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 * 3. 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/counter.h> 55 #include <sys/dirent.h> 56 #include <sys/event.h> 57 #include <sys/eventhandler.h> 58 #include <sys/extattr.h> 59 #include <sys/file.h> 60 #include <sys/fcntl.h> 61 #include <sys/jail.h> 62 #include <sys/kdb.h> 63 #include <sys/kernel.h> 64 #include <sys/kthread.h> 65 #include <sys/lockf.h> 66 #include <sys/malloc.h> 67 #include <sys/mount.h> 68 #include <sys/namei.h> 69 #include <sys/pctrie.h> 70 #include <sys/priv.h> 71 #include <sys/reboot.h> 72 #include <sys/refcount.h> 73 #include <sys/rwlock.h> 74 #include <sys/sched.h> 75 #include <sys/sleepqueue.h> 76 #include <sys/smp.h> 77 #include <sys/stat.h> 78 #include <sys/sysctl.h> 79 #include <sys/syslog.h> 80 #include <sys/vmmeter.h> 81 #include <sys/vnode.h> 82 #include <sys/watchdog.h> 83 84 #include <machine/stdarg.h> 85 86 #include <security/mac/mac_framework.h> 87 88 #include <vm/vm.h> 89 #include <vm/vm_object.h> 90 #include <vm/vm_extern.h> 91 #include <vm/pmap.h> 92 #include <vm/vm_map.h> 93 #include <vm/vm_page.h> 94 #include <vm/vm_kern.h> 95 #include <vm/uma.h> 96 97 #ifdef DDB 98 #include <ddb/ddb.h> 99 #endif 100 101 static void delmntque(struct vnode *vp); 102 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, 103 int slpflag, int slptimeo); 104 static void syncer_shutdown(void *arg, int howto); 105 static int vtryrecycle(struct vnode *vp); 106 static void v_init_counters(struct vnode *); 107 static void v_incr_usecount(struct vnode *); 108 static void v_incr_usecount_locked(struct vnode *); 109 static void v_incr_devcount(struct vnode *); 110 static void v_decr_devcount(struct vnode *); 111 static void vgonel(struct vnode *); 112 static void vfs_knllock(void *arg); 113 static void vfs_knlunlock(void *arg); 114 static void vfs_knl_assert_locked(void *arg); 115 static void vfs_knl_assert_unlocked(void *arg); 116 static void vnlru_return_batches(struct vfsops *mnt_op); 117 static void destroy_vpollinfo(struct vpollinfo *vi); 118 119 /* 120 * Number of vnodes in existence. Increased whenever getnewvnode() 121 * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode. 122 */ 123 static unsigned long numvnodes; 124 125 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, 126 "Number of vnodes in existence"); 127 128 static counter_u64_t vnodes_created; 129 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created, 130 "Number of vnodes created by getnewvnode"); 131 132 static u_long mnt_free_list_batch = 128; 133 SYSCTL_ULONG(_vfs, OID_AUTO, mnt_free_list_batch, CTLFLAG_RW, 134 &mnt_free_list_batch, 0, "Limit of vnodes held on mnt's free list"); 135 136 /* 137 * Conversion tables for conversion from vnode types to inode formats 138 * and back. 139 */ 140 enum vtype iftovt_tab[16] = { 141 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 142 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 143 }; 144 int vttoif_tab[10] = { 145 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 146 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT 147 }; 148 149 /* 150 * List of vnodes that are ready for recycling. 151 */ 152 static TAILQ_HEAD(freelst, vnode) vnode_free_list; 153 154 /* 155 * "Free" vnode target. Free vnodes are rarely completely free, but are 156 * just ones that are cheap to recycle. Usually they are for files which 157 * have been stat'd but not read; these usually have inode and namecache 158 * data attached to them. This target is the preferred minimum size of a 159 * sub-cache consisting mostly of such files. The system balances the size 160 * of this sub-cache with its complement to try to prevent either from 161 * thrashing while the other is relatively inactive. The targets express 162 * a preference for the best balance. 163 * 164 * "Above" this target there are 2 further targets (watermarks) related 165 * to recyling of free vnodes. In the best-operating case, the cache is 166 * exactly full, the free list has size between vlowat and vhiwat above the 167 * free target, and recycling from it and normal use maintains this state. 168 * Sometimes the free list is below vlowat or even empty, but this state 169 * is even better for immediate use provided the cache is not full. 170 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free 171 * ones) to reach one of these states. The watermarks are currently hard- 172 * coded as 4% and 9% of the available space higher. These and the default 173 * of 25% for wantfreevnodes are too large if the memory size is large. 174 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim 175 * whenever vnlru_proc() becomes active. 176 */ 177 static u_long wantfreevnodes; 178 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, 179 &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes"); 180 static u_long freevnodes; 181 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, 182 &freevnodes, 0, "Number of \"free\" vnodes"); 183 184 static counter_u64_t recycles_count; 185 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count, 186 "Number of vnodes recycled to meet vnode cache targets"); 187 188 /* 189 * Various variables used for debugging the new implementation of 190 * reassignbuf(). 191 * XXX these are probably of (very) limited utility now. 192 */ 193 static int reassignbufcalls; 194 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, 195 "Number of calls to reassignbuf"); 196 197 static counter_u64_t free_owe_inact; 198 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact, 199 "Number of times free vnodes kept on active list due to VFS " 200 "owing inactivation"); 201 202 /* To keep more than one thread at a time from running vfs_getnewfsid */ 203 static struct mtx mntid_mtx; 204 205 /* 206 * Lock for any access to the following: 207 * vnode_free_list 208 * numvnodes 209 * freevnodes 210 */ 211 static struct mtx vnode_free_list_mtx; 212 213 /* Publicly exported FS */ 214 struct nfs_public nfs_pub; 215 216 static uma_zone_t buf_trie_zone; 217 218 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 219 static uma_zone_t vnode_zone; 220 static uma_zone_t vnodepoll_zone; 221 222 /* 223 * The workitem queue. 224 * 225 * It is useful to delay writes of file data and filesystem metadata 226 * for tens of seconds so that quickly created and deleted files need 227 * not waste disk bandwidth being created and removed. To realize this, 228 * we append vnodes to a "workitem" queue. When running with a soft 229 * updates implementation, most pending metadata dependencies should 230 * not wait for more than a few seconds. Thus, mounted on block devices 231 * are delayed only about a half the time that file data is delayed. 232 * Similarly, directory updates are more critical, so are only delayed 233 * about a third the time that file data is delayed. Thus, there are 234 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 235 * one each second (driven off the filesystem syncer process). The 236 * syncer_delayno variable indicates the next queue that is to be processed. 237 * Items that need to be processed soon are placed in this queue: 238 * 239 * syncer_workitem_pending[syncer_delayno] 240 * 241 * A delay of fifteen seconds is done by placing the request fifteen 242 * entries later in the queue: 243 * 244 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 245 * 246 */ 247 static int syncer_delayno; 248 static long syncer_mask; 249 LIST_HEAD(synclist, bufobj); 250 static struct synclist *syncer_workitem_pending; 251 /* 252 * The sync_mtx protects: 253 * bo->bo_synclist 254 * sync_vnode_count 255 * syncer_delayno 256 * syncer_state 257 * syncer_workitem_pending 258 * syncer_worklist_len 259 * rushjob 260 */ 261 static struct mtx sync_mtx; 262 static struct cv sync_wakeup; 263 264 #define SYNCER_MAXDELAY 32 265 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 266 static int syncdelay = 30; /* max time to delay syncing data */ 267 static int filedelay = 30; /* time to delay syncing files */ 268 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, 269 "Time to delay syncing files (in seconds)"); 270 static int dirdelay = 29; /* time to delay syncing directories */ 271 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, 272 "Time to delay syncing directories (in seconds)"); 273 static int metadelay = 28; /* time to delay syncing metadata */ 274 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, 275 "Time to delay syncing metadata (in seconds)"); 276 static int rushjob; /* number of slots to run ASAP */ 277 static int stat_rush_requests; /* number of times I/O speeded up */ 278 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, 279 "Number of times I/O speeded up (rush requests)"); 280 281 /* 282 * When shutting down the syncer, run it at four times normal speed. 283 */ 284 #define SYNCER_SHUTDOWN_SPEEDUP 4 285 static int sync_vnode_count; 286 static int syncer_worklist_len; 287 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 288 syncer_state; 289 290 /* Target for maximum number of vnodes. */ 291 int desiredvnodes; 292 static int gapvnodes; /* gap between wanted and desired */ 293 static int vhiwat; /* enough extras after expansion */ 294 static int vlowat; /* minimal extras before expansion */ 295 static int vstir; /* nonzero to stir non-free vnodes */ 296 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */ 297 298 static int 299 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS) 300 { 301 int error, old_desiredvnodes; 302 303 old_desiredvnodes = desiredvnodes; 304 if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0) 305 return (error); 306 if (old_desiredvnodes != desiredvnodes) { 307 wantfreevnodes = desiredvnodes / 4; 308 /* XXX locking seems to be incomplete. */ 309 vfs_hash_changesize(desiredvnodes); 310 cache_changesize(desiredvnodes); 311 } 312 return (0); 313 } 314 315 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes, 316 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0, 317 sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes"); 318 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 319 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)"); 320 static int vnlru_nowhere; 321 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 322 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 323 324 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */ 325 static int vnsz2log; 326 327 /* 328 * Support for the bufobj clean & dirty pctrie. 329 */ 330 static void * 331 buf_trie_alloc(struct pctrie *ptree) 332 { 333 334 return uma_zalloc(buf_trie_zone, M_NOWAIT); 335 } 336 337 static void 338 buf_trie_free(struct pctrie *ptree, void *node) 339 { 340 341 uma_zfree(buf_trie_zone, node); 342 } 343 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free); 344 345 /* 346 * Initialize the vnode management data structures. 347 * 348 * Reevaluate the following cap on the number of vnodes after the physical 349 * memory size exceeds 512GB. In the limit, as the physical memory size 350 * grows, the ratio of the memory size in KB to to vnodes approaches 64:1. 351 */ 352 #ifndef MAXVNODES_MAX 353 #define MAXVNODES_MAX (512 * 1024 * 1024 / 64) /* 8M */ 354 #endif 355 356 /* 357 * Initialize a vnode as it first enters the zone. 358 */ 359 static int 360 vnode_init(void *mem, int size, int flags) 361 { 362 struct vnode *vp; 363 struct bufobj *bo; 364 365 vp = mem; 366 bzero(vp, size); 367 /* 368 * Setup locks. 369 */ 370 vp->v_vnlock = &vp->v_lock; 371 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 372 /* 373 * By default, don't allow shared locks unless filesystems opt-in. 374 */ 375 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT, 376 LK_NOSHARE | LK_IS_VNODE); 377 /* 378 * Initialize bufobj. 379 */ 380 bo = &vp->v_bufobj; 381 rw_init(BO_LOCKPTR(bo), "bufobj interlock"); 382 bo->bo_private = vp; 383 TAILQ_INIT(&bo->bo_clean.bv_hd); 384 TAILQ_INIT(&bo->bo_dirty.bv_hd); 385 /* 386 * Initialize namecache. 387 */ 388 LIST_INIT(&vp->v_cache_src); 389 TAILQ_INIT(&vp->v_cache_dst); 390 /* 391 * Initialize rangelocks. 392 */ 393 rangelock_init(&vp->v_rl); 394 return (0); 395 } 396 397 /* 398 * Free a vnode when it is cleared from the zone. 399 */ 400 static void 401 vnode_fini(void *mem, int size) 402 { 403 struct vnode *vp; 404 struct bufobj *bo; 405 406 vp = mem; 407 rangelock_destroy(&vp->v_rl); 408 lockdestroy(vp->v_vnlock); 409 mtx_destroy(&vp->v_interlock); 410 bo = &vp->v_bufobj; 411 rw_destroy(BO_LOCKPTR(bo)); 412 } 413 414 /* 415 * Provide the size of NFS nclnode and NFS fh for calculation of the 416 * vnode memory consumption. The size is specified directly to 417 * eliminate dependency on NFS-private header. 418 * 419 * Other filesystems may use bigger or smaller (like UFS and ZFS) 420 * private inode data, but the NFS-based estimation is ample enough. 421 * Still, we care about differences in the size between 64- and 32-bit 422 * platforms. 423 * 424 * Namecache structure size is heuristically 425 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1. 426 */ 427 #ifdef _LP64 428 #define NFS_NCLNODE_SZ (528 + 64) 429 #define NC_SZ 148 430 #else 431 #define NFS_NCLNODE_SZ (360 + 32) 432 #define NC_SZ 92 433 #endif 434 435 static void 436 vntblinit(void *dummy __unused) 437 { 438 u_int i; 439 int physvnodes, virtvnodes; 440 441 /* 442 * Desiredvnodes is a function of the physical memory size and the 443 * kernel's heap size. Generally speaking, it scales with the 444 * physical memory size. The ratio of desiredvnodes to the physical 445 * memory size is 1:16 until desiredvnodes exceeds 98,304. 446 * Thereafter, the 447 * marginal ratio of desiredvnodes to the physical memory size is 448 * 1:64. However, desiredvnodes is limited by the kernel's heap 449 * size. The memory required by desiredvnodes vnodes and vm objects 450 * must not exceed 1/10th of the kernel's heap size. 451 */ 452 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 + 453 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64; 454 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) + 455 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ)); 456 desiredvnodes = min(physvnodes, virtvnodes); 457 if (desiredvnodes > MAXVNODES_MAX) { 458 if (bootverbose) 459 printf("Reducing kern.maxvnodes %d -> %d\n", 460 desiredvnodes, MAXVNODES_MAX); 461 desiredvnodes = MAXVNODES_MAX; 462 } 463 wantfreevnodes = desiredvnodes / 4; 464 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 465 TAILQ_INIT(&vnode_free_list); 466 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 467 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 468 vnode_init, vnode_fini, UMA_ALIGN_PTR, 0); 469 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 470 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 471 /* 472 * Preallocate enough nodes to support one-per buf so that 473 * we can not fail an insert. reassignbuf() callers can not 474 * tolerate the insertion failure. 475 */ 476 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(), 477 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, 478 UMA_ZONE_NOFREE | UMA_ZONE_VM); 479 uma_prealloc(buf_trie_zone, nbuf); 480 481 vnodes_created = counter_u64_alloc(M_WAITOK); 482 recycles_count = counter_u64_alloc(M_WAITOK); 483 free_owe_inact = counter_u64_alloc(M_WAITOK); 484 485 /* 486 * Initialize the filesystem syncer. 487 */ 488 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 489 &syncer_mask); 490 syncer_maxdelay = syncer_mask + 1; 491 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 492 cv_init(&sync_wakeup, "syncer"); 493 for (i = 1; i <= sizeof(struct vnode); i <<= 1) 494 vnsz2log++; 495 vnsz2log--; 496 } 497 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL); 498 499 500 /* 501 * Mark a mount point as busy. Used to synchronize access and to delay 502 * unmounting. Eventually, mountlist_mtx is not released on failure. 503 * 504 * vfs_busy() is a custom lock, it can block the caller. 505 * vfs_busy() only sleeps if the unmount is active on the mount point. 506 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any 507 * vnode belonging to mp. 508 * 509 * Lookup uses vfs_busy() to traverse mount points. 510 * root fs var fs 511 * / vnode lock A / vnode lock (/var) D 512 * /var vnode lock B /log vnode lock(/var/log) E 513 * vfs_busy lock C vfs_busy lock F 514 * 515 * Within each file system, the lock order is C->A->B and F->D->E. 516 * 517 * When traversing across mounts, the system follows that lock order: 518 * 519 * C->A->B 520 * | 521 * +->F->D->E 522 * 523 * The lookup() process for namei("/var") illustrates the process: 524 * VOP_LOOKUP() obtains B while A is held 525 * vfs_busy() obtains a shared lock on F while A and B are held 526 * vput() releases lock on B 527 * vput() releases lock on A 528 * VFS_ROOT() obtains lock on D while shared lock on F is held 529 * vfs_unbusy() releases shared lock on F 530 * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A. 531 * Attempt to lock A (instead of vp_crossmp) while D is held would 532 * violate the global order, causing deadlocks. 533 * 534 * dounmount() locks B while F is drained. 535 */ 536 int 537 vfs_busy(struct mount *mp, int flags) 538 { 539 540 MPASS((flags & ~MBF_MASK) == 0); 541 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags); 542 543 MNT_ILOCK(mp); 544 MNT_REF(mp); 545 /* 546 * If mount point is currently being unmounted, sleep until the 547 * mount point fate is decided. If thread doing the unmounting fails, 548 * it will clear MNTK_UNMOUNT flag before waking us up, indicating 549 * that this mount point has survived the unmount attempt and vfs_busy 550 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE 551 * flag in addition to MNTK_UNMOUNT, indicating that mount point is 552 * about to be really destroyed. vfs_busy needs to release its 553 * reference on the mount point in this case and return with ENOENT, 554 * telling the caller that mount mount it tried to busy is no longer 555 * valid. 556 */ 557 while (mp->mnt_kern_flag & MNTK_UNMOUNT) { 558 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) { 559 MNT_REL(mp); 560 MNT_IUNLOCK(mp); 561 CTR1(KTR_VFS, "%s: failed busying before sleeping", 562 __func__); 563 return (ENOENT); 564 } 565 if (flags & MBF_MNTLSTLOCK) 566 mtx_unlock(&mountlist_mtx); 567 mp->mnt_kern_flag |= MNTK_MWAIT; 568 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0); 569 if (flags & MBF_MNTLSTLOCK) 570 mtx_lock(&mountlist_mtx); 571 MNT_ILOCK(mp); 572 } 573 if (flags & MBF_MNTLSTLOCK) 574 mtx_unlock(&mountlist_mtx); 575 mp->mnt_lockref++; 576 MNT_IUNLOCK(mp); 577 return (0); 578 } 579 580 /* 581 * Free a busy filesystem. 582 */ 583 void 584 vfs_unbusy(struct mount *mp) 585 { 586 587 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 588 MNT_ILOCK(mp); 589 MNT_REL(mp); 590 KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref")); 591 mp->mnt_lockref--; 592 if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) { 593 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT); 594 CTR1(KTR_VFS, "%s: waking up waiters", __func__); 595 mp->mnt_kern_flag &= ~MNTK_DRAINING; 596 wakeup(&mp->mnt_lockref); 597 } 598 MNT_IUNLOCK(mp); 599 } 600 601 /* 602 * Lookup a mount point by filesystem identifier. 603 */ 604 struct mount * 605 vfs_getvfs(fsid_t *fsid) 606 { 607 struct mount *mp; 608 609 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 610 mtx_lock(&mountlist_mtx); 611 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 612 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 613 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 614 vfs_ref(mp); 615 mtx_unlock(&mountlist_mtx); 616 return (mp); 617 } 618 } 619 mtx_unlock(&mountlist_mtx); 620 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 621 return ((struct mount *) 0); 622 } 623 624 /* 625 * Lookup a mount point by filesystem identifier, busying it before 626 * returning. 627 * 628 * To avoid congestion on mountlist_mtx, implement simple direct-mapped 629 * cache for popular filesystem identifiers. The cache is lockess, using 630 * the fact that struct mount's are never freed. In worst case we may 631 * get pointer to unmounted or even different filesystem, so we have to 632 * check what we got, and go slow way if so. 633 */ 634 struct mount * 635 vfs_busyfs(fsid_t *fsid) 636 { 637 #define FSID_CACHE_SIZE 256 638 typedef struct mount * volatile vmp_t; 639 static vmp_t cache[FSID_CACHE_SIZE]; 640 struct mount *mp; 641 int error; 642 uint32_t hash; 643 644 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 645 hash = fsid->val[0] ^ fsid->val[1]; 646 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1); 647 mp = cache[hash]; 648 if (mp == NULL || 649 mp->mnt_stat.f_fsid.val[0] != fsid->val[0] || 650 mp->mnt_stat.f_fsid.val[1] != fsid->val[1]) 651 goto slow; 652 if (vfs_busy(mp, 0) != 0) { 653 cache[hash] = NULL; 654 goto slow; 655 } 656 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 657 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) 658 return (mp); 659 else 660 vfs_unbusy(mp); 661 662 slow: 663 mtx_lock(&mountlist_mtx); 664 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 665 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 666 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 667 error = vfs_busy(mp, MBF_MNTLSTLOCK); 668 if (error) { 669 cache[hash] = NULL; 670 mtx_unlock(&mountlist_mtx); 671 return (NULL); 672 } 673 cache[hash] = mp; 674 return (mp); 675 } 676 } 677 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 678 mtx_unlock(&mountlist_mtx); 679 return ((struct mount *) 0); 680 } 681 682 /* 683 * Check if a user can access privileged mount options. 684 */ 685 int 686 vfs_suser(struct mount *mp, struct thread *td) 687 { 688 int error; 689 690 /* 691 * If the thread is jailed, but this is not a jail-friendly file 692 * system, deny immediately. 693 */ 694 if (!(mp->mnt_vfc->vfc_flags & VFCF_JAIL) && jailed(td->td_ucred)) 695 return (EPERM); 696 697 /* 698 * If the file system was mounted outside the jail of the calling 699 * thread, deny immediately. 700 */ 701 if (prison_check(td->td_ucred, mp->mnt_cred) != 0) 702 return (EPERM); 703 704 /* 705 * If file system supports delegated administration, we don't check 706 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified 707 * by the file system itself. 708 * If this is not the user that did original mount, we check for 709 * the PRIV_VFS_MOUNT_OWNER privilege. 710 */ 711 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) && 712 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 713 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) 714 return (error); 715 } 716 return (0); 717 } 718 719 /* 720 * Get a new unique fsid. Try to make its val[0] unique, since this value 721 * will be used to create fake device numbers for stat(). Also try (but 722 * not so hard) make its val[0] unique mod 2^16, since some emulators only 723 * support 16-bit device numbers. We end up with unique val[0]'s for the 724 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 725 * 726 * Keep in mind that several mounts may be running in parallel. Starting 727 * the search one past where the previous search terminated is both a 728 * micro-optimization and a defense against returning the same fsid to 729 * different mounts. 730 */ 731 void 732 vfs_getnewfsid(struct mount *mp) 733 { 734 static uint16_t mntid_base; 735 struct mount *nmp; 736 fsid_t tfsid; 737 int mtype; 738 739 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 740 mtx_lock(&mntid_mtx); 741 mtype = mp->mnt_vfc->vfc_typenum; 742 tfsid.val[1] = mtype; 743 mtype = (mtype & 0xFF) << 24; 744 for (;;) { 745 tfsid.val[0] = makedev(255, 746 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 747 mntid_base++; 748 if ((nmp = vfs_getvfs(&tfsid)) == NULL) 749 break; 750 vfs_rel(nmp); 751 } 752 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 753 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 754 mtx_unlock(&mntid_mtx); 755 } 756 757 /* 758 * Knob to control the precision of file timestamps: 759 * 760 * 0 = seconds only; nanoseconds zeroed. 761 * 1 = seconds and nanoseconds, accurate within 1/HZ. 762 * 2 = seconds and nanoseconds, truncated to microseconds. 763 * >=3 = seconds and nanoseconds, maximum precision. 764 */ 765 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 766 767 static int timestamp_precision = TSP_USEC; 768 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 769 ×tamp_precision, 0, "File timestamp precision (0: seconds, " 770 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, " 771 "3+: sec + ns (max. precision))"); 772 773 /* 774 * Get a current timestamp. 775 */ 776 void 777 vfs_timestamp(struct timespec *tsp) 778 { 779 struct timeval tv; 780 781 switch (timestamp_precision) { 782 case TSP_SEC: 783 tsp->tv_sec = time_second; 784 tsp->tv_nsec = 0; 785 break; 786 case TSP_HZ: 787 getnanotime(tsp); 788 break; 789 case TSP_USEC: 790 microtime(&tv); 791 TIMEVAL_TO_TIMESPEC(&tv, tsp); 792 break; 793 case TSP_NSEC: 794 default: 795 nanotime(tsp); 796 break; 797 } 798 } 799 800 /* 801 * Set vnode attributes to VNOVAL 802 */ 803 void 804 vattr_null(struct vattr *vap) 805 { 806 807 vap->va_type = VNON; 808 vap->va_size = VNOVAL; 809 vap->va_bytes = VNOVAL; 810 vap->va_mode = VNOVAL; 811 vap->va_nlink = VNOVAL; 812 vap->va_uid = VNOVAL; 813 vap->va_gid = VNOVAL; 814 vap->va_fsid = VNOVAL; 815 vap->va_fileid = VNOVAL; 816 vap->va_blocksize = VNOVAL; 817 vap->va_rdev = VNOVAL; 818 vap->va_atime.tv_sec = VNOVAL; 819 vap->va_atime.tv_nsec = VNOVAL; 820 vap->va_mtime.tv_sec = VNOVAL; 821 vap->va_mtime.tv_nsec = VNOVAL; 822 vap->va_ctime.tv_sec = VNOVAL; 823 vap->va_ctime.tv_nsec = VNOVAL; 824 vap->va_birthtime.tv_sec = VNOVAL; 825 vap->va_birthtime.tv_nsec = VNOVAL; 826 vap->va_flags = VNOVAL; 827 vap->va_gen = VNOVAL; 828 vap->va_vaflags = 0; 829 } 830 831 /* 832 * This routine is called when we have too many vnodes. It attempts 833 * to free <count> vnodes and will potentially free vnodes that still 834 * have VM backing store (VM backing store is typically the cause 835 * of a vnode blowout so we want to do this). Therefore, this operation 836 * is not considered cheap. 837 * 838 * A number of conditions may prevent a vnode from being reclaimed. 839 * the buffer cache may have references on the vnode, a directory 840 * vnode may still have references due to the namei cache representing 841 * underlying files, or the vnode may be in active use. It is not 842 * desirable to reuse such vnodes. These conditions may cause the 843 * number of vnodes to reach some minimum value regardless of what 844 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 845 */ 846 static int 847 vlrureclaim(struct mount *mp, int reclaim_nc_src, int trigger) 848 { 849 struct vnode *vp; 850 int count, done, target; 851 852 done = 0; 853 vn_start_write(NULL, &mp, V_WAIT); 854 MNT_ILOCK(mp); 855 count = mp->mnt_nvnodelistsize; 856 target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1); 857 target = target / 10 + 1; 858 while (count != 0 && done < target) { 859 vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 860 while (vp != NULL && vp->v_type == VMARKER) 861 vp = TAILQ_NEXT(vp, v_nmntvnodes); 862 if (vp == NULL) 863 break; 864 /* 865 * XXX LRU is completely broken for non-free vnodes. First 866 * by calling here in mountpoint order, then by moving 867 * unselected vnodes to the end here, and most grossly by 868 * removing the vlruvp() function that was supposed to 869 * maintain the order. (This function was born broken 870 * since syncer problems prevented it doing anything.) The 871 * order is closer to LRC (C = Created). 872 * 873 * LRU reclaiming of vnodes seems to have last worked in 874 * FreeBSD-3 where LRU wasn't mentioned under any spelling. 875 * Then there was no hold count, and inactive vnodes were 876 * simply put on the free list in LRU order. The separate 877 * lists also break LRU. We prefer to reclaim from the 878 * free list for technical reasons. This tends to thrash 879 * the free list to keep very unrecently used held vnodes. 880 * The problem is mitigated by keeping the free list large. 881 */ 882 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 883 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 884 --count; 885 if (!VI_TRYLOCK(vp)) 886 goto next_iter; 887 /* 888 * If it's been deconstructed already, it's still 889 * referenced, or it exceeds the trigger, skip it. 890 * Also skip free vnodes. We are trying to make space 891 * to expand the free list, not reduce it. 892 */ 893 if (vp->v_usecount || 894 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 895 ((vp->v_iflag & VI_FREE) != 0) || 896 (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL && 897 vp->v_object->resident_page_count > trigger)) { 898 VI_UNLOCK(vp); 899 goto next_iter; 900 } 901 MNT_IUNLOCK(mp); 902 vholdl(vp); 903 if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) { 904 vdrop(vp); 905 goto next_iter_mntunlocked; 906 } 907 VI_LOCK(vp); 908 /* 909 * v_usecount may have been bumped after VOP_LOCK() dropped 910 * the vnode interlock and before it was locked again. 911 * 912 * It is not necessary to recheck VI_DOOMED because it can 913 * only be set by another thread that holds both the vnode 914 * lock and vnode interlock. If another thread has the 915 * vnode lock before we get to VOP_LOCK() and obtains the 916 * vnode interlock after VOP_LOCK() drops the vnode 917 * interlock, the other thread will be unable to drop the 918 * vnode lock before our VOP_LOCK() call fails. 919 */ 920 if (vp->v_usecount || 921 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 922 (vp->v_iflag & VI_FREE) != 0 || 923 (vp->v_object != NULL && 924 vp->v_object->resident_page_count > trigger)) { 925 VOP_UNLOCK(vp, LK_INTERLOCK); 926 vdrop(vp); 927 goto next_iter_mntunlocked; 928 } 929 KASSERT((vp->v_iflag & VI_DOOMED) == 0, 930 ("VI_DOOMED unexpectedly detected in vlrureclaim()")); 931 counter_u64_add(recycles_count, 1); 932 vgonel(vp); 933 VOP_UNLOCK(vp, 0); 934 vdropl(vp); 935 done++; 936 next_iter_mntunlocked: 937 if (!should_yield()) 938 goto relock_mnt; 939 goto yield; 940 next_iter: 941 if (!should_yield()) 942 continue; 943 MNT_IUNLOCK(mp); 944 yield: 945 kern_yield(PRI_USER); 946 relock_mnt: 947 MNT_ILOCK(mp); 948 } 949 MNT_IUNLOCK(mp); 950 vn_finished_write(mp); 951 return done; 952 } 953 954 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */ 955 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free, 956 0, 957 "limit on vnode free requests per call to the vnlru_free routine"); 958 959 /* 960 * Attempt to reduce the free list by the requested amount. 961 */ 962 static void 963 vnlru_free_locked(int count, struct vfsops *mnt_op) 964 { 965 struct vnode *vp; 966 struct mount *mp; 967 bool tried_batches; 968 969 tried_batches = false; 970 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 971 if (count > max_vnlru_free) 972 count = max_vnlru_free; 973 for (; count > 0; count--) { 974 vp = TAILQ_FIRST(&vnode_free_list); 975 /* 976 * The list can be modified while the free_list_mtx 977 * has been dropped and vp could be NULL here. 978 */ 979 if (vp == NULL) { 980 if (tried_batches) 981 break; 982 mtx_unlock(&vnode_free_list_mtx); 983 vnlru_return_batches(mnt_op); 984 tried_batches = true; 985 mtx_lock(&vnode_free_list_mtx); 986 continue; 987 } 988 989 VNASSERT(vp->v_op != NULL, vp, 990 ("vnlru_free: vnode already reclaimed.")); 991 KASSERT((vp->v_iflag & VI_FREE) != 0, 992 ("Removing vnode not on freelist")); 993 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 994 ("Mangling active vnode")); 995 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist); 996 997 /* 998 * Don't recycle if our vnode is from different type 999 * of mount point. Note that mp is type-safe, the 1000 * check does not reach unmapped address even if 1001 * vnode is reclaimed. 1002 * Don't recycle if we can't get the interlock without 1003 * blocking. 1004 */ 1005 if ((mnt_op != NULL && (mp = vp->v_mount) != NULL && 1006 mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) { 1007 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist); 1008 continue; 1009 } 1010 VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0, 1011 vp, ("vp inconsistent on freelist")); 1012 1013 /* 1014 * The clear of VI_FREE prevents activation of the 1015 * vnode. There is no sense in putting the vnode on 1016 * the mount point active list, only to remove it 1017 * later during recycling. Inline the relevant part 1018 * of vholdl(), to avoid triggering assertions or 1019 * activating. 1020 */ 1021 freevnodes--; 1022 vp->v_iflag &= ~VI_FREE; 1023 refcount_acquire(&vp->v_holdcnt); 1024 1025 mtx_unlock(&vnode_free_list_mtx); 1026 VI_UNLOCK(vp); 1027 vtryrecycle(vp); 1028 /* 1029 * If the recycled succeeded this vdrop will actually free 1030 * the vnode. If not it will simply place it back on 1031 * the free list. 1032 */ 1033 vdrop(vp); 1034 mtx_lock(&vnode_free_list_mtx); 1035 } 1036 } 1037 1038 void 1039 vnlru_free(int count, struct vfsops *mnt_op) 1040 { 1041 1042 mtx_lock(&vnode_free_list_mtx); 1043 vnlru_free_locked(count, mnt_op); 1044 mtx_unlock(&vnode_free_list_mtx); 1045 } 1046 1047 1048 /* XXX some names and initialization are bad for limits and watermarks. */ 1049 static int 1050 vspace(void) 1051 { 1052 int space; 1053 1054 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100); 1055 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */ 1056 vlowat = vhiwat / 2; 1057 if (numvnodes > desiredvnodes) 1058 return (0); 1059 space = desiredvnodes - numvnodes; 1060 if (freevnodes > wantfreevnodes) 1061 space += freevnodes - wantfreevnodes; 1062 return (space); 1063 } 1064 1065 static void 1066 vnlru_return_batch_locked(struct mount *mp) 1067 { 1068 struct vnode *vp; 1069 1070 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 1071 1072 if (mp->mnt_tmpfreevnodelistsize == 0) 1073 return; 1074 1075 TAILQ_FOREACH(vp, &mp->mnt_tmpfreevnodelist, v_actfreelist) { 1076 VNASSERT((vp->v_mflag & VMP_TMPMNTFREELIST) != 0, vp, 1077 ("vnode without VMP_TMPMNTFREELIST on mnt_tmpfreevnodelist")); 1078 vp->v_mflag &= ~VMP_TMPMNTFREELIST; 1079 } 1080 mtx_lock(&vnode_free_list_mtx); 1081 TAILQ_CONCAT(&vnode_free_list, &mp->mnt_tmpfreevnodelist, v_actfreelist); 1082 freevnodes += mp->mnt_tmpfreevnodelistsize; 1083 mtx_unlock(&vnode_free_list_mtx); 1084 mp->mnt_tmpfreevnodelistsize = 0; 1085 } 1086 1087 static void 1088 vnlru_return_batch(struct mount *mp) 1089 { 1090 1091 mtx_lock(&mp->mnt_listmtx); 1092 vnlru_return_batch_locked(mp); 1093 mtx_unlock(&mp->mnt_listmtx); 1094 } 1095 1096 static void 1097 vnlru_return_batches(struct vfsops *mnt_op) 1098 { 1099 struct mount *mp, *nmp; 1100 bool need_unbusy; 1101 1102 mtx_lock(&mountlist_mtx); 1103 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 1104 need_unbusy = false; 1105 if (mnt_op != NULL && mp->mnt_op != mnt_op) 1106 goto next; 1107 if (mp->mnt_tmpfreevnodelistsize == 0) 1108 goto next; 1109 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) == 0) { 1110 vnlru_return_batch(mp); 1111 need_unbusy = true; 1112 mtx_lock(&mountlist_mtx); 1113 } 1114 next: 1115 nmp = TAILQ_NEXT(mp, mnt_list); 1116 if (need_unbusy) 1117 vfs_unbusy(mp); 1118 } 1119 mtx_unlock(&mountlist_mtx); 1120 } 1121 1122 /* 1123 * Attempt to recycle vnodes in a context that is always safe to block. 1124 * Calling vlrurecycle() from the bowels of filesystem code has some 1125 * interesting deadlock problems. 1126 */ 1127 static struct proc *vnlruproc; 1128 static int vnlruproc_sig; 1129 1130 static void 1131 vnlru_proc(void) 1132 { 1133 struct mount *mp, *nmp; 1134 unsigned long ofreevnodes, onumvnodes; 1135 int done, force, reclaim_nc_src, trigger, usevnodes; 1136 1137 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc, 1138 SHUTDOWN_PRI_FIRST); 1139 1140 force = 0; 1141 for (;;) { 1142 kproc_suspend_check(vnlruproc); 1143 mtx_lock(&vnode_free_list_mtx); 1144 /* 1145 * If numvnodes is too large (due to desiredvnodes being 1146 * adjusted using its sysctl, or emergency growth), first 1147 * try to reduce it by discarding from the free list. 1148 */ 1149 if (numvnodes > desiredvnodes) 1150 vnlru_free_locked(numvnodes - desiredvnodes, NULL); 1151 /* 1152 * Sleep if the vnode cache is in a good state. This is 1153 * when it is not over-full and has space for about a 4% 1154 * or 9% expansion (by growing its size or inexcessively 1155 * reducing its free list). Otherwise, try to reclaim 1156 * space for a 10% expansion. 1157 */ 1158 if (vstir && force == 0) { 1159 force = 1; 1160 vstir = 0; 1161 } 1162 if (vspace() >= vlowat && force == 0) { 1163 vnlruproc_sig = 0; 1164 wakeup(&vnlruproc_sig); 1165 msleep(vnlruproc, &vnode_free_list_mtx, 1166 PVFS|PDROP, "vlruwt", hz); 1167 continue; 1168 } 1169 mtx_unlock(&vnode_free_list_mtx); 1170 done = 0; 1171 ofreevnodes = freevnodes; 1172 onumvnodes = numvnodes; 1173 /* 1174 * Calculate parameters for recycling. These are the same 1175 * throughout the loop to give some semblance of fairness. 1176 * The trigger point is to avoid recycling vnodes with lots 1177 * of resident pages. We aren't trying to free memory; we 1178 * are trying to recycle or at least free vnodes. 1179 */ 1180 if (numvnodes <= desiredvnodes) 1181 usevnodes = numvnodes - freevnodes; 1182 else 1183 usevnodes = numvnodes; 1184 if (usevnodes <= 0) 1185 usevnodes = 1; 1186 /* 1187 * The trigger value is is chosen to give a conservatively 1188 * large value to ensure that it alone doesn't prevent 1189 * making progress. The value can easily be so large that 1190 * it is effectively infinite in some congested and 1191 * misconfigured cases, and this is necessary. Normally 1192 * it is about 8 to 100 (pages), which is quite large. 1193 */ 1194 trigger = vm_cnt.v_page_count * 2 / usevnodes; 1195 if (force < 2) 1196 trigger = vsmalltrigger; 1197 reclaim_nc_src = force >= 3; 1198 mtx_lock(&mountlist_mtx); 1199 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 1200 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) { 1201 nmp = TAILQ_NEXT(mp, mnt_list); 1202 continue; 1203 } 1204 done += vlrureclaim(mp, reclaim_nc_src, trigger); 1205 mtx_lock(&mountlist_mtx); 1206 nmp = TAILQ_NEXT(mp, mnt_list); 1207 vfs_unbusy(mp); 1208 } 1209 mtx_unlock(&mountlist_mtx); 1210 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes) 1211 uma_reclaim(); 1212 if (done == 0) { 1213 if (force == 0 || force == 1) { 1214 force = 2; 1215 continue; 1216 } 1217 if (force == 2) { 1218 force = 3; 1219 continue; 1220 } 1221 force = 0; 1222 vnlru_nowhere++; 1223 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 1224 } else 1225 kern_yield(PRI_USER); 1226 /* 1227 * After becoming active to expand above low water, keep 1228 * active until above high water. 1229 */ 1230 force = vspace() < vhiwat; 1231 } 1232 } 1233 1234 static struct kproc_desc vnlru_kp = { 1235 "vnlru", 1236 vnlru_proc, 1237 &vnlruproc 1238 }; 1239 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, 1240 &vnlru_kp); 1241 1242 /* 1243 * Routines having to do with the management of the vnode table. 1244 */ 1245 1246 /* 1247 * Try to recycle a freed vnode. We abort if anyone picks up a reference 1248 * before we actually vgone(). This function must be called with the vnode 1249 * held to prevent the vnode from being returned to the free list midway 1250 * through vgone(). 1251 */ 1252 static int 1253 vtryrecycle(struct vnode *vp) 1254 { 1255 struct mount *vnmp; 1256 1257 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 1258 VNASSERT(vp->v_holdcnt, vp, 1259 ("vtryrecycle: Recycling vp %p without a reference.", vp)); 1260 /* 1261 * This vnode may found and locked via some other list, if so we 1262 * can't recycle it yet. 1263 */ 1264 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 1265 CTR2(KTR_VFS, 1266 "%s: impossible to recycle, vp %p lock is already held", 1267 __func__, vp); 1268 return (EWOULDBLOCK); 1269 } 1270 /* 1271 * Don't recycle if its filesystem is being suspended. 1272 */ 1273 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 1274 VOP_UNLOCK(vp, 0); 1275 CTR2(KTR_VFS, 1276 "%s: impossible to recycle, cannot start the write for %p", 1277 __func__, vp); 1278 return (EBUSY); 1279 } 1280 /* 1281 * If we got this far, we need to acquire the interlock and see if 1282 * anyone picked up this vnode from another list. If not, we will 1283 * mark it with DOOMED via vgonel() so that anyone who does find it 1284 * will skip over it. 1285 */ 1286 VI_LOCK(vp); 1287 if (vp->v_usecount) { 1288 VOP_UNLOCK(vp, LK_INTERLOCK); 1289 vn_finished_write(vnmp); 1290 CTR2(KTR_VFS, 1291 "%s: impossible to recycle, %p is already referenced", 1292 __func__, vp); 1293 return (EBUSY); 1294 } 1295 if ((vp->v_iflag & VI_DOOMED) == 0) { 1296 counter_u64_add(recycles_count, 1); 1297 vgonel(vp); 1298 } 1299 VOP_UNLOCK(vp, LK_INTERLOCK); 1300 vn_finished_write(vnmp); 1301 return (0); 1302 } 1303 1304 static void 1305 vcheckspace(void) 1306 { 1307 1308 if (vspace() < vlowat && vnlruproc_sig == 0) { 1309 vnlruproc_sig = 1; 1310 wakeup(vnlruproc); 1311 } 1312 } 1313 1314 /* 1315 * Wait if necessary for space for a new vnode. 1316 */ 1317 static int 1318 getnewvnode_wait(int suspended) 1319 { 1320 1321 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 1322 if (numvnodes >= desiredvnodes) { 1323 if (suspended) { 1324 /* 1325 * The file system is being suspended. We cannot 1326 * risk a deadlock here, so allow allocation of 1327 * another vnode even if this would give too many. 1328 */ 1329 return (0); 1330 } 1331 if (vnlruproc_sig == 0) { 1332 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 1333 wakeup(vnlruproc); 1334 } 1335 msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS, 1336 "vlruwk", hz); 1337 } 1338 /* Post-adjust like the pre-adjust in getnewvnode(). */ 1339 if (numvnodes + 1 > desiredvnodes && freevnodes > 1) 1340 vnlru_free_locked(1, NULL); 1341 return (numvnodes >= desiredvnodes ? ENFILE : 0); 1342 } 1343 1344 /* 1345 * This hack is fragile, and probably not needed any more now that the 1346 * watermark handling works. 1347 */ 1348 void 1349 getnewvnode_reserve(u_int count) 1350 { 1351 struct thread *td; 1352 1353 /* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */ 1354 /* XXX no longer so quick, but this part is not racy. */ 1355 mtx_lock(&vnode_free_list_mtx); 1356 if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes) 1357 vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes, 1358 freevnodes - wantfreevnodes), NULL); 1359 mtx_unlock(&vnode_free_list_mtx); 1360 1361 td = curthread; 1362 /* First try to be quick and racy. */ 1363 if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) { 1364 td->td_vp_reserv += count; 1365 vcheckspace(); /* XXX no longer so quick, but more racy */ 1366 return; 1367 } else 1368 atomic_subtract_long(&numvnodes, count); 1369 1370 mtx_lock(&vnode_free_list_mtx); 1371 while (count > 0) { 1372 if (getnewvnode_wait(0) == 0) { 1373 count--; 1374 td->td_vp_reserv++; 1375 atomic_add_long(&numvnodes, 1); 1376 } 1377 } 1378 vcheckspace(); 1379 mtx_unlock(&vnode_free_list_mtx); 1380 } 1381 1382 /* 1383 * This hack is fragile, especially if desiredvnodes or wantvnodes are 1384 * misconfgured or changed significantly. Reducing desiredvnodes below 1385 * the reserved amount should cause bizarre behaviour like reducing it 1386 * below the number of active vnodes -- the system will try to reduce 1387 * numvnodes to match, but should fail, so the subtraction below should 1388 * not overflow. 1389 */ 1390 void 1391 getnewvnode_drop_reserve(void) 1392 { 1393 struct thread *td; 1394 1395 td = curthread; 1396 atomic_subtract_long(&numvnodes, td->td_vp_reserv); 1397 td->td_vp_reserv = 0; 1398 } 1399 1400 /* 1401 * Return the next vnode from the free list. 1402 */ 1403 int 1404 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, 1405 struct vnode **vpp) 1406 { 1407 struct vnode *vp; 1408 struct thread *td; 1409 struct lock_object *lo; 1410 static int cyclecount; 1411 int error; 1412 1413 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag); 1414 vp = NULL; 1415 td = curthread; 1416 if (td->td_vp_reserv > 0) { 1417 td->td_vp_reserv -= 1; 1418 goto alloc; 1419 } 1420 mtx_lock(&vnode_free_list_mtx); 1421 if (numvnodes < desiredvnodes) 1422 cyclecount = 0; 1423 else if (cyclecount++ >= freevnodes) { 1424 cyclecount = 0; 1425 vstir = 1; 1426 } 1427 /* 1428 * Grow the vnode cache if it will not be above its target max 1429 * after growing. Otherwise, if the free list is nonempty, try 1430 * to reclaim 1 item from it before growing the cache (possibly 1431 * above its target max if the reclamation failed or is delayed). 1432 * Otherwise, wait for some space. In all cases, schedule 1433 * vnlru_proc() if we are getting short of space. The watermarks 1434 * should be chosen so that we never wait or even reclaim from 1435 * the free list to below its target minimum. 1436 */ 1437 if (numvnodes + 1 <= desiredvnodes) 1438 ; 1439 else if (freevnodes > 0) 1440 vnlru_free_locked(1, NULL); 1441 else { 1442 error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag & 1443 MNTK_SUSPEND)); 1444 #if 0 /* XXX Not all VFS_VGET/ffs_vget callers check returns. */ 1445 if (error != 0) { 1446 mtx_unlock(&vnode_free_list_mtx); 1447 return (error); 1448 } 1449 #endif 1450 } 1451 vcheckspace(); 1452 atomic_add_long(&numvnodes, 1); 1453 mtx_unlock(&vnode_free_list_mtx); 1454 alloc: 1455 counter_u64_add(vnodes_created, 1); 1456 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK); 1457 /* 1458 * Locks are given the generic name "vnode" when created. 1459 * Follow the historic practice of using the filesystem 1460 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc. 1461 * 1462 * Locks live in a witness group keyed on their name. Thus, 1463 * when a lock is renamed, it must also move from the witness 1464 * group of its old name to the witness group of its new name. 1465 * 1466 * The change only needs to be made when the vnode moves 1467 * from one filesystem type to another. We ensure that each 1468 * filesystem use a single static name pointer for its tag so 1469 * that we can compare pointers rather than doing a strcmp(). 1470 */ 1471 lo = &vp->v_vnlock->lock_object; 1472 if (lo->lo_name != tag) { 1473 lo->lo_name = tag; 1474 WITNESS_DESTROY(lo); 1475 WITNESS_INIT(lo, tag); 1476 } 1477 /* 1478 * By default, don't allow shared locks unless filesystems opt-in. 1479 */ 1480 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE; 1481 /* 1482 * Finalize various vnode identity bits. 1483 */ 1484 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp)); 1485 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp)); 1486 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp)); 1487 vp->v_type = VNON; 1488 vp->v_tag = tag; 1489 vp->v_op = vops; 1490 v_init_counters(vp); 1491 vp->v_bufobj.bo_ops = &buf_ops_bio; 1492 #ifdef DIAGNOSTIC 1493 if (mp == NULL && vops != &dead_vnodeops) 1494 printf("NULL mp in getnewvnode(9), tag %s\n", tag); 1495 #endif 1496 #ifdef MAC 1497 mac_vnode_init(vp); 1498 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 1499 mac_vnode_associate_singlelabel(mp, vp); 1500 #endif 1501 if (mp != NULL) { 1502 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize; 1503 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) 1504 vp->v_vflag |= VV_NOKNOTE; 1505 } 1506 1507 /* 1508 * For the filesystems which do not use vfs_hash_insert(), 1509 * still initialize v_hash to have vfs_hash_index() useful. 1510 * E.g., nullfs uses vfs_hash_index() on the lower vnode for 1511 * its own hashing. 1512 */ 1513 vp->v_hash = (uintptr_t)vp >> vnsz2log; 1514 1515 *vpp = vp; 1516 return (0); 1517 } 1518 1519 /* 1520 * Delete from old mount point vnode list, if on one. 1521 */ 1522 static void 1523 delmntque(struct vnode *vp) 1524 { 1525 struct mount *mp; 1526 int active; 1527 1528 mp = vp->v_mount; 1529 if (mp == NULL) 1530 return; 1531 MNT_ILOCK(mp); 1532 VI_LOCK(vp); 1533 KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize, 1534 ("Active vnode list size %d > Vnode list size %d", 1535 mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize)); 1536 active = vp->v_iflag & VI_ACTIVE; 1537 vp->v_iflag &= ~VI_ACTIVE; 1538 if (active) { 1539 mtx_lock(&mp->mnt_listmtx); 1540 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist); 1541 mp->mnt_activevnodelistsize--; 1542 mtx_unlock(&mp->mnt_listmtx); 1543 } 1544 vp->v_mount = NULL; 1545 VI_UNLOCK(vp); 1546 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 1547 ("bad mount point vnode list size")); 1548 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1549 mp->mnt_nvnodelistsize--; 1550 MNT_REL(mp); 1551 MNT_IUNLOCK(mp); 1552 } 1553 1554 static void 1555 insmntque_stddtr(struct vnode *vp, void *dtr_arg) 1556 { 1557 1558 vp->v_data = NULL; 1559 vp->v_op = &dead_vnodeops; 1560 vgone(vp); 1561 vput(vp); 1562 } 1563 1564 /* 1565 * Insert into list of vnodes for the new mount point, if available. 1566 */ 1567 int 1568 insmntque1(struct vnode *vp, struct mount *mp, 1569 void (*dtr)(struct vnode *, void *), void *dtr_arg) 1570 { 1571 1572 KASSERT(vp->v_mount == NULL, 1573 ("insmntque: vnode already on per mount vnode list")); 1574 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 1575 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); 1576 1577 /* 1578 * We acquire the vnode interlock early to ensure that the 1579 * vnode cannot be recycled by another process releasing a 1580 * holdcnt on it before we get it on both the vnode list 1581 * and the active vnode list. The mount mutex protects only 1582 * manipulation of the vnode list and the vnode freelist 1583 * mutex protects only manipulation of the active vnode list. 1584 * Hence the need to hold the vnode interlock throughout. 1585 */ 1586 MNT_ILOCK(mp); 1587 VI_LOCK(vp); 1588 if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 && 1589 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || 1590 mp->mnt_nvnodelistsize == 0)) && 1591 (vp->v_vflag & VV_FORCEINSMQ) == 0) { 1592 VI_UNLOCK(vp); 1593 MNT_IUNLOCK(mp); 1594 if (dtr != NULL) 1595 dtr(vp, dtr_arg); 1596 return (EBUSY); 1597 } 1598 vp->v_mount = mp; 1599 MNT_REF(mp); 1600 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1601 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 1602 ("neg mount point vnode list size")); 1603 mp->mnt_nvnodelistsize++; 1604 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 1605 ("Activating already active vnode")); 1606 vp->v_iflag |= VI_ACTIVE; 1607 mtx_lock(&mp->mnt_listmtx); 1608 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist); 1609 mp->mnt_activevnodelistsize++; 1610 mtx_unlock(&mp->mnt_listmtx); 1611 VI_UNLOCK(vp); 1612 MNT_IUNLOCK(mp); 1613 return (0); 1614 } 1615 1616 int 1617 insmntque(struct vnode *vp, struct mount *mp) 1618 { 1619 1620 return (insmntque1(vp, mp, insmntque_stddtr, NULL)); 1621 } 1622 1623 /* 1624 * Flush out and invalidate all buffers associated with a bufobj 1625 * Called with the underlying object locked. 1626 */ 1627 int 1628 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) 1629 { 1630 int error; 1631 1632 BO_LOCK(bo); 1633 if (flags & V_SAVE) { 1634 error = bufobj_wwait(bo, slpflag, slptimeo); 1635 if (error) { 1636 BO_UNLOCK(bo); 1637 return (error); 1638 } 1639 if (bo->bo_dirty.bv_cnt > 0) { 1640 BO_UNLOCK(bo); 1641 if ((error = BO_SYNC(bo, MNT_WAIT)) != 0) 1642 return (error); 1643 /* 1644 * XXX We could save a lock/unlock if this was only 1645 * enabled under INVARIANTS 1646 */ 1647 BO_LOCK(bo); 1648 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 1649 panic("vinvalbuf: dirty bufs"); 1650 } 1651 } 1652 /* 1653 * If you alter this loop please notice that interlock is dropped and 1654 * reacquired in flushbuflist. Special care is needed to ensure that 1655 * no race conditions occur from this. 1656 */ 1657 do { 1658 error = flushbuflist(&bo->bo_clean, 1659 flags, bo, slpflag, slptimeo); 1660 if (error == 0 && !(flags & V_CLEANONLY)) 1661 error = flushbuflist(&bo->bo_dirty, 1662 flags, bo, slpflag, slptimeo); 1663 if (error != 0 && error != EAGAIN) { 1664 BO_UNLOCK(bo); 1665 return (error); 1666 } 1667 } while (error != 0); 1668 1669 /* 1670 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1671 * have write I/O in-progress but if there is a VM object then the 1672 * VM object can also have read-I/O in-progress. 1673 */ 1674 do { 1675 bufobj_wwait(bo, 0, 0); 1676 if ((flags & V_VMIO) == 0) { 1677 BO_UNLOCK(bo); 1678 if (bo->bo_object != NULL) { 1679 VM_OBJECT_WLOCK(bo->bo_object); 1680 vm_object_pip_wait(bo->bo_object, "bovlbx"); 1681 VM_OBJECT_WUNLOCK(bo->bo_object); 1682 } 1683 BO_LOCK(bo); 1684 } 1685 } while (bo->bo_numoutput > 0); 1686 BO_UNLOCK(bo); 1687 1688 /* 1689 * Destroy the copy in the VM cache, too. 1690 */ 1691 if (bo->bo_object != NULL && 1692 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) { 1693 VM_OBJECT_WLOCK(bo->bo_object); 1694 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? 1695 OBJPR_CLEANONLY : 0); 1696 VM_OBJECT_WUNLOCK(bo->bo_object); 1697 } 1698 1699 #ifdef INVARIANTS 1700 BO_LOCK(bo); 1701 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 && 1702 (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0)) 1703 panic("vinvalbuf: flush failed"); 1704 BO_UNLOCK(bo); 1705 #endif 1706 return (0); 1707 } 1708 1709 /* 1710 * Flush out and invalidate all buffers associated with a vnode. 1711 * Called with the underlying object locked. 1712 */ 1713 int 1714 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 1715 { 1716 1717 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 1718 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1719 if (vp->v_object != NULL && vp->v_object->handle != vp) 1720 return (0); 1721 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); 1722 } 1723 1724 /* 1725 * Flush out buffers on the specified list. 1726 * 1727 */ 1728 static int 1729 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 1730 int slptimeo) 1731 { 1732 struct buf *bp, *nbp; 1733 int retval, error; 1734 daddr_t lblkno; 1735 b_xflags_t xflags; 1736 1737 ASSERT_BO_WLOCKED(bo); 1738 1739 retval = 0; 1740 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 1741 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1742 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1743 continue; 1744 } 1745 lblkno = 0; 1746 xflags = 0; 1747 if (nbp != NULL) { 1748 lblkno = nbp->b_lblkno; 1749 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN); 1750 } 1751 retval = EAGAIN; 1752 error = BUF_TIMELOCK(bp, 1753 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), 1754 "flushbuf", slpflag, slptimeo); 1755 if (error) { 1756 BO_LOCK(bo); 1757 return (error != ENOLCK ? error : EAGAIN); 1758 } 1759 KASSERT(bp->b_bufobj == bo, 1760 ("bp %p wrong b_bufobj %p should be %p", 1761 bp, bp->b_bufobj, bo)); 1762 /* 1763 * XXX Since there are no node locks for NFS, I 1764 * believe there is a slight chance that a delayed 1765 * write will occur while sleeping just above, so 1766 * check for it. 1767 */ 1768 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1769 (flags & V_SAVE)) { 1770 bremfree(bp); 1771 bp->b_flags |= B_ASYNC; 1772 bwrite(bp); 1773 BO_LOCK(bo); 1774 return (EAGAIN); /* XXX: why not loop ? */ 1775 } 1776 bremfree(bp); 1777 bp->b_flags |= (B_INVAL | B_RELBUF); 1778 bp->b_flags &= ~B_ASYNC; 1779 brelse(bp); 1780 BO_LOCK(bo); 1781 nbp = gbincore(bo, lblkno); 1782 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1783 != xflags) 1784 break; /* nbp invalid */ 1785 } 1786 return (retval); 1787 } 1788 1789 int 1790 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn) 1791 { 1792 struct buf *bp; 1793 int error; 1794 daddr_t lblkno; 1795 1796 ASSERT_BO_LOCKED(bo); 1797 1798 for (lblkno = startn;;) { 1799 again: 1800 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno); 1801 if (bp == NULL || bp->b_lblkno >= endn || 1802 bp->b_lblkno < startn) 1803 break; 1804 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 1805 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0); 1806 if (error != 0) { 1807 BO_RLOCK(bo); 1808 if (error == ENOLCK) 1809 goto again; 1810 return (error); 1811 } 1812 KASSERT(bp->b_bufobj == bo, 1813 ("bp %p wrong b_bufobj %p should be %p", 1814 bp, bp->b_bufobj, bo)); 1815 lblkno = bp->b_lblkno + 1; 1816 if ((bp->b_flags & B_MANAGED) == 0) 1817 bremfree(bp); 1818 bp->b_flags |= B_RELBUF; 1819 /* 1820 * In the VMIO case, use the B_NOREUSE flag to hint that the 1821 * pages backing each buffer in the range are unlikely to be 1822 * reused. Dirty buffers will have the hint applied once 1823 * they've been written. 1824 */ 1825 if (bp->b_vp->v_object != NULL) 1826 bp->b_flags |= B_NOREUSE; 1827 brelse(bp); 1828 BO_RLOCK(bo); 1829 } 1830 return (0); 1831 } 1832 1833 /* 1834 * Truncate a file's buffer and pages to a specified length. This 1835 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1836 * sync activity. 1837 */ 1838 int 1839 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize) 1840 { 1841 struct buf *bp, *nbp; 1842 int anyfreed; 1843 int trunclbn; 1844 struct bufobj *bo; 1845 1846 CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__, 1847 vp, cred, blksize, (uintmax_t)length); 1848 1849 /* 1850 * Round up to the *next* lbn. 1851 */ 1852 trunclbn = howmany(length, blksize); 1853 1854 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1855 restart: 1856 bo = &vp->v_bufobj; 1857 BO_LOCK(bo); 1858 anyfreed = 1; 1859 for (;anyfreed;) { 1860 anyfreed = 0; 1861 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 1862 if (bp->b_lblkno < trunclbn) 1863 continue; 1864 if (BUF_LOCK(bp, 1865 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1866 BO_LOCKPTR(bo)) == ENOLCK) 1867 goto restart; 1868 1869 bremfree(bp); 1870 bp->b_flags |= (B_INVAL | B_RELBUF); 1871 bp->b_flags &= ~B_ASYNC; 1872 brelse(bp); 1873 anyfreed = 1; 1874 1875 BO_LOCK(bo); 1876 if (nbp != NULL && 1877 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1878 (nbp->b_vp != vp) || 1879 (nbp->b_flags & B_DELWRI))) { 1880 BO_UNLOCK(bo); 1881 goto restart; 1882 } 1883 } 1884 1885 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1886 if (bp->b_lblkno < trunclbn) 1887 continue; 1888 if (BUF_LOCK(bp, 1889 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1890 BO_LOCKPTR(bo)) == ENOLCK) 1891 goto restart; 1892 bremfree(bp); 1893 bp->b_flags |= (B_INVAL | B_RELBUF); 1894 bp->b_flags &= ~B_ASYNC; 1895 brelse(bp); 1896 anyfreed = 1; 1897 1898 BO_LOCK(bo); 1899 if (nbp != NULL && 1900 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1901 (nbp->b_vp != vp) || 1902 (nbp->b_flags & B_DELWRI) == 0)) { 1903 BO_UNLOCK(bo); 1904 goto restart; 1905 } 1906 } 1907 } 1908 1909 if (length > 0) { 1910 restartsync: 1911 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1912 if (bp->b_lblkno > 0) 1913 continue; 1914 /* 1915 * Since we hold the vnode lock this should only 1916 * fail if we're racing with the buf daemon. 1917 */ 1918 if (BUF_LOCK(bp, 1919 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1920 BO_LOCKPTR(bo)) == ENOLCK) { 1921 goto restart; 1922 } 1923 VNASSERT((bp->b_flags & B_DELWRI), vp, 1924 ("buf(%p) on dirty queue without DELWRI", bp)); 1925 1926 bremfree(bp); 1927 bawrite(bp); 1928 BO_LOCK(bo); 1929 goto restartsync; 1930 } 1931 } 1932 1933 bufobj_wwait(bo, 0, 0); 1934 BO_UNLOCK(bo); 1935 vnode_pager_setsize(vp, length); 1936 1937 return (0); 1938 } 1939 1940 static void 1941 buf_vlist_remove(struct buf *bp) 1942 { 1943 struct bufv *bv; 1944 1945 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1946 ASSERT_BO_WLOCKED(bp->b_bufobj); 1947 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) != 1948 (BX_VNDIRTY|BX_VNCLEAN), 1949 ("buf_vlist_remove: Buf %p is on two lists", bp)); 1950 if (bp->b_xflags & BX_VNDIRTY) 1951 bv = &bp->b_bufobj->bo_dirty; 1952 else 1953 bv = &bp->b_bufobj->bo_clean; 1954 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno); 1955 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 1956 bv->bv_cnt--; 1957 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1958 } 1959 1960 /* 1961 * Add the buffer to the sorted clean or dirty block list. 1962 * 1963 * NOTE: xflags is passed as a constant, optimizing this inline function! 1964 */ 1965 static void 1966 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 1967 { 1968 struct bufv *bv; 1969 struct buf *n; 1970 int error; 1971 1972 ASSERT_BO_WLOCKED(bo); 1973 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0, 1974 ("dead bo %p", bo)); 1975 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1976 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 1977 bp->b_xflags |= xflags; 1978 if (xflags & BX_VNDIRTY) 1979 bv = &bo->bo_dirty; 1980 else 1981 bv = &bo->bo_clean; 1982 1983 /* 1984 * Keep the list ordered. Optimize empty list insertion. Assume 1985 * we tend to grow at the tail so lookup_le should usually be cheaper 1986 * than _ge. 1987 */ 1988 if (bv->bv_cnt == 0 || 1989 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno) 1990 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 1991 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL) 1992 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs); 1993 else 1994 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs); 1995 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp); 1996 if (error) 1997 panic("buf_vlist_add: Preallocated nodes insufficient."); 1998 bv->bv_cnt++; 1999 } 2000 2001 /* 2002 * Look up a buffer using the buffer tries. 2003 */ 2004 struct buf * 2005 gbincore(struct bufobj *bo, daddr_t lblkno) 2006 { 2007 struct buf *bp; 2008 2009 ASSERT_BO_LOCKED(bo); 2010 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno); 2011 if (bp != NULL) 2012 return (bp); 2013 return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno); 2014 } 2015 2016 /* 2017 * Associate a buffer with a vnode. 2018 */ 2019 void 2020 bgetvp(struct vnode *vp, struct buf *bp) 2021 { 2022 struct bufobj *bo; 2023 2024 bo = &vp->v_bufobj; 2025 ASSERT_BO_WLOCKED(bo); 2026 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 2027 2028 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 2029 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 2030 ("bgetvp: bp already attached! %p", bp)); 2031 2032 vhold(vp); 2033 bp->b_vp = vp; 2034 bp->b_bufobj = bo; 2035 /* 2036 * Insert onto list for new vnode. 2037 */ 2038 buf_vlist_add(bp, bo, BX_VNCLEAN); 2039 } 2040 2041 /* 2042 * Disassociate a buffer from a vnode. 2043 */ 2044 void 2045 brelvp(struct buf *bp) 2046 { 2047 struct bufobj *bo; 2048 struct vnode *vp; 2049 2050 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2051 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 2052 2053 /* 2054 * Delete from old vnode list, if on one. 2055 */ 2056 vp = bp->b_vp; /* XXX */ 2057 bo = bp->b_bufobj; 2058 BO_LOCK(bo); 2059 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2060 buf_vlist_remove(bp); 2061 else 2062 panic("brelvp: Buffer %p not on queue.", bp); 2063 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2064 bo->bo_flag &= ~BO_ONWORKLST; 2065 mtx_lock(&sync_mtx); 2066 LIST_REMOVE(bo, bo_synclist); 2067 syncer_worklist_len--; 2068 mtx_unlock(&sync_mtx); 2069 } 2070 bp->b_vp = NULL; 2071 bp->b_bufobj = NULL; 2072 BO_UNLOCK(bo); 2073 vdrop(vp); 2074 } 2075 2076 /* 2077 * Add an item to the syncer work queue. 2078 */ 2079 static void 2080 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 2081 { 2082 int slot; 2083 2084 ASSERT_BO_WLOCKED(bo); 2085 2086 mtx_lock(&sync_mtx); 2087 if (bo->bo_flag & BO_ONWORKLST) 2088 LIST_REMOVE(bo, bo_synclist); 2089 else { 2090 bo->bo_flag |= BO_ONWORKLST; 2091 syncer_worklist_len++; 2092 } 2093 2094 if (delay > syncer_maxdelay - 2) 2095 delay = syncer_maxdelay - 2; 2096 slot = (syncer_delayno + delay) & syncer_mask; 2097 2098 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 2099 mtx_unlock(&sync_mtx); 2100 } 2101 2102 static int 2103 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 2104 { 2105 int error, len; 2106 2107 mtx_lock(&sync_mtx); 2108 len = syncer_worklist_len - sync_vnode_count; 2109 mtx_unlock(&sync_mtx); 2110 error = SYSCTL_OUT(req, &len, sizeof(len)); 2111 return (error); 2112 } 2113 2114 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0, 2115 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 2116 2117 static struct proc *updateproc; 2118 static void sched_sync(void); 2119 static struct kproc_desc up_kp = { 2120 "syncer", 2121 sched_sync, 2122 &updateproc 2123 }; 2124 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); 2125 2126 static int 2127 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) 2128 { 2129 struct vnode *vp; 2130 struct mount *mp; 2131 2132 *bo = LIST_FIRST(slp); 2133 if (*bo == NULL) 2134 return (0); 2135 vp = bo2vnode(*bo); 2136 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) 2137 return (1); 2138 /* 2139 * We use vhold in case the vnode does not 2140 * successfully sync. vhold prevents the vnode from 2141 * going away when we unlock the sync_mtx so that 2142 * we can acquire the vnode interlock. 2143 */ 2144 vholdl(vp); 2145 mtx_unlock(&sync_mtx); 2146 VI_UNLOCK(vp); 2147 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2148 vdrop(vp); 2149 mtx_lock(&sync_mtx); 2150 return (*bo == LIST_FIRST(slp)); 2151 } 2152 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2153 (void) VOP_FSYNC(vp, MNT_LAZY, td); 2154 VOP_UNLOCK(vp, 0); 2155 vn_finished_write(mp); 2156 BO_LOCK(*bo); 2157 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { 2158 /* 2159 * Put us back on the worklist. The worklist 2160 * routine will remove us from our current 2161 * position and then add us back in at a later 2162 * position. 2163 */ 2164 vn_syncer_add_to_worklist(*bo, syncdelay); 2165 } 2166 BO_UNLOCK(*bo); 2167 vdrop(vp); 2168 mtx_lock(&sync_mtx); 2169 return (0); 2170 } 2171 2172 static int first_printf = 1; 2173 2174 /* 2175 * System filesystem synchronizer daemon. 2176 */ 2177 static void 2178 sched_sync(void) 2179 { 2180 struct synclist *next, *slp; 2181 struct bufobj *bo; 2182 long starttime; 2183 struct thread *td = curthread; 2184 int last_work_seen; 2185 int net_worklist_len; 2186 int syncer_final_iter; 2187 int error; 2188 2189 last_work_seen = 0; 2190 syncer_final_iter = 0; 2191 syncer_state = SYNCER_RUNNING; 2192 starttime = time_uptime; 2193 td->td_pflags |= TDP_NORUNNINGBUF; 2194 2195 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 2196 SHUTDOWN_PRI_LAST); 2197 2198 mtx_lock(&sync_mtx); 2199 for (;;) { 2200 if (syncer_state == SYNCER_FINAL_DELAY && 2201 syncer_final_iter == 0) { 2202 mtx_unlock(&sync_mtx); 2203 kproc_suspend_check(td->td_proc); 2204 mtx_lock(&sync_mtx); 2205 } 2206 net_worklist_len = syncer_worklist_len - sync_vnode_count; 2207 if (syncer_state != SYNCER_RUNNING && 2208 starttime != time_uptime) { 2209 if (first_printf) { 2210 printf("\nSyncing disks, vnodes remaining... "); 2211 first_printf = 0; 2212 } 2213 printf("%d ", net_worklist_len); 2214 } 2215 starttime = time_uptime; 2216 2217 /* 2218 * Push files whose dirty time has expired. Be careful 2219 * of interrupt race on slp queue. 2220 * 2221 * Skip over empty worklist slots when shutting down. 2222 */ 2223 do { 2224 slp = &syncer_workitem_pending[syncer_delayno]; 2225 syncer_delayno += 1; 2226 if (syncer_delayno == syncer_maxdelay) 2227 syncer_delayno = 0; 2228 next = &syncer_workitem_pending[syncer_delayno]; 2229 /* 2230 * If the worklist has wrapped since the 2231 * it was emptied of all but syncer vnodes, 2232 * switch to the FINAL_DELAY state and run 2233 * for one more second. 2234 */ 2235 if (syncer_state == SYNCER_SHUTTING_DOWN && 2236 net_worklist_len == 0 && 2237 last_work_seen == syncer_delayno) { 2238 syncer_state = SYNCER_FINAL_DELAY; 2239 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 2240 } 2241 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 2242 syncer_worklist_len > 0); 2243 2244 /* 2245 * Keep track of the last time there was anything 2246 * on the worklist other than syncer vnodes. 2247 * Return to the SHUTTING_DOWN state if any 2248 * new work appears. 2249 */ 2250 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 2251 last_work_seen = syncer_delayno; 2252 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 2253 syncer_state = SYNCER_SHUTTING_DOWN; 2254 while (!LIST_EMPTY(slp)) { 2255 error = sync_vnode(slp, &bo, td); 2256 if (error == 1) { 2257 LIST_REMOVE(bo, bo_synclist); 2258 LIST_INSERT_HEAD(next, bo, bo_synclist); 2259 continue; 2260 } 2261 2262 if (first_printf == 0) { 2263 /* 2264 * Drop the sync mutex, because some watchdog 2265 * drivers need to sleep while patting 2266 */ 2267 mtx_unlock(&sync_mtx); 2268 wdog_kern_pat(WD_LASTVAL); 2269 mtx_lock(&sync_mtx); 2270 } 2271 2272 } 2273 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 2274 syncer_final_iter--; 2275 /* 2276 * The variable rushjob allows the kernel to speed up the 2277 * processing of the filesystem syncer process. A rushjob 2278 * value of N tells the filesystem syncer to process the next 2279 * N seconds worth of work on its queue ASAP. Currently rushjob 2280 * is used by the soft update code to speed up the filesystem 2281 * syncer process when the incore state is getting so far 2282 * ahead of the disk that the kernel memory pool is being 2283 * threatened with exhaustion. 2284 */ 2285 if (rushjob > 0) { 2286 rushjob -= 1; 2287 continue; 2288 } 2289 /* 2290 * Just sleep for a short period of time between 2291 * iterations when shutting down to allow some I/O 2292 * to happen. 2293 * 2294 * If it has taken us less than a second to process the 2295 * current work, then wait. Otherwise start right over 2296 * again. We can still lose time if any single round 2297 * takes more than two seconds, but it does not really 2298 * matter as we are just trying to generally pace the 2299 * filesystem activity. 2300 */ 2301 if (syncer_state != SYNCER_RUNNING || 2302 time_uptime == starttime) { 2303 thread_lock(td); 2304 sched_prio(td, PPAUSE); 2305 thread_unlock(td); 2306 } 2307 if (syncer_state != SYNCER_RUNNING) 2308 cv_timedwait(&sync_wakeup, &sync_mtx, 2309 hz / SYNCER_SHUTDOWN_SPEEDUP); 2310 else if (time_uptime == starttime) 2311 cv_timedwait(&sync_wakeup, &sync_mtx, hz); 2312 } 2313 } 2314 2315 /* 2316 * Request the syncer daemon to speed up its work. 2317 * We never push it to speed up more than half of its 2318 * normal turn time, otherwise it could take over the cpu. 2319 */ 2320 int 2321 speedup_syncer(void) 2322 { 2323 int ret = 0; 2324 2325 mtx_lock(&sync_mtx); 2326 if (rushjob < syncdelay / 2) { 2327 rushjob += 1; 2328 stat_rush_requests += 1; 2329 ret = 1; 2330 } 2331 mtx_unlock(&sync_mtx); 2332 cv_broadcast(&sync_wakeup); 2333 return (ret); 2334 } 2335 2336 /* 2337 * Tell the syncer to speed up its work and run though its work 2338 * list several times, then tell it to shut down. 2339 */ 2340 static void 2341 syncer_shutdown(void *arg, int howto) 2342 { 2343 2344 if (howto & RB_NOSYNC) 2345 return; 2346 mtx_lock(&sync_mtx); 2347 syncer_state = SYNCER_SHUTTING_DOWN; 2348 rushjob = 0; 2349 mtx_unlock(&sync_mtx); 2350 cv_broadcast(&sync_wakeup); 2351 kproc_shutdown(arg, howto); 2352 } 2353 2354 void 2355 syncer_suspend(void) 2356 { 2357 2358 syncer_shutdown(updateproc, 0); 2359 } 2360 2361 void 2362 syncer_resume(void) 2363 { 2364 2365 mtx_lock(&sync_mtx); 2366 first_printf = 1; 2367 syncer_state = SYNCER_RUNNING; 2368 mtx_unlock(&sync_mtx); 2369 cv_broadcast(&sync_wakeup); 2370 kproc_resume(updateproc); 2371 } 2372 2373 /* 2374 * Reassign a buffer from one vnode to another. 2375 * Used to assign file specific control information 2376 * (indirect blocks) to the vnode to which they belong. 2377 */ 2378 void 2379 reassignbuf(struct buf *bp) 2380 { 2381 struct vnode *vp; 2382 struct bufobj *bo; 2383 int delay; 2384 #ifdef INVARIANTS 2385 struct bufv *bv; 2386 #endif 2387 2388 vp = bp->b_vp; 2389 bo = bp->b_bufobj; 2390 ++reassignbufcalls; 2391 2392 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 2393 bp, bp->b_vp, bp->b_flags); 2394 /* 2395 * B_PAGING flagged buffers cannot be reassigned because their vp 2396 * is not fully linked in. 2397 */ 2398 if (bp->b_flags & B_PAGING) 2399 panic("cannot reassign paging buffer"); 2400 2401 /* 2402 * Delete from old vnode list, if on one. 2403 */ 2404 BO_LOCK(bo); 2405 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2406 buf_vlist_remove(bp); 2407 else 2408 panic("reassignbuf: Buffer %p not on queue.", bp); 2409 /* 2410 * If dirty, put on list of dirty buffers; otherwise insert onto list 2411 * of clean buffers. 2412 */ 2413 if (bp->b_flags & B_DELWRI) { 2414 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 2415 switch (vp->v_type) { 2416 case VDIR: 2417 delay = dirdelay; 2418 break; 2419 case VCHR: 2420 delay = metadelay; 2421 break; 2422 default: 2423 delay = filedelay; 2424 } 2425 vn_syncer_add_to_worklist(bo, delay); 2426 } 2427 buf_vlist_add(bp, bo, BX_VNDIRTY); 2428 } else { 2429 buf_vlist_add(bp, bo, BX_VNCLEAN); 2430 2431 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2432 mtx_lock(&sync_mtx); 2433 LIST_REMOVE(bo, bo_synclist); 2434 syncer_worklist_len--; 2435 mtx_unlock(&sync_mtx); 2436 bo->bo_flag &= ~BO_ONWORKLST; 2437 } 2438 } 2439 #ifdef INVARIANTS 2440 bv = &bo->bo_clean; 2441 bp = TAILQ_FIRST(&bv->bv_hd); 2442 KASSERT(bp == NULL || bp->b_bufobj == bo, 2443 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2444 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2445 KASSERT(bp == NULL || bp->b_bufobj == bo, 2446 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2447 bv = &bo->bo_dirty; 2448 bp = TAILQ_FIRST(&bv->bv_hd); 2449 KASSERT(bp == NULL || bp->b_bufobj == bo, 2450 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2451 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2452 KASSERT(bp == NULL || bp->b_bufobj == bo, 2453 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2454 #endif 2455 BO_UNLOCK(bo); 2456 } 2457 2458 /* 2459 * A temporary hack until refcount_* APIs are sorted out. 2460 */ 2461 static __inline int 2462 vfs_refcount_acquire_if_not_zero(volatile u_int *count) 2463 { 2464 u_int old; 2465 2466 old = *count; 2467 for (;;) { 2468 if (old == 0) 2469 return (0); 2470 if (atomic_fcmpset_int(count, &old, old + 1)) 2471 return (1); 2472 } 2473 } 2474 2475 static __inline int 2476 vfs_refcount_release_if_not_last(volatile u_int *count) 2477 { 2478 u_int old; 2479 2480 old = *count; 2481 for (;;) { 2482 if (old == 1) 2483 return (0); 2484 if (atomic_fcmpset_int(count, &old, old - 1)) 2485 return (1); 2486 } 2487 } 2488 2489 static void 2490 v_init_counters(struct vnode *vp) 2491 { 2492 2493 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0, 2494 vp, ("%s called for an initialized vnode", __FUNCTION__)); 2495 ASSERT_VI_UNLOCKED(vp, __FUNCTION__); 2496 2497 refcount_init(&vp->v_holdcnt, 1); 2498 refcount_init(&vp->v_usecount, 1); 2499 } 2500 2501 static void 2502 v_incr_usecount_locked(struct vnode *vp) 2503 { 2504 2505 ASSERT_VI_LOCKED(vp, __func__); 2506 if ((vp->v_iflag & VI_OWEINACT) != 0) { 2507 VNASSERT(vp->v_usecount == 0, vp, 2508 ("vnode with usecount and VI_OWEINACT set")); 2509 vp->v_iflag &= ~VI_OWEINACT; 2510 } 2511 refcount_acquire(&vp->v_usecount); 2512 v_incr_devcount(vp); 2513 } 2514 2515 /* 2516 * Increment the use count on the vnode, taking care to reference 2517 * the driver's usecount if this is a chardev. 2518 */ 2519 static void 2520 v_incr_usecount(struct vnode *vp) 2521 { 2522 2523 ASSERT_VI_UNLOCKED(vp, __func__); 2524 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2525 2526 if (vp->v_type != VCHR && 2527 vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) { 2528 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2529 ("vnode with usecount and VI_OWEINACT set")); 2530 } else { 2531 VI_LOCK(vp); 2532 v_incr_usecount_locked(vp); 2533 VI_UNLOCK(vp); 2534 } 2535 } 2536 2537 /* 2538 * Increment si_usecount of the associated device, if any. 2539 */ 2540 static void 2541 v_incr_devcount(struct vnode *vp) 2542 { 2543 2544 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2545 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2546 dev_lock(); 2547 vp->v_rdev->si_usecount++; 2548 dev_unlock(); 2549 } 2550 } 2551 2552 /* 2553 * Decrement si_usecount of the associated device, if any. 2554 */ 2555 static void 2556 v_decr_devcount(struct vnode *vp) 2557 { 2558 2559 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2560 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2561 dev_lock(); 2562 vp->v_rdev->si_usecount--; 2563 dev_unlock(); 2564 } 2565 } 2566 2567 /* 2568 * Grab a particular vnode from the free list, increment its 2569 * reference count and lock it. VI_DOOMED is set if the vnode 2570 * is being destroyed. Only callers who specify LK_RETRY will 2571 * see doomed vnodes. If inactive processing was delayed in 2572 * vput try to do it here. 2573 * 2574 * Notes on lockless counter manipulation: 2575 * _vhold, vputx and other routines make various decisions based 2576 * on either holdcnt or usecount being 0. As long as either counter 2577 * is not transitioning 0->1 nor 1->0, the manipulation can be done 2578 * with atomic operations. Otherwise the interlock is taken covering 2579 * both the atomic and additional actions. 2580 */ 2581 int 2582 vget(struct vnode *vp, int flags, struct thread *td) 2583 { 2584 int error, oweinact; 2585 2586 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 2587 ("vget: invalid lock operation")); 2588 2589 if ((flags & LK_INTERLOCK) != 0) 2590 ASSERT_VI_LOCKED(vp, __func__); 2591 else 2592 ASSERT_VI_UNLOCKED(vp, __func__); 2593 if ((flags & LK_VNHELD) != 0) 2594 VNASSERT((vp->v_holdcnt > 0), vp, 2595 ("vget: LK_VNHELD passed but vnode not held")); 2596 2597 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 2598 2599 if ((flags & LK_VNHELD) == 0) 2600 _vhold(vp, (flags & LK_INTERLOCK) != 0); 2601 2602 if ((error = vn_lock(vp, flags)) != 0) { 2603 vdrop(vp); 2604 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, 2605 vp); 2606 return (error); 2607 } 2608 if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0) 2609 panic("vget: vn_lock failed to return ENOENT\n"); 2610 /* 2611 * We don't guarantee that any particular close will 2612 * trigger inactive processing so just make a best effort 2613 * here at preventing a reference to a removed file. If 2614 * we don't succeed no harm is done. 2615 * 2616 * Upgrade our holdcnt to a usecount. 2617 */ 2618 if (vp->v_type == VCHR || 2619 !vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) { 2620 VI_LOCK(vp); 2621 if ((vp->v_iflag & VI_OWEINACT) == 0) { 2622 oweinact = 0; 2623 } else { 2624 oweinact = 1; 2625 vp->v_iflag &= ~VI_OWEINACT; 2626 } 2627 refcount_acquire(&vp->v_usecount); 2628 v_incr_devcount(vp); 2629 if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE && 2630 (flags & LK_NOWAIT) == 0) 2631 vinactive(vp, td); 2632 VI_UNLOCK(vp); 2633 } 2634 return (0); 2635 } 2636 2637 /* 2638 * Increase the reference (use) and hold count of a vnode. 2639 * This will also remove the vnode from the free list if it is presently free. 2640 */ 2641 void 2642 vref(struct vnode *vp) 2643 { 2644 2645 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2646 _vhold(vp, false); 2647 v_incr_usecount(vp); 2648 } 2649 2650 void 2651 vrefl(struct vnode *vp) 2652 { 2653 2654 ASSERT_VI_LOCKED(vp, __func__); 2655 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2656 _vhold(vp, true); 2657 v_incr_usecount_locked(vp); 2658 } 2659 2660 void 2661 vrefact(struct vnode *vp) 2662 { 2663 2664 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2665 if (__predict_false(vp->v_type == VCHR)) { 2666 VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp, 2667 ("%s: wrong ref counts", __func__)); 2668 vref(vp); 2669 return; 2670 } 2671 #ifdef INVARIANTS 2672 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 2673 VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__)); 2674 old = atomic_fetchadd_int(&vp->v_usecount, 1); 2675 VNASSERT(old > 0, vp, ("%s: wrong use count", __func__)); 2676 #else 2677 refcount_acquire(&vp->v_holdcnt); 2678 refcount_acquire(&vp->v_usecount); 2679 #endif 2680 } 2681 2682 /* 2683 * Return reference count of a vnode. 2684 * 2685 * The results of this call are only guaranteed when some mechanism is used to 2686 * stop other processes from gaining references to the vnode. This may be the 2687 * case if the caller holds the only reference. This is also useful when stale 2688 * data is acceptable as race conditions may be accounted for by some other 2689 * means. 2690 */ 2691 int 2692 vrefcnt(struct vnode *vp) 2693 { 2694 2695 return (vp->v_usecount); 2696 } 2697 2698 #define VPUTX_VRELE 1 2699 #define VPUTX_VPUT 2 2700 #define VPUTX_VUNREF 3 2701 2702 /* 2703 * Decrement the use and hold counts for a vnode. 2704 * 2705 * See an explanation near vget() as to why atomic operation is safe. 2706 */ 2707 static void 2708 vputx(struct vnode *vp, int func) 2709 { 2710 int error; 2711 2712 KASSERT(vp != NULL, ("vputx: null vp")); 2713 if (func == VPUTX_VUNREF) 2714 ASSERT_VOP_LOCKED(vp, "vunref"); 2715 else if (func == VPUTX_VPUT) 2716 ASSERT_VOP_LOCKED(vp, "vput"); 2717 else 2718 KASSERT(func == VPUTX_VRELE, ("vputx: wrong func")); 2719 ASSERT_VI_UNLOCKED(vp, __func__); 2720 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2721 2722 if (vp->v_type != VCHR && 2723 vfs_refcount_release_if_not_last(&vp->v_usecount)) { 2724 if (func == VPUTX_VPUT) 2725 VOP_UNLOCK(vp, 0); 2726 vdrop(vp); 2727 return; 2728 } 2729 2730 VI_LOCK(vp); 2731 2732 /* 2733 * We want to hold the vnode until the inactive finishes to 2734 * prevent vgone() races. We drop the use count here and the 2735 * hold count below when we're done. 2736 */ 2737 if (!refcount_release(&vp->v_usecount) || 2738 (vp->v_iflag & VI_DOINGINACT)) { 2739 if (func == VPUTX_VPUT) 2740 VOP_UNLOCK(vp, 0); 2741 v_decr_devcount(vp); 2742 vdropl(vp); 2743 return; 2744 } 2745 2746 v_decr_devcount(vp); 2747 2748 error = 0; 2749 2750 if (vp->v_usecount != 0) { 2751 vn_printf(vp, "vputx: usecount not zero for vnode "); 2752 panic("vputx: usecount not zero"); 2753 } 2754 2755 CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp); 2756 2757 /* 2758 * We must call VOP_INACTIVE with the node locked. Mark 2759 * as VI_DOINGINACT to avoid recursion. 2760 */ 2761 vp->v_iflag |= VI_OWEINACT; 2762 switch (func) { 2763 case VPUTX_VRELE: 2764 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 2765 VI_LOCK(vp); 2766 break; 2767 case VPUTX_VPUT: 2768 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 2769 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | 2770 LK_NOWAIT); 2771 VI_LOCK(vp); 2772 } 2773 break; 2774 case VPUTX_VUNREF: 2775 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 2776 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); 2777 VI_LOCK(vp); 2778 } 2779 break; 2780 } 2781 VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp, 2782 ("vnode with usecount and VI_OWEINACT set")); 2783 if (error == 0) { 2784 if (vp->v_iflag & VI_OWEINACT) 2785 vinactive(vp, curthread); 2786 if (func != VPUTX_VUNREF) 2787 VOP_UNLOCK(vp, 0); 2788 } 2789 vdropl(vp); 2790 } 2791 2792 /* 2793 * Vnode put/release. 2794 * If count drops to zero, call inactive routine and return to freelist. 2795 */ 2796 void 2797 vrele(struct vnode *vp) 2798 { 2799 2800 vputx(vp, VPUTX_VRELE); 2801 } 2802 2803 /* 2804 * Release an already locked vnode. This give the same effects as 2805 * unlock+vrele(), but takes less time and avoids releasing and 2806 * re-aquiring the lock (as vrele() acquires the lock internally.) 2807 */ 2808 void 2809 vput(struct vnode *vp) 2810 { 2811 2812 vputx(vp, VPUTX_VPUT); 2813 } 2814 2815 /* 2816 * Release an exclusively locked vnode. Do not unlock the vnode lock. 2817 */ 2818 void 2819 vunref(struct vnode *vp) 2820 { 2821 2822 vputx(vp, VPUTX_VUNREF); 2823 } 2824 2825 /* 2826 * Increase the hold count and activate if this is the first reference. 2827 */ 2828 void 2829 _vhold(struct vnode *vp, bool locked) 2830 { 2831 struct mount *mp; 2832 2833 if (locked) 2834 ASSERT_VI_LOCKED(vp, __func__); 2835 else 2836 ASSERT_VI_UNLOCKED(vp, __func__); 2837 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2838 if (!locked && vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt)) { 2839 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 2840 ("_vhold: vnode with holdcnt is free")); 2841 return; 2842 } 2843 2844 if (!locked) 2845 VI_LOCK(vp); 2846 if ((vp->v_iflag & VI_FREE) == 0) { 2847 refcount_acquire(&vp->v_holdcnt); 2848 if (!locked) 2849 VI_UNLOCK(vp); 2850 return; 2851 } 2852 VNASSERT(vp->v_holdcnt == 0, vp, 2853 ("%s: wrong hold count", __func__)); 2854 VNASSERT(vp->v_op != NULL, vp, 2855 ("%s: vnode already reclaimed.", __func__)); 2856 /* 2857 * Remove a vnode from the free list, mark it as in use, 2858 * and put it on the active list. 2859 */ 2860 mp = vp->v_mount; 2861 mtx_lock(&mp->mnt_listmtx); 2862 if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) { 2863 TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist); 2864 mp->mnt_tmpfreevnodelistsize--; 2865 vp->v_mflag &= ~VMP_TMPMNTFREELIST; 2866 } else { 2867 mtx_lock(&vnode_free_list_mtx); 2868 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist); 2869 freevnodes--; 2870 mtx_unlock(&vnode_free_list_mtx); 2871 } 2872 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 2873 ("Activating already active vnode")); 2874 vp->v_iflag &= ~VI_FREE; 2875 vp->v_iflag |= VI_ACTIVE; 2876 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist); 2877 mp->mnt_activevnodelistsize++; 2878 mtx_unlock(&mp->mnt_listmtx); 2879 refcount_acquire(&vp->v_holdcnt); 2880 if (!locked) 2881 VI_UNLOCK(vp); 2882 } 2883 2884 /* 2885 * Drop the hold count of the vnode. If this is the last reference to 2886 * the vnode we place it on the free list unless it has been vgone'd 2887 * (marked VI_DOOMED) in which case we will free it. 2888 * 2889 * Because the vnode vm object keeps a hold reference on the vnode if 2890 * there is at least one resident non-cached page, the vnode cannot 2891 * leave the active list without the page cleanup done. 2892 */ 2893 void 2894 _vdrop(struct vnode *vp, bool locked) 2895 { 2896 struct bufobj *bo; 2897 struct mount *mp; 2898 int active; 2899 2900 if (locked) 2901 ASSERT_VI_LOCKED(vp, __func__); 2902 else 2903 ASSERT_VI_UNLOCKED(vp, __func__); 2904 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2905 if ((int)vp->v_holdcnt <= 0) 2906 panic("vdrop: holdcnt %d", vp->v_holdcnt); 2907 if (vfs_refcount_release_if_not_last(&vp->v_holdcnt)) { 2908 if (locked) 2909 VI_UNLOCK(vp); 2910 return; 2911 } 2912 2913 if (!locked) 2914 VI_LOCK(vp); 2915 if (refcount_release(&vp->v_holdcnt) == 0) { 2916 VI_UNLOCK(vp); 2917 return; 2918 } 2919 if ((vp->v_iflag & VI_DOOMED) == 0) { 2920 /* 2921 * Mark a vnode as free: remove it from its active list 2922 * and put it up for recycling on the freelist. 2923 */ 2924 VNASSERT(vp->v_op != NULL, vp, 2925 ("vdropl: vnode already reclaimed.")); 2926 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 2927 ("vnode already free")); 2928 VNASSERT(vp->v_holdcnt == 0, vp, 2929 ("vdropl: freeing when we shouldn't")); 2930 active = vp->v_iflag & VI_ACTIVE; 2931 if ((vp->v_iflag & VI_OWEINACT) == 0) { 2932 vp->v_iflag &= ~VI_ACTIVE; 2933 mp = vp->v_mount; 2934 mtx_lock(&mp->mnt_listmtx); 2935 if (active) { 2936 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, 2937 v_actfreelist); 2938 mp->mnt_activevnodelistsize--; 2939 } 2940 TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist, vp, 2941 v_actfreelist); 2942 mp->mnt_tmpfreevnodelistsize++; 2943 vp->v_iflag |= VI_FREE; 2944 vp->v_mflag |= VMP_TMPMNTFREELIST; 2945 VI_UNLOCK(vp); 2946 if (mp->mnt_tmpfreevnodelistsize >= mnt_free_list_batch) 2947 vnlru_return_batch_locked(mp); 2948 mtx_unlock(&mp->mnt_listmtx); 2949 } else { 2950 VI_UNLOCK(vp); 2951 counter_u64_add(free_owe_inact, 1); 2952 } 2953 return; 2954 } 2955 /* 2956 * The vnode has been marked for destruction, so free it. 2957 * 2958 * The vnode will be returned to the zone where it will 2959 * normally remain until it is needed for another vnode. We 2960 * need to cleanup (or verify that the cleanup has already 2961 * been done) any residual data left from its current use 2962 * so as not to contaminate the freshly allocated vnode. 2963 */ 2964 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp); 2965 atomic_subtract_long(&numvnodes, 1); 2966 bo = &vp->v_bufobj; 2967 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 2968 ("cleaned vnode still on the free list.")); 2969 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 2970 VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count")); 2971 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 2972 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 2973 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 2974 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 2975 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp, 2976 ("clean blk trie not empty")); 2977 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 2978 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp, 2979 ("dirty blk trie not empty")); 2980 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); 2981 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); 2982 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for ..")); 2983 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp, 2984 ("Dangling rangelock waiters")); 2985 VI_UNLOCK(vp); 2986 #ifdef MAC 2987 mac_vnode_destroy(vp); 2988 #endif 2989 if (vp->v_pollinfo != NULL) { 2990 destroy_vpollinfo(vp->v_pollinfo); 2991 vp->v_pollinfo = NULL; 2992 } 2993 #ifdef INVARIANTS 2994 /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */ 2995 vp->v_op = NULL; 2996 #endif 2997 bzero(&vp->v_un, sizeof(vp->v_un)); 2998 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 2999 vp->v_iflag = 0; 3000 vp->v_vflag = 0; 3001 bo->bo_flag = 0; 3002 uma_zfree(vnode_zone, vp); 3003 } 3004 3005 /* 3006 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 3007 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 3008 * OWEINACT tracks whether a vnode missed a call to inactive due to a 3009 * failed lock upgrade. 3010 */ 3011 void 3012 vinactive(struct vnode *vp, struct thread *td) 3013 { 3014 struct vm_object *obj; 3015 3016 ASSERT_VOP_ELOCKED(vp, "vinactive"); 3017 ASSERT_VI_LOCKED(vp, "vinactive"); 3018 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, 3019 ("vinactive: recursed on VI_DOINGINACT")); 3020 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3021 vp->v_iflag |= VI_DOINGINACT; 3022 vp->v_iflag &= ~VI_OWEINACT; 3023 VI_UNLOCK(vp); 3024 /* 3025 * Before moving off the active list, we must be sure that any 3026 * modified pages are converted into the vnode's dirty 3027 * buffers, since these will no longer be checked once the 3028 * vnode is on the inactive list. 3029 * 3030 * The write-out of the dirty pages is asynchronous. At the 3031 * point that VOP_INACTIVE() is called, there could still be 3032 * pending I/O and dirty pages in the object. 3033 */ 3034 obj = vp->v_object; 3035 if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0) { 3036 VM_OBJECT_WLOCK(obj); 3037 vm_object_page_clean(obj, 0, 0, 0); 3038 VM_OBJECT_WUNLOCK(obj); 3039 } 3040 VOP_INACTIVE(vp, td); 3041 VI_LOCK(vp); 3042 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, 3043 ("vinactive: lost VI_DOINGINACT")); 3044 vp->v_iflag &= ~VI_DOINGINACT; 3045 } 3046 3047 /* 3048 * Remove any vnodes in the vnode table belonging to mount point mp. 3049 * 3050 * If FORCECLOSE is not specified, there should not be any active ones, 3051 * return error if any are found (nb: this is a user error, not a 3052 * system error). If FORCECLOSE is specified, detach any active vnodes 3053 * that are found. 3054 * 3055 * If WRITECLOSE is set, only flush out regular file vnodes open for 3056 * writing. 3057 * 3058 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 3059 * 3060 * `rootrefs' specifies the base reference count for the root vnode 3061 * of this filesystem. The root vnode is considered busy if its 3062 * v_usecount exceeds this value. On a successful return, vflush(, td) 3063 * will call vrele() on the root vnode exactly rootrefs times. 3064 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 3065 * be zero. 3066 */ 3067 #ifdef DIAGNOSTIC 3068 static int busyprt = 0; /* print out busy vnodes */ 3069 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 3070 #endif 3071 3072 int 3073 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 3074 { 3075 struct vnode *vp, *mvp, *rootvp = NULL; 3076 struct vattr vattr; 3077 int busy = 0, error; 3078 3079 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 3080 rootrefs, flags); 3081 if (rootrefs > 0) { 3082 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 3083 ("vflush: bad args")); 3084 /* 3085 * Get the filesystem root vnode. We can vput() it 3086 * immediately, since with rootrefs > 0, it won't go away. 3087 */ 3088 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 3089 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 3090 __func__, error); 3091 return (error); 3092 } 3093 vput(rootvp); 3094 } 3095 loop: 3096 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 3097 vholdl(vp); 3098 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 3099 if (error) { 3100 vdrop(vp); 3101 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3102 goto loop; 3103 } 3104 /* 3105 * Skip over a vnodes marked VV_SYSTEM. 3106 */ 3107 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 3108 VOP_UNLOCK(vp, 0); 3109 vdrop(vp); 3110 continue; 3111 } 3112 /* 3113 * If WRITECLOSE is set, flush out unlinked but still open 3114 * files (even if open only for reading) and regular file 3115 * vnodes open for writing. 3116 */ 3117 if (flags & WRITECLOSE) { 3118 if (vp->v_object != NULL) { 3119 VM_OBJECT_WLOCK(vp->v_object); 3120 vm_object_page_clean(vp->v_object, 0, 0, 0); 3121 VM_OBJECT_WUNLOCK(vp->v_object); 3122 } 3123 error = VOP_FSYNC(vp, MNT_WAIT, td); 3124 if (error != 0) { 3125 VOP_UNLOCK(vp, 0); 3126 vdrop(vp); 3127 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3128 return (error); 3129 } 3130 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 3131 VI_LOCK(vp); 3132 3133 if ((vp->v_type == VNON || 3134 (error == 0 && vattr.va_nlink > 0)) && 3135 (vp->v_writecount == 0 || vp->v_type != VREG)) { 3136 VOP_UNLOCK(vp, 0); 3137 vdropl(vp); 3138 continue; 3139 } 3140 } else 3141 VI_LOCK(vp); 3142 /* 3143 * With v_usecount == 0, all we need to do is clear out the 3144 * vnode data structures and we are done. 3145 * 3146 * If FORCECLOSE is set, forcibly close the vnode. 3147 */ 3148 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 3149 vgonel(vp); 3150 } else { 3151 busy++; 3152 #ifdef DIAGNOSTIC 3153 if (busyprt) 3154 vn_printf(vp, "vflush: busy vnode "); 3155 #endif 3156 } 3157 VOP_UNLOCK(vp, 0); 3158 vdropl(vp); 3159 } 3160 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 3161 /* 3162 * If just the root vnode is busy, and if its refcount 3163 * is equal to `rootrefs', then go ahead and kill it. 3164 */ 3165 VI_LOCK(rootvp); 3166 KASSERT(busy > 0, ("vflush: not busy")); 3167 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 3168 ("vflush: usecount %d < rootrefs %d", 3169 rootvp->v_usecount, rootrefs)); 3170 if (busy == 1 && rootvp->v_usecount == rootrefs) { 3171 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 3172 vgone(rootvp); 3173 VOP_UNLOCK(rootvp, 0); 3174 busy = 0; 3175 } else 3176 VI_UNLOCK(rootvp); 3177 } 3178 if (busy) { 3179 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 3180 busy); 3181 return (EBUSY); 3182 } 3183 for (; rootrefs > 0; rootrefs--) 3184 vrele(rootvp); 3185 return (0); 3186 } 3187 3188 /* 3189 * Recycle an unused vnode to the front of the free list. 3190 */ 3191 int 3192 vrecycle(struct vnode *vp) 3193 { 3194 int recycled; 3195 3196 VI_LOCK(vp); 3197 recycled = vrecyclel(vp); 3198 VI_UNLOCK(vp); 3199 return (recycled); 3200 } 3201 3202 /* 3203 * vrecycle, with the vp interlock held. 3204 */ 3205 int 3206 vrecyclel(struct vnode *vp) 3207 { 3208 int recycled; 3209 3210 ASSERT_VOP_ELOCKED(vp, __func__); 3211 ASSERT_VI_LOCKED(vp, __func__); 3212 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3213 recycled = 0; 3214 if (vp->v_usecount == 0) { 3215 recycled = 1; 3216 vgonel(vp); 3217 } 3218 return (recycled); 3219 } 3220 3221 /* 3222 * Eliminate all activity associated with a vnode 3223 * in preparation for reuse. 3224 */ 3225 void 3226 vgone(struct vnode *vp) 3227 { 3228 VI_LOCK(vp); 3229 vgonel(vp); 3230 VI_UNLOCK(vp); 3231 } 3232 3233 static void 3234 notify_lowervp_vfs_dummy(struct mount *mp __unused, 3235 struct vnode *lowervp __unused) 3236 { 3237 } 3238 3239 /* 3240 * Notify upper mounts about reclaimed or unlinked vnode. 3241 */ 3242 void 3243 vfs_notify_upper(struct vnode *vp, int event) 3244 { 3245 static struct vfsops vgonel_vfsops = { 3246 .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy, 3247 .vfs_unlink_lowervp = notify_lowervp_vfs_dummy, 3248 }; 3249 struct mount *mp, *ump, *mmp; 3250 3251 mp = vp->v_mount; 3252 if (mp == NULL) 3253 return; 3254 3255 MNT_ILOCK(mp); 3256 if (TAILQ_EMPTY(&mp->mnt_uppers)) 3257 goto unlock; 3258 MNT_IUNLOCK(mp); 3259 mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO); 3260 mmp->mnt_op = &vgonel_vfsops; 3261 mmp->mnt_kern_flag |= MNTK_MARKER; 3262 MNT_ILOCK(mp); 3263 mp->mnt_kern_flag |= MNTK_VGONE_UPPER; 3264 for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) { 3265 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) { 3266 ump = TAILQ_NEXT(ump, mnt_upper_link); 3267 continue; 3268 } 3269 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link); 3270 MNT_IUNLOCK(mp); 3271 switch (event) { 3272 case VFS_NOTIFY_UPPER_RECLAIM: 3273 VFS_RECLAIM_LOWERVP(ump, vp); 3274 break; 3275 case VFS_NOTIFY_UPPER_UNLINK: 3276 VFS_UNLINK_LOWERVP(ump, vp); 3277 break; 3278 default: 3279 KASSERT(0, ("invalid event %d", event)); 3280 break; 3281 } 3282 MNT_ILOCK(mp); 3283 ump = TAILQ_NEXT(mmp, mnt_upper_link); 3284 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link); 3285 } 3286 free(mmp, M_TEMP); 3287 mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER; 3288 if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) { 3289 mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER; 3290 wakeup(&mp->mnt_uppers); 3291 } 3292 unlock: 3293 MNT_IUNLOCK(mp); 3294 } 3295 3296 /* 3297 * vgone, with the vp interlock held. 3298 */ 3299 static void 3300 vgonel(struct vnode *vp) 3301 { 3302 struct thread *td; 3303 int oweinact; 3304 int active; 3305 struct mount *mp; 3306 3307 ASSERT_VOP_ELOCKED(vp, "vgonel"); 3308 ASSERT_VI_LOCKED(vp, "vgonel"); 3309 VNASSERT(vp->v_holdcnt, vp, 3310 ("vgonel: vp %p has no reference.", vp)); 3311 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3312 td = curthread; 3313 3314 /* 3315 * Don't vgonel if we're already doomed. 3316 */ 3317 if (vp->v_iflag & VI_DOOMED) 3318 return; 3319 vp->v_iflag |= VI_DOOMED; 3320 3321 /* 3322 * Check to see if the vnode is in use. If so, we have to call 3323 * VOP_CLOSE() and VOP_INACTIVE(). 3324 */ 3325 active = vp->v_usecount; 3326 oweinact = (vp->v_iflag & VI_OWEINACT); 3327 VI_UNLOCK(vp); 3328 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); 3329 3330 /* 3331 * If purging an active vnode, it must be closed and 3332 * deactivated before being reclaimed. 3333 */ 3334 if (active) 3335 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 3336 if (oweinact || active) { 3337 VI_LOCK(vp); 3338 if ((vp->v_iflag & VI_DOINGINACT) == 0) 3339 vinactive(vp, td); 3340 VI_UNLOCK(vp); 3341 } 3342 if (vp->v_type == VSOCK) 3343 vfs_unp_reclaim(vp); 3344 3345 /* 3346 * Clean out any buffers associated with the vnode. 3347 * If the flush fails, just toss the buffers. 3348 */ 3349 mp = NULL; 3350 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 3351 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 3352 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { 3353 while (vinvalbuf(vp, 0, 0, 0) != 0) 3354 ; 3355 } 3356 3357 BO_LOCK(&vp->v_bufobj); 3358 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && 3359 vp->v_bufobj.bo_dirty.bv_cnt == 0 && 3360 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && 3361 vp->v_bufobj.bo_clean.bv_cnt == 0, 3362 ("vp %p bufobj not invalidated", vp)); 3363 3364 /* 3365 * For VMIO bufobj, BO_DEAD is set in vm_object_terminate() 3366 * after the object's page queue is flushed. 3367 */ 3368 if (vp->v_bufobj.bo_object == NULL) 3369 vp->v_bufobj.bo_flag |= BO_DEAD; 3370 BO_UNLOCK(&vp->v_bufobj); 3371 3372 /* 3373 * Reclaim the vnode. 3374 */ 3375 if (VOP_RECLAIM(vp, td)) 3376 panic("vgone: cannot reclaim"); 3377 if (mp != NULL) 3378 vn_finished_secondary_write(mp); 3379 VNASSERT(vp->v_object == NULL, vp, 3380 ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag)); 3381 /* 3382 * Clear the advisory locks and wake up waiting threads. 3383 */ 3384 (void)VOP_ADVLOCKPURGE(vp); 3385 vp->v_lockf = NULL; 3386 /* 3387 * Delete from old mount point vnode list. 3388 */ 3389 delmntque(vp); 3390 cache_purge(vp); 3391 /* 3392 * Done with purge, reset to the standard lock and invalidate 3393 * the vnode. 3394 */ 3395 VI_LOCK(vp); 3396 vp->v_vnlock = &vp->v_lock; 3397 vp->v_op = &dead_vnodeops; 3398 vp->v_tag = "none"; 3399 vp->v_type = VBAD; 3400 } 3401 3402 /* 3403 * Calculate the total number of references to a special device. 3404 */ 3405 int 3406 vcount(struct vnode *vp) 3407 { 3408 int count; 3409 3410 dev_lock(); 3411 count = vp->v_rdev->si_usecount; 3412 dev_unlock(); 3413 return (count); 3414 } 3415 3416 /* 3417 * Same as above, but using the struct cdev *as argument 3418 */ 3419 int 3420 count_dev(struct cdev *dev) 3421 { 3422 int count; 3423 3424 dev_lock(); 3425 count = dev->si_usecount; 3426 dev_unlock(); 3427 return(count); 3428 } 3429 3430 /* 3431 * Print out a description of a vnode. 3432 */ 3433 static char *typename[] = 3434 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", 3435 "VMARKER"}; 3436 3437 void 3438 vn_printf(struct vnode *vp, const char *fmt, ...) 3439 { 3440 va_list ap; 3441 char buf[256], buf2[16]; 3442 u_long flags; 3443 3444 va_start(ap, fmt); 3445 vprintf(fmt, ap); 3446 va_end(ap); 3447 printf("%p: ", (void *)vp); 3448 printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]); 3449 printf(" usecount %d, writecount %d, refcount %d mountedhere %p\n", 3450 vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere); 3451 buf[0] = '\0'; 3452 buf[1] = '\0'; 3453 if (vp->v_vflag & VV_ROOT) 3454 strlcat(buf, "|VV_ROOT", sizeof(buf)); 3455 if (vp->v_vflag & VV_ISTTY) 3456 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 3457 if (vp->v_vflag & VV_NOSYNC) 3458 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 3459 if (vp->v_vflag & VV_ETERNALDEV) 3460 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 3461 if (vp->v_vflag & VV_CACHEDLABEL) 3462 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 3463 if (vp->v_vflag & VV_TEXT) 3464 strlcat(buf, "|VV_TEXT", sizeof(buf)); 3465 if (vp->v_vflag & VV_COPYONWRITE) 3466 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 3467 if (vp->v_vflag & VV_SYSTEM) 3468 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 3469 if (vp->v_vflag & VV_PROCDEP) 3470 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 3471 if (vp->v_vflag & VV_NOKNOTE) 3472 strlcat(buf, "|VV_NOKNOTE", sizeof(buf)); 3473 if (vp->v_vflag & VV_DELETED) 3474 strlcat(buf, "|VV_DELETED", sizeof(buf)); 3475 if (vp->v_vflag & VV_MD) 3476 strlcat(buf, "|VV_MD", sizeof(buf)); 3477 if (vp->v_vflag & VV_FORCEINSMQ) 3478 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 3479 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 3480 VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | 3481 VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ); 3482 if (flags != 0) { 3483 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 3484 strlcat(buf, buf2, sizeof(buf)); 3485 } 3486 if (vp->v_iflag & VI_MOUNT) 3487 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 3488 if (vp->v_iflag & VI_DOOMED) 3489 strlcat(buf, "|VI_DOOMED", sizeof(buf)); 3490 if (vp->v_iflag & VI_FREE) 3491 strlcat(buf, "|VI_FREE", sizeof(buf)); 3492 if (vp->v_iflag & VI_ACTIVE) 3493 strlcat(buf, "|VI_ACTIVE", sizeof(buf)); 3494 if (vp->v_iflag & VI_DOINGINACT) 3495 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 3496 if (vp->v_iflag & VI_OWEINACT) 3497 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 3498 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE | 3499 VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT); 3500 if (flags != 0) { 3501 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 3502 strlcat(buf, buf2, sizeof(buf)); 3503 } 3504 printf(" flags (%s)\n", buf + 1); 3505 if (mtx_owned(VI_MTX(vp))) 3506 printf(" VI_LOCKed"); 3507 if (vp->v_object != NULL) 3508 printf(" v_object %p ref %d pages %d " 3509 "cleanbuf %d dirtybuf %d\n", 3510 vp->v_object, vp->v_object->ref_count, 3511 vp->v_object->resident_page_count, 3512 vp->v_bufobj.bo_clean.bv_cnt, 3513 vp->v_bufobj.bo_dirty.bv_cnt); 3514 printf(" "); 3515 lockmgr_printinfo(vp->v_vnlock); 3516 if (vp->v_data != NULL) 3517 VOP_PRINT(vp); 3518 } 3519 3520 #ifdef DDB 3521 /* 3522 * List all of the locked vnodes in the system. 3523 * Called when debugging the kernel. 3524 */ 3525 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 3526 { 3527 struct mount *mp; 3528 struct vnode *vp; 3529 3530 /* 3531 * Note: because this is DDB, we can't obey the locking semantics 3532 * for these structures, which means we could catch an inconsistent 3533 * state and dereference a nasty pointer. Not much to be done 3534 * about that. 3535 */ 3536 db_printf("Locked vnodes\n"); 3537 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3538 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3539 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) 3540 vn_printf(vp, "vnode "); 3541 } 3542 } 3543 } 3544 3545 /* 3546 * Show details about the given vnode. 3547 */ 3548 DB_SHOW_COMMAND(vnode, db_show_vnode) 3549 { 3550 struct vnode *vp; 3551 3552 if (!have_addr) 3553 return; 3554 vp = (struct vnode *)addr; 3555 vn_printf(vp, "vnode "); 3556 } 3557 3558 /* 3559 * Show details about the given mount point. 3560 */ 3561 DB_SHOW_COMMAND(mount, db_show_mount) 3562 { 3563 struct mount *mp; 3564 struct vfsopt *opt; 3565 struct statfs *sp; 3566 struct vnode *vp; 3567 char buf[512]; 3568 uint64_t mflags; 3569 u_int flags; 3570 3571 if (!have_addr) { 3572 /* No address given, print short info about all mount points. */ 3573 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3574 db_printf("%p %s on %s (%s)\n", mp, 3575 mp->mnt_stat.f_mntfromname, 3576 mp->mnt_stat.f_mntonname, 3577 mp->mnt_stat.f_fstypename); 3578 if (db_pager_quit) 3579 break; 3580 } 3581 db_printf("\nMore info: show mount <addr>\n"); 3582 return; 3583 } 3584 3585 mp = (struct mount *)addr; 3586 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 3587 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 3588 3589 buf[0] = '\0'; 3590 mflags = mp->mnt_flag; 3591 #define MNT_FLAG(flag) do { \ 3592 if (mflags & (flag)) { \ 3593 if (buf[0] != '\0') \ 3594 strlcat(buf, ", ", sizeof(buf)); \ 3595 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 3596 mflags &= ~(flag); \ 3597 } \ 3598 } while (0) 3599 MNT_FLAG(MNT_RDONLY); 3600 MNT_FLAG(MNT_SYNCHRONOUS); 3601 MNT_FLAG(MNT_NOEXEC); 3602 MNT_FLAG(MNT_NOSUID); 3603 MNT_FLAG(MNT_NFS4ACLS); 3604 MNT_FLAG(MNT_UNION); 3605 MNT_FLAG(MNT_ASYNC); 3606 MNT_FLAG(MNT_SUIDDIR); 3607 MNT_FLAG(MNT_SOFTDEP); 3608 MNT_FLAG(MNT_NOSYMFOLLOW); 3609 MNT_FLAG(MNT_GJOURNAL); 3610 MNT_FLAG(MNT_MULTILABEL); 3611 MNT_FLAG(MNT_ACLS); 3612 MNT_FLAG(MNT_NOATIME); 3613 MNT_FLAG(MNT_NOCLUSTERR); 3614 MNT_FLAG(MNT_NOCLUSTERW); 3615 MNT_FLAG(MNT_SUJ); 3616 MNT_FLAG(MNT_EXRDONLY); 3617 MNT_FLAG(MNT_EXPORTED); 3618 MNT_FLAG(MNT_DEFEXPORTED); 3619 MNT_FLAG(MNT_EXPORTANON); 3620 MNT_FLAG(MNT_EXKERB); 3621 MNT_FLAG(MNT_EXPUBLIC); 3622 MNT_FLAG(MNT_LOCAL); 3623 MNT_FLAG(MNT_QUOTA); 3624 MNT_FLAG(MNT_ROOTFS); 3625 MNT_FLAG(MNT_USER); 3626 MNT_FLAG(MNT_IGNORE); 3627 MNT_FLAG(MNT_UPDATE); 3628 MNT_FLAG(MNT_DELEXPORT); 3629 MNT_FLAG(MNT_RELOAD); 3630 MNT_FLAG(MNT_FORCE); 3631 MNT_FLAG(MNT_SNAPSHOT); 3632 MNT_FLAG(MNT_BYFSID); 3633 #undef MNT_FLAG 3634 if (mflags != 0) { 3635 if (buf[0] != '\0') 3636 strlcat(buf, ", ", sizeof(buf)); 3637 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 3638 "0x%016jx", mflags); 3639 } 3640 db_printf(" mnt_flag = %s\n", buf); 3641 3642 buf[0] = '\0'; 3643 flags = mp->mnt_kern_flag; 3644 #define MNT_KERN_FLAG(flag) do { \ 3645 if (flags & (flag)) { \ 3646 if (buf[0] != '\0') \ 3647 strlcat(buf, ", ", sizeof(buf)); \ 3648 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 3649 flags &= ~(flag); \ 3650 } \ 3651 } while (0) 3652 MNT_KERN_FLAG(MNTK_UNMOUNTF); 3653 MNT_KERN_FLAG(MNTK_ASYNC); 3654 MNT_KERN_FLAG(MNTK_SOFTDEP); 3655 MNT_KERN_FLAG(MNTK_NOINSMNTQ); 3656 MNT_KERN_FLAG(MNTK_DRAINING); 3657 MNT_KERN_FLAG(MNTK_REFEXPIRE); 3658 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 3659 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 3660 MNT_KERN_FLAG(MNTK_NO_IOPF); 3661 MNT_KERN_FLAG(MNTK_VGONE_UPPER); 3662 MNT_KERN_FLAG(MNTK_VGONE_WAITER); 3663 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT); 3664 MNT_KERN_FLAG(MNTK_MARKER); 3665 MNT_KERN_FLAG(MNTK_USES_BCACHE); 3666 MNT_KERN_FLAG(MNTK_NOASYNC); 3667 MNT_KERN_FLAG(MNTK_UNMOUNT); 3668 MNT_KERN_FLAG(MNTK_MWAIT); 3669 MNT_KERN_FLAG(MNTK_SUSPEND); 3670 MNT_KERN_FLAG(MNTK_SUSPEND2); 3671 MNT_KERN_FLAG(MNTK_SUSPENDED); 3672 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 3673 MNT_KERN_FLAG(MNTK_NOKNOTE); 3674 #undef MNT_KERN_FLAG 3675 if (flags != 0) { 3676 if (buf[0] != '\0') 3677 strlcat(buf, ", ", sizeof(buf)); 3678 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 3679 "0x%08x", flags); 3680 } 3681 db_printf(" mnt_kern_flag = %s\n", buf); 3682 3683 db_printf(" mnt_opt = "); 3684 opt = TAILQ_FIRST(mp->mnt_opt); 3685 if (opt != NULL) { 3686 db_printf("%s", opt->name); 3687 opt = TAILQ_NEXT(opt, link); 3688 while (opt != NULL) { 3689 db_printf(", %s", opt->name); 3690 opt = TAILQ_NEXT(opt, link); 3691 } 3692 } 3693 db_printf("\n"); 3694 3695 sp = &mp->mnt_stat; 3696 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 3697 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 3698 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 3699 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 3700 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 3701 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 3702 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 3703 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 3704 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 3705 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 3706 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 3707 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 3708 3709 db_printf(" mnt_cred = { uid=%u ruid=%u", 3710 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 3711 if (jailed(mp->mnt_cred)) 3712 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 3713 db_printf(" }\n"); 3714 db_printf(" mnt_ref = %d\n", mp->mnt_ref); 3715 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 3716 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 3717 db_printf(" mnt_activevnodelistsize = %d\n", 3718 mp->mnt_activevnodelistsize); 3719 db_printf(" mnt_writeopcount = %d\n", mp->mnt_writeopcount); 3720 db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen); 3721 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 3722 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 3723 db_printf(" mnt_lockref = %d\n", mp->mnt_lockref); 3724 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 3725 db_printf(" mnt_secondary_accwrites = %d\n", 3726 mp->mnt_secondary_accwrites); 3727 db_printf(" mnt_gjprovider = %s\n", 3728 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 3729 3730 db_printf("\n\nList of active vnodes\n"); 3731 TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) { 3732 if (vp->v_type != VMARKER) { 3733 vn_printf(vp, "vnode "); 3734 if (db_pager_quit) 3735 break; 3736 } 3737 } 3738 db_printf("\n\nList of inactive vnodes\n"); 3739 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3740 if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) { 3741 vn_printf(vp, "vnode "); 3742 if (db_pager_quit) 3743 break; 3744 } 3745 } 3746 } 3747 #endif /* DDB */ 3748 3749 /* 3750 * Fill in a struct xvfsconf based on a struct vfsconf. 3751 */ 3752 static int 3753 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 3754 { 3755 struct xvfsconf xvfsp; 3756 3757 bzero(&xvfsp, sizeof(xvfsp)); 3758 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 3759 xvfsp.vfc_typenum = vfsp->vfc_typenum; 3760 xvfsp.vfc_refcount = vfsp->vfc_refcount; 3761 xvfsp.vfc_flags = vfsp->vfc_flags; 3762 /* 3763 * These are unused in userland, we keep them 3764 * to not break binary compatibility. 3765 */ 3766 xvfsp.vfc_vfsops = NULL; 3767 xvfsp.vfc_next = NULL; 3768 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 3769 } 3770 3771 #ifdef COMPAT_FREEBSD32 3772 struct xvfsconf32 { 3773 uint32_t vfc_vfsops; 3774 char vfc_name[MFSNAMELEN]; 3775 int32_t vfc_typenum; 3776 int32_t vfc_refcount; 3777 int32_t vfc_flags; 3778 uint32_t vfc_next; 3779 }; 3780 3781 static int 3782 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 3783 { 3784 struct xvfsconf32 xvfsp; 3785 3786 bzero(&xvfsp, sizeof(xvfsp)); 3787 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 3788 xvfsp.vfc_typenum = vfsp->vfc_typenum; 3789 xvfsp.vfc_refcount = vfsp->vfc_refcount; 3790 xvfsp.vfc_flags = vfsp->vfc_flags; 3791 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 3792 } 3793 #endif 3794 3795 /* 3796 * Top level filesystem related information gathering. 3797 */ 3798 static int 3799 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 3800 { 3801 struct vfsconf *vfsp; 3802 int error; 3803 3804 error = 0; 3805 vfsconf_slock(); 3806 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 3807 #ifdef COMPAT_FREEBSD32 3808 if (req->flags & SCTL_MASK32) 3809 error = vfsconf2x32(req, vfsp); 3810 else 3811 #endif 3812 error = vfsconf2x(req, vfsp); 3813 if (error) 3814 break; 3815 } 3816 vfsconf_sunlock(); 3817 return (error); 3818 } 3819 3820 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | 3821 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, 3822 "S,xvfsconf", "List of all configured filesystems"); 3823 3824 #ifndef BURN_BRIDGES 3825 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 3826 3827 static int 3828 vfs_sysctl(SYSCTL_HANDLER_ARGS) 3829 { 3830 int *name = (int *)arg1 - 1; /* XXX */ 3831 u_int namelen = arg2 + 1; /* XXX */ 3832 struct vfsconf *vfsp; 3833 3834 log(LOG_WARNING, "userland calling deprecated sysctl, " 3835 "please rebuild world\n"); 3836 3837 #if 1 || defined(COMPAT_PRELITE2) 3838 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 3839 if (namelen == 1) 3840 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 3841 #endif 3842 3843 switch (name[1]) { 3844 case VFS_MAXTYPENUM: 3845 if (namelen != 2) 3846 return (ENOTDIR); 3847 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 3848 case VFS_CONF: 3849 if (namelen != 3) 3850 return (ENOTDIR); /* overloaded */ 3851 vfsconf_slock(); 3852 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 3853 if (vfsp->vfc_typenum == name[2]) 3854 break; 3855 } 3856 vfsconf_sunlock(); 3857 if (vfsp == NULL) 3858 return (EOPNOTSUPP); 3859 #ifdef COMPAT_FREEBSD32 3860 if (req->flags & SCTL_MASK32) 3861 return (vfsconf2x32(req, vfsp)); 3862 else 3863 #endif 3864 return (vfsconf2x(req, vfsp)); 3865 } 3866 return (EOPNOTSUPP); 3867 } 3868 3869 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | 3870 CTLFLAG_MPSAFE, vfs_sysctl, 3871 "Generic filesystem"); 3872 3873 #if 1 || defined(COMPAT_PRELITE2) 3874 3875 static int 3876 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 3877 { 3878 int error; 3879 struct vfsconf *vfsp; 3880 struct ovfsconf ovfs; 3881 3882 vfsconf_slock(); 3883 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 3884 bzero(&ovfs, sizeof(ovfs)); 3885 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 3886 strcpy(ovfs.vfc_name, vfsp->vfc_name); 3887 ovfs.vfc_index = vfsp->vfc_typenum; 3888 ovfs.vfc_refcount = vfsp->vfc_refcount; 3889 ovfs.vfc_flags = vfsp->vfc_flags; 3890 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 3891 if (error != 0) { 3892 vfsconf_sunlock(); 3893 return (error); 3894 } 3895 } 3896 vfsconf_sunlock(); 3897 return (0); 3898 } 3899 3900 #endif /* 1 || COMPAT_PRELITE2 */ 3901 #endif /* !BURN_BRIDGES */ 3902 3903 #define KINFO_VNODESLOP 10 3904 #ifdef notyet 3905 /* 3906 * Dump vnode list (via sysctl). 3907 */ 3908 /* ARGSUSED */ 3909 static int 3910 sysctl_vnode(SYSCTL_HANDLER_ARGS) 3911 { 3912 struct xvnode *xvn; 3913 struct mount *mp; 3914 struct vnode *vp; 3915 int error, len, n; 3916 3917 /* 3918 * Stale numvnodes access is not fatal here. 3919 */ 3920 req->lock = 0; 3921 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 3922 if (!req->oldptr) 3923 /* Make an estimate */ 3924 return (SYSCTL_OUT(req, 0, len)); 3925 3926 error = sysctl_wire_old_buffer(req, 0); 3927 if (error != 0) 3928 return (error); 3929 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 3930 n = 0; 3931 mtx_lock(&mountlist_mtx); 3932 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3933 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) 3934 continue; 3935 MNT_ILOCK(mp); 3936 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3937 if (n == len) 3938 break; 3939 vref(vp); 3940 xvn[n].xv_size = sizeof *xvn; 3941 xvn[n].xv_vnode = vp; 3942 xvn[n].xv_id = 0; /* XXX compat */ 3943 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 3944 XV_COPY(usecount); 3945 XV_COPY(writecount); 3946 XV_COPY(holdcnt); 3947 XV_COPY(mount); 3948 XV_COPY(numoutput); 3949 XV_COPY(type); 3950 #undef XV_COPY 3951 xvn[n].xv_flag = vp->v_vflag; 3952 3953 switch (vp->v_type) { 3954 case VREG: 3955 case VDIR: 3956 case VLNK: 3957 break; 3958 case VBLK: 3959 case VCHR: 3960 if (vp->v_rdev == NULL) { 3961 vrele(vp); 3962 continue; 3963 } 3964 xvn[n].xv_dev = dev2udev(vp->v_rdev); 3965 break; 3966 case VSOCK: 3967 xvn[n].xv_socket = vp->v_socket; 3968 break; 3969 case VFIFO: 3970 xvn[n].xv_fifo = vp->v_fifoinfo; 3971 break; 3972 case VNON: 3973 case VBAD: 3974 default: 3975 /* shouldn't happen? */ 3976 vrele(vp); 3977 continue; 3978 } 3979 vrele(vp); 3980 ++n; 3981 } 3982 MNT_IUNLOCK(mp); 3983 mtx_lock(&mountlist_mtx); 3984 vfs_unbusy(mp); 3985 if (n == len) 3986 break; 3987 } 3988 mtx_unlock(&mountlist_mtx); 3989 3990 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 3991 free(xvn, M_TEMP); 3992 return (error); 3993 } 3994 3995 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD | 3996 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode", 3997 ""); 3998 #endif 3999 4000 static void 4001 unmount_or_warn(struct mount *mp) 4002 { 4003 int error; 4004 4005 error = dounmount(mp, MNT_FORCE, curthread); 4006 if (error != 0) { 4007 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); 4008 if (error == EBUSY) 4009 printf("BUSY)\n"); 4010 else 4011 printf("%d)\n", error); 4012 } 4013 } 4014 4015 /* 4016 * Unmount all filesystems. The list is traversed in reverse order 4017 * of mounting to avoid dependencies. 4018 */ 4019 void 4020 vfs_unmountall(void) 4021 { 4022 struct mount *mp, *tmp; 4023 4024 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 4025 4026 /* 4027 * Since this only runs when rebooting, it is not interlocked. 4028 */ 4029 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { 4030 vfs_ref(mp); 4031 4032 /* 4033 * Forcibly unmounting "/dev" before "/" would prevent clean 4034 * unmount of the latter. 4035 */ 4036 if (mp == rootdevmp) 4037 continue; 4038 4039 unmount_or_warn(mp); 4040 } 4041 4042 if (rootdevmp != NULL) 4043 unmount_or_warn(rootdevmp); 4044 } 4045 4046 /* 4047 * perform msync on all vnodes under a mount point 4048 * the mount point must be locked. 4049 */ 4050 void 4051 vfs_msync(struct mount *mp, int flags) 4052 { 4053 struct vnode *vp, *mvp; 4054 struct vm_object *obj; 4055 4056 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 4057 4058 vnlru_return_batch(mp); 4059 4060 MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) { 4061 obj = vp->v_object; 4062 if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 && 4063 (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) { 4064 if (!vget(vp, 4065 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 4066 curthread)) { 4067 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 4068 vput(vp); 4069 continue; 4070 } 4071 4072 obj = vp->v_object; 4073 if (obj != NULL) { 4074 VM_OBJECT_WLOCK(obj); 4075 vm_object_page_clean(obj, 0, 0, 4076 flags == MNT_WAIT ? 4077 OBJPC_SYNC : OBJPC_NOSYNC); 4078 VM_OBJECT_WUNLOCK(obj); 4079 } 4080 vput(vp); 4081 } 4082 } else 4083 VI_UNLOCK(vp); 4084 } 4085 } 4086 4087 static void 4088 destroy_vpollinfo_free(struct vpollinfo *vi) 4089 { 4090 4091 knlist_destroy(&vi->vpi_selinfo.si_note); 4092 mtx_destroy(&vi->vpi_lock); 4093 uma_zfree(vnodepoll_zone, vi); 4094 } 4095 4096 static void 4097 destroy_vpollinfo(struct vpollinfo *vi) 4098 { 4099 4100 knlist_clear(&vi->vpi_selinfo.si_note, 1); 4101 seldrain(&vi->vpi_selinfo); 4102 destroy_vpollinfo_free(vi); 4103 } 4104 4105 /* 4106 * Initialize per-vnode helper structure to hold poll-related state. 4107 */ 4108 void 4109 v_addpollinfo(struct vnode *vp) 4110 { 4111 struct vpollinfo *vi; 4112 4113 if (vp->v_pollinfo != NULL) 4114 return; 4115 vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO); 4116 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 4117 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 4118 vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked); 4119 VI_LOCK(vp); 4120 if (vp->v_pollinfo != NULL) { 4121 VI_UNLOCK(vp); 4122 destroy_vpollinfo_free(vi); 4123 return; 4124 } 4125 vp->v_pollinfo = vi; 4126 VI_UNLOCK(vp); 4127 } 4128 4129 /* 4130 * Record a process's interest in events which might happen to 4131 * a vnode. Because poll uses the historic select-style interface 4132 * internally, this routine serves as both the ``check for any 4133 * pending events'' and the ``record my interest in future events'' 4134 * functions. (These are done together, while the lock is held, 4135 * to avoid race conditions.) 4136 */ 4137 int 4138 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 4139 { 4140 4141 v_addpollinfo(vp); 4142 mtx_lock(&vp->v_pollinfo->vpi_lock); 4143 if (vp->v_pollinfo->vpi_revents & events) { 4144 /* 4145 * This leaves events we are not interested 4146 * in available for the other process which 4147 * which presumably had requested them 4148 * (otherwise they would never have been 4149 * recorded). 4150 */ 4151 events &= vp->v_pollinfo->vpi_revents; 4152 vp->v_pollinfo->vpi_revents &= ~events; 4153 4154 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4155 return (events); 4156 } 4157 vp->v_pollinfo->vpi_events |= events; 4158 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 4159 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4160 return (0); 4161 } 4162 4163 /* 4164 * Routine to create and manage a filesystem syncer vnode. 4165 */ 4166 #define sync_close ((int (*)(struct vop_close_args *))nullop) 4167 static int sync_fsync(struct vop_fsync_args *); 4168 static int sync_inactive(struct vop_inactive_args *); 4169 static int sync_reclaim(struct vop_reclaim_args *); 4170 4171 static struct vop_vector sync_vnodeops = { 4172 .vop_bypass = VOP_EOPNOTSUPP, 4173 .vop_close = sync_close, /* close */ 4174 .vop_fsync = sync_fsync, /* fsync */ 4175 .vop_inactive = sync_inactive, /* inactive */ 4176 .vop_reclaim = sync_reclaim, /* reclaim */ 4177 .vop_lock1 = vop_stdlock, /* lock */ 4178 .vop_unlock = vop_stdunlock, /* unlock */ 4179 .vop_islocked = vop_stdislocked, /* islocked */ 4180 }; 4181 4182 /* 4183 * Create a new filesystem syncer vnode for the specified mount point. 4184 */ 4185 void 4186 vfs_allocate_syncvnode(struct mount *mp) 4187 { 4188 struct vnode *vp; 4189 struct bufobj *bo; 4190 static long start, incr, next; 4191 int error; 4192 4193 /* Allocate a new vnode */ 4194 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 4195 if (error != 0) 4196 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 4197 vp->v_type = VNON; 4198 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4199 vp->v_vflag |= VV_FORCEINSMQ; 4200 error = insmntque(vp, mp); 4201 if (error != 0) 4202 panic("vfs_allocate_syncvnode: insmntque() failed"); 4203 vp->v_vflag &= ~VV_FORCEINSMQ; 4204 VOP_UNLOCK(vp, 0); 4205 /* 4206 * Place the vnode onto the syncer worklist. We attempt to 4207 * scatter them about on the list so that they will go off 4208 * at evenly distributed times even if all the filesystems 4209 * are mounted at once. 4210 */ 4211 next += incr; 4212 if (next == 0 || next > syncer_maxdelay) { 4213 start /= 2; 4214 incr /= 2; 4215 if (start == 0) { 4216 start = syncer_maxdelay / 2; 4217 incr = syncer_maxdelay; 4218 } 4219 next = start; 4220 } 4221 bo = &vp->v_bufobj; 4222 BO_LOCK(bo); 4223 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 4224 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 4225 mtx_lock(&sync_mtx); 4226 sync_vnode_count++; 4227 if (mp->mnt_syncer == NULL) { 4228 mp->mnt_syncer = vp; 4229 vp = NULL; 4230 } 4231 mtx_unlock(&sync_mtx); 4232 BO_UNLOCK(bo); 4233 if (vp != NULL) { 4234 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4235 vgone(vp); 4236 vput(vp); 4237 } 4238 } 4239 4240 void 4241 vfs_deallocate_syncvnode(struct mount *mp) 4242 { 4243 struct vnode *vp; 4244 4245 mtx_lock(&sync_mtx); 4246 vp = mp->mnt_syncer; 4247 if (vp != NULL) 4248 mp->mnt_syncer = NULL; 4249 mtx_unlock(&sync_mtx); 4250 if (vp != NULL) 4251 vrele(vp); 4252 } 4253 4254 /* 4255 * Do a lazy sync of the filesystem. 4256 */ 4257 static int 4258 sync_fsync(struct vop_fsync_args *ap) 4259 { 4260 struct vnode *syncvp = ap->a_vp; 4261 struct mount *mp = syncvp->v_mount; 4262 int error, save; 4263 struct bufobj *bo; 4264 4265 /* 4266 * We only need to do something if this is a lazy evaluation. 4267 */ 4268 if (ap->a_waitfor != MNT_LAZY) 4269 return (0); 4270 4271 /* 4272 * Move ourselves to the back of the sync list. 4273 */ 4274 bo = &syncvp->v_bufobj; 4275 BO_LOCK(bo); 4276 vn_syncer_add_to_worklist(bo, syncdelay); 4277 BO_UNLOCK(bo); 4278 4279 /* 4280 * Walk the list of vnodes pushing all that are dirty and 4281 * not already on the sync list. 4282 */ 4283 if (vfs_busy(mp, MBF_NOWAIT) != 0) 4284 return (0); 4285 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 4286 vfs_unbusy(mp); 4287 return (0); 4288 } 4289 save = curthread_pflags_set(TDP_SYNCIO); 4290 vfs_msync(mp, MNT_NOWAIT); 4291 error = VFS_SYNC(mp, MNT_LAZY); 4292 curthread_pflags_restore(save); 4293 vn_finished_write(mp); 4294 vfs_unbusy(mp); 4295 return (error); 4296 } 4297 4298 /* 4299 * The syncer vnode is no referenced. 4300 */ 4301 static int 4302 sync_inactive(struct vop_inactive_args *ap) 4303 { 4304 4305 vgone(ap->a_vp); 4306 return (0); 4307 } 4308 4309 /* 4310 * The syncer vnode is no longer needed and is being decommissioned. 4311 * 4312 * Modifications to the worklist must be protected by sync_mtx. 4313 */ 4314 static int 4315 sync_reclaim(struct vop_reclaim_args *ap) 4316 { 4317 struct vnode *vp = ap->a_vp; 4318 struct bufobj *bo; 4319 4320 bo = &vp->v_bufobj; 4321 BO_LOCK(bo); 4322 mtx_lock(&sync_mtx); 4323 if (vp->v_mount->mnt_syncer == vp) 4324 vp->v_mount->mnt_syncer = NULL; 4325 if (bo->bo_flag & BO_ONWORKLST) { 4326 LIST_REMOVE(bo, bo_synclist); 4327 syncer_worklist_len--; 4328 sync_vnode_count--; 4329 bo->bo_flag &= ~BO_ONWORKLST; 4330 } 4331 mtx_unlock(&sync_mtx); 4332 BO_UNLOCK(bo); 4333 4334 return (0); 4335 } 4336 4337 /* 4338 * Check if vnode represents a disk device 4339 */ 4340 int 4341 vn_isdisk(struct vnode *vp, int *errp) 4342 { 4343 int error; 4344 4345 if (vp->v_type != VCHR) { 4346 error = ENOTBLK; 4347 goto out; 4348 } 4349 error = 0; 4350 dev_lock(); 4351 if (vp->v_rdev == NULL) 4352 error = ENXIO; 4353 else if (vp->v_rdev->si_devsw == NULL) 4354 error = ENXIO; 4355 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 4356 error = ENOTBLK; 4357 dev_unlock(); 4358 out: 4359 if (errp != NULL) 4360 *errp = error; 4361 return (error == 0); 4362 } 4363 4364 /* 4365 * Common filesystem object access control check routine. Accepts a 4366 * vnode's type, "mode", uid and gid, requested access mode, credentials, 4367 * and optional call-by-reference privused argument allowing vaccess() 4368 * to indicate to the caller whether privilege was used to satisfy the 4369 * request (obsoleted). Returns 0 on success, or an errno on failure. 4370 */ 4371 int 4372 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 4373 accmode_t accmode, struct ucred *cred, int *privused) 4374 { 4375 accmode_t dac_granted; 4376 accmode_t priv_granted; 4377 4378 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 4379 ("invalid bit in accmode")); 4380 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 4381 ("VAPPEND without VWRITE")); 4382 4383 /* 4384 * Look for a normal, non-privileged way to access the file/directory 4385 * as requested. If it exists, go with that. 4386 */ 4387 4388 if (privused != NULL) 4389 *privused = 0; 4390 4391 dac_granted = 0; 4392 4393 /* Check the owner. */ 4394 if (cred->cr_uid == file_uid) { 4395 dac_granted |= VADMIN; 4396 if (file_mode & S_IXUSR) 4397 dac_granted |= VEXEC; 4398 if (file_mode & S_IRUSR) 4399 dac_granted |= VREAD; 4400 if (file_mode & S_IWUSR) 4401 dac_granted |= (VWRITE | VAPPEND); 4402 4403 if ((accmode & dac_granted) == accmode) 4404 return (0); 4405 4406 goto privcheck; 4407 } 4408 4409 /* Otherwise, check the groups (first match) */ 4410 if (groupmember(file_gid, cred)) { 4411 if (file_mode & S_IXGRP) 4412 dac_granted |= VEXEC; 4413 if (file_mode & S_IRGRP) 4414 dac_granted |= VREAD; 4415 if (file_mode & S_IWGRP) 4416 dac_granted |= (VWRITE | VAPPEND); 4417 4418 if ((accmode & dac_granted) == accmode) 4419 return (0); 4420 4421 goto privcheck; 4422 } 4423 4424 /* Otherwise, check everyone else. */ 4425 if (file_mode & S_IXOTH) 4426 dac_granted |= VEXEC; 4427 if (file_mode & S_IROTH) 4428 dac_granted |= VREAD; 4429 if (file_mode & S_IWOTH) 4430 dac_granted |= (VWRITE | VAPPEND); 4431 if ((accmode & dac_granted) == accmode) 4432 return (0); 4433 4434 privcheck: 4435 /* 4436 * Build a privilege mask to determine if the set of privileges 4437 * satisfies the requirements when combined with the granted mask 4438 * from above. For each privilege, if the privilege is required, 4439 * bitwise or the request type onto the priv_granted mask. 4440 */ 4441 priv_granted = 0; 4442 4443 if (type == VDIR) { 4444 /* 4445 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 4446 * requests, instead of PRIV_VFS_EXEC. 4447 */ 4448 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 4449 !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0)) 4450 priv_granted |= VEXEC; 4451 } else { 4452 /* 4453 * Ensure that at least one execute bit is on. Otherwise, 4454 * a privileged user will always succeed, and we don't want 4455 * this to happen unless the file really is executable. 4456 */ 4457 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 4458 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 4459 !priv_check_cred(cred, PRIV_VFS_EXEC, 0)) 4460 priv_granted |= VEXEC; 4461 } 4462 4463 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 4464 !priv_check_cred(cred, PRIV_VFS_READ, 0)) 4465 priv_granted |= VREAD; 4466 4467 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 4468 !priv_check_cred(cred, PRIV_VFS_WRITE, 0)) 4469 priv_granted |= (VWRITE | VAPPEND); 4470 4471 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 4472 !priv_check_cred(cred, PRIV_VFS_ADMIN, 0)) 4473 priv_granted |= VADMIN; 4474 4475 if ((accmode & (priv_granted | dac_granted)) == accmode) { 4476 /* XXX audit: privilege used */ 4477 if (privused != NULL) 4478 *privused = 1; 4479 return (0); 4480 } 4481 4482 return ((accmode & VADMIN) ? EPERM : EACCES); 4483 } 4484 4485 /* 4486 * Credential check based on process requesting service, and per-attribute 4487 * permissions. 4488 */ 4489 int 4490 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 4491 struct thread *td, accmode_t accmode) 4492 { 4493 4494 /* 4495 * Kernel-invoked always succeeds. 4496 */ 4497 if (cred == NOCRED) 4498 return (0); 4499 4500 /* 4501 * Do not allow privileged processes in jail to directly manipulate 4502 * system attributes. 4503 */ 4504 switch (attrnamespace) { 4505 case EXTATTR_NAMESPACE_SYSTEM: 4506 /* Potentially should be: return (EPERM); */ 4507 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0)); 4508 case EXTATTR_NAMESPACE_USER: 4509 return (VOP_ACCESS(vp, accmode, cred, td)); 4510 default: 4511 return (EPERM); 4512 } 4513 } 4514 4515 #ifdef DEBUG_VFS_LOCKS 4516 /* 4517 * This only exists to suppress warnings from unlocked specfs accesses. It is 4518 * no longer ok to have an unlocked VFS. 4519 */ 4520 #define IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL || \ 4521 (vp)->v_type == VCHR || (vp)->v_type == VBAD) 4522 4523 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 4524 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, 4525 "Drop into debugger on lock violation"); 4526 4527 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 4528 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 4529 0, "Check for interlock across VOPs"); 4530 4531 int vfs_badlock_print = 1; /* Print lock violations. */ 4532 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 4533 0, "Print lock violations"); 4534 4535 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */ 4536 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode, 4537 0, "Print vnode details on lock violations"); 4538 4539 #ifdef KDB 4540 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 4541 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, 4542 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations"); 4543 #endif 4544 4545 static void 4546 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 4547 { 4548 4549 #ifdef KDB 4550 if (vfs_badlock_backtrace) 4551 kdb_backtrace(); 4552 #endif 4553 if (vfs_badlock_vnode) 4554 vn_printf(vp, "vnode "); 4555 if (vfs_badlock_print) 4556 printf("%s: %p %s\n", str, (void *)vp, msg); 4557 if (vfs_badlock_ddb) 4558 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 4559 } 4560 4561 void 4562 assert_vi_locked(struct vnode *vp, const char *str) 4563 { 4564 4565 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 4566 vfs_badlock("interlock is not locked but should be", str, vp); 4567 } 4568 4569 void 4570 assert_vi_unlocked(struct vnode *vp, const char *str) 4571 { 4572 4573 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 4574 vfs_badlock("interlock is locked but should not be", str, vp); 4575 } 4576 4577 void 4578 assert_vop_locked(struct vnode *vp, const char *str) 4579 { 4580 int locked; 4581 4582 if (!IGNORE_LOCK(vp)) { 4583 locked = VOP_ISLOCKED(vp); 4584 if (locked == 0 || locked == LK_EXCLOTHER) 4585 vfs_badlock("is not locked but should be", str, vp); 4586 } 4587 } 4588 4589 void 4590 assert_vop_unlocked(struct vnode *vp, const char *str) 4591 { 4592 4593 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE) 4594 vfs_badlock("is locked but should not be", str, vp); 4595 } 4596 4597 void 4598 assert_vop_elocked(struct vnode *vp, const char *str) 4599 { 4600 4601 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 4602 vfs_badlock("is not exclusive locked but should be", str, vp); 4603 } 4604 #endif /* DEBUG_VFS_LOCKS */ 4605 4606 void 4607 vop_rename_fail(struct vop_rename_args *ap) 4608 { 4609 4610 if (ap->a_tvp != NULL) 4611 vput(ap->a_tvp); 4612 if (ap->a_tdvp == ap->a_tvp) 4613 vrele(ap->a_tdvp); 4614 else 4615 vput(ap->a_tdvp); 4616 vrele(ap->a_fdvp); 4617 vrele(ap->a_fvp); 4618 } 4619 4620 void 4621 vop_rename_pre(void *ap) 4622 { 4623 struct vop_rename_args *a = ap; 4624 4625 #ifdef DEBUG_VFS_LOCKS 4626 if (a->a_tvp) 4627 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 4628 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 4629 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 4630 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 4631 4632 /* Check the source (from). */ 4633 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 4634 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 4635 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 4636 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 4637 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 4638 4639 /* Check the target. */ 4640 if (a->a_tvp) 4641 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 4642 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 4643 #endif 4644 if (a->a_tdvp != a->a_fdvp) 4645 vhold(a->a_fdvp); 4646 if (a->a_tvp != a->a_fvp) 4647 vhold(a->a_fvp); 4648 vhold(a->a_tdvp); 4649 if (a->a_tvp) 4650 vhold(a->a_tvp); 4651 } 4652 4653 #ifdef DEBUG_VFS_LOCKS 4654 void 4655 vop_strategy_pre(void *ap) 4656 { 4657 struct vop_strategy_args *a; 4658 struct buf *bp; 4659 4660 a = ap; 4661 bp = a->a_bp; 4662 4663 /* 4664 * Cluster ops lock their component buffers but not the IO container. 4665 */ 4666 if ((bp->b_flags & B_CLUSTER) != 0) 4667 return; 4668 4669 if (panicstr == NULL && !BUF_ISLOCKED(bp)) { 4670 if (vfs_badlock_print) 4671 printf( 4672 "VOP_STRATEGY: bp is not locked but should be\n"); 4673 if (vfs_badlock_ddb) 4674 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 4675 } 4676 } 4677 4678 void 4679 vop_lock_pre(void *ap) 4680 { 4681 struct vop_lock1_args *a = ap; 4682 4683 if ((a->a_flags & LK_INTERLOCK) == 0) 4684 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 4685 else 4686 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 4687 } 4688 4689 void 4690 vop_lock_post(void *ap, int rc) 4691 { 4692 struct vop_lock1_args *a = ap; 4693 4694 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 4695 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) 4696 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 4697 } 4698 4699 void 4700 vop_unlock_pre(void *ap) 4701 { 4702 struct vop_unlock_args *a = ap; 4703 4704 if (a->a_flags & LK_INTERLOCK) 4705 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 4706 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 4707 } 4708 4709 void 4710 vop_unlock_post(void *ap, int rc) 4711 { 4712 struct vop_unlock_args *a = ap; 4713 4714 if (a->a_flags & LK_INTERLOCK) 4715 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 4716 } 4717 #endif 4718 4719 void 4720 vop_create_post(void *ap, int rc) 4721 { 4722 struct vop_create_args *a = ap; 4723 4724 if (!rc) 4725 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4726 } 4727 4728 void 4729 vop_deleteextattr_post(void *ap, int rc) 4730 { 4731 struct vop_deleteextattr_args *a = ap; 4732 4733 if (!rc) 4734 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 4735 } 4736 4737 void 4738 vop_link_post(void *ap, int rc) 4739 { 4740 struct vop_link_args *a = ap; 4741 4742 if (!rc) { 4743 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK); 4744 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE); 4745 } 4746 } 4747 4748 void 4749 vop_mkdir_post(void *ap, int rc) 4750 { 4751 struct vop_mkdir_args *a = ap; 4752 4753 if (!rc) 4754 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 4755 } 4756 4757 void 4758 vop_mknod_post(void *ap, int rc) 4759 { 4760 struct vop_mknod_args *a = ap; 4761 4762 if (!rc) 4763 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4764 } 4765 4766 void 4767 vop_reclaim_post(void *ap, int rc) 4768 { 4769 struct vop_reclaim_args *a = ap; 4770 4771 if (!rc) 4772 VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE); 4773 } 4774 4775 void 4776 vop_remove_post(void *ap, int rc) 4777 { 4778 struct vop_remove_args *a = ap; 4779 4780 if (!rc) { 4781 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4782 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 4783 } 4784 } 4785 4786 void 4787 vop_rename_post(void *ap, int rc) 4788 { 4789 struct vop_rename_args *a = ap; 4790 long hint; 4791 4792 if (!rc) { 4793 hint = NOTE_WRITE; 4794 if (a->a_fdvp == a->a_tdvp) { 4795 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) 4796 hint |= NOTE_LINK; 4797 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 4798 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 4799 } else { 4800 hint |= NOTE_EXTEND; 4801 if (a->a_fvp->v_type == VDIR) 4802 hint |= NOTE_LINK; 4803 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 4804 4805 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && 4806 a->a_tvp->v_type == VDIR) 4807 hint &= ~NOTE_LINK; 4808 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 4809 } 4810 4811 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 4812 if (a->a_tvp) 4813 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 4814 } 4815 if (a->a_tdvp != a->a_fdvp) 4816 vdrop(a->a_fdvp); 4817 if (a->a_tvp != a->a_fvp) 4818 vdrop(a->a_fvp); 4819 vdrop(a->a_tdvp); 4820 if (a->a_tvp) 4821 vdrop(a->a_tvp); 4822 } 4823 4824 void 4825 vop_rmdir_post(void *ap, int rc) 4826 { 4827 struct vop_rmdir_args *a = ap; 4828 4829 if (!rc) { 4830 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 4831 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 4832 } 4833 } 4834 4835 void 4836 vop_setattr_post(void *ap, int rc) 4837 { 4838 struct vop_setattr_args *a = ap; 4839 4840 if (!rc) 4841 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 4842 } 4843 4844 void 4845 vop_setextattr_post(void *ap, int rc) 4846 { 4847 struct vop_setextattr_args *a = ap; 4848 4849 if (!rc) 4850 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 4851 } 4852 4853 void 4854 vop_symlink_post(void *ap, int rc) 4855 { 4856 struct vop_symlink_args *a = ap; 4857 4858 if (!rc) 4859 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4860 } 4861 4862 void 4863 vop_open_post(void *ap, int rc) 4864 { 4865 struct vop_open_args *a = ap; 4866 4867 if (!rc) 4868 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); 4869 } 4870 4871 void 4872 vop_close_post(void *ap, int rc) 4873 { 4874 struct vop_close_args *a = ap; 4875 4876 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ 4877 (a->a_vp->v_iflag & VI_DOOMED) == 0)) { 4878 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 4879 NOTE_CLOSE_WRITE : NOTE_CLOSE); 4880 } 4881 } 4882 4883 void 4884 vop_read_post(void *ap, int rc) 4885 { 4886 struct vop_read_args *a = ap; 4887 4888 if (!rc) 4889 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 4890 } 4891 4892 void 4893 vop_readdir_post(void *ap, int rc) 4894 { 4895 struct vop_readdir_args *a = ap; 4896 4897 if (!rc) 4898 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 4899 } 4900 4901 static struct knlist fs_knlist; 4902 4903 static void 4904 vfs_event_init(void *arg) 4905 { 4906 knlist_init_mtx(&fs_knlist, NULL); 4907 } 4908 /* XXX - correct order? */ 4909 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 4910 4911 void 4912 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 4913 { 4914 4915 KNOTE_UNLOCKED(&fs_knlist, event); 4916 } 4917 4918 static int filt_fsattach(struct knote *kn); 4919 static void filt_fsdetach(struct knote *kn); 4920 static int filt_fsevent(struct knote *kn, long hint); 4921 4922 struct filterops fs_filtops = { 4923 .f_isfd = 0, 4924 .f_attach = filt_fsattach, 4925 .f_detach = filt_fsdetach, 4926 .f_event = filt_fsevent 4927 }; 4928 4929 static int 4930 filt_fsattach(struct knote *kn) 4931 { 4932 4933 kn->kn_flags |= EV_CLEAR; 4934 knlist_add(&fs_knlist, kn, 0); 4935 return (0); 4936 } 4937 4938 static void 4939 filt_fsdetach(struct knote *kn) 4940 { 4941 4942 knlist_remove(&fs_knlist, kn, 0); 4943 } 4944 4945 static int 4946 filt_fsevent(struct knote *kn, long hint) 4947 { 4948 4949 kn->kn_fflags |= hint; 4950 return (kn->kn_fflags != 0); 4951 } 4952 4953 static int 4954 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 4955 { 4956 struct vfsidctl vc; 4957 int error; 4958 struct mount *mp; 4959 4960 error = SYSCTL_IN(req, &vc, sizeof(vc)); 4961 if (error) 4962 return (error); 4963 if (vc.vc_vers != VFS_CTL_VERS1) 4964 return (EINVAL); 4965 mp = vfs_getvfs(&vc.vc_fsid); 4966 if (mp == NULL) 4967 return (ENOENT); 4968 /* ensure that a specific sysctl goes to the right filesystem. */ 4969 if (strcmp(vc.vc_fstypename, "*") != 0 && 4970 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 4971 vfs_rel(mp); 4972 return (EINVAL); 4973 } 4974 VCTLTOREQ(&vc, req); 4975 error = VFS_SYSCTL(mp, vc.vc_op, req); 4976 vfs_rel(mp); 4977 return (error); 4978 } 4979 4980 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR, 4981 NULL, 0, sysctl_vfs_ctl, "", 4982 "Sysctl by fsid"); 4983 4984 /* 4985 * Function to initialize a va_filerev field sensibly. 4986 * XXX: Wouldn't a random number make a lot more sense ?? 4987 */ 4988 u_quad_t 4989 init_va_filerev(void) 4990 { 4991 struct bintime bt; 4992 4993 getbinuptime(&bt); 4994 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 4995 } 4996 4997 static int filt_vfsread(struct knote *kn, long hint); 4998 static int filt_vfswrite(struct knote *kn, long hint); 4999 static int filt_vfsvnode(struct knote *kn, long hint); 5000 static void filt_vfsdetach(struct knote *kn); 5001 static struct filterops vfsread_filtops = { 5002 .f_isfd = 1, 5003 .f_detach = filt_vfsdetach, 5004 .f_event = filt_vfsread 5005 }; 5006 static struct filterops vfswrite_filtops = { 5007 .f_isfd = 1, 5008 .f_detach = filt_vfsdetach, 5009 .f_event = filt_vfswrite 5010 }; 5011 static struct filterops vfsvnode_filtops = { 5012 .f_isfd = 1, 5013 .f_detach = filt_vfsdetach, 5014 .f_event = filt_vfsvnode 5015 }; 5016 5017 static void 5018 vfs_knllock(void *arg) 5019 { 5020 struct vnode *vp = arg; 5021 5022 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5023 } 5024 5025 static void 5026 vfs_knlunlock(void *arg) 5027 { 5028 struct vnode *vp = arg; 5029 5030 VOP_UNLOCK(vp, 0); 5031 } 5032 5033 static void 5034 vfs_knl_assert_locked(void *arg) 5035 { 5036 #ifdef DEBUG_VFS_LOCKS 5037 struct vnode *vp = arg; 5038 5039 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 5040 #endif 5041 } 5042 5043 static void 5044 vfs_knl_assert_unlocked(void *arg) 5045 { 5046 #ifdef DEBUG_VFS_LOCKS 5047 struct vnode *vp = arg; 5048 5049 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 5050 #endif 5051 } 5052 5053 int 5054 vfs_kqfilter(struct vop_kqfilter_args *ap) 5055 { 5056 struct vnode *vp = ap->a_vp; 5057 struct knote *kn = ap->a_kn; 5058 struct knlist *knl; 5059 5060 switch (kn->kn_filter) { 5061 case EVFILT_READ: 5062 kn->kn_fop = &vfsread_filtops; 5063 break; 5064 case EVFILT_WRITE: 5065 kn->kn_fop = &vfswrite_filtops; 5066 break; 5067 case EVFILT_VNODE: 5068 kn->kn_fop = &vfsvnode_filtops; 5069 break; 5070 default: 5071 return (EINVAL); 5072 } 5073 5074 kn->kn_hook = (caddr_t)vp; 5075 5076 v_addpollinfo(vp); 5077 if (vp->v_pollinfo == NULL) 5078 return (ENOMEM); 5079 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 5080 vhold(vp); 5081 knlist_add(knl, kn, 0); 5082 5083 return (0); 5084 } 5085 5086 /* 5087 * Detach knote from vnode 5088 */ 5089 static void 5090 filt_vfsdetach(struct knote *kn) 5091 { 5092 struct vnode *vp = (struct vnode *)kn->kn_hook; 5093 5094 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 5095 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 5096 vdrop(vp); 5097 } 5098 5099 /*ARGSUSED*/ 5100 static int 5101 filt_vfsread(struct knote *kn, long hint) 5102 { 5103 struct vnode *vp = (struct vnode *)kn->kn_hook; 5104 struct vattr va; 5105 int res; 5106 5107 /* 5108 * filesystem is gone, so set the EOF flag and schedule 5109 * the knote for deletion. 5110 */ 5111 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 5112 VI_LOCK(vp); 5113 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 5114 VI_UNLOCK(vp); 5115 return (1); 5116 } 5117 5118 if (VOP_GETATTR(vp, &va, curthread->td_ucred)) 5119 return (0); 5120 5121 VI_LOCK(vp); 5122 kn->kn_data = va.va_size - kn->kn_fp->f_offset; 5123 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; 5124 VI_UNLOCK(vp); 5125 return (res); 5126 } 5127 5128 /*ARGSUSED*/ 5129 static int 5130 filt_vfswrite(struct knote *kn, long hint) 5131 { 5132 struct vnode *vp = (struct vnode *)kn->kn_hook; 5133 5134 VI_LOCK(vp); 5135 5136 /* 5137 * filesystem is gone, so set the EOF flag and schedule 5138 * the knote for deletion. 5139 */ 5140 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) 5141 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 5142 5143 kn->kn_data = 0; 5144 VI_UNLOCK(vp); 5145 return (1); 5146 } 5147 5148 static int 5149 filt_vfsvnode(struct knote *kn, long hint) 5150 { 5151 struct vnode *vp = (struct vnode *)kn->kn_hook; 5152 int res; 5153 5154 VI_LOCK(vp); 5155 if (kn->kn_sfflags & hint) 5156 kn->kn_fflags |= hint; 5157 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 5158 kn->kn_flags |= EV_EOF; 5159 VI_UNLOCK(vp); 5160 return (1); 5161 } 5162 res = (kn->kn_fflags != 0); 5163 VI_UNLOCK(vp); 5164 return (res); 5165 } 5166 5167 int 5168 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 5169 { 5170 int error; 5171 5172 if (dp->d_reclen > ap->a_uio->uio_resid) 5173 return (ENAMETOOLONG); 5174 error = uiomove(dp, dp->d_reclen, ap->a_uio); 5175 if (error) { 5176 if (ap->a_ncookies != NULL) { 5177 if (ap->a_cookies != NULL) 5178 free(ap->a_cookies, M_TEMP); 5179 ap->a_cookies = NULL; 5180 *ap->a_ncookies = 0; 5181 } 5182 return (error); 5183 } 5184 if (ap->a_ncookies == NULL) 5185 return (0); 5186 5187 KASSERT(ap->a_cookies, 5188 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 5189 5190 *ap->a_cookies = realloc(*ap->a_cookies, 5191 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); 5192 (*ap->a_cookies)[*ap->a_ncookies] = off; 5193 *ap->a_ncookies += 1; 5194 return (0); 5195 } 5196 5197 /* 5198 * Mark for update the access time of the file if the filesystem 5199 * supports VOP_MARKATIME. This functionality is used by execve and 5200 * mmap, so we want to avoid the I/O implied by directly setting 5201 * va_atime for the sake of efficiency. 5202 */ 5203 void 5204 vfs_mark_atime(struct vnode *vp, struct ucred *cred) 5205 { 5206 struct mount *mp; 5207 5208 mp = vp->v_mount; 5209 ASSERT_VOP_LOCKED(vp, "vfs_mark_atime"); 5210 if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) 5211 (void)VOP_MARKATIME(vp); 5212 } 5213 5214 /* 5215 * The purpose of this routine is to remove granularity from accmode_t, 5216 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 5217 * VADMIN and VAPPEND. 5218 * 5219 * If it returns 0, the caller is supposed to continue with the usual 5220 * access checks using 'accmode' as modified by this routine. If it 5221 * returns nonzero value, the caller is supposed to return that value 5222 * as errno. 5223 * 5224 * Note that after this routine runs, accmode may be zero. 5225 */ 5226 int 5227 vfs_unixify_accmode(accmode_t *accmode) 5228 { 5229 /* 5230 * There is no way to specify explicit "deny" rule using 5231 * file mode or POSIX.1e ACLs. 5232 */ 5233 if (*accmode & VEXPLICIT_DENY) { 5234 *accmode = 0; 5235 return (0); 5236 } 5237 5238 /* 5239 * None of these can be translated into usual access bits. 5240 * Also, the common case for NFSv4 ACLs is to not contain 5241 * either of these bits. Caller should check for VWRITE 5242 * on the containing directory instead. 5243 */ 5244 if (*accmode & (VDELETE_CHILD | VDELETE)) 5245 return (EPERM); 5246 5247 if (*accmode & VADMIN_PERMS) { 5248 *accmode &= ~VADMIN_PERMS; 5249 *accmode |= VADMIN; 5250 } 5251 5252 /* 5253 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 5254 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 5255 */ 5256 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 5257 5258 return (0); 5259 } 5260 5261 /* 5262 * These are helper functions for filesystems to traverse all 5263 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 5264 * 5265 * This interface replaces MNT_VNODE_FOREACH. 5266 */ 5267 5268 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker"); 5269 5270 struct vnode * 5271 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 5272 { 5273 struct vnode *vp; 5274 5275 if (should_yield()) 5276 kern_yield(PRI_USER); 5277 MNT_ILOCK(mp); 5278 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5279 vp = TAILQ_NEXT(*mvp, v_nmntvnodes); 5280 while (vp != NULL && (vp->v_type == VMARKER || 5281 (vp->v_iflag & VI_DOOMED) != 0)) 5282 vp = TAILQ_NEXT(vp, v_nmntvnodes); 5283 5284 /* Check if we are done */ 5285 if (vp == NULL) { 5286 __mnt_vnode_markerfree_all(mvp, mp); 5287 /* MNT_IUNLOCK(mp); -- done in above function */ 5288 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 5289 return (NULL); 5290 } 5291 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 5292 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 5293 VI_LOCK(vp); 5294 MNT_IUNLOCK(mp); 5295 return (vp); 5296 } 5297 5298 struct vnode * 5299 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 5300 { 5301 struct vnode *vp; 5302 5303 *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 5304 MNT_ILOCK(mp); 5305 MNT_REF(mp); 5306 (*mvp)->v_type = VMARKER; 5307 5308 vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 5309 while (vp != NULL && (vp->v_type == VMARKER || 5310 (vp->v_iflag & VI_DOOMED) != 0)) 5311 vp = TAILQ_NEXT(vp, v_nmntvnodes); 5312 5313 /* Check if we are done */ 5314 if (vp == NULL) { 5315 MNT_REL(mp); 5316 MNT_IUNLOCK(mp); 5317 free(*mvp, M_VNODE_MARKER); 5318 *mvp = NULL; 5319 return (NULL); 5320 } 5321 (*mvp)->v_mount = mp; 5322 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 5323 VI_LOCK(vp); 5324 MNT_IUNLOCK(mp); 5325 return (vp); 5326 } 5327 5328 5329 void 5330 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 5331 { 5332 5333 if (*mvp == NULL) { 5334 MNT_IUNLOCK(mp); 5335 return; 5336 } 5337 5338 mtx_assert(MNT_MTX(mp), MA_OWNED); 5339 5340 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5341 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 5342 MNT_REL(mp); 5343 MNT_IUNLOCK(mp); 5344 free(*mvp, M_VNODE_MARKER); 5345 *mvp = NULL; 5346 } 5347 5348 /* 5349 * These are helper functions for filesystems to traverse their 5350 * active vnodes. See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h 5351 */ 5352 static void 5353 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp) 5354 { 5355 5356 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5357 5358 MNT_ILOCK(mp); 5359 MNT_REL(mp); 5360 MNT_IUNLOCK(mp); 5361 free(*mvp, M_VNODE_MARKER); 5362 *mvp = NULL; 5363 } 5364 5365 /* 5366 * Relock the mp mount vnode list lock with the vp vnode interlock in the 5367 * conventional lock order during mnt_vnode_next_active iteration. 5368 * 5369 * On entry, the mount vnode list lock is held and the vnode interlock is not. 5370 * The list lock is dropped and reacquired. On success, both locks are held. 5371 * On failure, the mount vnode list lock is held but the vnode interlock is 5372 * not, and the procedure may have yielded. 5373 */ 5374 static bool 5375 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp, 5376 struct vnode *vp) 5377 { 5378 const struct vnode *tmp; 5379 bool held, ret; 5380 5381 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER && 5382 TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp, 5383 ("%s: bad marker", __func__)); 5384 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp, 5385 ("%s: inappropriate vnode", __func__)); 5386 ASSERT_VI_UNLOCKED(vp, __func__); 5387 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 5388 5389 ret = false; 5390 5391 TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist); 5392 TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist); 5393 5394 /* 5395 * Use a hold to prevent vp from disappearing while the mount vnode 5396 * list lock is dropped and reacquired. Normally a hold would be 5397 * acquired with vhold(), but that might try to acquire the vnode 5398 * interlock, which would be a LOR with the mount vnode list lock. 5399 */ 5400 held = vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt); 5401 mtx_unlock(&mp->mnt_listmtx); 5402 if (!held) 5403 goto abort; 5404 VI_LOCK(vp); 5405 if (!vfs_refcount_release_if_not_last(&vp->v_holdcnt)) { 5406 vdropl(vp); 5407 goto abort; 5408 } 5409 mtx_lock(&mp->mnt_listmtx); 5410 5411 /* 5412 * Determine whether the vnode is still the next one after the marker, 5413 * excepting any other markers. If the vnode has not been doomed by 5414 * vgone() then the hold should have ensured that it remained on the 5415 * active list. If it has been doomed but is still on the active list, 5416 * don't abort, but rather skip over it (avoid spinning on doomed 5417 * vnodes). 5418 */ 5419 tmp = mvp; 5420 do { 5421 tmp = TAILQ_NEXT(tmp, v_actfreelist); 5422 } while (tmp != NULL && tmp->v_type == VMARKER); 5423 if (tmp != vp) { 5424 mtx_unlock(&mp->mnt_listmtx); 5425 VI_UNLOCK(vp); 5426 goto abort; 5427 } 5428 5429 ret = true; 5430 goto out; 5431 abort: 5432 maybe_yield(); 5433 mtx_lock(&mp->mnt_listmtx); 5434 out: 5435 if (ret) 5436 ASSERT_VI_LOCKED(vp, __func__); 5437 else 5438 ASSERT_VI_UNLOCKED(vp, __func__); 5439 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 5440 return (ret); 5441 } 5442 5443 static struct vnode * 5444 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp) 5445 { 5446 struct vnode *vp, *nvp; 5447 5448 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 5449 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5450 restart: 5451 vp = TAILQ_NEXT(*mvp, v_actfreelist); 5452 while (vp != NULL) { 5453 if (vp->v_type == VMARKER) { 5454 vp = TAILQ_NEXT(vp, v_actfreelist); 5455 continue; 5456 } 5457 /* 5458 * Try-lock because this is the wrong lock order. If that does 5459 * not succeed, drop the mount vnode list lock and try to 5460 * reacquire it and the vnode interlock in the right order. 5461 */ 5462 if (!VI_TRYLOCK(vp) && 5463 !mnt_vnode_next_active_relock(*mvp, mp, vp)) 5464 goto restart; 5465 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 5466 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 5467 ("alien vnode on the active list %p %p", vp, mp)); 5468 if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0) 5469 break; 5470 nvp = TAILQ_NEXT(vp, v_actfreelist); 5471 VI_UNLOCK(vp); 5472 vp = nvp; 5473 } 5474 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist); 5475 5476 /* Check if we are done */ 5477 if (vp == NULL) { 5478 mtx_unlock(&mp->mnt_listmtx); 5479 mnt_vnode_markerfree_active(mvp, mp); 5480 return (NULL); 5481 } 5482 TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist); 5483 mtx_unlock(&mp->mnt_listmtx); 5484 ASSERT_VI_LOCKED(vp, "active iter"); 5485 KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp)); 5486 return (vp); 5487 } 5488 5489 struct vnode * 5490 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp) 5491 { 5492 5493 if (should_yield()) 5494 kern_yield(PRI_USER); 5495 mtx_lock(&mp->mnt_listmtx); 5496 return (mnt_vnode_next_active(mvp, mp)); 5497 } 5498 5499 struct vnode * 5500 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp) 5501 { 5502 struct vnode *vp; 5503 5504 *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 5505 MNT_ILOCK(mp); 5506 MNT_REF(mp); 5507 MNT_IUNLOCK(mp); 5508 (*mvp)->v_type = VMARKER; 5509 (*mvp)->v_mount = mp; 5510 5511 mtx_lock(&mp->mnt_listmtx); 5512 vp = TAILQ_FIRST(&mp->mnt_activevnodelist); 5513 if (vp == NULL) { 5514 mtx_unlock(&mp->mnt_listmtx); 5515 mnt_vnode_markerfree_active(mvp, mp); 5516 return (NULL); 5517 } 5518 TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist); 5519 return (mnt_vnode_next_active(mvp, mp)); 5520 } 5521 5522 void 5523 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp) 5524 { 5525 5526 if (*mvp == NULL) 5527 return; 5528 5529 mtx_lock(&mp->mnt_listmtx); 5530 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist); 5531 mtx_unlock(&mp->mnt_listmtx); 5532 mnt_vnode_markerfree_active(mvp, mp); 5533 } 5534