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 ms, " 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 MAC 1493 mac_vnode_init(vp); 1494 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 1495 mac_vnode_associate_singlelabel(mp, vp); 1496 else if (mp == NULL && vops != &dead_vnodeops) 1497 printf("NULL mp in getnewvnode()\n"); 1498 #endif 1499 if (mp != NULL) { 1500 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize; 1501 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) 1502 vp->v_vflag |= VV_NOKNOTE; 1503 } 1504 1505 /* 1506 * For the filesystems which do not use vfs_hash_insert(), 1507 * still initialize v_hash to have vfs_hash_index() useful. 1508 * E.g., nullfs uses vfs_hash_index() on the lower vnode for 1509 * its own hashing. 1510 */ 1511 vp->v_hash = (uintptr_t)vp >> vnsz2log; 1512 1513 *vpp = vp; 1514 return (0); 1515 } 1516 1517 /* 1518 * Delete from old mount point vnode list, if on one. 1519 */ 1520 static void 1521 delmntque(struct vnode *vp) 1522 { 1523 struct mount *mp; 1524 int active; 1525 1526 mp = vp->v_mount; 1527 if (mp == NULL) 1528 return; 1529 MNT_ILOCK(mp); 1530 VI_LOCK(vp); 1531 KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize, 1532 ("Active vnode list size %d > Vnode list size %d", 1533 mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize)); 1534 active = vp->v_iflag & VI_ACTIVE; 1535 vp->v_iflag &= ~VI_ACTIVE; 1536 if (active) { 1537 mtx_lock(&mp->mnt_listmtx); 1538 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist); 1539 mp->mnt_activevnodelistsize--; 1540 mtx_unlock(&mp->mnt_listmtx); 1541 } 1542 vp->v_mount = NULL; 1543 VI_UNLOCK(vp); 1544 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 1545 ("bad mount point vnode list size")); 1546 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1547 mp->mnt_nvnodelistsize--; 1548 MNT_REL(mp); 1549 MNT_IUNLOCK(mp); 1550 } 1551 1552 static void 1553 insmntque_stddtr(struct vnode *vp, void *dtr_arg) 1554 { 1555 1556 vp->v_data = NULL; 1557 vp->v_op = &dead_vnodeops; 1558 vgone(vp); 1559 vput(vp); 1560 } 1561 1562 /* 1563 * Insert into list of vnodes for the new mount point, if available. 1564 */ 1565 int 1566 insmntque1(struct vnode *vp, struct mount *mp, 1567 void (*dtr)(struct vnode *, void *), void *dtr_arg) 1568 { 1569 1570 KASSERT(vp->v_mount == NULL, 1571 ("insmntque: vnode already on per mount vnode list")); 1572 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 1573 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); 1574 1575 /* 1576 * We acquire the vnode interlock early to ensure that the 1577 * vnode cannot be recycled by another process releasing a 1578 * holdcnt on it before we get it on both the vnode list 1579 * and the active vnode list. The mount mutex protects only 1580 * manipulation of the vnode list and the vnode freelist 1581 * mutex protects only manipulation of the active vnode list. 1582 * Hence the need to hold the vnode interlock throughout. 1583 */ 1584 MNT_ILOCK(mp); 1585 VI_LOCK(vp); 1586 if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 && 1587 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || 1588 mp->mnt_nvnodelistsize == 0)) && 1589 (vp->v_vflag & VV_FORCEINSMQ) == 0) { 1590 VI_UNLOCK(vp); 1591 MNT_IUNLOCK(mp); 1592 if (dtr != NULL) 1593 dtr(vp, dtr_arg); 1594 return (EBUSY); 1595 } 1596 vp->v_mount = mp; 1597 MNT_REF(mp); 1598 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1599 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 1600 ("neg mount point vnode list size")); 1601 mp->mnt_nvnodelistsize++; 1602 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 1603 ("Activating already active vnode")); 1604 vp->v_iflag |= VI_ACTIVE; 1605 mtx_lock(&mp->mnt_listmtx); 1606 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist); 1607 mp->mnt_activevnodelistsize++; 1608 mtx_unlock(&mp->mnt_listmtx); 1609 VI_UNLOCK(vp); 1610 MNT_IUNLOCK(mp); 1611 return (0); 1612 } 1613 1614 int 1615 insmntque(struct vnode *vp, struct mount *mp) 1616 { 1617 1618 return (insmntque1(vp, mp, insmntque_stddtr, NULL)); 1619 } 1620 1621 /* 1622 * Flush out and invalidate all buffers associated with a bufobj 1623 * Called with the underlying object locked. 1624 */ 1625 int 1626 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) 1627 { 1628 int error; 1629 1630 BO_LOCK(bo); 1631 if (flags & V_SAVE) { 1632 error = bufobj_wwait(bo, slpflag, slptimeo); 1633 if (error) { 1634 BO_UNLOCK(bo); 1635 return (error); 1636 } 1637 if (bo->bo_dirty.bv_cnt > 0) { 1638 BO_UNLOCK(bo); 1639 if ((error = BO_SYNC(bo, MNT_WAIT)) != 0) 1640 return (error); 1641 /* 1642 * XXX We could save a lock/unlock if this was only 1643 * enabled under INVARIANTS 1644 */ 1645 BO_LOCK(bo); 1646 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 1647 panic("vinvalbuf: dirty bufs"); 1648 } 1649 } 1650 /* 1651 * If you alter this loop please notice that interlock is dropped and 1652 * reacquired in flushbuflist. Special care is needed to ensure that 1653 * no race conditions occur from this. 1654 */ 1655 do { 1656 error = flushbuflist(&bo->bo_clean, 1657 flags, bo, slpflag, slptimeo); 1658 if (error == 0 && !(flags & V_CLEANONLY)) 1659 error = flushbuflist(&bo->bo_dirty, 1660 flags, bo, slpflag, slptimeo); 1661 if (error != 0 && error != EAGAIN) { 1662 BO_UNLOCK(bo); 1663 return (error); 1664 } 1665 } while (error != 0); 1666 1667 /* 1668 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1669 * have write I/O in-progress but if there is a VM object then the 1670 * VM object can also have read-I/O in-progress. 1671 */ 1672 do { 1673 bufobj_wwait(bo, 0, 0); 1674 BO_UNLOCK(bo); 1675 if (bo->bo_object != NULL) { 1676 VM_OBJECT_WLOCK(bo->bo_object); 1677 vm_object_pip_wait(bo->bo_object, "bovlbx"); 1678 VM_OBJECT_WUNLOCK(bo->bo_object); 1679 } 1680 BO_LOCK(bo); 1681 } while (bo->bo_numoutput > 0); 1682 BO_UNLOCK(bo); 1683 1684 /* 1685 * Destroy the copy in the VM cache, too. 1686 */ 1687 if (bo->bo_object != NULL && 1688 (flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0) { 1689 VM_OBJECT_WLOCK(bo->bo_object); 1690 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? 1691 OBJPR_CLEANONLY : 0); 1692 VM_OBJECT_WUNLOCK(bo->bo_object); 1693 } 1694 1695 #ifdef INVARIANTS 1696 BO_LOCK(bo); 1697 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0 && 1698 (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0)) 1699 panic("vinvalbuf: flush failed"); 1700 BO_UNLOCK(bo); 1701 #endif 1702 return (0); 1703 } 1704 1705 /* 1706 * Flush out and invalidate all buffers associated with a vnode. 1707 * Called with the underlying object locked. 1708 */ 1709 int 1710 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 1711 { 1712 1713 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 1714 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1715 if (vp->v_object != NULL && vp->v_object->handle != vp) 1716 return (0); 1717 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); 1718 } 1719 1720 /* 1721 * Flush out buffers on the specified list. 1722 * 1723 */ 1724 static int 1725 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 1726 int slptimeo) 1727 { 1728 struct buf *bp, *nbp; 1729 int retval, error; 1730 daddr_t lblkno; 1731 b_xflags_t xflags; 1732 1733 ASSERT_BO_WLOCKED(bo); 1734 1735 retval = 0; 1736 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 1737 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1738 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1739 continue; 1740 } 1741 lblkno = 0; 1742 xflags = 0; 1743 if (nbp != NULL) { 1744 lblkno = nbp->b_lblkno; 1745 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN); 1746 } 1747 retval = EAGAIN; 1748 error = BUF_TIMELOCK(bp, 1749 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), 1750 "flushbuf", slpflag, slptimeo); 1751 if (error) { 1752 BO_LOCK(bo); 1753 return (error != ENOLCK ? error : EAGAIN); 1754 } 1755 KASSERT(bp->b_bufobj == bo, 1756 ("bp %p wrong b_bufobj %p should be %p", 1757 bp, bp->b_bufobj, bo)); 1758 /* 1759 * XXX Since there are no node locks for NFS, I 1760 * believe there is a slight chance that a delayed 1761 * write will occur while sleeping just above, so 1762 * check for it. 1763 */ 1764 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1765 (flags & V_SAVE)) { 1766 bremfree(bp); 1767 bp->b_flags |= B_ASYNC; 1768 bwrite(bp); 1769 BO_LOCK(bo); 1770 return (EAGAIN); /* XXX: why not loop ? */ 1771 } 1772 bremfree(bp); 1773 bp->b_flags |= (B_INVAL | B_RELBUF); 1774 bp->b_flags &= ~B_ASYNC; 1775 brelse(bp); 1776 BO_LOCK(bo); 1777 nbp = gbincore(bo, lblkno); 1778 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1779 != xflags) 1780 break; /* nbp invalid */ 1781 } 1782 return (retval); 1783 } 1784 1785 int 1786 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn) 1787 { 1788 struct buf *bp; 1789 int error; 1790 daddr_t lblkno; 1791 1792 ASSERT_BO_LOCKED(bo); 1793 1794 for (lblkno = startn;;) { 1795 again: 1796 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno); 1797 if (bp == NULL || bp->b_lblkno >= endn || 1798 bp->b_lblkno < startn) 1799 break; 1800 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 1801 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0); 1802 if (error != 0) { 1803 BO_RLOCK(bo); 1804 if (error == ENOLCK) 1805 goto again; 1806 return (error); 1807 } 1808 KASSERT(bp->b_bufobj == bo, 1809 ("bp %p wrong b_bufobj %p should be %p", 1810 bp, bp->b_bufobj, bo)); 1811 lblkno = bp->b_lblkno + 1; 1812 if ((bp->b_flags & B_MANAGED) == 0) 1813 bremfree(bp); 1814 bp->b_flags |= B_RELBUF; 1815 /* 1816 * In the VMIO case, use the B_NOREUSE flag to hint that the 1817 * pages backing each buffer in the range are unlikely to be 1818 * reused. Dirty buffers will have the hint applied once 1819 * they've been written. 1820 */ 1821 if (bp->b_vp->v_object != NULL) 1822 bp->b_flags |= B_NOREUSE; 1823 brelse(bp); 1824 BO_RLOCK(bo); 1825 } 1826 return (0); 1827 } 1828 1829 /* 1830 * Truncate a file's buffer and pages to a specified length. This 1831 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1832 * sync activity. 1833 */ 1834 int 1835 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize) 1836 { 1837 struct buf *bp, *nbp; 1838 int anyfreed; 1839 int trunclbn; 1840 struct bufobj *bo; 1841 1842 CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__, 1843 vp, cred, blksize, (uintmax_t)length); 1844 1845 /* 1846 * Round up to the *next* lbn. 1847 */ 1848 trunclbn = howmany(length, blksize); 1849 1850 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1851 restart: 1852 bo = &vp->v_bufobj; 1853 BO_LOCK(bo); 1854 anyfreed = 1; 1855 for (;anyfreed;) { 1856 anyfreed = 0; 1857 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 1858 if (bp->b_lblkno < trunclbn) 1859 continue; 1860 if (BUF_LOCK(bp, 1861 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1862 BO_LOCKPTR(bo)) == ENOLCK) 1863 goto restart; 1864 1865 bremfree(bp); 1866 bp->b_flags |= (B_INVAL | B_RELBUF); 1867 bp->b_flags &= ~B_ASYNC; 1868 brelse(bp); 1869 anyfreed = 1; 1870 1871 BO_LOCK(bo); 1872 if (nbp != NULL && 1873 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1874 (nbp->b_vp != vp) || 1875 (nbp->b_flags & B_DELWRI))) { 1876 BO_UNLOCK(bo); 1877 goto restart; 1878 } 1879 } 1880 1881 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1882 if (bp->b_lblkno < trunclbn) 1883 continue; 1884 if (BUF_LOCK(bp, 1885 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1886 BO_LOCKPTR(bo)) == ENOLCK) 1887 goto restart; 1888 bremfree(bp); 1889 bp->b_flags |= (B_INVAL | B_RELBUF); 1890 bp->b_flags &= ~B_ASYNC; 1891 brelse(bp); 1892 anyfreed = 1; 1893 1894 BO_LOCK(bo); 1895 if (nbp != NULL && 1896 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1897 (nbp->b_vp != vp) || 1898 (nbp->b_flags & B_DELWRI) == 0)) { 1899 BO_UNLOCK(bo); 1900 goto restart; 1901 } 1902 } 1903 } 1904 1905 if (length > 0) { 1906 restartsync: 1907 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1908 if (bp->b_lblkno > 0) 1909 continue; 1910 /* 1911 * Since we hold the vnode lock this should only 1912 * fail if we're racing with the buf daemon. 1913 */ 1914 if (BUF_LOCK(bp, 1915 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1916 BO_LOCKPTR(bo)) == ENOLCK) { 1917 goto restart; 1918 } 1919 VNASSERT((bp->b_flags & B_DELWRI), vp, 1920 ("buf(%p) on dirty queue without DELWRI", bp)); 1921 1922 bremfree(bp); 1923 bawrite(bp); 1924 BO_LOCK(bo); 1925 goto restartsync; 1926 } 1927 } 1928 1929 bufobj_wwait(bo, 0, 0); 1930 BO_UNLOCK(bo); 1931 vnode_pager_setsize(vp, length); 1932 1933 return (0); 1934 } 1935 1936 static void 1937 buf_vlist_remove(struct buf *bp) 1938 { 1939 struct bufv *bv; 1940 1941 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1942 ASSERT_BO_WLOCKED(bp->b_bufobj); 1943 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) != 1944 (BX_VNDIRTY|BX_VNCLEAN), 1945 ("buf_vlist_remove: Buf %p is on two lists", bp)); 1946 if (bp->b_xflags & BX_VNDIRTY) 1947 bv = &bp->b_bufobj->bo_dirty; 1948 else 1949 bv = &bp->b_bufobj->bo_clean; 1950 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno); 1951 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 1952 bv->bv_cnt--; 1953 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1954 } 1955 1956 /* 1957 * Add the buffer to the sorted clean or dirty block list. 1958 * 1959 * NOTE: xflags is passed as a constant, optimizing this inline function! 1960 */ 1961 static void 1962 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 1963 { 1964 struct bufv *bv; 1965 struct buf *n; 1966 int error; 1967 1968 ASSERT_BO_WLOCKED(bo); 1969 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0, 1970 ("dead bo %p", bo)); 1971 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1972 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 1973 bp->b_xflags |= xflags; 1974 if (xflags & BX_VNDIRTY) 1975 bv = &bo->bo_dirty; 1976 else 1977 bv = &bo->bo_clean; 1978 1979 /* 1980 * Keep the list ordered. Optimize empty list insertion. Assume 1981 * we tend to grow at the tail so lookup_le should usually be cheaper 1982 * than _ge. 1983 */ 1984 if (bv->bv_cnt == 0 || 1985 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno) 1986 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 1987 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL) 1988 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs); 1989 else 1990 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs); 1991 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp); 1992 if (error) 1993 panic("buf_vlist_add: Preallocated nodes insufficient."); 1994 bv->bv_cnt++; 1995 } 1996 1997 /* 1998 * Look up a buffer using the buffer tries. 1999 */ 2000 struct buf * 2001 gbincore(struct bufobj *bo, daddr_t lblkno) 2002 { 2003 struct buf *bp; 2004 2005 ASSERT_BO_LOCKED(bo); 2006 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno); 2007 if (bp != NULL) 2008 return (bp); 2009 return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno); 2010 } 2011 2012 /* 2013 * Associate a buffer with a vnode. 2014 */ 2015 void 2016 bgetvp(struct vnode *vp, struct buf *bp) 2017 { 2018 struct bufobj *bo; 2019 2020 bo = &vp->v_bufobj; 2021 ASSERT_BO_WLOCKED(bo); 2022 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 2023 2024 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 2025 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 2026 ("bgetvp: bp already attached! %p", bp)); 2027 2028 vhold(vp); 2029 bp->b_vp = vp; 2030 bp->b_bufobj = bo; 2031 /* 2032 * Insert onto list for new vnode. 2033 */ 2034 buf_vlist_add(bp, bo, BX_VNCLEAN); 2035 } 2036 2037 /* 2038 * Disassociate a buffer from a vnode. 2039 */ 2040 void 2041 brelvp(struct buf *bp) 2042 { 2043 struct bufobj *bo; 2044 struct vnode *vp; 2045 2046 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2047 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 2048 2049 /* 2050 * Delete from old vnode list, if on one. 2051 */ 2052 vp = bp->b_vp; /* XXX */ 2053 bo = bp->b_bufobj; 2054 BO_LOCK(bo); 2055 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2056 buf_vlist_remove(bp); 2057 else 2058 panic("brelvp: Buffer %p not on queue.", bp); 2059 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2060 bo->bo_flag &= ~BO_ONWORKLST; 2061 mtx_lock(&sync_mtx); 2062 LIST_REMOVE(bo, bo_synclist); 2063 syncer_worklist_len--; 2064 mtx_unlock(&sync_mtx); 2065 } 2066 bp->b_vp = NULL; 2067 bp->b_bufobj = NULL; 2068 BO_UNLOCK(bo); 2069 vdrop(vp); 2070 } 2071 2072 /* 2073 * Add an item to the syncer work queue. 2074 */ 2075 static void 2076 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 2077 { 2078 int slot; 2079 2080 ASSERT_BO_WLOCKED(bo); 2081 2082 mtx_lock(&sync_mtx); 2083 if (bo->bo_flag & BO_ONWORKLST) 2084 LIST_REMOVE(bo, bo_synclist); 2085 else { 2086 bo->bo_flag |= BO_ONWORKLST; 2087 syncer_worklist_len++; 2088 } 2089 2090 if (delay > syncer_maxdelay - 2) 2091 delay = syncer_maxdelay - 2; 2092 slot = (syncer_delayno + delay) & syncer_mask; 2093 2094 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 2095 mtx_unlock(&sync_mtx); 2096 } 2097 2098 static int 2099 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 2100 { 2101 int error, len; 2102 2103 mtx_lock(&sync_mtx); 2104 len = syncer_worklist_len - sync_vnode_count; 2105 mtx_unlock(&sync_mtx); 2106 error = SYSCTL_OUT(req, &len, sizeof(len)); 2107 return (error); 2108 } 2109 2110 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0, 2111 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 2112 2113 static struct proc *updateproc; 2114 static void sched_sync(void); 2115 static struct kproc_desc up_kp = { 2116 "syncer", 2117 sched_sync, 2118 &updateproc 2119 }; 2120 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); 2121 2122 static int 2123 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) 2124 { 2125 struct vnode *vp; 2126 struct mount *mp; 2127 2128 *bo = LIST_FIRST(slp); 2129 if (*bo == NULL) 2130 return (0); 2131 vp = bo2vnode(*bo); 2132 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) 2133 return (1); 2134 /* 2135 * We use vhold in case the vnode does not 2136 * successfully sync. vhold prevents the vnode from 2137 * going away when we unlock the sync_mtx so that 2138 * we can acquire the vnode interlock. 2139 */ 2140 vholdl(vp); 2141 mtx_unlock(&sync_mtx); 2142 VI_UNLOCK(vp); 2143 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2144 vdrop(vp); 2145 mtx_lock(&sync_mtx); 2146 return (*bo == LIST_FIRST(slp)); 2147 } 2148 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2149 (void) VOP_FSYNC(vp, MNT_LAZY, td); 2150 VOP_UNLOCK(vp, 0); 2151 vn_finished_write(mp); 2152 BO_LOCK(*bo); 2153 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { 2154 /* 2155 * Put us back on the worklist. The worklist 2156 * routine will remove us from our current 2157 * position and then add us back in at a later 2158 * position. 2159 */ 2160 vn_syncer_add_to_worklist(*bo, syncdelay); 2161 } 2162 BO_UNLOCK(*bo); 2163 vdrop(vp); 2164 mtx_lock(&sync_mtx); 2165 return (0); 2166 } 2167 2168 static int first_printf = 1; 2169 2170 /* 2171 * System filesystem synchronizer daemon. 2172 */ 2173 static void 2174 sched_sync(void) 2175 { 2176 struct synclist *next, *slp; 2177 struct bufobj *bo; 2178 long starttime; 2179 struct thread *td = curthread; 2180 int last_work_seen; 2181 int net_worklist_len; 2182 int syncer_final_iter; 2183 int error; 2184 2185 last_work_seen = 0; 2186 syncer_final_iter = 0; 2187 syncer_state = SYNCER_RUNNING; 2188 starttime = time_uptime; 2189 td->td_pflags |= TDP_NORUNNINGBUF; 2190 2191 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 2192 SHUTDOWN_PRI_LAST); 2193 2194 mtx_lock(&sync_mtx); 2195 for (;;) { 2196 if (syncer_state == SYNCER_FINAL_DELAY && 2197 syncer_final_iter == 0) { 2198 mtx_unlock(&sync_mtx); 2199 kproc_suspend_check(td->td_proc); 2200 mtx_lock(&sync_mtx); 2201 } 2202 net_worklist_len = syncer_worklist_len - sync_vnode_count; 2203 if (syncer_state != SYNCER_RUNNING && 2204 starttime != time_uptime) { 2205 if (first_printf) { 2206 printf("\nSyncing disks, vnodes remaining... "); 2207 first_printf = 0; 2208 } 2209 printf("%d ", net_worklist_len); 2210 } 2211 starttime = time_uptime; 2212 2213 /* 2214 * Push files whose dirty time has expired. Be careful 2215 * of interrupt race on slp queue. 2216 * 2217 * Skip over empty worklist slots when shutting down. 2218 */ 2219 do { 2220 slp = &syncer_workitem_pending[syncer_delayno]; 2221 syncer_delayno += 1; 2222 if (syncer_delayno == syncer_maxdelay) 2223 syncer_delayno = 0; 2224 next = &syncer_workitem_pending[syncer_delayno]; 2225 /* 2226 * If the worklist has wrapped since the 2227 * it was emptied of all but syncer vnodes, 2228 * switch to the FINAL_DELAY state and run 2229 * for one more second. 2230 */ 2231 if (syncer_state == SYNCER_SHUTTING_DOWN && 2232 net_worklist_len == 0 && 2233 last_work_seen == syncer_delayno) { 2234 syncer_state = SYNCER_FINAL_DELAY; 2235 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 2236 } 2237 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 2238 syncer_worklist_len > 0); 2239 2240 /* 2241 * Keep track of the last time there was anything 2242 * on the worklist other than syncer vnodes. 2243 * Return to the SHUTTING_DOWN state if any 2244 * new work appears. 2245 */ 2246 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 2247 last_work_seen = syncer_delayno; 2248 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 2249 syncer_state = SYNCER_SHUTTING_DOWN; 2250 while (!LIST_EMPTY(slp)) { 2251 error = sync_vnode(slp, &bo, td); 2252 if (error == 1) { 2253 LIST_REMOVE(bo, bo_synclist); 2254 LIST_INSERT_HEAD(next, bo, bo_synclist); 2255 continue; 2256 } 2257 2258 if (first_printf == 0) { 2259 /* 2260 * Drop the sync mutex, because some watchdog 2261 * drivers need to sleep while patting 2262 */ 2263 mtx_unlock(&sync_mtx); 2264 wdog_kern_pat(WD_LASTVAL); 2265 mtx_lock(&sync_mtx); 2266 } 2267 2268 } 2269 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 2270 syncer_final_iter--; 2271 /* 2272 * The variable rushjob allows the kernel to speed up the 2273 * processing of the filesystem syncer process. A rushjob 2274 * value of N tells the filesystem syncer to process the next 2275 * N seconds worth of work on its queue ASAP. Currently rushjob 2276 * is used by the soft update code to speed up the filesystem 2277 * syncer process when the incore state is getting so far 2278 * ahead of the disk that the kernel memory pool is being 2279 * threatened with exhaustion. 2280 */ 2281 if (rushjob > 0) { 2282 rushjob -= 1; 2283 continue; 2284 } 2285 /* 2286 * Just sleep for a short period of time between 2287 * iterations when shutting down to allow some I/O 2288 * to happen. 2289 * 2290 * If it has taken us less than a second to process the 2291 * current work, then wait. Otherwise start right over 2292 * again. We can still lose time if any single round 2293 * takes more than two seconds, but it does not really 2294 * matter as we are just trying to generally pace the 2295 * filesystem activity. 2296 */ 2297 if (syncer_state != SYNCER_RUNNING || 2298 time_uptime == starttime) { 2299 thread_lock(td); 2300 sched_prio(td, PPAUSE); 2301 thread_unlock(td); 2302 } 2303 if (syncer_state != SYNCER_RUNNING) 2304 cv_timedwait(&sync_wakeup, &sync_mtx, 2305 hz / SYNCER_SHUTDOWN_SPEEDUP); 2306 else if (time_uptime == starttime) 2307 cv_timedwait(&sync_wakeup, &sync_mtx, hz); 2308 } 2309 } 2310 2311 /* 2312 * Request the syncer daemon to speed up its work. 2313 * We never push it to speed up more than half of its 2314 * normal turn time, otherwise it could take over the cpu. 2315 */ 2316 int 2317 speedup_syncer(void) 2318 { 2319 int ret = 0; 2320 2321 mtx_lock(&sync_mtx); 2322 if (rushjob < syncdelay / 2) { 2323 rushjob += 1; 2324 stat_rush_requests += 1; 2325 ret = 1; 2326 } 2327 mtx_unlock(&sync_mtx); 2328 cv_broadcast(&sync_wakeup); 2329 return (ret); 2330 } 2331 2332 /* 2333 * Tell the syncer to speed up its work and run though its work 2334 * list several times, then tell it to shut down. 2335 */ 2336 static void 2337 syncer_shutdown(void *arg, int howto) 2338 { 2339 2340 if (howto & RB_NOSYNC) 2341 return; 2342 mtx_lock(&sync_mtx); 2343 syncer_state = SYNCER_SHUTTING_DOWN; 2344 rushjob = 0; 2345 mtx_unlock(&sync_mtx); 2346 cv_broadcast(&sync_wakeup); 2347 kproc_shutdown(arg, howto); 2348 } 2349 2350 void 2351 syncer_suspend(void) 2352 { 2353 2354 syncer_shutdown(updateproc, 0); 2355 } 2356 2357 void 2358 syncer_resume(void) 2359 { 2360 2361 mtx_lock(&sync_mtx); 2362 first_printf = 1; 2363 syncer_state = SYNCER_RUNNING; 2364 mtx_unlock(&sync_mtx); 2365 cv_broadcast(&sync_wakeup); 2366 kproc_resume(updateproc); 2367 } 2368 2369 /* 2370 * Reassign a buffer from one vnode to another. 2371 * Used to assign file specific control information 2372 * (indirect blocks) to the vnode to which they belong. 2373 */ 2374 void 2375 reassignbuf(struct buf *bp) 2376 { 2377 struct vnode *vp; 2378 struct bufobj *bo; 2379 int delay; 2380 #ifdef INVARIANTS 2381 struct bufv *bv; 2382 #endif 2383 2384 vp = bp->b_vp; 2385 bo = bp->b_bufobj; 2386 ++reassignbufcalls; 2387 2388 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 2389 bp, bp->b_vp, bp->b_flags); 2390 /* 2391 * B_PAGING flagged buffers cannot be reassigned because their vp 2392 * is not fully linked in. 2393 */ 2394 if (bp->b_flags & B_PAGING) 2395 panic("cannot reassign paging buffer"); 2396 2397 /* 2398 * Delete from old vnode list, if on one. 2399 */ 2400 BO_LOCK(bo); 2401 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2402 buf_vlist_remove(bp); 2403 else 2404 panic("reassignbuf: Buffer %p not on queue.", bp); 2405 /* 2406 * If dirty, put on list of dirty buffers; otherwise insert onto list 2407 * of clean buffers. 2408 */ 2409 if (bp->b_flags & B_DELWRI) { 2410 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 2411 switch (vp->v_type) { 2412 case VDIR: 2413 delay = dirdelay; 2414 break; 2415 case VCHR: 2416 delay = metadelay; 2417 break; 2418 default: 2419 delay = filedelay; 2420 } 2421 vn_syncer_add_to_worklist(bo, delay); 2422 } 2423 buf_vlist_add(bp, bo, BX_VNDIRTY); 2424 } else { 2425 buf_vlist_add(bp, bo, BX_VNCLEAN); 2426 2427 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2428 mtx_lock(&sync_mtx); 2429 LIST_REMOVE(bo, bo_synclist); 2430 syncer_worklist_len--; 2431 mtx_unlock(&sync_mtx); 2432 bo->bo_flag &= ~BO_ONWORKLST; 2433 } 2434 } 2435 #ifdef INVARIANTS 2436 bv = &bo->bo_clean; 2437 bp = TAILQ_FIRST(&bv->bv_hd); 2438 KASSERT(bp == NULL || bp->b_bufobj == bo, 2439 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2440 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2441 KASSERT(bp == NULL || bp->b_bufobj == bo, 2442 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2443 bv = &bo->bo_dirty; 2444 bp = TAILQ_FIRST(&bv->bv_hd); 2445 KASSERT(bp == NULL || bp->b_bufobj == bo, 2446 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2447 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2448 KASSERT(bp == NULL || bp->b_bufobj == bo, 2449 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2450 #endif 2451 BO_UNLOCK(bo); 2452 } 2453 2454 /* 2455 * A temporary hack until refcount_* APIs are sorted out. 2456 */ 2457 static __inline int 2458 vfs_refcount_acquire_if_not_zero(volatile u_int *count) 2459 { 2460 u_int old; 2461 2462 for (;;) { 2463 old = *count; 2464 if (old == 0) 2465 return (0); 2466 if (atomic_cmpset_int(count, old, old + 1)) 2467 return (1); 2468 } 2469 } 2470 2471 static __inline int 2472 vfs_refcount_release_if_not_last(volatile u_int *count) 2473 { 2474 u_int old; 2475 2476 for (;;) { 2477 old = *count; 2478 if (old == 1) 2479 return (0); 2480 if (atomic_cmpset_int(count, old, old - 1)) 2481 return (1); 2482 } 2483 } 2484 2485 static void 2486 v_init_counters(struct vnode *vp) 2487 { 2488 2489 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0, 2490 vp, ("%s called for an initialized vnode", __FUNCTION__)); 2491 ASSERT_VI_UNLOCKED(vp, __FUNCTION__); 2492 2493 refcount_init(&vp->v_holdcnt, 1); 2494 refcount_init(&vp->v_usecount, 1); 2495 } 2496 2497 static void 2498 v_incr_usecount_locked(struct vnode *vp) 2499 { 2500 2501 ASSERT_VI_LOCKED(vp, __func__); 2502 if ((vp->v_iflag & VI_OWEINACT) != 0) { 2503 VNASSERT(vp->v_usecount == 0, vp, 2504 ("vnode with usecount and VI_OWEINACT set")); 2505 vp->v_iflag &= ~VI_OWEINACT; 2506 } 2507 refcount_acquire(&vp->v_usecount); 2508 v_incr_devcount(vp); 2509 } 2510 2511 /* 2512 * Increment the use count on the vnode, taking care to reference 2513 * the driver's usecount if this is a chardev. 2514 */ 2515 static void 2516 v_incr_usecount(struct vnode *vp) 2517 { 2518 2519 ASSERT_VI_UNLOCKED(vp, __func__); 2520 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2521 2522 if (vp->v_type != VCHR && 2523 vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) { 2524 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2525 ("vnode with usecount and VI_OWEINACT set")); 2526 } else { 2527 VI_LOCK(vp); 2528 v_incr_usecount_locked(vp); 2529 VI_UNLOCK(vp); 2530 } 2531 } 2532 2533 /* 2534 * Increment si_usecount of the associated device, if any. 2535 */ 2536 static void 2537 v_incr_devcount(struct vnode *vp) 2538 { 2539 2540 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2541 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2542 dev_lock(); 2543 vp->v_rdev->si_usecount++; 2544 dev_unlock(); 2545 } 2546 } 2547 2548 /* 2549 * Decrement si_usecount of the associated device, if any. 2550 */ 2551 static void 2552 v_decr_devcount(struct vnode *vp) 2553 { 2554 2555 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2556 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2557 dev_lock(); 2558 vp->v_rdev->si_usecount--; 2559 dev_unlock(); 2560 } 2561 } 2562 2563 /* 2564 * Grab a particular vnode from the free list, increment its 2565 * reference count and lock it. VI_DOOMED is set if the vnode 2566 * is being destroyed. Only callers who specify LK_RETRY will 2567 * see doomed vnodes. If inactive processing was delayed in 2568 * vput try to do it here. 2569 * 2570 * Notes on lockless counter manipulation: 2571 * _vhold, vputx and other routines make various decisions based 2572 * on either holdcnt or usecount being 0. As long as either counter 2573 * is not transitioning 0->1 nor 1->0, the manipulation can be done 2574 * with atomic operations. Otherwise the interlock is taken covering 2575 * both the atomic and additional actions. 2576 */ 2577 int 2578 vget(struct vnode *vp, int flags, struct thread *td) 2579 { 2580 int error, oweinact; 2581 2582 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 2583 ("vget: invalid lock operation")); 2584 2585 if ((flags & LK_INTERLOCK) != 0) 2586 ASSERT_VI_LOCKED(vp, __func__); 2587 else 2588 ASSERT_VI_UNLOCKED(vp, __func__); 2589 if ((flags & LK_VNHELD) != 0) 2590 VNASSERT((vp->v_holdcnt > 0), vp, 2591 ("vget: LK_VNHELD passed but vnode not held")); 2592 2593 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 2594 2595 if ((flags & LK_VNHELD) == 0) 2596 _vhold(vp, (flags & LK_INTERLOCK) != 0); 2597 2598 if ((error = vn_lock(vp, flags)) != 0) { 2599 vdrop(vp); 2600 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, 2601 vp); 2602 return (error); 2603 } 2604 if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0) 2605 panic("vget: vn_lock failed to return ENOENT\n"); 2606 /* 2607 * We don't guarantee that any particular close will 2608 * trigger inactive processing so just make a best effort 2609 * here at preventing a reference to a removed file. If 2610 * we don't succeed no harm is done. 2611 * 2612 * Upgrade our holdcnt to a usecount. 2613 */ 2614 if (vp->v_type == VCHR || 2615 !vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) { 2616 VI_LOCK(vp); 2617 if ((vp->v_iflag & VI_OWEINACT) == 0) { 2618 oweinact = 0; 2619 } else { 2620 oweinact = 1; 2621 vp->v_iflag &= ~VI_OWEINACT; 2622 } 2623 refcount_acquire(&vp->v_usecount); 2624 v_incr_devcount(vp); 2625 if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE && 2626 (flags & LK_NOWAIT) == 0) 2627 vinactive(vp, td); 2628 VI_UNLOCK(vp); 2629 } 2630 return (0); 2631 } 2632 2633 /* 2634 * Increase the reference (use) and hold count of a vnode. 2635 * This will also remove the vnode from the free list if it is presently free. 2636 */ 2637 void 2638 vref(struct vnode *vp) 2639 { 2640 2641 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2642 _vhold(vp, false); 2643 v_incr_usecount(vp); 2644 } 2645 2646 void 2647 vrefl(struct vnode *vp) 2648 { 2649 2650 ASSERT_VI_LOCKED(vp, __func__); 2651 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2652 _vhold(vp, true); 2653 v_incr_usecount_locked(vp); 2654 } 2655 2656 void 2657 vrefact(struct vnode *vp) 2658 { 2659 2660 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2661 if (__predict_false(vp->v_type == VCHR)) { 2662 VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp, 2663 ("%s: wrong ref counts", __func__)); 2664 vref(vp); 2665 return; 2666 } 2667 #ifdef INVARIANTS 2668 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 2669 VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__)); 2670 old = atomic_fetchadd_int(&vp->v_usecount, 1); 2671 VNASSERT(old > 0, vp, ("%s: wrong use count", __func__)); 2672 #else 2673 refcount_acquire(&vp->v_holdcnt); 2674 refcount_acquire(&vp->v_usecount); 2675 #endif 2676 } 2677 2678 /* 2679 * Return reference count of a vnode. 2680 * 2681 * The results of this call are only guaranteed when some mechanism is used to 2682 * stop other processes from gaining references to the vnode. This may be the 2683 * case if the caller holds the only reference. This is also useful when stale 2684 * data is acceptable as race conditions may be accounted for by some other 2685 * means. 2686 */ 2687 int 2688 vrefcnt(struct vnode *vp) 2689 { 2690 2691 return (vp->v_usecount); 2692 } 2693 2694 #define VPUTX_VRELE 1 2695 #define VPUTX_VPUT 2 2696 #define VPUTX_VUNREF 3 2697 2698 /* 2699 * Decrement the use and hold counts for a vnode. 2700 * 2701 * See an explanation near vget() as to why atomic operation is safe. 2702 */ 2703 static void 2704 vputx(struct vnode *vp, int func) 2705 { 2706 int error; 2707 2708 KASSERT(vp != NULL, ("vputx: null vp")); 2709 if (func == VPUTX_VUNREF) 2710 ASSERT_VOP_LOCKED(vp, "vunref"); 2711 else if (func == VPUTX_VPUT) 2712 ASSERT_VOP_LOCKED(vp, "vput"); 2713 else 2714 KASSERT(func == VPUTX_VRELE, ("vputx: wrong func")); 2715 ASSERT_VI_UNLOCKED(vp, __func__); 2716 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2717 2718 if (vp->v_type != VCHR && 2719 vfs_refcount_release_if_not_last(&vp->v_usecount)) { 2720 if (func == VPUTX_VPUT) 2721 VOP_UNLOCK(vp, 0); 2722 vdrop(vp); 2723 return; 2724 } 2725 2726 VI_LOCK(vp); 2727 2728 /* 2729 * We want to hold the vnode until the inactive finishes to 2730 * prevent vgone() races. We drop the use count here and the 2731 * hold count below when we're done. 2732 */ 2733 if (!refcount_release(&vp->v_usecount) || 2734 (vp->v_iflag & VI_DOINGINACT)) { 2735 if (func == VPUTX_VPUT) 2736 VOP_UNLOCK(vp, 0); 2737 v_decr_devcount(vp); 2738 vdropl(vp); 2739 return; 2740 } 2741 2742 v_decr_devcount(vp); 2743 2744 error = 0; 2745 2746 if (vp->v_usecount != 0) { 2747 vn_printf(vp, "vputx: usecount not zero for vnode "); 2748 panic("vputx: usecount not zero"); 2749 } 2750 2751 CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp); 2752 2753 /* 2754 * We must call VOP_INACTIVE with the node locked. Mark 2755 * as VI_DOINGINACT to avoid recursion. 2756 */ 2757 vp->v_iflag |= VI_OWEINACT; 2758 switch (func) { 2759 case VPUTX_VRELE: 2760 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 2761 VI_LOCK(vp); 2762 break; 2763 case VPUTX_VPUT: 2764 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 2765 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | 2766 LK_NOWAIT); 2767 VI_LOCK(vp); 2768 } 2769 break; 2770 case VPUTX_VUNREF: 2771 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 2772 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); 2773 VI_LOCK(vp); 2774 } 2775 break; 2776 } 2777 VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp, 2778 ("vnode with usecount and VI_OWEINACT set")); 2779 if (error == 0) { 2780 if (vp->v_iflag & VI_OWEINACT) 2781 vinactive(vp, curthread); 2782 if (func != VPUTX_VUNREF) 2783 VOP_UNLOCK(vp, 0); 2784 } 2785 vdropl(vp); 2786 } 2787 2788 /* 2789 * Vnode put/release. 2790 * If count drops to zero, call inactive routine and return to freelist. 2791 */ 2792 void 2793 vrele(struct vnode *vp) 2794 { 2795 2796 vputx(vp, VPUTX_VRELE); 2797 } 2798 2799 /* 2800 * Release an already locked vnode. This give the same effects as 2801 * unlock+vrele(), but takes less time and avoids releasing and 2802 * re-aquiring the lock (as vrele() acquires the lock internally.) 2803 */ 2804 void 2805 vput(struct vnode *vp) 2806 { 2807 2808 vputx(vp, VPUTX_VPUT); 2809 } 2810 2811 /* 2812 * Release an exclusively locked vnode. Do not unlock the vnode lock. 2813 */ 2814 void 2815 vunref(struct vnode *vp) 2816 { 2817 2818 vputx(vp, VPUTX_VUNREF); 2819 } 2820 2821 /* 2822 * Increase the hold count and activate if this is the first reference. 2823 */ 2824 void 2825 _vhold(struct vnode *vp, bool locked) 2826 { 2827 struct mount *mp; 2828 2829 if (locked) 2830 ASSERT_VI_LOCKED(vp, __func__); 2831 else 2832 ASSERT_VI_UNLOCKED(vp, __func__); 2833 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2834 if (!locked && vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt)) { 2835 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 2836 ("_vhold: vnode with holdcnt is free")); 2837 return; 2838 } 2839 2840 if (!locked) 2841 VI_LOCK(vp); 2842 if ((vp->v_iflag & VI_FREE) == 0) { 2843 refcount_acquire(&vp->v_holdcnt); 2844 if (!locked) 2845 VI_UNLOCK(vp); 2846 return; 2847 } 2848 VNASSERT(vp->v_holdcnt == 0, vp, 2849 ("%s: wrong hold count", __func__)); 2850 VNASSERT(vp->v_op != NULL, vp, 2851 ("%s: vnode already reclaimed.", __func__)); 2852 /* 2853 * Remove a vnode from the free list, mark it as in use, 2854 * and put it on the active list. 2855 */ 2856 mp = vp->v_mount; 2857 mtx_lock(&mp->mnt_listmtx); 2858 if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) { 2859 TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist); 2860 mp->mnt_tmpfreevnodelistsize--; 2861 vp->v_mflag &= ~VMP_TMPMNTFREELIST; 2862 } else { 2863 mtx_lock(&vnode_free_list_mtx); 2864 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist); 2865 freevnodes--; 2866 mtx_unlock(&vnode_free_list_mtx); 2867 } 2868 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 2869 ("Activating already active vnode")); 2870 vp->v_iflag &= ~VI_FREE; 2871 vp->v_iflag |= VI_ACTIVE; 2872 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist); 2873 mp->mnt_activevnodelistsize++; 2874 mtx_unlock(&mp->mnt_listmtx); 2875 refcount_acquire(&vp->v_holdcnt); 2876 if (!locked) 2877 VI_UNLOCK(vp); 2878 } 2879 2880 /* 2881 * Drop the hold count of the vnode. If this is the last reference to 2882 * the vnode we place it on the free list unless it has been vgone'd 2883 * (marked VI_DOOMED) in which case we will free it. 2884 * 2885 * Because the vnode vm object keeps a hold reference on the vnode if 2886 * there is at least one resident non-cached page, the vnode cannot 2887 * leave the active list without the page cleanup done. 2888 */ 2889 void 2890 _vdrop(struct vnode *vp, bool locked) 2891 { 2892 struct bufobj *bo; 2893 struct mount *mp; 2894 int active; 2895 2896 if (locked) 2897 ASSERT_VI_LOCKED(vp, __func__); 2898 else 2899 ASSERT_VI_UNLOCKED(vp, __func__); 2900 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2901 if ((int)vp->v_holdcnt <= 0) 2902 panic("vdrop: holdcnt %d", vp->v_holdcnt); 2903 if (vfs_refcount_release_if_not_last(&vp->v_holdcnt)) { 2904 if (locked) 2905 VI_UNLOCK(vp); 2906 return; 2907 } 2908 2909 if (!locked) 2910 VI_LOCK(vp); 2911 if (refcount_release(&vp->v_holdcnt) == 0) { 2912 VI_UNLOCK(vp); 2913 return; 2914 } 2915 if ((vp->v_iflag & VI_DOOMED) == 0) { 2916 /* 2917 * Mark a vnode as free: remove it from its active list 2918 * and put it up for recycling on the freelist. 2919 */ 2920 VNASSERT(vp->v_op != NULL, vp, 2921 ("vdropl: vnode already reclaimed.")); 2922 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 2923 ("vnode already free")); 2924 VNASSERT(vp->v_holdcnt == 0, vp, 2925 ("vdropl: freeing when we shouldn't")); 2926 active = vp->v_iflag & VI_ACTIVE; 2927 if ((vp->v_iflag & VI_OWEINACT) == 0) { 2928 vp->v_iflag &= ~VI_ACTIVE; 2929 mp = vp->v_mount; 2930 mtx_lock(&mp->mnt_listmtx); 2931 if (active) { 2932 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, 2933 v_actfreelist); 2934 mp->mnt_activevnodelistsize--; 2935 } 2936 TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist, vp, 2937 v_actfreelist); 2938 mp->mnt_tmpfreevnodelistsize++; 2939 vp->v_iflag |= VI_FREE; 2940 vp->v_mflag |= VMP_TMPMNTFREELIST; 2941 VI_UNLOCK(vp); 2942 if (mp->mnt_tmpfreevnodelistsize >= mnt_free_list_batch) 2943 vnlru_return_batch_locked(mp); 2944 mtx_unlock(&mp->mnt_listmtx); 2945 } else { 2946 VI_UNLOCK(vp); 2947 counter_u64_add(free_owe_inact, 1); 2948 } 2949 return; 2950 } 2951 /* 2952 * The vnode has been marked for destruction, so free it. 2953 * 2954 * The vnode will be returned to the zone where it will 2955 * normally remain until it is needed for another vnode. We 2956 * need to cleanup (or verify that the cleanup has already 2957 * been done) any residual data left from its current use 2958 * so as not to contaminate the freshly allocated vnode. 2959 */ 2960 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp); 2961 atomic_subtract_long(&numvnodes, 1); 2962 bo = &vp->v_bufobj; 2963 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 2964 ("cleaned vnode still on the free list.")); 2965 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 2966 VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count")); 2967 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 2968 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 2969 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 2970 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 2971 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp, 2972 ("clean blk trie not empty")); 2973 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 2974 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp, 2975 ("dirty blk trie not empty")); 2976 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); 2977 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); 2978 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for ..")); 2979 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp, 2980 ("Dangling rangelock waiters")); 2981 VI_UNLOCK(vp); 2982 #ifdef MAC 2983 mac_vnode_destroy(vp); 2984 #endif 2985 if (vp->v_pollinfo != NULL) { 2986 destroy_vpollinfo(vp->v_pollinfo); 2987 vp->v_pollinfo = NULL; 2988 } 2989 #ifdef INVARIANTS 2990 /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */ 2991 vp->v_op = NULL; 2992 #endif 2993 bzero(&vp->v_un, sizeof(vp->v_un)); 2994 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 2995 vp->v_iflag = 0; 2996 vp->v_vflag = 0; 2997 bo->bo_flag = 0; 2998 uma_zfree(vnode_zone, vp); 2999 } 3000 3001 /* 3002 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 3003 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 3004 * OWEINACT tracks whether a vnode missed a call to inactive due to a 3005 * failed lock upgrade. 3006 */ 3007 void 3008 vinactive(struct vnode *vp, struct thread *td) 3009 { 3010 struct vm_object *obj; 3011 3012 ASSERT_VOP_ELOCKED(vp, "vinactive"); 3013 ASSERT_VI_LOCKED(vp, "vinactive"); 3014 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, 3015 ("vinactive: recursed on VI_DOINGINACT")); 3016 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3017 vp->v_iflag |= VI_DOINGINACT; 3018 vp->v_iflag &= ~VI_OWEINACT; 3019 VI_UNLOCK(vp); 3020 /* 3021 * Before moving off the active list, we must be sure that any 3022 * modified pages are converted into the vnode's dirty 3023 * buffers, since these will no longer be checked once the 3024 * vnode is on the inactive list. 3025 * 3026 * The write-out of the dirty pages is asynchronous. At the 3027 * point that VOP_INACTIVE() is called, there could still be 3028 * pending I/O and dirty pages in the object. 3029 */ 3030 obj = vp->v_object; 3031 if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0) { 3032 VM_OBJECT_WLOCK(obj); 3033 vm_object_page_clean(obj, 0, 0, 0); 3034 VM_OBJECT_WUNLOCK(obj); 3035 } 3036 VOP_INACTIVE(vp, td); 3037 VI_LOCK(vp); 3038 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, 3039 ("vinactive: lost VI_DOINGINACT")); 3040 vp->v_iflag &= ~VI_DOINGINACT; 3041 } 3042 3043 /* 3044 * Remove any vnodes in the vnode table belonging to mount point mp. 3045 * 3046 * If FORCECLOSE is not specified, there should not be any active ones, 3047 * return error if any are found (nb: this is a user error, not a 3048 * system error). If FORCECLOSE is specified, detach any active vnodes 3049 * that are found. 3050 * 3051 * If WRITECLOSE is set, only flush out regular file vnodes open for 3052 * writing. 3053 * 3054 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 3055 * 3056 * `rootrefs' specifies the base reference count for the root vnode 3057 * of this filesystem. The root vnode is considered busy if its 3058 * v_usecount exceeds this value. On a successful return, vflush(, td) 3059 * will call vrele() on the root vnode exactly rootrefs times. 3060 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 3061 * be zero. 3062 */ 3063 #ifdef DIAGNOSTIC 3064 static int busyprt = 0; /* print out busy vnodes */ 3065 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 3066 #endif 3067 3068 int 3069 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 3070 { 3071 struct vnode *vp, *mvp, *rootvp = NULL; 3072 struct vattr vattr; 3073 int busy = 0, error; 3074 3075 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 3076 rootrefs, flags); 3077 if (rootrefs > 0) { 3078 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 3079 ("vflush: bad args")); 3080 /* 3081 * Get the filesystem root vnode. We can vput() it 3082 * immediately, since with rootrefs > 0, it won't go away. 3083 */ 3084 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 3085 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 3086 __func__, error); 3087 return (error); 3088 } 3089 vput(rootvp); 3090 } 3091 loop: 3092 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 3093 vholdl(vp); 3094 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 3095 if (error) { 3096 vdrop(vp); 3097 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3098 goto loop; 3099 } 3100 /* 3101 * Skip over a vnodes marked VV_SYSTEM. 3102 */ 3103 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 3104 VOP_UNLOCK(vp, 0); 3105 vdrop(vp); 3106 continue; 3107 } 3108 /* 3109 * If WRITECLOSE is set, flush out unlinked but still open 3110 * files (even if open only for reading) and regular file 3111 * vnodes open for writing. 3112 */ 3113 if (flags & WRITECLOSE) { 3114 if (vp->v_object != NULL) { 3115 VM_OBJECT_WLOCK(vp->v_object); 3116 vm_object_page_clean(vp->v_object, 0, 0, 0); 3117 VM_OBJECT_WUNLOCK(vp->v_object); 3118 } 3119 error = VOP_FSYNC(vp, MNT_WAIT, td); 3120 if (error != 0) { 3121 VOP_UNLOCK(vp, 0); 3122 vdrop(vp); 3123 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3124 return (error); 3125 } 3126 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 3127 VI_LOCK(vp); 3128 3129 if ((vp->v_type == VNON || 3130 (error == 0 && vattr.va_nlink > 0)) && 3131 (vp->v_writecount == 0 || vp->v_type != VREG)) { 3132 VOP_UNLOCK(vp, 0); 3133 vdropl(vp); 3134 continue; 3135 } 3136 } else 3137 VI_LOCK(vp); 3138 /* 3139 * With v_usecount == 0, all we need to do is clear out the 3140 * vnode data structures and we are done. 3141 * 3142 * If FORCECLOSE is set, forcibly close the vnode. 3143 */ 3144 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 3145 vgonel(vp); 3146 } else { 3147 busy++; 3148 #ifdef DIAGNOSTIC 3149 if (busyprt) 3150 vn_printf(vp, "vflush: busy vnode "); 3151 #endif 3152 } 3153 VOP_UNLOCK(vp, 0); 3154 vdropl(vp); 3155 } 3156 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 3157 /* 3158 * If just the root vnode is busy, and if its refcount 3159 * is equal to `rootrefs', then go ahead and kill it. 3160 */ 3161 VI_LOCK(rootvp); 3162 KASSERT(busy > 0, ("vflush: not busy")); 3163 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 3164 ("vflush: usecount %d < rootrefs %d", 3165 rootvp->v_usecount, rootrefs)); 3166 if (busy == 1 && rootvp->v_usecount == rootrefs) { 3167 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 3168 vgone(rootvp); 3169 VOP_UNLOCK(rootvp, 0); 3170 busy = 0; 3171 } else 3172 VI_UNLOCK(rootvp); 3173 } 3174 if (busy) { 3175 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 3176 busy); 3177 return (EBUSY); 3178 } 3179 for (; rootrefs > 0; rootrefs--) 3180 vrele(rootvp); 3181 return (0); 3182 } 3183 3184 /* 3185 * Recycle an unused vnode to the front of the free list. 3186 */ 3187 int 3188 vrecycle(struct vnode *vp) 3189 { 3190 int recycled; 3191 3192 VI_LOCK(vp); 3193 recycled = vrecyclel(vp); 3194 VI_UNLOCK(vp); 3195 return (recycled); 3196 } 3197 3198 /* 3199 * vrecycle, with the vp interlock held. 3200 */ 3201 int 3202 vrecyclel(struct vnode *vp) 3203 { 3204 int recycled; 3205 3206 ASSERT_VOP_ELOCKED(vp, __func__); 3207 ASSERT_VI_LOCKED(vp, __func__); 3208 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3209 recycled = 0; 3210 if (vp->v_usecount == 0) { 3211 recycled = 1; 3212 vgonel(vp); 3213 } 3214 return (recycled); 3215 } 3216 3217 /* 3218 * Eliminate all activity associated with a vnode 3219 * in preparation for reuse. 3220 */ 3221 void 3222 vgone(struct vnode *vp) 3223 { 3224 VI_LOCK(vp); 3225 vgonel(vp); 3226 VI_UNLOCK(vp); 3227 } 3228 3229 static void 3230 notify_lowervp_vfs_dummy(struct mount *mp __unused, 3231 struct vnode *lowervp __unused) 3232 { 3233 } 3234 3235 /* 3236 * Notify upper mounts about reclaimed or unlinked vnode. 3237 */ 3238 void 3239 vfs_notify_upper(struct vnode *vp, int event) 3240 { 3241 static struct vfsops vgonel_vfsops = { 3242 .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy, 3243 .vfs_unlink_lowervp = notify_lowervp_vfs_dummy, 3244 }; 3245 struct mount *mp, *ump, *mmp; 3246 3247 mp = vp->v_mount; 3248 if (mp == NULL) 3249 return; 3250 3251 MNT_ILOCK(mp); 3252 if (TAILQ_EMPTY(&mp->mnt_uppers)) 3253 goto unlock; 3254 MNT_IUNLOCK(mp); 3255 mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO); 3256 mmp->mnt_op = &vgonel_vfsops; 3257 mmp->mnt_kern_flag |= MNTK_MARKER; 3258 MNT_ILOCK(mp); 3259 mp->mnt_kern_flag |= MNTK_VGONE_UPPER; 3260 for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) { 3261 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) { 3262 ump = TAILQ_NEXT(ump, mnt_upper_link); 3263 continue; 3264 } 3265 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link); 3266 MNT_IUNLOCK(mp); 3267 switch (event) { 3268 case VFS_NOTIFY_UPPER_RECLAIM: 3269 VFS_RECLAIM_LOWERVP(ump, vp); 3270 break; 3271 case VFS_NOTIFY_UPPER_UNLINK: 3272 VFS_UNLINK_LOWERVP(ump, vp); 3273 break; 3274 default: 3275 KASSERT(0, ("invalid event %d", event)); 3276 break; 3277 } 3278 MNT_ILOCK(mp); 3279 ump = TAILQ_NEXT(mmp, mnt_upper_link); 3280 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link); 3281 } 3282 free(mmp, M_TEMP); 3283 mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER; 3284 if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) { 3285 mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER; 3286 wakeup(&mp->mnt_uppers); 3287 } 3288 unlock: 3289 MNT_IUNLOCK(mp); 3290 } 3291 3292 /* 3293 * vgone, with the vp interlock held. 3294 */ 3295 static void 3296 vgonel(struct vnode *vp) 3297 { 3298 struct thread *td; 3299 int oweinact; 3300 int active; 3301 struct mount *mp; 3302 3303 ASSERT_VOP_ELOCKED(vp, "vgonel"); 3304 ASSERT_VI_LOCKED(vp, "vgonel"); 3305 VNASSERT(vp->v_holdcnt, vp, 3306 ("vgonel: vp %p has no reference.", vp)); 3307 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3308 td = curthread; 3309 3310 /* 3311 * Don't vgonel if we're already doomed. 3312 */ 3313 if (vp->v_iflag & VI_DOOMED) 3314 return; 3315 vp->v_iflag |= VI_DOOMED; 3316 3317 /* 3318 * Check to see if the vnode is in use. If so, we have to call 3319 * VOP_CLOSE() and VOP_INACTIVE(). 3320 */ 3321 active = vp->v_usecount; 3322 oweinact = (vp->v_iflag & VI_OWEINACT); 3323 VI_UNLOCK(vp); 3324 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); 3325 3326 /* 3327 * If purging an active vnode, it must be closed and 3328 * deactivated before being reclaimed. 3329 */ 3330 if (active) 3331 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 3332 if (oweinact || active) { 3333 VI_LOCK(vp); 3334 if ((vp->v_iflag & VI_DOINGINACT) == 0) 3335 vinactive(vp, td); 3336 VI_UNLOCK(vp); 3337 } 3338 if (vp->v_type == VSOCK) 3339 vfs_unp_reclaim(vp); 3340 3341 /* 3342 * Clean out any buffers associated with the vnode. 3343 * If the flush fails, just toss the buffers. 3344 */ 3345 mp = NULL; 3346 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 3347 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 3348 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { 3349 while (vinvalbuf(vp, 0, 0, 0) != 0) 3350 ; 3351 } 3352 3353 BO_LOCK(&vp->v_bufobj); 3354 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && 3355 vp->v_bufobj.bo_dirty.bv_cnt == 0 && 3356 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && 3357 vp->v_bufobj.bo_clean.bv_cnt == 0, 3358 ("vp %p bufobj not invalidated", vp)); 3359 3360 /* 3361 * For VMIO bufobj, BO_DEAD is set in vm_object_terminate() 3362 * after the object's page queue is flushed. 3363 */ 3364 if (vp->v_bufobj.bo_object == NULL) 3365 vp->v_bufobj.bo_flag |= BO_DEAD; 3366 BO_UNLOCK(&vp->v_bufobj); 3367 3368 /* 3369 * Reclaim the vnode. 3370 */ 3371 if (VOP_RECLAIM(vp, td)) 3372 panic("vgone: cannot reclaim"); 3373 if (mp != NULL) 3374 vn_finished_secondary_write(mp); 3375 VNASSERT(vp->v_object == NULL, vp, 3376 ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag)); 3377 /* 3378 * Clear the advisory locks and wake up waiting threads. 3379 */ 3380 (void)VOP_ADVLOCKPURGE(vp); 3381 vp->v_lockf = NULL; 3382 /* 3383 * Delete from old mount point vnode list. 3384 */ 3385 delmntque(vp); 3386 cache_purge(vp); 3387 /* 3388 * Done with purge, reset to the standard lock and invalidate 3389 * the vnode. 3390 */ 3391 VI_LOCK(vp); 3392 vp->v_vnlock = &vp->v_lock; 3393 vp->v_op = &dead_vnodeops; 3394 vp->v_tag = "none"; 3395 vp->v_type = VBAD; 3396 } 3397 3398 /* 3399 * Calculate the total number of references to a special device. 3400 */ 3401 int 3402 vcount(struct vnode *vp) 3403 { 3404 int count; 3405 3406 dev_lock(); 3407 count = vp->v_rdev->si_usecount; 3408 dev_unlock(); 3409 return (count); 3410 } 3411 3412 /* 3413 * Same as above, but using the struct cdev *as argument 3414 */ 3415 int 3416 count_dev(struct cdev *dev) 3417 { 3418 int count; 3419 3420 dev_lock(); 3421 count = dev->si_usecount; 3422 dev_unlock(); 3423 return(count); 3424 } 3425 3426 /* 3427 * Print out a description of a vnode. 3428 */ 3429 static char *typename[] = 3430 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", 3431 "VMARKER"}; 3432 3433 void 3434 vn_printf(struct vnode *vp, const char *fmt, ...) 3435 { 3436 va_list ap; 3437 char buf[256], buf2[16]; 3438 u_long flags; 3439 3440 va_start(ap, fmt); 3441 vprintf(fmt, ap); 3442 va_end(ap); 3443 printf("%p: ", (void *)vp); 3444 printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]); 3445 printf(" usecount %d, writecount %d, refcount %d mountedhere %p\n", 3446 vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere); 3447 buf[0] = '\0'; 3448 buf[1] = '\0'; 3449 if (vp->v_vflag & VV_ROOT) 3450 strlcat(buf, "|VV_ROOT", sizeof(buf)); 3451 if (vp->v_vflag & VV_ISTTY) 3452 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 3453 if (vp->v_vflag & VV_NOSYNC) 3454 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 3455 if (vp->v_vflag & VV_ETERNALDEV) 3456 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 3457 if (vp->v_vflag & VV_CACHEDLABEL) 3458 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 3459 if (vp->v_vflag & VV_TEXT) 3460 strlcat(buf, "|VV_TEXT", sizeof(buf)); 3461 if (vp->v_vflag & VV_COPYONWRITE) 3462 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 3463 if (vp->v_vflag & VV_SYSTEM) 3464 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 3465 if (vp->v_vflag & VV_PROCDEP) 3466 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 3467 if (vp->v_vflag & VV_NOKNOTE) 3468 strlcat(buf, "|VV_NOKNOTE", sizeof(buf)); 3469 if (vp->v_vflag & VV_DELETED) 3470 strlcat(buf, "|VV_DELETED", sizeof(buf)); 3471 if (vp->v_vflag & VV_MD) 3472 strlcat(buf, "|VV_MD", sizeof(buf)); 3473 if (vp->v_vflag & VV_FORCEINSMQ) 3474 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 3475 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 3476 VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | 3477 VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ); 3478 if (flags != 0) { 3479 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 3480 strlcat(buf, buf2, sizeof(buf)); 3481 } 3482 if (vp->v_iflag & VI_MOUNT) 3483 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 3484 if (vp->v_iflag & VI_DOOMED) 3485 strlcat(buf, "|VI_DOOMED", sizeof(buf)); 3486 if (vp->v_iflag & VI_FREE) 3487 strlcat(buf, "|VI_FREE", sizeof(buf)); 3488 if (vp->v_iflag & VI_ACTIVE) 3489 strlcat(buf, "|VI_ACTIVE", sizeof(buf)); 3490 if (vp->v_iflag & VI_DOINGINACT) 3491 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 3492 if (vp->v_iflag & VI_OWEINACT) 3493 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 3494 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE | 3495 VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT); 3496 if (flags != 0) { 3497 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 3498 strlcat(buf, buf2, sizeof(buf)); 3499 } 3500 printf(" flags (%s)\n", buf + 1); 3501 if (mtx_owned(VI_MTX(vp))) 3502 printf(" VI_LOCKed"); 3503 if (vp->v_object != NULL) 3504 printf(" v_object %p ref %d pages %d " 3505 "cleanbuf %d dirtybuf %d\n", 3506 vp->v_object, vp->v_object->ref_count, 3507 vp->v_object->resident_page_count, 3508 vp->v_bufobj.bo_clean.bv_cnt, 3509 vp->v_bufobj.bo_dirty.bv_cnt); 3510 printf(" "); 3511 lockmgr_printinfo(vp->v_vnlock); 3512 if (vp->v_data != NULL) 3513 VOP_PRINT(vp); 3514 } 3515 3516 #ifdef DDB 3517 /* 3518 * List all of the locked vnodes in the system. 3519 * Called when debugging the kernel. 3520 */ 3521 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 3522 { 3523 struct mount *mp; 3524 struct vnode *vp; 3525 3526 /* 3527 * Note: because this is DDB, we can't obey the locking semantics 3528 * for these structures, which means we could catch an inconsistent 3529 * state and dereference a nasty pointer. Not much to be done 3530 * about that. 3531 */ 3532 db_printf("Locked vnodes\n"); 3533 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3534 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3535 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) 3536 vn_printf(vp, "vnode "); 3537 } 3538 } 3539 } 3540 3541 /* 3542 * Show details about the given vnode. 3543 */ 3544 DB_SHOW_COMMAND(vnode, db_show_vnode) 3545 { 3546 struct vnode *vp; 3547 3548 if (!have_addr) 3549 return; 3550 vp = (struct vnode *)addr; 3551 vn_printf(vp, "vnode "); 3552 } 3553 3554 /* 3555 * Show details about the given mount point. 3556 */ 3557 DB_SHOW_COMMAND(mount, db_show_mount) 3558 { 3559 struct mount *mp; 3560 struct vfsopt *opt; 3561 struct statfs *sp; 3562 struct vnode *vp; 3563 char buf[512]; 3564 uint64_t mflags; 3565 u_int flags; 3566 3567 if (!have_addr) { 3568 /* No address given, print short info about all mount points. */ 3569 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3570 db_printf("%p %s on %s (%s)\n", mp, 3571 mp->mnt_stat.f_mntfromname, 3572 mp->mnt_stat.f_mntonname, 3573 mp->mnt_stat.f_fstypename); 3574 if (db_pager_quit) 3575 break; 3576 } 3577 db_printf("\nMore info: show mount <addr>\n"); 3578 return; 3579 } 3580 3581 mp = (struct mount *)addr; 3582 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 3583 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 3584 3585 buf[0] = '\0'; 3586 mflags = mp->mnt_flag; 3587 #define MNT_FLAG(flag) do { \ 3588 if (mflags & (flag)) { \ 3589 if (buf[0] != '\0') \ 3590 strlcat(buf, ", ", sizeof(buf)); \ 3591 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 3592 mflags &= ~(flag); \ 3593 } \ 3594 } while (0) 3595 MNT_FLAG(MNT_RDONLY); 3596 MNT_FLAG(MNT_SYNCHRONOUS); 3597 MNT_FLAG(MNT_NOEXEC); 3598 MNT_FLAG(MNT_NOSUID); 3599 MNT_FLAG(MNT_NFS4ACLS); 3600 MNT_FLAG(MNT_UNION); 3601 MNT_FLAG(MNT_ASYNC); 3602 MNT_FLAG(MNT_SUIDDIR); 3603 MNT_FLAG(MNT_SOFTDEP); 3604 MNT_FLAG(MNT_NOSYMFOLLOW); 3605 MNT_FLAG(MNT_GJOURNAL); 3606 MNT_FLAG(MNT_MULTILABEL); 3607 MNT_FLAG(MNT_ACLS); 3608 MNT_FLAG(MNT_NOATIME); 3609 MNT_FLAG(MNT_NOCLUSTERR); 3610 MNT_FLAG(MNT_NOCLUSTERW); 3611 MNT_FLAG(MNT_SUJ); 3612 MNT_FLAG(MNT_EXRDONLY); 3613 MNT_FLAG(MNT_EXPORTED); 3614 MNT_FLAG(MNT_DEFEXPORTED); 3615 MNT_FLAG(MNT_EXPORTANON); 3616 MNT_FLAG(MNT_EXKERB); 3617 MNT_FLAG(MNT_EXPUBLIC); 3618 MNT_FLAG(MNT_LOCAL); 3619 MNT_FLAG(MNT_QUOTA); 3620 MNT_FLAG(MNT_ROOTFS); 3621 MNT_FLAG(MNT_USER); 3622 MNT_FLAG(MNT_IGNORE); 3623 MNT_FLAG(MNT_UPDATE); 3624 MNT_FLAG(MNT_DELEXPORT); 3625 MNT_FLAG(MNT_RELOAD); 3626 MNT_FLAG(MNT_FORCE); 3627 MNT_FLAG(MNT_SNAPSHOT); 3628 MNT_FLAG(MNT_BYFSID); 3629 #undef MNT_FLAG 3630 if (mflags != 0) { 3631 if (buf[0] != '\0') 3632 strlcat(buf, ", ", sizeof(buf)); 3633 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 3634 "0x%016jx", mflags); 3635 } 3636 db_printf(" mnt_flag = %s\n", buf); 3637 3638 buf[0] = '\0'; 3639 flags = mp->mnt_kern_flag; 3640 #define MNT_KERN_FLAG(flag) do { \ 3641 if (flags & (flag)) { \ 3642 if (buf[0] != '\0') \ 3643 strlcat(buf, ", ", sizeof(buf)); \ 3644 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 3645 flags &= ~(flag); \ 3646 } \ 3647 } while (0) 3648 MNT_KERN_FLAG(MNTK_UNMOUNTF); 3649 MNT_KERN_FLAG(MNTK_ASYNC); 3650 MNT_KERN_FLAG(MNTK_SOFTDEP); 3651 MNT_KERN_FLAG(MNTK_NOINSMNTQ); 3652 MNT_KERN_FLAG(MNTK_DRAINING); 3653 MNT_KERN_FLAG(MNTK_REFEXPIRE); 3654 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 3655 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 3656 MNT_KERN_FLAG(MNTK_NO_IOPF); 3657 MNT_KERN_FLAG(MNTK_VGONE_UPPER); 3658 MNT_KERN_FLAG(MNTK_VGONE_WAITER); 3659 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT); 3660 MNT_KERN_FLAG(MNTK_MARKER); 3661 MNT_KERN_FLAG(MNTK_USES_BCACHE); 3662 MNT_KERN_FLAG(MNTK_NOASYNC); 3663 MNT_KERN_FLAG(MNTK_UNMOUNT); 3664 MNT_KERN_FLAG(MNTK_MWAIT); 3665 MNT_KERN_FLAG(MNTK_SUSPEND); 3666 MNT_KERN_FLAG(MNTK_SUSPEND2); 3667 MNT_KERN_FLAG(MNTK_SUSPENDED); 3668 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 3669 MNT_KERN_FLAG(MNTK_NOKNOTE); 3670 #undef MNT_KERN_FLAG 3671 if (flags != 0) { 3672 if (buf[0] != '\0') 3673 strlcat(buf, ", ", sizeof(buf)); 3674 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 3675 "0x%08x", flags); 3676 } 3677 db_printf(" mnt_kern_flag = %s\n", buf); 3678 3679 db_printf(" mnt_opt = "); 3680 opt = TAILQ_FIRST(mp->mnt_opt); 3681 if (opt != NULL) { 3682 db_printf("%s", opt->name); 3683 opt = TAILQ_NEXT(opt, link); 3684 while (opt != NULL) { 3685 db_printf(", %s", opt->name); 3686 opt = TAILQ_NEXT(opt, link); 3687 } 3688 } 3689 db_printf("\n"); 3690 3691 sp = &mp->mnt_stat; 3692 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 3693 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 3694 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 3695 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 3696 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 3697 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 3698 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 3699 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 3700 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 3701 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 3702 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 3703 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 3704 3705 db_printf(" mnt_cred = { uid=%u ruid=%u", 3706 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 3707 if (jailed(mp->mnt_cred)) 3708 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 3709 db_printf(" }\n"); 3710 db_printf(" mnt_ref = %d\n", mp->mnt_ref); 3711 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 3712 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 3713 db_printf(" mnt_activevnodelistsize = %d\n", 3714 mp->mnt_activevnodelistsize); 3715 db_printf(" mnt_writeopcount = %d\n", mp->mnt_writeopcount); 3716 db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen); 3717 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 3718 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 3719 db_printf(" mnt_lockref = %d\n", mp->mnt_lockref); 3720 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 3721 db_printf(" mnt_secondary_accwrites = %d\n", 3722 mp->mnt_secondary_accwrites); 3723 db_printf(" mnt_gjprovider = %s\n", 3724 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 3725 3726 db_printf("\n\nList of active vnodes\n"); 3727 TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) { 3728 if (vp->v_type != VMARKER) { 3729 vn_printf(vp, "vnode "); 3730 if (db_pager_quit) 3731 break; 3732 } 3733 } 3734 db_printf("\n\nList of inactive vnodes\n"); 3735 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3736 if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) { 3737 vn_printf(vp, "vnode "); 3738 if (db_pager_quit) 3739 break; 3740 } 3741 } 3742 } 3743 #endif /* DDB */ 3744 3745 /* 3746 * Fill in a struct xvfsconf based on a struct vfsconf. 3747 */ 3748 static int 3749 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 3750 { 3751 struct xvfsconf xvfsp; 3752 3753 bzero(&xvfsp, sizeof(xvfsp)); 3754 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 3755 xvfsp.vfc_typenum = vfsp->vfc_typenum; 3756 xvfsp.vfc_refcount = vfsp->vfc_refcount; 3757 xvfsp.vfc_flags = vfsp->vfc_flags; 3758 /* 3759 * These are unused in userland, we keep them 3760 * to not break binary compatibility. 3761 */ 3762 xvfsp.vfc_vfsops = NULL; 3763 xvfsp.vfc_next = NULL; 3764 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 3765 } 3766 3767 #ifdef COMPAT_FREEBSD32 3768 struct xvfsconf32 { 3769 uint32_t vfc_vfsops; 3770 char vfc_name[MFSNAMELEN]; 3771 int32_t vfc_typenum; 3772 int32_t vfc_refcount; 3773 int32_t vfc_flags; 3774 uint32_t vfc_next; 3775 }; 3776 3777 static int 3778 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 3779 { 3780 struct xvfsconf32 xvfsp; 3781 3782 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 3783 xvfsp.vfc_typenum = vfsp->vfc_typenum; 3784 xvfsp.vfc_refcount = vfsp->vfc_refcount; 3785 xvfsp.vfc_flags = vfsp->vfc_flags; 3786 xvfsp.vfc_vfsops = 0; 3787 xvfsp.vfc_next = 0; 3788 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 3789 } 3790 #endif 3791 3792 /* 3793 * Top level filesystem related information gathering. 3794 */ 3795 static int 3796 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 3797 { 3798 struct vfsconf *vfsp; 3799 int error; 3800 3801 error = 0; 3802 vfsconf_slock(); 3803 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 3804 #ifdef COMPAT_FREEBSD32 3805 if (req->flags & SCTL_MASK32) 3806 error = vfsconf2x32(req, vfsp); 3807 else 3808 #endif 3809 error = vfsconf2x(req, vfsp); 3810 if (error) 3811 break; 3812 } 3813 vfsconf_sunlock(); 3814 return (error); 3815 } 3816 3817 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | 3818 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, 3819 "S,xvfsconf", "List of all configured filesystems"); 3820 3821 #ifndef BURN_BRIDGES 3822 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 3823 3824 static int 3825 vfs_sysctl(SYSCTL_HANDLER_ARGS) 3826 { 3827 int *name = (int *)arg1 - 1; /* XXX */ 3828 u_int namelen = arg2 + 1; /* XXX */ 3829 struct vfsconf *vfsp; 3830 3831 log(LOG_WARNING, "userland calling deprecated sysctl, " 3832 "please rebuild world\n"); 3833 3834 #if 1 || defined(COMPAT_PRELITE2) 3835 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 3836 if (namelen == 1) 3837 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 3838 #endif 3839 3840 switch (name[1]) { 3841 case VFS_MAXTYPENUM: 3842 if (namelen != 2) 3843 return (ENOTDIR); 3844 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 3845 case VFS_CONF: 3846 if (namelen != 3) 3847 return (ENOTDIR); /* overloaded */ 3848 vfsconf_slock(); 3849 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 3850 if (vfsp->vfc_typenum == name[2]) 3851 break; 3852 } 3853 vfsconf_sunlock(); 3854 if (vfsp == NULL) 3855 return (EOPNOTSUPP); 3856 #ifdef COMPAT_FREEBSD32 3857 if (req->flags & SCTL_MASK32) 3858 return (vfsconf2x32(req, vfsp)); 3859 else 3860 #endif 3861 return (vfsconf2x(req, vfsp)); 3862 } 3863 return (EOPNOTSUPP); 3864 } 3865 3866 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | 3867 CTLFLAG_MPSAFE, vfs_sysctl, 3868 "Generic filesystem"); 3869 3870 #if 1 || defined(COMPAT_PRELITE2) 3871 3872 static int 3873 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 3874 { 3875 int error; 3876 struct vfsconf *vfsp; 3877 struct ovfsconf ovfs; 3878 3879 vfsconf_slock(); 3880 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 3881 bzero(&ovfs, sizeof(ovfs)); 3882 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 3883 strcpy(ovfs.vfc_name, vfsp->vfc_name); 3884 ovfs.vfc_index = vfsp->vfc_typenum; 3885 ovfs.vfc_refcount = vfsp->vfc_refcount; 3886 ovfs.vfc_flags = vfsp->vfc_flags; 3887 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 3888 if (error != 0) { 3889 vfsconf_sunlock(); 3890 return (error); 3891 } 3892 } 3893 vfsconf_sunlock(); 3894 return (0); 3895 } 3896 3897 #endif /* 1 || COMPAT_PRELITE2 */ 3898 #endif /* !BURN_BRIDGES */ 3899 3900 #define KINFO_VNODESLOP 10 3901 #ifdef notyet 3902 /* 3903 * Dump vnode list (via sysctl). 3904 */ 3905 /* ARGSUSED */ 3906 static int 3907 sysctl_vnode(SYSCTL_HANDLER_ARGS) 3908 { 3909 struct xvnode *xvn; 3910 struct mount *mp; 3911 struct vnode *vp; 3912 int error, len, n; 3913 3914 /* 3915 * Stale numvnodes access is not fatal here. 3916 */ 3917 req->lock = 0; 3918 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 3919 if (!req->oldptr) 3920 /* Make an estimate */ 3921 return (SYSCTL_OUT(req, 0, len)); 3922 3923 error = sysctl_wire_old_buffer(req, 0); 3924 if (error != 0) 3925 return (error); 3926 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 3927 n = 0; 3928 mtx_lock(&mountlist_mtx); 3929 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3930 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) 3931 continue; 3932 MNT_ILOCK(mp); 3933 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3934 if (n == len) 3935 break; 3936 vref(vp); 3937 xvn[n].xv_size = sizeof *xvn; 3938 xvn[n].xv_vnode = vp; 3939 xvn[n].xv_id = 0; /* XXX compat */ 3940 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 3941 XV_COPY(usecount); 3942 XV_COPY(writecount); 3943 XV_COPY(holdcnt); 3944 XV_COPY(mount); 3945 XV_COPY(numoutput); 3946 XV_COPY(type); 3947 #undef XV_COPY 3948 xvn[n].xv_flag = vp->v_vflag; 3949 3950 switch (vp->v_type) { 3951 case VREG: 3952 case VDIR: 3953 case VLNK: 3954 break; 3955 case VBLK: 3956 case VCHR: 3957 if (vp->v_rdev == NULL) { 3958 vrele(vp); 3959 continue; 3960 } 3961 xvn[n].xv_dev = dev2udev(vp->v_rdev); 3962 break; 3963 case VSOCK: 3964 xvn[n].xv_socket = vp->v_socket; 3965 break; 3966 case VFIFO: 3967 xvn[n].xv_fifo = vp->v_fifoinfo; 3968 break; 3969 case VNON: 3970 case VBAD: 3971 default: 3972 /* shouldn't happen? */ 3973 vrele(vp); 3974 continue; 3975 } 3976 vrele(vp); 3977 ++n; 3978 } 3979 MNT_IUNLOCK(mp); 3980 mtx_lock(&mountlist_mtx); 3981 vfs_unbusy(mp); 3982 if (n == len) 3983 break; 3984 } 3985 mtx_unlock(&mountlist_mtx); 3986 3987 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 3988 free(xvn, M_TEMP); 3989 return (error); 3990 } 3991 3992 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD | 3993 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode", 3994 ""); 3995 #endif 3996 3997 static void 3998 unmount_or_warn(struct mount *mp) 3999 { 4000 int error; 4001 4002 error = dounmount(mp, MNT_FORCE, curthread); 4003 if (error != 0) { 4004 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); 4005 if (error == EBUSY) 4006 printf("BUSY)\n"); 4007 else 4008 printf("%d)\n", error); 4009 } 4010 } 4011 4012 /* 4013 * Unmount all filesystems. The list is traversed in reverse order 4014 * of mounting to avoid dependencies. 4015 */ 4016 void 4017 vfs_unmountall(void) 4018 { 4019 struct mount *mp, *tmp; 4020 4021 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 4022 4023 /* 4024 * Since this only runs when rebooting, it is not interlocked. 4025 */ 4026 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { 4027 vfs_ref(mp); 4028 4029 /* 4030 * Forcibly unmounting "/dev" before "/" would prevent clean 4031 * unmount of the latter. 4032 */ 4033 if (mp == rootdevmp) 4034 continue; 4035 4036 unmount_or_warn(mp); 4037 } 4038 4039 if (rootdevmp != NULL) 4040 unmount_or_warn(rootdevmp); 4041 } 4042 4043 /* 4044 * perform msync on all vnodes under a mount point 4045 * the mount point must be locked. 4046 */ 4047 void 4048 vfs_msync(struct mount *mp, int flags) 4049 { 4050 struct vnode *vp, *mvp; 4051 struct vm_object *obj; 4052 4053 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 4054 4055 vnlru_return_batch(mp); 4056 4057 MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) { 4058 obj = vp->v_object; 4059 if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 && 4060 (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) { 4061 if (!vget(vp, 4062 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 4063 curthread)) { 4064 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 4065 vput(vp); 4066 continue; 4067 } 4068 4069 obj = vp->v_object; 4070 if (obj != NULL) { 4071 VM_OBJECT_WLOCK(obj); 4072 vm_object_page_clean(obj, 0, 0, 4073 flags == MNT_WAIT ? 4074 OBJPC_SYNC : OBJPC_NOSYNC); 4075 VM_OBJECT_WUNLOCK(obj); 4076 } 4077 vput(vp); 4078 } 4079 } else 4080 VI_UNLOCK(vp); 4081 } 4082 } 4083 4084 static void 4085 destroy_vpollinfo_free(struct vpollinfo *vi) 4086 { 4087 4088 knlist_destroy(&vi->vpi_selinfo.si_note); 4089 mtx_destroy(&vi->vpi_lock); 4090 uma_zfree(vnodepoll_zone, vi); 4091 } 4092 4093 static void 4094 destroy_vpollinfo(struct vpollinfo *vi) 4095 { 4096 4097 knlist_clear(&vi->vpi_selinfo.si_note, 1); 4098 seldrain(&vi->vpi_selinfo); 4099 destroy_vpollinfo_free(vi); 4100 } 4101 4102 /* 4103 * Initialize per-vnode helper structure to hold poll-related state. 4104 */ 4105 void 4106 v_addpollinfo(struct vnode *vp) 4107 { 4108 struct vpollinfo *vi; 4109 4110 if (vp->v_pollinfo != NULL) 4111 return; 4112 vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO); 4113 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 4114 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 4115 vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked); 4116 VI_LOCK(vp); 4117 if (vp->v_pollinfo != NULL) { 4118 VI_UNLOCK(vp); 4119 destroy_vpollinfo_free(vi); 4120 return; 4121 } 4122 vp->v_pollinfo = vi; 4123 VI_UNLOCK(vp); 4124 } 4125 4126 /* 4127 * Record a process's interest in events which might happen to 4128 * a vnode. Because poll uses the historic select-style interface 4129 * internally, this routine serves as both the ``check for any 4130 * pending events'' and the ``record my interest in future events'' 4131 * functions. (These are done together, while the lock is held, 4132 * to avoid race conditions.) 4133 */ 4134 int 4135 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 4136 { 4137 4138 v_addpollinfo(vp); 4139 mtx_lock(&vp->v_pollinfo->vpi_lock); 4140 if (vp->v_pollinfo->vpi_revents & events) { 4141 /* 4142 * This leaves events we are not interested 4143 * in available for the other process which 4144 * which presumably had requested them 4145 * (otherwise they would never have been 4146 * recorded). 4147 */ 4148 events &= vp->v_pollinfo->vpi_revents; 4149 vp->v_pollinfo->vpi_revents &= ~events; 4150 4151 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4152 return (events); 4153 } 4154 vp->v_pollinfo->vpi_events |= events; 4155 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 4156 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4157 return (0); 4158 } 4159 4160 /* 4161 * Routine to create and manage a filesystem syncer vnode. 4162 */ 4163 #define sync_close ((int (*)(struct vop_close_args *))nullop) 4164 static int sync_fsync(struct vop_fsync_args *); 4165 static int sync_inactive(struct vop_inactive_args *); 4166 static int sync_reclaim(struct vop_reclaim_args *); 4167 4168 static struct vop_vector sync_vnodeops = { 4169 .vop_bypass = VOP_EOPNOTSUPP, 4170 .vop_close = sync_close, /* close */ 4171 .vop_fsync = sync_fsync, /* fsync */ 4172 .vop_inactive = sync_inactive, /* inactive */ 4173 .vop_reclaim = sync_reclaim, /* reclaim */ 4174 .vop_lock1 = vop_stdlock, /* lock */ 4175 .vop_unlock = vop_stdunlock, /* unlock */ 4176 .vop_islocked = vop_stdislocked, /* islocked */ 4177 }; 4178 4179 /* 4180 * Create a new filesystem syncer vnode for the specified mount point. 4181 */ 4182 void 4183 vfs_allocate_syncvnode(struct mount *mp) 4184 { 4185 struct vnode *vp; 4186 struct bufobj *bo; 4187 static long start, incr, next; 4188 int error; 4189 4190 /* Allocate a new vnode */ 4191 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 4192 if (error != 0) 4193 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 4194 vp->v_type = VNON; 4195 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4196 vp->v_vflag |= VV_FORCEINSMQ; 4197 error = insmntque(vp, mp); 4198 if (error != 0) 4199 panic("vfs_allocate_syncvnode: insmntque() failed"); 4200 vp->v_vflag &= ~VV_FORCEINSMQ; 4201 VOP_UNLOCK(vp, 0); 4202 /* 4203 * Place the vnode onto the syncer worklist. We attempt to 4204 * scatter them about on the list so that they will go off 4205 * at evenly distributed times even if all the filesystems 4206 * are mounted at once. 4207 */ 4208 next += incr; 4209 if (next == 0 || next > syncer_maxdelay) { 4210 start /= 2; 4211 incr /= 2; 4212 if (start == 0) { 4213 start = syncer_maxdelay / 2; 4214 incr = syncer_maxdelay; 4215 } 4216 next = start; 4217 } 4218 bo = &vp->v_bufobj; 4219 BO_LOCK(bo); 4220 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 4221 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 4222 mtx_lock(&sync_mtx); 4223 sync_vnode_count++; 4224 if (mp->mnt_syncer == NULL) { 4225 mp->mnt_syncer = vp; 4226 vp = NULL; 4227 } 4228 mtx_unlock(&sync_mtx); 4229 BO_UNLOCK(bo); 4230 if (vp != NULL) { 4231 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4232 vgone(vp); 4233 vput(vp); 4234 } 4235 } 4236 4237 void 4238 vfs_deallocate_syncvnode(struct mount *mp) 4239 { 4240 struct vnode *vp; 4241 4242 mtx_lock(&sync_mtx); 4243 vp = mp->mnt_syncer; 4244 if (vp != NULL) 4245 mp->mnt_syncer = NULL; 4246 mtx_unlock(&sync_mtx); 4247 if (vp != NULL) 4248 vrele(vp); 4249 } 4250 4251 /* 4252 * Do a lazy sync of the filesystem. 4253 */ 4254 static int 4255 sync_fsync(struct vop_fsync_args *ap) 4256 { 4257 struct vnode *syncvp = ap->a_vp; 4258 struct mount *mp = syncvp->v_mount; 4259 int error, save; 4260 struct bufobj *bo; 4261 4262 /* 4263 * We only need to do something if this is a lazy evaluation. 4264 */ 4265 if (ap->a_waitfor != MNT_LAZY) 4266 return (0); 4267 4268 /* 4269 * Move ourselves to the back of the sync list. 4270 */ 4271 bo = &syncvp->v_bufobj; 4272 BO_LOCK(bo); 4273 vn_syncer_add_to_worklist(bo, syncdelay); 4274 BO_UNLOCK(bo); 4275 4276 /* 4277 * Walk the list of vnodes pushing all that are dirty and 4278 * not already on the sync list. 4279 */ 4280 if (vfs_busy(mp, MBF_NOWAIT) != 0) 4281 return (0); 4282 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 4283 vfs_unbusy(mp); 4284 return (0); 4285 } 4286 save = curthread_pflags_set(TDP_SYNCIO); 4287 vfs_msync(mp, MNT_NOWAIT); 4288 error = VFS_SYNC(mp, MNT_LAZY); 4289 curthread_pflags_restore(save); 4290 vn_finished_write(mp); 4291 vfs_unbusy(mp); 4292 return (error); 4293 } 4294 4295 /* 4296 * The syncer vnode is no referenced. 4297 */ 4298 static int 4299 sync_inactive(struct vop_inactive_args *ap) 4300 { 4301 4302 vgone(ap->a_vp); 4303 return (0); 4304 } 4305 4306 /* 4307 * The syncer vnode is no longer needed and is being decommissioned. 4308 * 4309 * Modifications to the worklist must be protected by sync_mtx. 4310 */ 4311 static int 4312 sync_reclaim(struct vop_reclaim_args *ap) 4313 { 4314 struct vnode *vp = ap->a_vp; 4315 struct bufobj *bo; 4316 4317 bo = &vp->v_bufobj; 4318 BO_LOCK(bo); 4319 mtx_lock(&sync_mtx); 4320 if (vp->v_mount->mnt_syncer == vp) 4321 vp->v_mount->mnt_syncer = NULL; 4322 if (bo->bo_flag & BO_ONWORKLST) { 4323 LIST_REMOVE(bo, bo_synclist); 4324 syncer_worklist_len--; 4325 sync_vnode_count--; 4326 bo->bo_flag &= ~BO_ONWORKLST; 4327 } 4328 mtx_unlock(&sync_mtx); 4329 BO_UNLOCK(bo); 4330 4331 return (0); 4332 } 4333 4334 /* 4335 * Check if vnode represents a disk device 4336 */ 4337 int 4338 vn_isdisk(struct vnode *vp, int *errp) 4339 { 4340 int error; 4341 4342 if (vp->v_type != VCHR) { 4343 error = ENOTBLK; 4344 goto out; 4345 } 4346 error = 0; 4347 dev_lock(); 4348 if (vp->v_rdev == NULL) 4349 error = ENXIO; 4350 else if (vp->v_rdev->si_devsw == NULL) 4351 error = ENXIO; 4352 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 4353 error = ENOTBLK; 4354 dev_unlock(); 4355 out: 4356 if (errp != NULL) 4357 *errp = error; 4358 return (error == 0); 4359 } 4360 4361 /* 4362 * Common filesystem object access control check routine. Accepts a 4363 * vnode's type, "mode", uid and gid, requested access mode, credentials, 4364 * and optional call-by-reference privused argument allowing vaccess() 4365 * to indicate to the caller whether privilege was used to satisfy the 4366 * request (obsoleted). Returns 0 on success, or an errno on failure. 4367 */ 4368 int 4369 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 4370 accmode_t accmode, struct ucred *cred, int *privused) 4371 { 4372 accmode_t dac_granted; 4373 accmode_t priv_granted; 4374 4375 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 4376 ("invalid bit in accmode")); 4377 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 4378 ("VAPPEND without VWRITE")); 4379 4380 /* 4381 * Look for a normal, non-privileged way to access the file/directory 4382 * as requested. If it exists, go with that. 4383 */ 4384 4385 if (privused != NULL) 4386 *privused = 0; 4387 4388 dac_granted = 0; 4389 4390 /* Check the owner. */ 4391 if (cred->cr_uid == file_uid) { 4392 dac_granted |= VADMIN; 4393 if (file_mode & S_IXUSR) 4394 dac_granted |= VEXEC; 4395 if (file_mode & S_IRUSR) 4396 dac_granted |= VREAD; 4397 if (file_mode & S_IWUSR) 4398 dac_granted |= (VWRITE | VAPPEND); 4399 4400 if ((accmode & dac_granted) == accmode) 4401 return (0); 4402 4403 goto privcheck; 4404 } 4405 4406 /* Otherwise, check the groups (first match) */ 4407 if (groupmember(file_gid, cred)) { 4408 if (file_mode & S_IXGRP) 4409 dac_granted |= VEXEC; 4410 if (file_mode & S_IRGRP) 4411 dac_granted |= VREAD; 4412 if (file_mode & S_IWGRP) 4413 dac_granted |= (VWRITE | VAPPEND); 4414 4415 if ((accmode & dac_granted) == accmode) 4416 return (0); 4417 4418 goto privcheck; 4419 } 4420 4421 /* Otherwise, check everyone else. */ 4422 if (file_mode & S_IXOTH) 4423 dac_granted |= VEXEC; 4424 if (file_mode & S_IROTH) 4425 dac_granted |= VREAD; 4426 if (file_mode & S_IWOTH) 4427 dac_granted |= (VWRITE | VAPPEND); 4428 if ((accmode & dac_granted) == accmode) 4429 return (0); 4430 4431 privcheck: 4432 /* 4433 * Build a privilege mask to determine if the set of privileges 4434 * satisfies the requirements when combined with the granted mask 4435 * from above. For each privilege, if the privilege is required, 4436 * bitwise or the request type onto the priv_granted mask. 4437 */ 4438 priv_granted = 0; 4439 4440 if (type == VDIR) { 4441 /* 4442 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 4443 * requests, instead of PRIV_VFS_EXEC. 4444 */ 4445 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 4446 !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0)) 4447 priv_granted |= VEXEC; 4448 } else { 4449 /* 4450 * Ensure that at least one execute bit is on. Otherwise, 4451 * a privileged user will always succeed, and we don't want 4452 * this to happen unless the file really is executable. 4453 */ 4454 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 4455 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 4456 !priv_check_cred(cred, PRIV_VFS_EXEC, 0)) 4457 priv_granted |= VEXEC; 4458 } 4459 4460 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 4461 !priv_check_cred(cred, PRIV_VFS_READ, 0)) 4462 priv_granted |= VREAD; 4463 4464 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 4465 !priv_check_cred(cred, PRIV_VFS_WRITE, 0)) 4466 priv_granted |= (VWRITE | VAPPEND); 4467 4468 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 4469 !priv_check_cred(cred, PRIV_VFS_ADMIN, 0)) 4470 priv_granted |= VADMIN; 4471 4472 if ((accmode & (priv_granted | dac_granted)) == accmode) { 4473 /* XXX audit: privilege used */ 4474 if (privused != NULL) 4475 *privused = 1; 4476 return (0); 4477 } 4478 4479 return ((accmode & VADMIN) ? EPERM : EACCES); 4480 } 4481 4482 /* 4483 * Credential check based on process requesting service, and per-attribute 4484 * permissions. 4485 */ 4486 int 4487 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 4488 struct thread *td, accmode_t accmode) 4489 { 4490 4491 /* 4492 * Kernel-invoked always succeeds. 4493 */ 4494 if (cred == NOCRED) 4495 return (0); 4496 4497 /* 4498 * Do not allow privileged processes in jail to directly manipulate 4499 * system attributes. 4500 */ 4501 switch (attrnamespace) { 4502 case EXTATTR_NAMESPACE_SYSTEM: 4503 /* Potentially should be: return (EPERM); */ 4504 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0)); 4505 case EXTATTR_NAMESPACE_USER: 4506 return (VOP_ACCESS(vp, accmode, cred, td)); 4507 default: 4508 return (EPERM); 4509 } 4510 } 4511 4512 #ifdef DEBUG_VFS_LOCKS 4513 /* 4514 * This only exists to suppress warnings from unlocked specfs accesses. It is 4515 * no longer ok to have an unlocked VFS. 4516 */ 4517 #define IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL || \ 4518 (vp)->v_type == VCHR || (vp)->v_type == VBAD) 4519 4520 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 4521 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, 4522 "Drop into debugger on lock violation"); 4523 4524 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 4525 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 4526 0, "Check for interlock across VOPs"); 4527 4528 int vfs_badlock_print = 1; /* Print lock violations. */ 4529 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 4530 0, "Print lock violations"); 4531 4532 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */ 4533 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode, 4534 0, "Print vnode details on lock violations"); 4535 4536 #ifdef KDB 4537 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 4538 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, 4539 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations"); 4540 #endif 4541 4542 static void 4543 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 4544 { 4545 4546 #ifdef KDB 4547 if (vfs_badlock_backtrace) 4548 kdb_backtrace(); 4549 #endif 4550 if (vfs_badlock_vnode) 4551 vn_printf(vp, "vnode "); 4552 if (vfs_badlock_print) 4553 printf("%s: %p %s\n", str, (void *)vp, msg); 4554 if (vfs_badlock_ddb) 4555 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 4556 } 4557 4558 void 4559 assert_vi_locked(struct vnode *vp, const char *str) 4560 { 4561 4562 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 4563 vfs_badlock("interlock is not locked but should be", str, vp); 4564 } 4565 4566 void 4567 assert_vi_unlocked(struct vnode *vp, const char *str) 4568 { 4569 4570 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 4571 vfs_badlock("interlock is locked but should not be", str, vp); 4572 } 4573 4574 void 4575 assert_vop_locked(struct vnode *vp, const char *str) 4576 { 4577 int locked; 4578 4579 if (!IGNORE_LOCK(vp)) { 4580 locked = VOP_ISLOCKED(vp); 4581 if (locked == 0 || locked == LK_EXCLOTHER) 4582 vfs_badlock("is not locked but should be", str, vp); 4583 } 4584 } 4585 4586 void 4587 assert_vop_unlocked(struct vnode *vp, const char *str) 4588 { 4589 4590 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE) 4591 vfs_badlock("is locked but should not be", str, vp); 4592 } 4593 4594 void 4595 assert_vop_elocked(struct vnode *vp, const char *str) 4596 { 4597 4598 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 4599 vfs_badlock("is not exclusive locked but should be", str, vp); 4600 } 4601 #endif /* DEBUG_VFS_LOCKS */ 4602 4603 void 4604 vop_rename_fail(struct vop_rename_args *ap) 4605 { 4606 4607 if (ap->a_tvp != NULL) 4608 vput(ap->a_tvp); 4609 if (ap->a_tdvp == ap->a_tvp) 4610 vrele(ap->a_tdvp); 4611 else 4612 vput(ap->a_tdvp); 4613 vrele(ap->a_fdvp); 4614 vrele(ap->a_fvp); 4615 } 4616 4617 void 4618 vop_rename_pre(void *ap) 4619 { 4620 struct vop_rename_args *a = ap; 4621 4622 #ifdef DEBUG_VFS_LOCKS 4623 if (a->a_tvp) 4624 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 4625 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 4626 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 4627 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 4628 4629 /* Check the source (from). */ 4630 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 4631 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 4632 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 4633 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 4634 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 4635 4636 /* Check the target. */ 4637 if (a->a_tvp) 4638 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 4639 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 4640 #endif 4641 if (a->a_tdvp != a->a_fdvp) 4642 vhold(a->a_fdvp); 4643 if (a->a_tvp != a->a_fvp) 4644 vhold(a->a_fvp); 4645 vhold(a->a_tdvp); 4646 if (a->a_tvp) 4647 vhold(a->a_tvp); 4648 } 4649 4650 #ifdef DEBUG_VFS_LOCKS 4651 void 4652 vop_strategy_pre(void *ap) 4653 { 4654 struct vop_strategy_args *a; 4655 struct buf *bp; 4656 4657 a = ap; 4658 bp = a->a_bp; 4659 4660 /* 4661 * Cluster ops lock their component buffers but not the IO container. 4662 */ 4663 if ((bp->b_flags & B_CLUSTER) != 0) 4664 return; 4665 4666 if (panicstr == NULL && !BUF_ISLOCKED(bp)) { 4667 if (vfs_badlock_print) 4668 printf( 4669 "VOP_STRATEGY: bp is not locked but should be\n"); 4670 if (vfs_badlock_ddb) 4671 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 4672 } 4673 } 4674 4675 void 4676 vop_lock_pre(void *ap) 4677 { 4678 struct vop_lock1_args *a = ap; 4679 4680 if ((a->a_flags & LK_INTERLOCK) == 0) 4681 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 4682 else 4683 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 4684 } 4685 4686 void 4687 vop_lock_post(void *ap, int rc) 4688 { 4689 struct vop_lock1_args *a = ap; 4690 4691 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 4692 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) 4693 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 4694 } 4695 4696 void 4697 vop_unlock_pre(void *ap) 4698 { 4699 struct vop_unlock_args *a = ap; 4700 4701 if (a->a_flags & LK_INTERLOCK) 4702 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 4703 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 4704 } 4705 4706 void 4707 vop_unlock_post(void *ap, int rc) 4708 { 4709 struct vop_unlock_args *a = ap; 4710 4711 if (a->a_flags & LK_INTERLOCK) 4712 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 4713 } 4714 #endif 4715 4716 void 4717 vop_create_post(void *ap, int rc) 4718 { 4719 struct vop_create_args *a = ap; 4720 4721 if (!rc) 4722 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4723 } 4724 4725 void 4726 vop_deleteextattr_post(void *ap, int rc) 4727 { 4728 struct vop_deleteextattr_args *a = ap; 4729 4730 if (!rc) 4731 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 4732 } 4733 4734 void 4735 vop_link_post(void *ap, int rc) 4736 { 4737 struct vop_link_args *a = ap; 4738 4739 if (!rc) { 4740 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK); 4741 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE); 4742 } 4743 } 4744 4745 void 4746 vop_mkdir_post(void *ap, int rc) 4747 { 4748 struct vop_mkdir_args *a = ap; 4749 4750 if (!rc) 4751 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 4752 } 4753 4754 void 4755 vop_mknod_post(void *ap, int rc) 4756 { 4757 struct vop_mknod_args *a = ap; 4758 4759 if (!rc) 4760 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4761 } 4762 4763 void 4764 vop_reclaim_post(void *ap, int rc) 4765 { 4766 struct vop_reclaim_args *a = ap; 4767 4768 if (!rc) 4769 VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE); 4770 } 4771 4772 void 4773 vop_remove_post(void *ap, int rc) 4774 { 4775 struct vop_remove_args *a = ap; 4776 4777 if (!rc) { 4778 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4779 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 4780 } 4781 } 4782 4783 void 4784 vop_rename_post(void *ap, int rc) 4785 { 4786 struct vop_rename_args *a = ap; 4787 long hint; 4788 4789 if (!rc) { 4790 hint = NOTE_WRITE; 4791 if (a->a_fdvp == a->a_tdvp) { 4792 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) 4793 hint |= NOTE_LINK; 4794 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 4795 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 4796 } else { 4797 hint |= NOTE_EXTEND; 4798 if (a->a_fvp->v_type == VDIR) 4799 hint |= NOTE_LINK; 4800 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 4801 4802 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && 4803 a->a_tvp->v_type == VDIR) 4804 hint &= ~NOTE_LINK; 4805 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 4806 } 4807 4808 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 4809 if (a->a_tvp) 4810 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 4811 } 4812 if (a->a_tdvp != a->a_fdvp) 4813 vdrop(a->a_fdvp); 4814 if (a->a_tvp != a->a_fvp) 4815 vdrop(a->a_fvp); 4816 vdrop(a->a_tdvp); 4817 if (a->a_tvp) 4818 vdrop(a->a_tvp); 4819 } 4820 4821 void 4822 vop_rmdir_post(void *ap, int rc) 4823 { 4824 struct vop_rmdir_args *a = ap; 4825 4826 if (!rc) { 4827 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 4828 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 4829 } 4830 } 4831 4832 void 4833 vop_setattr_post(void *ap, int rc) 4834 { 4835 struct vop_setattr_args *a = ap; 4836 4837 if (!rc) 4838 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 4839 } 4840 4841 void 4842 vop_setextattr_post(void *ap, int rc) 4843 { 4844 struct vop_setextattr_args *a = ap; 4845 4846 if (!rc) 4847 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 4848 } 4849 4850 void 4851 vop_symlink_post(void *ap, int rc) 4852 { 4853 struct vop_symlink_args *a = ap; 4854 4855 if (!rc) 4856 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4857 } 4858 4859 void 4860 vop_open_post(void *ap, int rc) 4861 { 4862 struct vop_open_args *a = ap; 4863 4864 if (!rc) 4865 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); 4866 } 4867 4868 void 4869 vop_close_post(void *ap, int rc) 4870 { 4871 struct vop_close_args *a = ap; 4872 4873 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ 4874 (a->a_vp->v_iflag & VI_DOOMED) == 0)) { 4875 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 4876 NOTE_CLOSE_WRITE : NOTE_CLOSE); 4877 } 4878 } 4879 4880 void 4881 vop_read_post(void *ap, int rc) 4882 { 4883 struct vop_read_args *a = ap; 4884 4885 if (!rc) 4886 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 4887 } 4888 4889 void 4890 vop_readdir_post(void *ap, int rc) 4891 { 4892 struct vop_readdir_args *a = ap; 4893 4894 if (!rc) 4895 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 4896 } 4897 4898 static struct knlist fs_knlist; 4899 4900 static void 4901 vfs_event_init(void *arg) 4902 { 4903 knlist_init_mtx(&fs_knlist, NULL); 4904 } 4905 /* XXX - correct order? */ 4906 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 4907 4908 void 4909 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 4910 { 4911 4912 KNOTE_UNLOCKED(&fs_knlist, event); 4913 } 4914 4915 static int filt_fsattach(struct knote *kn); 4916 static void filt_fsdetach(struct knote *kn); 4917 static int filt_fsevent(struct knote *kn, long hint); 4918 4919 struct filterops fs_filtops = { 4920 .f_isfd = 0, 4921 .f_attach = filt_fsattach, 4922 .f_detach = filt_fsdetach, 4923 .f_event = filt_fsevent 4924 }; 4925 4926 static int 4927 filt_fsattach(struct knote *kn) 4928 { 4929 4930 kn->kn_flags |= EV_CLEAR; 4931 knlist_add(&fs_knlist, kn, 0); 4932 return (0); 4933 } 4934 4935 static void 4936 filt_fsdetach(struct knote *kn) 4937 { 4938 4939 knlist_remove(&fs_knlist, kn, 0); 4940 } 4941 4942 static int 4943 filt_fsevent(struct knote *kn, long hint) 4944 { 4945 4946 kn->kn_fflags |= hint; 4947 return (kn->kn_fflags != 0); 4948 } 4949 4950 static int 4951 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 4952 { 4953 struct vfsidctl vc; 4954 int error; 4955 struct mount *mp; 4956 4957 error = SYSCTL_IN(req, &vc, sizeof(vc)); 4958 if (error) 4959 return (error); 4960 if (vc.vc_vers != VFS_CTL_VERS1) 4961 return (EINVAL); 4962 mp = vfs_getvfs(&vc.vc_fsid); 4963 if (mp == NULL) 4964 return (ENOENT); 4965 /* ensure that a specific sysctl goes to the right filesystem. */ 4966 if (strcmp(vc.vc_fstypename, "*") != 0 && 4967 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 4968 vfs_rel(mp); 4969 return (EINVAL); 4970 } 4971 VCTLTOREQ(&vc, req); 4972 error = VFS_SYSCTL(mp, vc.vc_op, req); 4973 vfs_rel(mp); 4974 return (error); 4975 } 4976 4977 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR, 4978 NULL, 0, sysctl_vfs_ctl, "", 4979 "Sysctl by fsid"); 4980 4981 /* 4982 * Function to initialize a va_filerev field sensibly. 4983 * XXX: Wouldn't a random number make a lot more sense ?? 4984 */ 4985 u_quad_t 4986 init_va_filerev(void) 4987 { 4988 struct bintime bt; 4989 4990 getbinuptime(&bt); 4991 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 4992 } 4993 4994 static int filt_vfsread(struct knote *kn, long hint); 4995 static int filt_vfswrite(struct knote *kn, long hint); 4996 static int filt_vfsvnode(struct knote *kn, long hint); 4997 static void filt_vfsdetach(struct knote *kn); 4998 static struct filterops vfsread_filtops = { 4999 .f_isfd = 1, 5000 .f_detach = filt_vfsdetach, 5001 .f_event = filt_vfsread 5002 }; 5003 static struct filterops vfswrite_filtops = { 5004 .f_isfd = 1, 5005 .f_detach = filt_vfsdetach, 5006 .f_event = filt_vfswrite 5007 }; 5008 static struct filterops vfsvnode_filtops = { 5009 .f_isfd = 1, 5010 .f_detach = filt_vfsdetach, 5011 .f_event = filt_vfsvnode 5012 }; 5013 5014 static void 5015 vfs_knllock(void *arg) 5016 { 5017 struct vnode *vp = arg; 5018 5019 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5020 } 5021 5022 static void 5023 vfs_knlunlock(void *arg) 5024 { 5025 struct vnode *vp = arg; 5026 5027 VOP_UNLOCK(vp, 0); 5028 } 5029 5030 static void 5031 vfs_knl_assert_locked(void *arg) 5032 { 5033 #ifdef DEBUG_VFS_LOCKS 5034 struct vnode *vp = arg; 5035 5036 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 5037 #endif 5038 } 5039 5040 static void 5041 vfs_knl_assert_unlocked(void *arg) 5042 { 5043 #ifdef DEBUG_VFS_LOCKS 5044 struct vnode *vp = arg; 5045 5046 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 5047 #endif 5048 } 5049 5050 int 5051 vfs_kqfilter(struct vop_kqfilter_args *ap) 5052 { 5053 struct vnode *vp = ap->a_vp; 5054 struct knote *kn = ap->a_kn; 5055 struct knlist *knl; 5056 5057 switch (kn->kn_filter) { 5058 case EVFILT_READ: 5059 kn->kn_fop = &vfsread_filtops; 5060 break; 5061 case EVFILT_WRITE: 5062 kn->kn_fop = &vfswrite_filtops; 5063 break; 5064 case EVFILT_VNODE: 5065 kn->kn_fop = &vfsvnode_filtops; 5066 break; 5067 default: 5068 return (EINVAL); 5069 } 5070 5071 kn->kn_hook = (caddr_t)vp; 5072 5073 v_addpollinfo(vp); 5074 if (vp->v_pollinfo == NULL) 5075 return (ENOMEM); 5076 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 5077 vhold(vp); 5078 knlist_add(knl, kn, 0); 5079 5080 return (0); 5081 } 5082 5083 /* 5084 * Detach knote from vnode 5085 */ 5086 static void 5087 filt_vfsdetach(struct knote *kn) 5088 { 5089 struct vnode *vp = (struct vnode *)kn->kn_hook; 5090 5091 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 5092 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 5093 vdrop(vp); 5094 } 5095 5096 /*ARGSUSED*/ 5097 static int 5098 filt_vfsread(struct knote *kn, long hint) 5099 { 5100 struct vnode *vp = (struct vnode *)kn->kn_hook; 5101 struct vattr va; 5102 int res; 5103 5104 /* 5105 * filesystem is gone, so set the EOF flag and schedule 5106 * the knote for deletion. 5107 */ 5108 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 5109 VI_LOCK(vp); 5110 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 5111 VI_UNLOCK(vp); 5112 return (1); 5113 } 5114 5115 if (VOP_GETATTR(vp, &va, curthread->td_ucred)) 5116 return (0); 5117 5118 VI_LOCK(vp); 5119 kn->kn_data = va.va_size - kn->kn_fp->f_offset; 5120 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; 5121 VI_UNLOCK(vp); 5122 return (res); 5123 } 5124 5125 /*ARGSUSED*/ 5126 static int 5127 filt_vfswrite(struct knote *kn, long hint) 5128 { 5129 struct vnode *vp = (struct vnode *)kn->kn_hook; 5130 5131 VI_LOCK(vp); 5132 5133 /* 5134 * filesystem is gone, so set the EOF flag and schedule 5135 * the knote for deletion. 5136 */ 5137 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) 5138 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 5139 5140 kn->kn_data = 0; 5141 VI_UNLOCK(vp); 5142 return (1); 5143 } 5144 5145 static int 5146 filt_vfsvnode(struct knote *kn, long hint) 5147 { 5148 struct vnode *vp = (struct vnode *)kn->kn_hook; 5149 int res; 5150 5151 VI_LOCK(vp); 5152 if (kn->kn_sfflags & hint) 5153 kn->kn_fflags |= hint; 5154 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 5155 kn->kn_flags |= EV_EOF; 5156 VI_UNLOCK(vp); 5157 return (1); 5158 } 5159 res = (kn->kn_fflags != 0); 5160 VI_UNLOCK(vp); 5161 return (res); 5162 } 5163 5164 int 5165 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 5166 { 5167 int error; 5168 5169 if (dp->d_reclen > ap->a_uio->uio_resid) 5170 return (ENAMETOOLONG); 5171 error = uiomove(dp, dp->d_reclen, ap->a_uio); 5172 if (error) { 5173 if (ap->a_ncookies != NULL) { 5174 if (ap->a_cookies != NULL) 5175 free(ap->a_cookies, M_TEMP); 5176 ap->a_cookies = NULL; 5177 *ap->a_ncookies = 0; 5178 } 5179 return (error); 5180 } 5181 if (ap->a_ncookies == NULL) 5182 return (0); 5183 5184 KASSERT(ap->a_cookies, 5185 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 5186 5187 *ap->a_cookies = realloc(*ap->a_cookies, 5188 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); 5189 (*ap->a_cookies)[*ap->a_ncookies] = off; 5190 *ap->a_ncookies += 1; 5191 return (0); 5192 } 5193 5194 /* 5195 * Mark for update the access time of the file if the filesystem 5196 * supports VOP_MARKATIME. This functionality is used by execve and 5197 * mmap, so we want to avoid the I/O implied by directly setting 5198 * va_atime for the sake of efficiency. 5199 */ 5200 void 5201 vfs_mark_atime(struct vnode *vp, struct ucred *cred) 5202 { 5203 struct mount *mp; 5204 5205 mp = vp->v_mount; 5206 ASSERT_VOP_LOCKED(vp, "vfs_mark_atime"); 5207 if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) 5208 (void)VOP_MARKATIME(vp); 5209 } 5210 5211 /* 5212 * The purpose of this routine is to remove granularity from accmode_t, 5213 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 5214 * VADMIN and VAPPEND. 5215 * 5216 * If it returns 0, the caller is supposed to continue with the usual 5217 * access checks using 'accmode' as modified by this routine. If it 5218 * returns nonzero value, the caller is supposed to return that value 5219 * as errno. 5220 * 5221 * Note that after this routine runs, accmode may be zero. 5222 */ 5223 int 5224 vfs_unixify_accmode(accmode_t *accmode) 5225 { 5226 /* 5227 * There is no way to specify explicit "deny" rule using 5228 * file mode or POSIX.1e ACLs. 5229 */ 5230 if (*accmode & VEXPLICIT_DENY) { 5231 *accmode = 0; 5232 return (0); 5233 } 5234 5235 /* 5236 * None of these can be translated into usual access bits. 5237 * Also, the common case for NFSv4 ACLs is to not contain 5238 * either of these bits. Caller should check for VWRITE 5239 * on the containing directory instead. 5240 */ 5241 if (*accmode & (VDELETE_CHILD | VDELETE)) 5242 return (EPERM); 5243 5244 if (*accmode & VADMIN_PERMS) { 5245 *accmode &= ~VADMIN_PERMS; 5246 *accmode |= VADMIN; 5247 } 5248 5249 /* 5250 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 5251 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 5252 */ 5253 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 5254 5255 return (0); 5256 } 5257 5258 /* 5259 * These are helper functions for filesystems to traverse all 5260 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 5261 * 5262 * This interface replaces MNT_VNODE_FOREACH. 5263 */ 5264 5265 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker"); 5266 5267 struct vnode * 5268 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 5269 { 5270 struct vnode *vp; 5271 5272 if (should_yield()) 5273 kern_yield(PRI_USER); 5274 MNT_ILOCK(mp); 5275 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5276 vp = TAILQ_NEXT(*mvp, v_nmntvnodes); 5277 while (vp != NULL && (vp->v_type == VMARKER || 5278 (vp->v_iflag & VI_DOOMED) != 0)) 5279 vp = TAILQ_NEXT(vp, v_nmntvnodes); 5280 5281 /* Check if we are done */ 5282 if (vp == NULL) { 5283 __mnt_vnode_markerfree_all(mvp, mp); 5284 /* MNT_IUNLOCK(mp); -- done in above function */ 5285 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 5286 return (NULL); 5287 } 5288 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 5289 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 5290 VI_LOCK(vp); 5291 MNT_IUNLOCK(mp); 5292 return (vp); 5293 } 5294 5295 struct vnode * 5296 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 5297 { 5298 struct vnode *vp; 5299 5300 *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 5301 MNT_ILOCK(mp); 5302 MNT_REF(mp); 5303 (*mvp)->v_type = VMARKER; 5304 5305 vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 5306 while (vp != NULL && (vp->v_type == VMARKER || 5307 (vp->v_iflag & VI_DOOMED) != 0)) 5308 vp = TAILQ_NEXT(vp, v_nmntvnodes); 5309 5310 /* Check if we are done */ 5311 if (vp == NULL) { 5312 MNT_REL(mp); 5313 MNT_IUNLOCK(mp); 5314 free(*mvp, M_VNODE_MARKER); 5315 *mvp = NULL; 5316 return (NULL); 5317 } 5318 (*mvp)->v_mount = mp; 5319 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 5320 VI_LOCK(vp); 5321 MNT_IUNLOCK(mp); 5322 return (vp); 5323 } 5324 5325 5326 void 5327 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 5328 { 5329 5330 if (*mvp == NULL) { 5331 MNT_IUNLOCK(mp); 5332 return; 5333 } 5334 5335 mtx_assert(MNT_MTX(mp), MA_OWNED); 5336 5337 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5338 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 5339 MNT_REL(mp); 5340 MNT_IUNLOCK(mp); 5341 free(*mvp, M_VNODE_MARKER); 5342 *mvp = NULL; 5343 } 5344 5345 /* 5346 * These are helper functions for filesystems to traverse their 5347 * active vnodes. See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h 5348 */ 5349 static void 5350 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp) 5351 { 5352 5353 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5354 5355 MNT_ILOCK(mp); 5356 MNT_REL(mp); 5357 MNT_IUNLOCK(mp); 5358 free(*mvp, M_VNODE_MARKER); 5359 *mvp = NULL; 5360 } 5361 5362 static struct vnode * 5363 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp) 5364 { 5365 struct vnode *vp, *nvp; 5366 5367 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 5368 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 5369 restart: 5370 vp = TAILQ_NEXT(*mvp, v_actfreelist); 5371 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist); 5372 while (vp != NULL) { 5373 if (vp->v_type == VMARKER) { 5374 vp = TAILQ_NEXT(vp, v_actfreelist); 5375 continue; 5376 } 5377 if (!VI_TRYLOCK(vp)) { 5378 if (mp_ncpus == 1 || should_yield()) { 5379 TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist); 5380 mtx_unlock(&mp->mnt_listmtx); 5381 pause("vnacti", 1); 5382 mtx_lock(&mp->mnt_listmtx); 5383 goto restart; 5384 } 5385 continue; 5386 } 5387 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 5388 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 5389 ("alien vnode on the active list %p %p", vp, mp)); 5390 if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0) 5391 break; 5392 nvp = TAILQ_NEXT(vp, v_actfreelist); 5393 VI_UNLOCK(vp); 5394 vp = nvp; 5395 } 5396 5397 /* Check if we are done */ 5398 if (vp == NULL) { 5399 mtx_unlock(&mp->mnt_listmtx); 5400 mnt_vnode_markerfree_active(mvp, mp); 5401 return (NULL); 5402 } 5403 TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist); 5404 mtx_unlock(&mp->mnt_listmtx); 5405 ASSERT_VI_LOCKED(vp, "active iter"); 5406 KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp)); 5407 return (vp); 5408 } 5409 5410 struct vnode * 5411 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp) 5412 { 5413 5414 if (should_yield()) 5415 kern_yield(PRI_USER); 5416 mtx_lock(&mp->mnt_listmtx); 5417 return (mnt_vnode_next_active(mvp, mp)); 5418 } 5419 5420 struct vnode * 5421 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp) 5422 { 5423 struct vnode *vp; 5424 5425 *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 5426 MNT_ILOCK(mp); 5427 MNT_REF(mp); 5428 MNT_IUNLOCK(mp); 5429 (*mvp)->v_type = VMARKER; 5430 (*mvp)->v_mount = mp; 5431 5432 mtx_lock(&mp->mnt_listmtx); 5433 vp = TAILQ_FIRST(&mp->mnt_activevnodelist); 5434 if (vp == NULL) { 5435 mtx_unlock(&mp->mnt_listmtx); 5436 mnt_vnode_markerfree_active(mvp, mp); 5437 return (NULL); 5438 } 5439 TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist); 5440 return (mnt_vnode_next_active(mvp, mp)); 5441 } 5442 5443 void 5444 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp) 5445 { 5446 5447 if (*mvp == NULL) 5448 return; 5449 5450 mtx_lock(&mp->mnt_listmtx); 5451 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist); 5452 mtx_unlock(&mp->mnt_listmtx); 5453 mnt_vnode_markerfree_active(mvp, mp); 5454 } 5455