified mount point. 3995 */ 3996 void 3997 vfs_allocate_syncvnode(struct mount *mp) 3998 { 3999 struct vnode *vp; 4000 struct bufobj *bo; 4001 static long start, incr, next; 4002 int error; 4003 4004 /* Allocate a new vnode */ 4005 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 4006 if (error != 0) 4007 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 4008 vp->v_type = VNON; 4009 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4010 vp->v_vflag |= VV_FORCEINSMQ; 4011 error = insmntque(vp, mp); 4012 if (error != 0) 4013 panic("vfs_allocate_syncvnode: insmntque() failed"); 4014 vp->v_vflag &= ~VV_FORCEINSMQ; 4015 VOP_UNLOCK(vp, 0); 4016 /* 4017 * Place the vnode onto the syncer worklist. We attempt to 4018 * scatter them about on the list so that they will go off 4019 * at evenly distributed times even if all the filesystems 4020 * are mounted at once. 4021 */ 4022 next += incr; 4023 if (next == 0 || next > syncer_maxdelay) { 4024 start /= 2; 4025 incr /= 2; 4026 if (start == 0) { 4027 start = syncer_maxdelay / 2; 4028 incr = syncer_maxdelay; 4029 } 4030 next = start; 4031 } 4032 bo = &vp->v_bufobj; 4033 BO_LOCK(bo); 4034 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 4035 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 4036 mtx_lock(&sync_mtx); 4037 sync_vnode_count++; 4038 if (mp->mnt_syncer == NULL) { 4039 mp->mnt_syncer = vp; 4040 vp = NULL; 4041 } 4042 mtx_unlock(&sync_mtx); 4043 BO_UNLOCK(bo); 4044 if (vp != NULL) { 4045 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4046 vgone(vp); 4047 vput(vp); 4048 } 4049 } 4050 4051 void 4052 vfs_deallocate_syncvnode(struct mount *mp) 4053 { 4054 struct vnode *vp; 4055 4056 mtx_lock(&sync_mtx); 4057 vp = mp->mnt_syncer; 4058 if (vp != NULL) 4059 mp->mnt_syncer = NULL; 4060 mtx_unlock(&sync_mtx); 4061 if (vp != NULL) 4062 vrele(vp); 4063 } 4064 4065 /* 4066 * Do a lazy sync of the filesystem. 4067 */ 4068 static int 4069 sync_fsync(struct vop_fsync_args *ap) 4070 { 4071 struct vnode *syncvp = ap->a_vp; 4072 struct mount *mp = syncvp->v_mount; 4073 int error, save; 4074 struct bufobj *bo; 4075 4076 /* 4077 * We only need to do something if this is a lazy evaluation. 4078 */ 4079 if (ap->a_waitfor != MNT_LAZY) 4080 return (0); 4081 4082 /* 4083 * Move ourselves to the back of the sync list. 4084 */ 4085 bo = &syncvp->v_bufobj; 4086 BO_LOCK(bo); 4087 vn_syncer_add_to_worklist(bo, syncdelay); 4088 BO_UNLOCK(bo); 4089 4090 /* 4091 * Walk the list of vnodes pushing all that are dirty and 4092 * not already on the sync list. 4093 */ 4094 if (vfs_busy(mp, MBF_NOWAIT) != 0) 4095 return (0); 4096 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 4097 vfs_unbusy(mp); 4098 return (0); 4099 } 4100 save = curthread_pflags_set(TDP_SYNCIO); 4101 vfs_msync(mp, MNT_NOWAIT); 4102 error = VFS_SYNC(mp, MNT_LAZY); 4103 curthread_pflags_restore(save); 4104 vn_finished_write(mp); 4105 vfs_unbusy(mp); 4106 return (error); 4107 } 4108 4109 /* 4110 * The syncer vnode is no referenced. 4111 */ 4112 static int 4113 sync_inactive(struct vop_inactive_args *ap) 4114 { 4115 4116 vgone(ap->a_vp); 4117 return (0); 4118 } 4119 4120 /* 4121 * The syncer vnode is no longer needed and is being decommissioned. 4122 * 4123 * Modifications to the worklist must be protected by sync_mtx. 4124 */ 4125 static int 4126 sync_reclaim(struct vop_reclaim_args *ap) 4127 { 4128 struct vnode *vp = ap->a_vp; 4129 struct bufobj *bo; 4130 4131 bo = &vp->v_bufobj; 4132 BO_LOCK(bo); 4133 mtx_lock(&sync_mtx); 4134 if (vp->v_mount->mnt_syncer == vp) 4135 vp->v_mount->mnt_syncer = NULL; 4136 if (bo->bo_flag & BO_ONWORKLST) { 4137 LIST_REMOVE(bo, bo_synclist); 4138 syncer_worklist_len--; 4139 sync_vnode_count--; 4140 bo->bo_flag &= ~BO_ONWORKLST; 4141 } 4142 mtx_unlock(&sync_mtx); 4143 BO_UNLOCK(bo); 4144 4145 return (0); 4146 } 4147 4148 /* 4149 * Check if vnode represents a disk device 4150 */ 4151 int 4152 vn_isdisk(struct vnode *vp, int *errp) 4153 { 4154 int error; 4155 4156 if (vp->v_type != VCHR) { 4157 error = ENOTBLK; 4158 goto out; 4159 } 4160 error = 0; 4161 dev_lock(); 4162 if (vp->v_rdev == NULL) 4163 error = ENXIO; 4164 else if (vp->v_rdev->si_devsw == NULL) 4165 error = ENXIO; 4166 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 4167 error = ENOTBLK; 4168 dev_unlock(); 4169 out: 4170 if (errp != NULL) 4171 *errp = error; 4172 return (error == 0); 4173 } 4174 4175 /* 4176 * Common filesystem object access control check routine. Accepts a 4177 * vnode's type, "mode", uid and gid, requested access mode, credentials, 4178 * and optional call-by-reference privused argument allowing vaccess() 4179 * to indicate to the caller whether privilege was used to satisfy the 4180 * request (obsoleted). Returns 0 on success, or an errno on failure. 4181 */ 4182 int 4183 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 4184 accmode_t accmode, struct ucred *cred, int *privused) 4185 { 4186 accmode_t dac_granted; 4187 accmode_t priv_granted; 4188 4189 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 4190 ("invalid bit in accmode")); 4191 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 4192 ("VAPPEND without VWRITE")); 4193 4194 /* 4195 * Look for a normal, non-privileged way to access the file/directory 4196 * as requested. If it exists, go with that. 4197 */ 4198 4199 if (privused != NULL) 4200 *privused = 0; 4201 4202 dac_granted = 0; 4203 4204 /* Check the owner. */ 4205 if (cred->cr_uid == file_uid) { 4206 dac_granted |= VADMIN; 4207 if (file_mode & S_IXUSR) 4208 dac_granted |= VEXEC; 4209 if (file_mode & S_IRUSR) 4210 dac_granted |= VREAD; 4211 if (file_mode & S_IWUSR) 4212 dac_granted |= (VWRITE | VAPPEND); 4213 4214 if ((accmode & dac_granted) == accmode) 4215 return (0); 4216 4217 goto privcheck; 4218 } 4219 4220 /* Otherwise, check the groups (first match) */ 4221 if (groupmember(file_gid, cred)) { 4222 if (file_mode & S_IXGRP) 4223 dac_granted |= VEXEC; 4224 if (file_mode & S_IRGRP) 4225 dac_granted |= VREAD; 4226 if (file_mode & S_IWGRP) 4227 dac_granted |= (VWRITE | VAPPEND); 4228 4229 if ((accmode & dac_granted) == accmode) 4230 return (0); 4231 4232 goto privcheck; 4233 } 4234 4235 /* Otherwise, check everyone else. */ 4236 if (file_mode & S_IXOTH) 4237 dac_granted |= VEXEC; 4238 if (file_mode & S_IROTH) 4239 dac_granted |= VREAD; 4240 if (file_mode & S_IWOTH) 4241 dac_granted |= (VWRITE | VAPPEND); 4242 if ((accmode & dac_granted) == accmode) 4243 return (0); 4244 4245 privcheck: 4246 /* 4247 * Build a privilege mask to determine if the set of privileges 4248 * satisfies the requirements when combined with the granted mask 4249 * from above. For each privilege, if the privilege is required, 4250 * bitwise or the request type onto the priv_granted mask. 4251 */ 4252 priv_granted = 0; 4253 4254 if (type == VDIR) { 4255 /* 4256 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 4257 * requests, instead of PRIV_VFS_EXEC. 4258 */ 4259 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 4260 !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0)) 4261 priv_granted |= VEXEC; 4262 } else { 4263 /* 4264 * Ensure that at least one execute bit is on. Otherwise, 4265 * a privileged user will always succeed, and we don't want 4266 * this to happen unless the file really is executable. 4267 */ 4268 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 4269 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 4270 !priv_check_cred(cred, PRIV_VFS_EXEC, 0)) 4271 priv_granted |= VEXEC; 4272 } 4273 4274 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 4275 !priv_check_cred(cred, PRIV_VFS_READ, 0)) 4276 priv_granted |= VREAD; 4277 4278 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 4279 !priv_check_cred(cred, PRIV_VFS_WRITE, 0)) 4280 priv_granted |= (VWRITE | VAPPEND); 4281 4282 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 4283 !priv_check_cred(cred, PRIV_VFS_ADMIN, 0)) 4284 priv_granted |= VADMIN; 4285 4286 if ((accmode & (priv_granted | dac_granted)) == accmode) { 4287 /* XXX audit: privilege used */ 4288 if (privused != NULL) 4289 *privused = 1; 4290 return (0); 4291 } 4292 4293 return ((accmode & VADMIN) ? EPERM : EACCES); 4294 } 4295 4296 /* 4297 * Credential check based on process requesting service, and per-attribute 4298 * permissions. 4299 */ 4300 int 4301 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 4302 struct thread *td, accmode_t accmode) 4303 { 4304 4305 /* 4306 * Kernel-invoked always succeeds. 4307 */ 4308 if (cred == NOCRED) 4309 return (0); 4310 4311 /* 4312 * Do not allow privileged processes in jail to directly manipulate 4313 * system attributes. 4314 */ 4315 switch (attrnamespace) { 4316 case EXTATTR_NAMESPACE_SYSTEM: 4317 /* Potentially should be: return (EPERM); */ 4318 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0)); 4319 case EXTATTR_NAMESPACE_USER: 4320 return (VOP_ACCESS(vp, accmode, cred, td)); 4321 default: 4322 return (EPERM); 4323 } 4324 } 4325 4326 #ifdef DEBUG_VFS_LOCKS 4327 /* 4328 * This only exists to supress warnings from unlocked specfs accesses. It is 4329 * no longer ok to have an unlocked VFS. 4330 */ 4331 #define IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL || \ 4332 (vp)->v_type == VCHR || (vp)->v_type == VBAD) 4333 4334 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 4335 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, 4336 "Drop into debugger on lock violation"); 4337 4338 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 4339 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 4340 0, "Check for interlock across VOPs"); 4341 4342 int vfs_badlock_print = 1; /* Print lock violations. */ 4343 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 4344 0, "Print lock violations"); 4345 4346 #ifdef KDB 4347 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 4348 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, 4349 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations"); 4350 #endif 4351 4352 static void 4353 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 4354 { 4355 4356 #ifdef KDB 4357 if (vfs_badlock_backtrace) 4358 kdb_backtrace(); 4359 #endif 4360 if (vfs_badlock_print) 4361 printf("%s: %p %s\n", str, (void *)vp, msg); 4362 if (vfs_badlock_ddb) 4363 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 4364 } 4365 4366 void 4367 assert_vi_locked(struct vnode *vp, const char *str) 4368 { 4369 4370 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 4371 vfs_badlock("interlock is not locked but should be", str, vp); 4372 } 4373 4374 void 4375 assert_vi_unlocked(struct vnode *vp, const char *str) 4376 { 4377 4378 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 4379 vfs_badlock("interlock is locked but should not be", str, vp); 4380 } 4381 4382 void 4383 assert_vop_locked(struct vnode *vp, const char *str) 4384 { 4385 int locked; 4386 4387 if (!IGNORE_LOCK(vp)) { 4388 locked = VOP_ISLOCKED(vp); 4389 if (locked == 0 || locked == LK_EXCLOTHER) 4390 vfs_badlock("is not locked but should be", str, vp); 4391 } 4392 } 4393 4394 void 4395 assert_vop_unlocked(struct vnode *vp, const char *str) 4396 { 4397 4398 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE) 4399 vfs_badlock("is locked but should not be", str, vp); 4400 } 4401 4402 void 4403 assert_vop_elocked(struct vnode *vp, const char *str) 4404 { 4405 4406 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 4407 vfs_badlock("is not exclusive locked but should be", str, vp); 4408 } 4409 4410 #if 0 4411 void 4412 assert_vop_elocked_other(struct vnode *vp, const char *str) 4413 { 4414 4415 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLOTHER) 4416 vfs_badlock("is not exclusive locked by another thread", 4417 str, vp); 4418 } 4419 4420 void 4421 assert_vop_slocked(struct vnode *vp, const char *str) 4422 { 4423 4424 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_SHARED) 4425 vfs_badlock("is not locked shared but should be", str, vp); 4426 } 4427 #endif /* 0 */ 4428 #endif /* DEBUG_VFS_LOCKS */ 4429 4430 void 4431 vop_rename_fail(struct vop_rename_args *ap) 4432 { 4433 4434 if (ap->a_tvp != NULL) 4435 vput(ap->a_tvp); 4436 if (ap->a_tdvp == ap->a_tvp) 4437 vrele(ap->a_tdvp); 4438 else 4439 vput(ap->a_tdvp); 4440 vrele(ap->a_fdvp); 4441 vrele(ap->a_fvp); 4442 } 4443 4444 void 4445 vop_rename_pre(void *ap) 4446 { 4447 struct vop_rename_args *a = ap; 4448 4449 #ifdef DEBUG_VFS_LOCKS 4450 if (a->a_tvp) 4451 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 4452 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 4453 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 4454 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 4455 4456 /* Check the source (from). */ 4457 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 4458 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 4459 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 4460 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 4461 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 4462 4463 /* Check the target. */ 4464 if (a->a_tvp) 4465 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 4466 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 4467 #endif 4468 if (a->a_tdvp != a->a_fdvp) 4469 vhold(a->a_fdvp); 4470 if (a->a_tvp != a->a_fvp) 4471 vhold(a->a_fvp); 4472 vhold(a->a_tdvp); 4473 if (a->a_tvp) 4474 vhold(a->a_tvp); 4475 } 4476 4477 void 4478 vop_strategy_pre(void *ap) 4479 { 4480 #ifdef DEBUG_VFS_LOCKS 4481 struct vop_strategy_args *a; 4482 struct buf *bp; 4483 4484 a = ap; 4485 bp = a->a_bp; 4486 4487 /* 4488 * Cluster ops lock their component buffers but not the IO container. 4489 */ 4490 if ((bp->b_flags & B_CLUSTER) != 0) 4491 return; 4492 4493 if (panicstr == NULL && !BUF_ISLOCKED(bp)) { 4494 if (vfs_badlock_print) 4495 printf( 4496 "VOP_STRATEGY: bp is not locked but should be\n"); 4497 if (vfs_badlock_ddb) 4498 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 4499 } 4500 #endif 4501 } 4502 4503 void 4504 vop_lock_pre(void *ap) 4505 { 4506 #ifdef DEBUG_VFS_LOCKS 4507 struct vop_lock1_args *a = ap; 4508 4509 if ((a->a_flags & LK_INTERLOCK) == 0) 4510 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 4511 else 4512 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 4513 #endif 4514 } 4515 4516 void 4517 vop_lock_post(void *ap, int rc) 4518 { 4519 #ifdef DEBUG_VFS_LOCKS 4520 struct vop_lock1_args *a = ap; 4521 4522 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 4523 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) 4524 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 4525 #endif 4526 } 4527 4528 void 4529 vop_unlock_pre(void *ap) 4530 { 4531 #ifdef DEBUG_VFS_LOCKS 4532 struct vop_unlock_args *a = ap; 4533 4534 if (a->a_flags & LK_INTERLOCK) 4535 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 4536 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 4537 #endif 4538 } 4539 4540 void 4541 vop_unlock_post(void *ap, int rc) 4542 { 4543 #ifdef DEBUG_VFS_LOCKS 4544 struct vop_unlock_args *a = ap; 4545 4546 if (a->a_flags & LK_INTERLOCK) 4547 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 4548 #endif 4549 } 4550 4551 void 4552 vop_create_post(void *ap, int rc) 4553 { 4554 struct vop_create_args *a = ap; 4555 4556 if (!rc) 4557 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4558 } 4559 4560 void 4561 vop_deleteextattr_post(void *ap, int rc) 4562 { 4563 struct vop_deleteextattr_args *a = ap; 4564 4565 if (!rc) 4566 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 4567 } 4568 4569 void 4570 vop_link_post(void *ap, int rc) 4571 { 4572 struct vop_link_args *a = ap; 4573 4574 if (!rc) { 4575 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK); 4576 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE); 4577 } 4578 } 4579 4580 void 4581 vop_mkdir_post(void *ap, int rc) 4582 { 4583 struct vop_mkdir_args *a = ap; 4584 4585 if (!rc) 4586 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 4587 } 4588 4589 void 4590 vop_mknod_post(void *ap, int rc) 4591 { 4592 struct vop_mknod_args *a = ap; 4593 4594 if (!rc) 4595 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4596 } 4597 4598 void 4599 vop_reclaim_post(void *ap, int rc) 4600 { 4601 struct vop_reclaim_args *a = ap; 4602 4603 if (!rc) 4604 VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE); 4605 } 4606 4607 void 4608 vop_remove_post(void *ap, int rc) 4609 { 4610 struct vop_remove_args *a = ap; 4611 4612 if (!rc) { 4613 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4614 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 4615 } 4616 } 4617 4618 void 4619 vop_rename_post(void *ap, int rc) 4620 { 4621 struct vop_rename_args *a = ap; 4622 4623 if (!rc) { 4624 VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE); 4625 VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE); 4626 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 4627 if (a->a_tvp) 4628 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 4629 } 4630 if (a->a_tdvp != a->a_fdvp) 4631 vdrop(a->a_fdvp); 4632 if (a->a_tvp != a->a_fvp) 4633 vdrop(a->a_fvp); 4634 vdrop(a->a_tdvp); 4635 if (a->a_tvp) 4636 vdrop(a->a_tvp); 4637 } 4638 4639 void 4640 vop_rmdir_post(void *ap, int rc) 4641 { 4642 struct vop_rmdir_args *a = ap; 4643 4644 if (!rc) { 4645 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 4646 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 4647 } 4648 } 4649 4650 void 4651 vop_setattr_post(void *ap, int rc) 4652 { 4653 struct vop_setattr_args *a = ap; 4654 4655 if (!rc) 4656 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 4657 } 4658 4659 void 4660 vop_setextattr_post(void *ap, int rc) 4661 { 4662 struct vop_setextattr_args *a = ap; 4663 4664 if (!rc) 4665 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 4666 } 4667 4668 void 4669 vop_symlink_post(void *ap, int rc) 4670 { 4671 struct vop_symlink_args *a = ap; 4672 4673 if (!rc) 4674 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4675 } 4676 4677 static struct knlist fs_knlist; 4678 4679 static void 4680 vfs_event_init(void *arg) 4681 { 4682 knlist_init_mtx(&fs_knlist, NULL); 4683 } 4684 /* XXX - correct order? */ 4685 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 4686 4687 void 4688 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 4689 { 4690 4691 KNOTE_UNLOCKED(&fs_knlist, event); 4692 } 4693 4694 static int filt_fsattach(struct knote *kn); 4695 static void filt_fsdetach(struct knote *kn); 4696 static int filt_fsevent(struct knote *kn, long hint); 4697 4698 struct filterops fs_filtops = { 4699 .f_isfd = 0, 4700 .f_attach = filt_fsattach, 4701 .f_detach = filt_fsdetach, 4702 .f_event = filt_fsevent 4703 }; 4704 4705 static int 4706 filt_fsattach(struct knote *kn) 4707 { 4708 4709 kn->kn_flags |= EV_CLEAR; 4710 knlist_add(&fs_knlist, kn, 0); 4711 return (0); 4712 } 4713 4714 static void 4715 filt_fsdetach(struct knote *kn) 4716 { 4717 4718 knlist_remove(&fs_knlist, kn, 0); 4719 } 4720 4721 static int 4722 filt_fsevent(struct knote *kn, long hint) 4723 { 4724 4725 kn->kn_fflags |= hint; 4726 return (kn->kn_fflags != 0); 4727 } 4728 4729 static int 4730 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 4731 { 4732 struct vfsidctl vc; 4733 int error; 4734 struct mount *mp; 4735 4736 error = SYSCTL_IN(req, &vc, sizeof(vc)); 4737 if (error) 4738 return (error); 4739 if (vc.vc_vers != VFS_CTL_VERS1) 4740 return (EINVAL); 4741 mp = vfs_getvfs(&vc.vc_fsid); 4742 if (mp == NULL) 4743 return (ENOENT); 4744 /* ensure that a specific sysctl goes to the right filesystem. */ 4745 if (strcmp(vc.vc_fstypename, "*") != 0 && 4746 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 4747 vfs_rel(mp); 4748 return (EINVAL); 4749 } 4750 VCTLTOREQ(&vc, req); 4751 error = VFS_SYSCTL(mp, vc.vc_op, req); 4752 vfs_rel(mp); 4753 return (error); 4754 } 4755 4756 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR, 4757 NULL, 0, sysctl_vfs_ctl, "", 4758 "Sysctl by fsid"); 4759 4760 /* 4761 * Function to initialize a va_filerev field sensibly. 4762 * XXX: Wouldn't a random number make a lot more sense ?? 4763 */ 4764 u_quad_t 4765 init_va_filerev(void) 4766 { 4767 struct bintime bt; 4768 4769 getbinuptime(&bt); 4770 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 4771 } 4772 4773 static int filt_vfsread(struct knote *kn, long hint); 4774 static int filt_vfswrite(struct knote *kn, long hint); 4775 static int filt_vfsvnode(struct knote *kn, long hint); 4776 static void filt_vfsdetach(struct knote *kn); 4777 static struct filterops vfsread_filtops = { 4778 .f_isfd = 1, 4779 .f_detach = filt_vfsdetach, 4780 .f_event = filt_vfsread 4781 }; 4782 static struct filterops vfswrite_filtops = { 4783 .f_isfd = 1, 4784 .f_detach = filt_vfsdetach, 4785 .f_event = filt_vfswrite 4786 }; 4787 static struct filterops vfsvnode_filtops = { 4788 .f_isfd = 1, 4789 .f_detach = filt_vfsdetach, 4790 .f_event = filt_vfsvnode 4791 }; 4792 4793 static void 4794 vfs_knllock(void *arg) 4795 { 4796 struct vnode *vp = arg; 4797 4798 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4799 } 4800 4801 static void 4802 vfs_knlunlock(void *arg) 4803 { 4804 struct vnode *vp = arg; 4805 4806 VOP_UNLOCK(vp, 0); 4807 } 4808 4809 static void 4810 vfs_knl_assert_locked(void *arg) 4811 { 4812 #ifdef DEBUG_VFS_LOCKS 4813 struct vnode *vp = arg; 4814 4815 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 4816 #endif 4817 } 4818 4819 static void 4820 vfs_knl_assert_unlocked(void *arg) 4821 { 4822 #ifdef DEBUG_VFS_LOCKS 4823 struct vnode *vp = arg; 4824 4825 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 4826 #endif 4827 } 4828 4829 int 4830 vfs_kqfilter(struct vop_kqfilter_args *ap) 4831 { 4832 struct vnode *vp = ap->a_vp; 4833 struct knote *kn = ap->a_kn; 4834 struct knlist *knl; 4835 4836 switch (kn->kn_filter) { 4837 case EVFILT_READ: 4838 kn->kn_fop = &vfsread_filtops; 4839 break; 4840 case EVFILT_WRITE: 4841 kn->kn_fop = &vfswrite_filtops; 4842 break; 4843 case EVFILT_VNODE: 4844 kn->kn_fop = &vfsvnode_filtops; 4845 break; 4846 default: 4847 return (EINVAL); 4848 } 4849 4850 kn->kn_hook = (caddr_t)vp; 4851 4852 v_addpollinfo(vp); 4853 if (vp->v_pollinfo == NULL) 4854 return (ENOMEM); 4855 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 4856 vhold(vp); 4857 knlist_add(knl, kn, 0); 4858 4859 return (0); 4860 } 4861 4862 /* 4863 * Detach knote from vnode 4864 */ 4865 static void 4866 filt_vfsdetach(struct knote *kn) 4867 { 4868 struct vnode *vp = (struct vnode *)kn->kn_hook; 4869 4870 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 4871 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 4872 vdrop(vp); 4873 } 4874 4875 /*ARGSUSED*/ 4876 static int 4877 filt_vfsread(struct knote *kn, long hint) 4878 { 4879 struct vnode *vp = (struct vnode *)kn->kn_hook; 4880 struct vattr va; 4881 int res; 4882 4883 /* 4884 * filesystem is gone, so set the EOF flag and schedule 4885 * the knote for deletion. 4886 */ 4887 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 4888 VI_LOCK(vp); 4889 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 4890 VI_UNLOCK(vp); 4891 return (1); 4892 } 4893 4894 if (VOP_GETATTR(vp, &va, curthread->td_ucred)) 4895 return (0); 4896 4897 VI_LOCK(vp); 4898 kn->kn_data = va.va_size - kn->kn_fp->f_offset; 4899 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; 4900 VI_UNLOCK(vp); 4901 return (res); 4902 } 4903 4904 /*ARGSUSED*/ 4905 static int 4906 filt_vfswrite(struct knote *kn, long hint) 4907 { 4908 struct vnode *vp = (struct vnode *)kn->kn_hook; 4909 4910 VI_LOCK(vp); 4911 4912 /* 4913 * filesystem is gone, so set the EOF flag and schedule 4914 * the knote for deletion. 4915 */ 4916 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) 4917 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 4918 4919 kn->kn_data = 0; 4920 VI_UNLOCK(vp); 4921 return (1); 4922 } 4923 4924 static int 4925 filt_vfsvnode(struct knote *kn, long hint) 4926 { 4927 struct vnode *vp = (struct vnode *)kn->kn_hook; 4928 int res; 4929 4930 VI_LOCK(vp); 4931 if (kn->kn_sfflags & hint) 4932 kn->kn_fflags |= hint; 4933 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 4934 kn->kn_flags |= EV_EOF; 4935 VI_UNLOCK(vp); 4936 return (1); 4937 } 4938 res = (kn->kn_fflags != 0); 4939 VI_UNLOCK(vp); 4940 return (res); 4941 } 4942 4943 int 4944 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 4945 { 4946 int error; 4947 4948 if (dp->d_reclen > ap->a_uio->uio_resid) 4949 return (ENAMETOOLONG); 4950 error = uiomove(dp, dp->d_reclen, ap->a_uio); 4951 if (error) { 4952 if (ap->a_ncookies != NULL) { 4953 if (ap->a_cookies != NULL) 4954 free(ap->a_cookies, M_TEMP); 4955 ap->a_cookies = NULL; 4956 *ap->a_ncookies = 0; 4957 } 4958 return (error); 4959 } 4960 if (ap->a_ncookies == NULL) 4961 return (0); 4962 4963 KASSERT(ap->a_cookies, 4964 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 4965 4966 *ap->a_cookies = realloc(*ap->a_cookies, 4967 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); 4968 (*ap->a_cookies)[*ap->a_ncookies] = off; 4969 return (0); 4970 } 4971 4972 /* 4973 * Mark for update the access time of the file if the filesystem 4974 * supports VOP_MARKATIME. This functionality is used by execve and 4975 * mmap, so we want to avoid the I/O implied by directly setting 4976 * va_atime for the sake of efficiency. 4977 */ 4978 void 4979 vfs_mark_atime(struct vnode *vp, struct ucred *cred) 4980 { 4981 struct mount *mp; 4982 4983 mp = vp->v_mount; 4984 ASSERT_VOP_LOCKED(vp, "vfs_mark_atime"); 4985 if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) 4986 (void)VOP_MARKATIME(vp); 4987 } 4988 4989 /* 4990 * The purpose of this routine is to remove granularity from accmode_t, 4991 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 4992 * VADMIN and VAPPEND. 4993 * 4994 * If it returns 0, the caller is supposed to continue with the usual 4995 * access checks using 'accmode' as modified by this routine. If it 4996 * returns nonzero value, the caller is supposed to return that value 4997 * as errno. 4998 * 4999 * Note that after this routine runs, accmode may be zero. 5000 */ 5001 int 5002 vfs_unixify_accmode(accmode_t *accmode) 5003 { 5004 /* 5005 * There is no way to specify explicit "deny" rule using 5006 * file mode or POSIX.1e ACLs. 5007 */ 5008 if (*accmode & VEXPLICIT_DENY) { 5009 *accmode = 0; 5010 return (0); 5011 } 5012 5013 /* 5014 * None of these can be translated into usual access bits. 5015 * Also, the common case for NFSv4 ACLs is to not contain 5016 * either of these bits. Caller should check for VWRITE 5017 * on the containing directory instead. 5018 */ 5019 if (*accmode & (VDELETE_CHILD | VDELETE)) 5020 return (EPERM); 5021 5022 if (*accmode & VADMIN_PERMS) { 5023 *accmode &= ~VADMIN_PERMS; 5024 *accmode |= VADMIN; 5025 } 5026 5027 /* 5028 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 5029 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 5030 */ 5031 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 5032 5033 return (0); 5034 } 5035 5036 /* 5037 * These are helper functions for filesystems to traverse all 5038 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 5039 * 5040 * This interface replaces MNT_VNODE_FOREACH. 5041 */ 5042 5043 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker"); 5044 5045 struct vnode * 5046 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 5047 { 5048 struct vnode *vp; 5049 5050 if (should_yield()) 5051 kern_yield(PRI_USER); 5052 MNT_ILOCK(mp); 5053 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5054 vp = TAILQ_NEXT(*mvp, v_nmntvnodes); 5055 while (vp != NULL && (vp->v_type == VMARKER || 5056 (vp->v_iflag & VI_DOOMED) != 0)) 5057 vp = TAILQ_NEXT(vp, v_nmntvnodes); 5058 5059 /* Check if we are done */ 5060 if (vp == NULL) { 5061 __mnt_vnode_markerfree_all(mvp, mp); 5062 /* MNT_IUNLOCK(mp); -- done in above function */ 5063 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 5064 return (NULL); 5065 } 5066 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 5067 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 5068 VI_LOCK(vp); 5069 MNT_IUNLOCK(mp); 5070 return (vp); 5071 } 5072 5073 struct vnode * 5074 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 5075 { 5076 struct vnode *vp; 5077 5078 *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 5079 MNT_ILOCK(mp); 5080 MNT_REF(mp); 5081 (*mvp)->v_type = VMARKER; 5082 5083 vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 5084 while (vp != NULL && (vp->v_type == VMARKER || 5085 (vp->v_iflag & VI_DOOMED) != 0)) 5086 vp = TAILQ_NEXT(vp, v_nmntvnodes); 5087 5088 /* Check if we are done */ 5089 if (vp == NULL) { 5090 MNT_REL(mp); 5091 MNT_IUNLOCK(mp); 5092 free(*mvp, M_VNODE_MARKER); 5093 *mvp = NULL; 5094 return (NULL); 5095 } 5096 (*mvp)->v_mount = mp; 5097 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 5098 VI_LOCK(vp); 5099 MNT_IUNLOCK(mp); 5100 return (vp); 5101 } 5102 5103 5104 void 5105 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 5106 { 5107 5108 if (*mvp == NULL) { 5109 MNT_IUNLOCK(mp); 5110 return; 5111 } 5112 5113 mtx_assert(MNT_MTX(mp), MA_OWNED); 5114 5115 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5116 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 5117 MNT_REL(mp); 5118 MNT_IUNLOCK(mp); 5119 free(*mvp, M_VNODE_MARKER); 5120 *mvp = NULL; 5121 } 5122 5123 /* 5124 * These are helper functions for filesystems to traverse their 5125 * active vnodes. See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h 5126 */ 5127 static void 5128 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp) 5129 { 5130 5131 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5132 5133 MNT_ILOCK(mp); 5134 MNT_REL(mp); 5135 MNT_IUNLOCK(mp); 5136 free(*mvp, M_VNODE_MARKER); 5137 *mvp = NULL; 5138 } 5139 5140 static struct vnode * 5141 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp) 5142 { 5143 struct vnode *vp, *nvp; 5144 5145 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 5146 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5147 restart: 5148 vp = TAILQ_NEXT(*mvp, v_actfreelist); 5149 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist); 5150 while (vp != NULL) { 5151 if (vp->v_type == VMARKER) { 5152 vp = TAILQ_NEXT(vp, v_actfreelist); 5153 continue; 5154 } 5155 if (!VI_TRYLOCK(vp)) { 5156 if (mp_ncpus == 1 || should_yield()) { 5157 TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist); 5158 mtx_unlock(&vnode_free_list_mtx); 5159 pause("vnacti", 1); 5160 mtx_lock(&vnode_free_list_mtx); 5161 goto restart; 5162 } 5163 continue; 5164 } 5165 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 5166 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 5167 ("alien vnode on the active list %p %p", vp, mp)); 5168 if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0) 5169 break; 5170 nvp = TAILQ_NEXT(vp, v_actfreelist); 5171 VI_UNLOCK(vp); 5172 vp = nvp; 5173 } 5174 5175 /* Check if we are done */ 5176 if (vp == NULL) { 5177 mtx_unlock(&vnode_free_list_mtx); 5178 mnt_vnode_markerfree_active(mvp, mp); 5179 return (NULL); 5180 } 5181 TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist); 5182 mtx_unlock(&vnode_free_list_mtx); 5183 ASSERT_VI_LOCKED(vp, "active iter"); 5184 KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp)); 5185 return (vp); 5186 } 5187 5188 struct vnode * 5189 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp) 5190 { 5191 5192 if (should_yield()) 5193 kern_yield(PRI_USER); 5194 mtx_lock(&vnode_free_list_mtx); 5195 return (mnt_vnode_next_active(mvp, mp)); 5196 } 5197 5198 struct vnode * 5199 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp) 5200 { 5201 struct vnode *vp; 5202 5203 *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 5204 MNT_ILOCK(mp); 5205 MNT_REF(mp); 5206 MNT_IUNLOCK(mp); 5207 (*mvp)->v_type = VMARKER; 5208 (*mvp)->v_mount = mp; 5209 5210 mtx_lock(&vnode_free_list_mtx); 5211 vp = TAILQ_FIRST(&mp->mnt_activevnodelist); 5212 if (vp == NULL) { 5213 mtx_unlock(&vnode_free_list_mtx); 5214 mnt_vnode_markerfree_active(mvp, mp); 5215 return (NULL); 5216 } 5217 TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist); 5218 return (mnt_vnode_next_active(mvp, mp)); 5219 } 5220 5221 void 5222 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp) 5223 { 5224 5225 if (*mvp == NULL) 5226 return; 5227 5228 mtx_lock(&vnode_free_list_mtx); 5229 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist); 5230 mtx_unlock(&vnode_free_list_mtx); 5231 mnt_vnode_markerfree_active(mvp, mp); 5232 } 5233