1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1989, 1993 5 * The Regents of the University of California. All rights reserved. 6 * (c) UNIX System Laboratories, Inc. 7 * All or some portions of this file are derived from material licensed 8 * to the University of California by American Telephone and Telegraph 9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 10 * the permission of UNIX System Laboratories, Inc. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 37 */ 38 39 /* 40 * External virtual filesystem routines 41 */ 42 43 #include <sys/cdefs.h> 44 __FBSDID("$FreeBSD$"); 45 46 #include "opt_ddb.h" 47 #include "opt_watchdog.h" 48 49 #include <sys/param.h> 50 #include <sys/systm.h> 51 #include <sys/bio.h> 52 #include <sys/buf.h> 53 #include <sys/capsicum.h> 54 #include <sys/condvar.h> 55 #include <sys/conf.h> 56 #include <sys/counter.h> 57 #include <sys/dirent.h> 58 #include <sys/event.h> 59 #include <sys/eventhandler.h> 60 #include <sys/extattr.h> 61 #include <sys/file.h> 62 #include <sys/fcntl.h> 63 #include <sys/jail.h> 64 #include <sys/kdb.h> 65 #include <sys/kernel.h> 66 #include <sys/kthread.h> 67 #include <sys/ktr.h> 68 #include <sys/lockf.h> 69 #include <sys/malloc.h> 70 #include <sys/mount.h> 71 #include <sys/namei.h> 72 #include <sys/pctrie.h> 73 #include <sys/priv.h> 74 #include <sys/reboot.h> 75 #include <sys/refcount.h> 76 #include <sys/rwlock.h> 77 #include <sys/sched.h> 78 #include <sys/sleepqueue.h> 79 #include <sys/smp.h> 80 #include <sys/stat.h> 81 #include <sys/sysctl.h> 82 #include <sys/syslog.h> 83 #include <sys/vmmeter.h> 84 #include <sys/vnode.h> 85 #include <sys/watchdog.h> 86 87 #include <machine/stdarg.h> 88 89 #include <security/mac/mac_framework.h> 90 91 #include <vm/vm.h> 92 #include <vm/vm_object.h> 93 #include <vm/vm_extern.h> 94 #include <vm/pmap.h> 95 #include <vm/vm_map.h> 96 #include <vm/vm_page.h> 97 #include <vm/vm_kern.h> 98 #include <vm/uma.h> 99 100 #ifdef DDB 101 #include <ddb/ddb.h> 102 #endif 103 104 static void delmntque(struct vnode *vp); 105 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, 106 int slpflag, int slptimeo); 107 static void syncer_shutdown(void *arg, int howto); 108 static int vtryrecycle(struct vnode *vp); 109 static void v_init_counters(struct vnode *); 110 static void v_incr_devcount(struct vnode *); 111 static void v_decr_devcount(struct vnode *); 112 static void vgonel(struct vnode *); 113 static void vfs_knllock(void *arg); 114 static void vfs_knlunlock(void *arg); 115 static void vfs_knl_assert_locked(void *arg); 116 static void vfs_knl_assert_unlocked(void *arg); 117 static void destroy_vpollinfo(struct vpollinfo *vi); 118 static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, 119 daddr_t startlbn, daddr_t endlbn); 120 static void vnlru_recalc(void); 121 122 /* 123 * These fences are intended for cases where some synchronization is 124 * needed between access of v_iflags and lockless vnode refcount (v_holdcnt 125 * and v_usecount) updates. Access to v_iflags is generally synchronized 126 * by the interlock, but we have some internal assertions that check vnode 127 * flags without acquiring the lock. Thus, these fences are INVARIANTS-only 128 * for now. 129 */ 130 #ifdef INVARIANTS 131 #define VNODE_REFCOUNT_FENCE_ACQ() atomic_thread_fence_acq() 132 #define VNODE_REFCOUNT_FENCE_REL() atomic_thread_fence_rel() 133 #else 134 #define VNODE_REFCOUNT_FENCE_ACQ() 135 #define VNODE_REFCOUNT_FENCE_REL() 136 #endif 137 138 /* 139 * Number of vnodes in existence. Increased whenever getnewvnode() 140 * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode. 141 */ 142 static u_long __exclusive_cache_line numvnodes; 143 144 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, 145 "Number of vnodes in existence"); 146 147 static counter_u64_t vnodes_created; 148 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created, 149 "Number of vnodes created by getnewvnode"); 150 151 /* 152 * Conversion tables for conversion from vnode types to inode formats 153 * and back. 154 */ 155 enum vtype iftovt_tab[16] = { 156 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 157 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON 158 }; 159 int vttoif_tab[10] = { 160 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 161 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT 162 }; 163 164 /* 165 * List of allocates vnodes in the system. 166 */ 167 static TAILQ_HEAD(freelst, vnode) vnode_list; 168 static struct vnode *vnode_list_free_marker; 169 static struct vnode *vnode_list_reclaim_marker; 170 171 /* 172 * "Free" vnode target. Free vnodes are rarely completely free, but are 173 * just ones that are cheap to recycle. Usually they are for files which 174 * have been stat'd but not read; these usually have inode and namecache 175 * data attached to them. This target is the preferred minimum size of a 176 * sub-cache consisting mostly of such files. The system balances the size 177 * of this sub-cache with its complement to try to prevent either from 178 * thrashing while the other is relatively inactive. The targets express 179 * a preference for the best balance. 180 * 181 * "Above" this target there are 2 further targets (watermarks) related 182 * to recyling of free vnodes. In the best-operating case, the cache is 183 * exactly full, the free list has size between vlowat and vhiwat above the 184 * free target, and recycling from it and normal use maintains this state. 185 * Sometimes the free list is below vlowat or even empty, but this state 186 * is even better for immediate use provided the cache is not full. 187 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free 188 * ones) to reach one of these states. The watermarks are currently hard- 189 * coded as 4% and 9% of the available space higher. These and the default 190 * of 25% for wantfreevnodes are too large if the memory size is large. 191 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim 192 * whenever vnlru_proc() becomes active. 193 */ 194 static long wantfreevnodes; 195 static long __exclusive_cache_line freevnodes; 196 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, 197 &freevnodes, 0, "Number of \"free\" vnodes"); 198 static long freevnodes_old; 199 200 static counter_u64_t recycles_count; 201 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count, 202 "Number of vnodes recycled to meet vnode cache targets"); 203 204 static counter_u64_t recycles_free_count; 205 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count, 206 "Number of free vnodes recycled to meet vnode cache targets"); 207 208 /* 209 * Various variables used for debugging the new implementation of 210 * reassignbuf(). 211 * XXX these are probably of (very) limited utility now. 212 */ 213 static int reassignbufcalls; 214 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW | CTLFLAG_STATS, 215 &reassignbufcalls, 0, "Number of calls to reassignbuf"); 216 217 static counter_u64_t deferred_inact; 218 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact, 219 "Number of times inactive processing was deferred"); 220 221 /* To keep more than one thread at a time from running vfs_getnewfsid */ 222 static struct mtx mntid_mtx; 223 224 /* 225 * Lock for any access to the following: 226 * vnode_list 227 * numvnodes 228 * freevnodes 229 */ 230 static struct mtx __exclusive_cache_line vnode_list_mtx; 231 232 /* Publicly exported FS */ 233 struct nfs_public nfs_pub; 234 235 static uma_zone_t buf_trie_zone; 236 237 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 238 static uma_zone_t vnode_zone; 239 static uma_zone_t vnodepoll_zone; 240 241 /* 242 * The workitem queue. 243 * 244 * It is useful to delay writes of file data and filesystem metadata 245 * for tens of seconds so that quickly created and deleted files need 246 * not waste disk bandwidth being created and removed. To realize this, 247 * we append vnodes to a "workitem" queue. When running with a soft 248 * updates implementation, most pending metadata dependencies should 249 * not wait for more than a few seconds. Thus, mounted on block devices 250 * are delayed only about a half the time that file data is delayed. 251 * Similarly, directory updates are more critical, so are only delayed 252 * about a third the time that file data is delayed. Thus, there are 253 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 254 * one each second (driven off the filesystem syncer process). The 255 * syncer_delayno variable indicates the next queue that is to be processed. 256 * Items that need to be processed soon are placed in this queue: 257 * 258 * syncer_workitem_pending[syncer_delayno] 259 * 260 * A delay of fifteen seconds is done by placing the request fifteen 261 * entries later in the queue: 262 * 263 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 264 * 265 */ 266 static int syncer_delayno; 267 static long syncer_mask; 268 LIST_HEAD(synclist, bufobj); 269 static struct synclist *syncer_workitem_pending; 270 /* 271 * The sync_mtx protects: 272 * bo->bo_synclist 273 * sync_vnode_count 274 * syncer_delayno 275 * syncer_state 276 * syncer_workitem_pending 277 * syncer_worklist_len 278 * rushjob 279 */ 280 static struct mtx sync_mtx; 281 static struct cv sync_wakeup; 282 283 #define SYNCER_MAXDELAY 32 284 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 285 static int syncdelay = 30; /* max time to delay syncing data */ 286 static int filedelay = 30; /* time to delay syncing files */ 287 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, 288 "Time to delay syncing files (in seconds)"); 289 static int dirdelay = 29; /* time to delay syncing directories */ 290 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, 291 "Time to delay syncing directories (in seconds)"); 292 static int metadelay = 28; /* time to delay syncing metadata */ 293 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, 294 "Time to delay syncing metadata (in seconds)"); 295 static int rushjob; /* number of slots to run ASAP */ 296 static int stat_rush_requests; /* number of times I/O speeded up */ 297 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, 298 "Number of times I/O speeded up (rush requests)"); 299 300 #define VDBATCH_SIZE 8 301 struct vdbatch { 302 u_int index; 303 long freevnodes; 304 struct mtx lock; 305 struct vnode *tab[VDBATCH_SIZE]; 306 }; 307 DPCPU_DEFINE_STATIC(struct vdbatch, vd); 308 309 static void vdbatch_dequeue(struct vnode *vp); 310 311 /* 312 * When shutting down the syncer, run it at four times normal speed. 313 */ 314 #define SYNCER_SHUTDOWN_SPEEDUP 4 315 static int sync_vnode_count; 316 static int syncer_worklist_len; 317 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 318 syncer_state; 319 320 /* Target for maximum number of vnodes. */ 321 u_long desiredvnodes; 322 static u_long gapvnodes; /* gap between wanted and desired */ 323 static u_long vhiwat; /* enough extras after expansion */ 324 static u_long vlowat; /* minimal extras before expansion */ 325 static u_long vstir; /* nonzero to stir non-free vnodes */ 326 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */ 327 328 static u_long vnlru_read_freevnodes(void); 329 330 /* 331 * Note that no attempt is made to sanitize these parameters. 332 */ 333 static int 334 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS) 335 { 336 u_long val; 337 int error; 338 339 val = desiredvnodes; 340 error = sysctl_handle_long(oidp, &val, 0, req); 341 if (error != 0 || req->newptr == NULL) 342 return (error); 343 344 if (val == desiredvnodes) 345 return (0); 346 mtx_lock(&vnode_list_mtx); 347 desiredvnodes = val; 348 wantfreevnodes = desiredvnodes / 4; 349 vnlru_recalc(); 350 mtx_unlock(&vnode_list_mtx); 351 /* 352 * XXX There is no protection against multiple threads changing 353 * desiredvnodes at the same time. Locking above only helps vnlru and 354 * getnewvnode. 355 */ 356 vfs_hash_changesize(desiredvnodes); 357 cache_changesize(desiredvnodes); 358 return (0); 359 } 360 361 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes, 362 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes, 363 "UL", "Target for maximum number of vnodes"); 364 365 static int 366 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS) 367 { 368 u_long val; 369 int error; 370 371 val = wantfreevnodes; 372 error = sysctl_handle_long(oidp, &val, 0, req); 373 if (error != 0 || req->newptr == NULL) 374 return (error); 375 376 if (val == wantfreevnodes) 377 return (0); 378 mtx_lock(&vnode_list_mtx); 379 wantfreevnodes = val; 380 vnlru_recalc(); 381 mtx_unlock(&vnode_list_mtx); 382 return (0); 383 } 384 385 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes, 386 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes, 387 "UL", "Target for minimum number of \"free\" vnodes"); 388 389 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 390 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)"); 391 static int vnlru_nowhere; 392 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 393 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 394 395 static int 396 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS) 397 { 398 struct vnode *vp; 399 struct nameidata nd; 400 char *buf; 401 unsigned long ndflags; 402 int error; 403 404 if (req->newptr == NULL) 405 return (EINVAL); 406 if (req->newlen >= PATH_MAX) 407 return (E2BIG); 408 409 buf = malloc(PATH_MAX, M_TEMP, M_WAITOK); 410 error = SYSCTL_IN(req, buf, req->newlen); 411 if (error != 0) 412 goto out; 413 414 buf[req->newlen] = '\0'; 415 416 ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME; 417 NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread); 418 if ((error = namei(&nd)) != 0) 419 goto out; 420 vp = nd.ni_vp; 421 422 if (VN_IS_DOOMED(vp)) { 423 /* 424 * This vnode is being recycled. Return != 0 to let the caller 425 * know that the sysctl had no effect. Return EAGAIN because a 426 * subsequent call will likely succeed (since namei will create 427 * a new vnode if necessary) 428 */ 429 error = EAGAIN; 430 goto putvnode; 431 } 432 433 counter_u64_add(recycles_count, 1); 434 vgone(vp); 435 putvnode: 436 NDFREE(&nd, 0); 437 out: 438 free(buf, M_TEMP); 439 return (error); 440 } 441 442 static int 443 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS) 444 { 445 struct thread *td = curthread; 446 struct vnode *vp; 447 struct file *fp; 448 int error; 449 int fd; 450 451 if (req->newptr == NULL) 452 return (EBADF); 453 454 error = sysctl_handle_int(oidp, &fd, 0, req); 455 if (error != 0) 456 return (error); 457 error = getvnode(curthread, fd, &cap_fcntl_rights, &fp); 458 if (error != 0) 459 return (error); 460 vp = fp->f_vnode; 461 462 error = vn_lock(vp, LK_EXCLUSIVE); 463 if (error != 0) 464 goto drop; 465 466 counter_u64_add(recycles_count, 1); 467 vgone(vp); 468 VOP_UNLOCK(vp); 469 drop: 470 fdrop(fp, td); 471 return (error); 472 } 473 474 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode, 475 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, 476 sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname"); 477 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode, 478 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, 479 sysctl_ftry_reclaim_vnode, "I", 480 "Try to reclaim a vnode by its file descriptor"); 481 482 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */ 483 static int vnsz2log; 484 485 /* 486 * Support for the bufobj clean & dirty pctrie. 487 */ 488 static void * 489 buf_trie_alloc(struct pctrie *ptree) 490 { 491 492 return uma_zalloc(buf_trie_zone, M_NOWAIT); 493 } 494 495 static void 496 buf_trie_free(struct pctrie *ptree, void *node) 497 { 498 499 uma_zfree(buf_trie_zone, node); 500 } 501 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free); 502 503 /* 504 * Initialize the vnode management data structures. 505 * 506 * Reevaluate the following cap on the number of vnodes after the physical 507 * memory size exceeds 512GB. In the limit, as the physical memory size 508 * grows, the ratio of the memory size in KB to vnodes approaches 64:1. 509 */ 510 #ifndef MAXVNODES_MAX 511 #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */ 512 #endif 513 514 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker"); 515 516 static struct vnode * 517 vn_alloc_marker(struct mount *mp) 518 { 519 struct vnode *vp; 520 521 vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 522 vp->v_type = VMARKER; 523 vp->v_mount = mp; 524 525 return (vp); 526 } 527 528 static void 529 vn_free_marker(struct vnode *vp) 530 { 531 532 MPASS(vp->v_type == VMARKER); 533 free(vp, M_VNODE_MARKER); 534 } 535 536 /* 537 * Initialize a vnode as it first enters the zone. 538 */ 539 static int 540 vnode_init(void *mem, int size, int flags) 541 { 542 struct vnode *vp; 543 544 vp = mem; 545 bzero(vp, size); 546 /* 547 * Setup locks. 548 */ 549 vp->v_vnlock = &vp->v_lock; 550 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 551 /* 552 * By default, don't allow shared locks unless filesystems opt-in. 553 */ 554 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT, 555 LK_NOSHARE | LK_IS_VNODE); 556 /* 557 * Initialize bufobj. 558 */ 559 bufobj_init(&vp->v_bufobj, vp); 560 /* 561 * Initialize namecache. 562 */ 563 LIST_INIT(&vp->v_cache_src); 564 TAILQ_INIT(&vp->v_cache_dst); 565 /* 566 * Initialize rangelocks. 567 */ 568 rangelock_init(&vp->v_rl); 569 570 vp->v_dbatchcpu = NOCPU; 571 572 mtx_lock(&vnode_list_mtx); 573 TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist); 574 mtx_unlock(&vnode_list_mtx); 575 return (0); 576 } 577 578 /* 579 * Free a vnode when it is cleared from the zone. 580 */ 581 static void 582 vnode_fini(void *mem, int size) 583 { 584 struct vnode *vp; 585 struct bufobj *bo; 586 587 vp = mem; 588 vdbatch_dequeue(vp); 589 mtx_lock(&vnode_list_mtx); 590 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 591 mtx_unlock(&vnode_list_mtx); 592 rangelock_destroy(&vp->v_rl); 593 lockdestroy(vp->v_vnlock); 594 mtx_destroy(&vp->v_interlock); 595 bo = &vp->v_bufobj; 596 rw_destroy(BO_LOCKPTR(bo)); 597 } 598 599 /* 600 * Provide the size of NFS nclnode and NFS fh for calculation of the 601 * vnode memory consumption. The size is specified directly to 602 * eliminate dependency on NFS-private header. 603 * 604 * Other filesystems may use bigger or smaller (like UFS and ZFS) 605 * private inode data, but the NFS-based estimation is ample enough. 606 * Still, we care about differences in the size between 64- and 32-bit 607 * platforms. 608 * 609 * Namecache structure size is heuristically 610 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1. 611 */ 612 #ifdef _LP64 613 #define NFS_NCLNODE_SZ (528 + 64) 614 #define NC_SZ 148 615 #else 616 #define NFS_NCLNODE_SZ (360 + 32) 617 #define NC_SZ 92 618 #endif 619 620 static void 621 vntblinit(void *dummy __unused) 622 { 623 struct vdbatch *vd; 624 int cpu, physvnodes, virtvnodes; 625 u_int i; 626 627 /* 628 * Desiredvnodes is a function of the physical memory size and the 629 * kernel's heap size. Generally speaking, it scales with the 630 * physical memory size. The ratio of desiredvnodes to the physical 631 * memory size is 1:16 until desiredvnodes exceeds 98,304. 632 * Thereafter, the 633 * marginal ratio of desiredvnodes to the physical memory size is 634 * 1:64. However, desiredvnodes is limited by the kernel's heap 635 * size. The memory required by desiredvnodes vnodes and vm objects 636 * must not exceed 1/10th of the kernel's heap size. 637 */ 638 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 + 639 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64; 640 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) + 641 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ)); 642 desiredvnodes = min(physvnodes, virtvnodes); 643 if (desiredvnodes > MAXVNODES_MAX) { 644 if (bootverbose) 645 printf("Reducing kern.maxvnodes %lu -> %lu\n", 646 desiredvnodes, MAXVNODES_MAX); 647 desiredvnodes = MAXVNODES_MAX; 648 } 649 wantfreevnodes = desiredvnodes / 4; 650 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 651 TAILQ_INIT(&vnode_list); 652 mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF); 653 /* 654 * The lock is taken to appease WITNESS. 655 */ 656 mtx_lock(&vnode_list_mtx); 657 vnlru_recalc(); 658 mtx_unlock(&vnode_list_mtx); 659 vnode_list_free_marker = vn_alloc_marker(NULL); 660 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist); 661 vnode_list_reclaim_marker = vn_alloc_marker(NULL); 662 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist); 663 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 664 vnode_init, vnode_fini, UMA_ALIGN_PTR, 0); 665 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 666 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 667 /* 668 * Preallocate enough nodes to support one-per buf so that 669 * we can not fail an insert. reassignbuf() callers can not 670 * tolerate the insertion failure. 671 */ 672 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(), 673 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, 674 UMA_ZONE_NOFREE | UMA_ZONE_VM); 675 uma_prealloc(buf_trie_zone, nbuf); 676 677 vnodes_created = counter_u64_alloc(M_WAITOK); 678 recycles_count = counter_u64_alloc(M_WAITOK); 679 recycles_free_count = counter_u64_alloc(M_WAITOK); 680 deferred_inact = counter_u64_alloc(M_WAITOK); 681 682 /* 683 * Initialize the filesystem syncer. 684 */ 685 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 686 &syncer_mask); 687 syncer_maxdelay = syncer_mask + 1; 688 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 689 cv_init(&sync_wakeup, "syncer"); 690 for (i = 1; i <= sizeof(struct vnode); i <<= 1) 691 vnsz2log++; 692 vnsz2log--; 693 694 CPU_FOREACH(cpu) { 695 vd = DPCPU_ID_PTR((cpu), vd); 696 bzero(vd, sizeof(*vd)); 697 mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF); 698 } 699 } 700 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL); 701 702 703 /* 704 * Mark a mount point as busy. Used to synchronize access and to delay 705 * unmounting. Eventually, mountlist_mtx is not released on failure. 706 * 707 * vfs_busy() is a custom lock, it can block the caller. 708 * vfs_busy() only sleeps if the unmount is active on the mount point. 709 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any 710 * vnode belonging to mp. 711 * 712 * Lookup uses vfs_busy() to traverse mount points. 713 * root fs var fs 714 * / vnode lock A / vnode lock (/var) D 715 * /var vnode lock B /log vnode lock(/var/log) E 716 * vfs_busy lock C vfs_busy lock F 717 * 718 * Within each file system, the lock order is C->A->B and F->D->E. 719 * 720 * When traversing across mounts, the system follows that lock order: 721 * 722 * C->A->B 723 * | 724 * +->F->D->E 725 * 726 * The lookup() process for namei("/var") illustrates the process: 727 * VOP_LOOKUP() obtains B while A is held 728 * vfs_busy() obtains a shared lock on F while A and B are held 729 * vput() releases lock on B 730 * vput() releases lock on A 731 * VFS_ROOT() obtains lock on D while shared lock on F is held 732 * vfs_unbusy() releases shared lock on F 733 * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A. 734 * Attempt to lock A (instead of vp_crossmp) while D is held would 735 * violate the global order, causing deadlocks. 736 * 737 * dounmount() locks B while F is drained. 738 */ 739 int 740 vfs_busy(struct mount *mp, int flags) 741 { 742 743 MPASS((flags & ~MBF_MASK) == 0); 744 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags); 745 746 if (vfs_op_thread_enter(mp)) { 747 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 748 MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0); 749 MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0); 750 vfs_mp_count_add_pcpu(mp, ref, 1); 751 vfs_mp_count_add_pcpu(mp, lockref, 1); 752 vfs_op_thread_exit(mp); 753 if (flags & MBF_MNTLSTLOCK) 754 mtx_unlock(&mountlist_mtx); 755 return (0); 756 } 757 758 MNT_ILOCK(mp); 759 vfs_assert_mount_counters(mp); 760 MNT_REF(mp); 761 /* 762 * If mount point is currently being unmounted, sleep until the 763 * mount point fate is decided. If thread doing the unmounting fails, 764 * it will clear MNTK_UNMOUNT flag before waking us up, indicating 765 * that this mount point has survived the unmount attempt and vfs_busy 766 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE 767 * flag in addition to MNTK_UNMOUNT, indicating that mount point is 768 * about to be really destroyed. vfs_busy needs to release its 769 * reference on the mount point in this case and return with ENOENT, 770 * telling the caller that mount mount it tried to busy is no longer 771 * valid. 772 */ 773 while (mp->mnt_kern_flag & MNTK_UNMOUNT) { 774 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) { 775 MNT_REL(mp); 776 MNT_IUNLOCK(mp); 777 CTR1(KTR_VFS, "%s: failed busying before sleeping", 778 __func__); 779 return (ENOENT); 780 } 781 if (flags & MBF_MNTLSTLOCK) 782 mtx_unlock(&mountlist_mtx); 783 mp->mnt_kern_flag |= MNTK_MWAIT; 784 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0); 785 if (flags & MBF_MNTLSTLOCK) 786 mtx_lock(&mountlist_mtx); 787 MNT_ILOCK(mp); 788 } 789 if (flags & MBF_MNTLSTLOCK) 790 mtx_unlock(&mountlist_mtx); 791 mp->mnt_lockref++; 792 MNT_IUNLOCK(mp); 793 return (0); 794 } 795 796 /* 797 * Free a busy filesystem. 798 */ 799 void 800 vfs_unbusy(struct mount *mp) 801 { 802 int c; 803 804 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 805 806 if (vfs_op_thread_enter(mp)) { 807 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 808 vfs_mp_count_sub_pcpu(mp, lockref, 1); 809 vfs_mp_count_sub_pcpu(mp, ref, 1); 810 vfs_op_thread_exit(mp); 811 return; 812 } 813 814 MNT_ILOCK(mp); 815 vfs_assert_mount_counters(mp); 816 MNT_REL(mp); 817 c = --mp->mnt_lockref; 818 if (mp->mnt_vfs_ops == 0) { 819 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 820 MNT_IUNLOCK(mp); 821 return; 822 } 823 if (c < 0) 824 vfs_dump_mount_counters(mp); 825 if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) { 826 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT); 827 CTR1(KTR_VFS, "%s: waking up waiters", __func__); 828 mp->mnt_kern_flag &= ~MNTK_DRAINING; 829 wakeup(&mp->mnt_lockref); 830 } 831 MNT_IUNLOCK(mp); 832 } 833 834 /* 835 * Lookup a mount point by filesystem identifier. 836 */ 837 struct mount * 838 vfs_getvfs(fsid_t *fsid) 839 { 840 struct mount *mp; 841 842 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 843 mtx_lock(&mountlist_mtx); 844 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 845 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 846 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 847 vfs_ref(mp); 848 mtx_unlock(&mountlist_mtx); 849 return (mp); 850 } 851 } 852 mtx_unlock(&mountlist_mtx); 853 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 854 return ((struct mount *) 0); 855 } 856 857 /* 858 * Lookup a mount point by filesystem identifier, busying it before 859 * returning. 860 * 861 * To avoid congestion on mountlist_mtx, implement simple direct-mapped 862 * cache for popular filesystem identifiers. The cache is lockess, using 863 * the fact that struct mount's are never freed. In worst case we may 864 * get pointer to unmounted or even different filesystem, so we have to 865 * check what we got, and go slow way if so. 866 */ 867 struct mount * 868 vfs_busyfs(fsid_t *fsid) 869 { 870 #define FSID_CACHE_SIZE 256 871 typedef struct mount * volatile vmp_t; 872 static vmp_t cache[FSID_CACHE_SIZE]; 873 struct mount *mp; 874 int error; 875 uint32_t hash; 876 877 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 878 hash = fsid->val[0] ^ fsid->val[1]; 879 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1); 880 mp = cache[hash]; 881 if (mp == NULL || 882 mp->mnt_stat.f_fsid.val[0] != fsid->val[0] || 883 mp->mnt_stat.f_fsid.val[1] != fsid->val[1]) 884 goto slow; 885 if (vfs_busy(mp, 0) != 0) { 886 cache[hash] = NULL; 887 goto slow; 888 } 889 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 890 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) 891 return (mp); 892 else 893 vfs_unbusy(mp); 894 895 slow: 896 mtx_lock(&mountlist_mtx); 897 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 898 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 899 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 900 error = vfs_busy(mp, MBF_MNTLSTLOCK); 901 if (error) { 902 cache[hash] = NULL; 903 mtx_unlock(&mountlist_mtx); 904 return (NULL); 905 } 906 cache[hash] = mp; 907 return (mp); 908 } 909 } 910 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 911 mtx_unlock(&mountlist_mtx); 912 return ((struct mount *) 0); 913 } 914 915 /* 916 * Check if a user can access privileged mount options. 917 */ 918 int 919 vfs_suser(struct mount *mp, struct thread *td) 920 { 921 int error; 922 923 if (jailed(td->td_ucred)) { 924 /* 925 * If the jail of the calling thread lacks permission for 926 * this type of file system, deny immediately. 927 */ 928 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag)) 929 return (EPERM); 930 931 /* 932 * If the file system was mounted outside the jail of the 933 * calling thread, deny immediately. 934 */ 935 if (prison_check(td->td_ucred, mp->mnt_cred) != 0) 936 return (EPERM); 937 } 938 939 /* 940 * If file system supports delegated administration, we don't check 941 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified 942 * by the file system itself. 943 * If this is not the user that did original mount, we check for 944 * the PRIV_VFS_MOUNT_OWNER privilege. 945 */ 946 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) && 947 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 948 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) 949 return (error); 950 } 951 return (0); 952 } 953 954 /* 955 * Get a new unique fsid. Try to make its val[0] unique, since this value 956 * will be used to create fake device numbers for stat(). Also try (but 957 * not so hard) make its val[0] unique mod 2^16, since some emulators only 958 * support 16-bit device numbers. We end up with unique val[0]'s for the 959 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 960 * 961 * Keep in mind that several mounts may be running in parallel. Starting 962 * the search one past where the previous search terminated is both a 963 * micro-optimization and a defense against returning the same fsid to 964 * different mounts. 965 */ 966 void 967 vfs_getnewfsid(struct mount *mp) 968 { 969 static uint16_t mntid_base; 970 struct mount *nmp; 971 fsid_t tfsid; 972 int mtype; 973 974 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 975 mtx_lock(&mntid_mtx); 976 mtype = mp->mnt_vfc->vfc_typenum; 977 tfsid.val[1] = mtype; 978 mtype = (mtype & 0xFF) << 24; 979 for (;;) { 980 tfsid.val[0] = makedev(255, 981 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 982 mntid_base++; 983 if ((nmp = vfs_getvfs(&tfsid)) == NULL) 984 break; 985 vfs_rel(nmp); 986 } 987 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 988 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 989 mtx_unlock(&mntid_mtx); 990 } 991 992 /* 993 * Knob to control the precision of file timestamps: 994 * 995 * 0 = seconds only; nanoseconds zeroed. 996 * 1 = seconds and nanoseconds, accurate within 1/HZ. 997 * 2 = seconds and nanoseconds, truncated to microseconds. 998 * >=3 = seconds and nanoseconds, maximum precision. 999 */ 1000 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 1001 1002 static int timestamp_precision = TSP_USEC; 1003 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 1004 ×tamp_precision, 0, "File timestamp precision (0: seconds, " 1005 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, " 1006 "3+: sec + ns (max. precision))"); 1007 1008 /* 1009 * Get a current timestamp. 1010 */ 1011 void 1012 vfs_timestamp(struct timespec *tsp) 1013 { 1014 struct timeval tv; 1015 1016 switch (timestamp_precision) { 1017 case TSP_SEC: 1018 tsp->tv_sec = time_second; 1019 tsp->tv_nsec = 0; 1020 break; 1021 case TSP_HZ: 1022 getnanotime(tsp); 1023 break; 1024 case TSP_USEC: 1025 microtime(&tv); 1026 TIMEVAL_TO_TIMESPEC(&tv, tsp); 1027 break; 1028 case TSP_NSEC: 1029 default: 1030 nanotime(tsp); 1031 break; 1032 } 1033 } 1034 1035 /* 1036 * Set vnode attributes to VNOVAL 1037 */ 1038 void 1039 vattr_null(struct vattr *vap) 1040 { 1041 1042 vap->va_type = VNON; 1043 vap->va_size = VNOVAL; 1044 vap->va_bytes = VNOVAL; 1045 vap->va_mode = VNOVAL; 1046 vap->va_nlink = VNOVAL; 1047 vap->va_uid = VNOVAL; 1048 vap->va_gid = VNOVAL; 1049 vap->va_fsid = VNOVAL; 1050 vap->va_fileid = VNOVAL; 1051 vap->va_blocksize = VNOVAL; 1052 vap->va_rdev = VNOVAL; 1053 vap->va_atime.tv_sec = VNOVAL; 1054 vap->va_atime.tv_nsec = VNOVAL; 1055 vap->va_mtime.tv_sec = VNOVAL; 1056 vap->va_mtime.tv_nsec = VNOVAL; 1057 vap->va_ctime.tv_sec = VNOVAL; 1058 vap->va_ctime.tv_nsec = VNOVAL; 1059 vap->va_birthtime.tv_sec = VNOVAL; 1060 vap->va_birthtime.tv_nsec = VNOVAL; 1061 vap->va_flags = VNOVAL; 1062 vap->va_gen = VNOVAL; 1063 vap->va_vaflags = 0; 1064 } 1065 1066 /* 1067 * Try to reduce the total number of vnodes. 1068 * 1069 * This routine (and its user) are buggy in at least the following ways: 1070 * - all parameters were picked years ago when RAM sizes were significantly 1071 * smaller 1072 * - it can pick vnodes based on pages used by the vm object, but filesystems 1073 * like ZFS don't use it making the pick broken 1074 * - since ZFS has its own aging policy it gets partially combated by this one 1075 * - a dedicated method should be provided for filesystems to let them decide 1076 * whether the vnode should be recycled 1077 * 1078 * This routine is called when we have too many vnodes. It attempts 1079 * to free <count> vnodes and will potentially free vnodes that still 1080 * have VM backing store (VM backing store is typically the cause 1081 * of a vnode blowout so we want to do this). Therefore, this operation 1082 * is not considered cheap. 1083 * 1084 * A number of conditions may prevent a vnode from being reclaimed. 1085 * the buffer cache may have references on the vnode, a directory 1086 * vnode may still have references due to the namei cache representing 1087 * underlying files, or the vnode may be in active use. It is not 1088 * desirable to reuse such vnodes. These conditions may cause the 1089 * number of vnodes to reach some minimum value regardless of what 1090 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 1091 * 1092 * @param reclaim_nc_src Only reclaim directories with outgoing namecache 1093 * entries if this argument is strue 1094 * @param trigger Only reclaim vnodes with fewer than this many resident 1095 * pages. 1096 * @param target How many vnodes to reclaim. 1097 * @return The number of vnodes that were reclaimed. 1098 */ 1099 static int 1100 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target) 1101 { 1102 struct vnode *vp, *mvp; 1103 struct mount *mp; 1104 u_long done; 1105 bool retried; 1106 1107 mtx_assert(&vnode_list_mtx, MA_OWNED); 1108 1109 retried = false; 1110 done = 0; 1111 1112 mvp = vnode_list_reclaim_marker; 1113 restart: 1114 vp = mvp; 1115 while (done < target) { 1116 vp = TAILQ_NEXT(vp, v_vnodelist); 1117 if (__predict_false(vp == NULL)) 1118 break; 1119 1120 if (__predict_false(vp->v_type == VMARKER)) 1121 continue; 1122 1123 /* 1124 * If it's been deconstructed already, it's still 1125 * referenced, or it exceeds the trigger, skip it. 1126 * Also skip free vnodes. We are trying to make space 1127 * to expand the free list, not reduce it. 1128 */ 1129 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 || 1130 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src))) 1131 goto next_iter; 1132 1133 if (vp->v_type == VBAD || vp->v_type == VNON) 1134 goto next_iter; 1135 1136 if (!VI_TRYLOCK(vp)) 1137 goto next_iter; 1138 1139 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 || 1140 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 1141 vp->v_type == VBAD || vp->v_type == VNON || 1142 (vp->v_object != NULL && 1143 vp->v_object->resident_page_count > trigger)) { 1144 VI_UNLOCK(vp); 1145 goto next_iter; 1146 } 1147 vholdl(vp); 1148 VI_UNLOCK(vp); 1149 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1150 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1151 mtx_unlock(&vnode_list_mtx); 1152 1153 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1154 vdrop(vp); 1155 goto next_iter_unlocked; 1156 } 1157 if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) { 1158 vdrop(vp); 1159 vn_finished_write(mp); 1160 goto next_iter_unlocked; 1161 } 1162 1163 VI_LOCK(vp); 1164 if (vp->v_usecount > 0 || 1165 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 1166 (vp->v_object != NULL && 1167 vp->v_object->resident_page_count > trigger)) { 1168 VOP_UNLOCK(vp); 1169 vdropl(vp); 1170 vn_finished_write(mp); 1171 goto next_iter_unlocked; 1172 } 1173 counter_u64_add(recycles_count, 1); 1174 vgonel(vp); 1175 VOP_UNLOCK(vp); 1176 vdropl(vp); 1177 vn_finished_write(mp); 1178 done++; 1179 next_iter_unlocked: 1180 if (should_yield()) 1181 kern_yield(PRI_USER); 1182 mtx_lock(&vnode_list_mtx); 1183 goto restart; 1184 next_iter: 1185 MPASS(vp->v_type != VMARKER); 1186 if (!should_yield()) 1187 continue; 1188 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1189 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1190 mtx_unlock(&vnode_list_mtx); 1191 kern_yield(PRI_USER); 1192 mtx_lock(&vnode_list_mtx); 1193 goto restart; 1194 } 1195 if (done == 0 && !retried) { 1196 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1197 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist); 1198 retried = true; 1199 goto restart; 1200 } 1201 return (done); 1202 } 1203 1204 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */ 1205 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free, 1206 0, 1207 "limit on vnode free requests per call to the vnlru_free routine"); 1208 1209 /* 1210 * Attempt to reduce the free list by the requested amount. 1211 */ 1212 static int 1213 vnlru_free_locked(int count, struct vfsops *mnt_op) 1214 { 1215 struct vnode *vp, *mvp; 1216 struct mount *mp; 1217 int ocount; 1218 1219 mtx_assert(&vnode_list_mtx, MA_OWNED); 1220 if (count > max_vnlru_free) 1221 count = max_vnlru_free; 1222 ocount = count; 1223 mvp = vnode_list_free_marker; 1224 restart: 1225 vp = mvp; 1226 while (count > 0) { 1227 vp = TAILQ_NEXT(vp, v_vnodelist); 1228 if (__predict_false(vp == NULL)) { 1229 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1230 TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist); 1231 break; 1232 } 1233 if (__predict_false(vp->v_type == VMARKER)) 1234 continue; 1235 1236 /* 1237 * Don't recycle if our vnode is from different type 1238 * of mount point. Note that mp is type-safe, the 1239 * check does not reach unmapped address even if 1240 * vnode is reclaimed. 1241 * Don't recycle if we can't get the interlock without 1242 * blocking. 1243 */ 1244 if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL && 1245 mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) { 1246 continue; 1247 } 1248 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1249 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1250 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) { 1251 VI_UNLOCK(vp); 1252 continue; 1253 } 1254 vholdl(vp); 1255 count--; 1256 mtx_unlock(&vnode_list_mtx); 1257 VI_UNLOCK(vp); 1258 vtryrecycle(vp); 1259 vdrop(vp); 1260 mtx_lock(&vnode_list_mtx); 1261 goto restart; 1262 } 1263 return (ocount - count); 1264 } 1265 1266 void 1267 vnlru_free(int count, struct vfsops *mnt_op) 1268 { 1269 1270 mtx_lock(&vnode_list_mtx); 1271 vnlru_free_locked(count, mnt_op); 1272 mtx_unlock(&vnode_list_mtx); 1273 } 1274 1275 static void 1276 vnlru_recalc(void) 1277 { 1278 1279 mtx_assert(&vnode_list_mtx, MA_OWNED); 1280 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100); 1281 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */ 1282 vlowat = vhiwat / 2; 1283 } 1284 1285 /* 1286 * Attempt to recycle vnodes in a context that is always safe to block. 1287 * Calling vlrurecycle() from the bowels of filesystem code has some 1288 * interesting deadlock problems. 1289 */ 1290 static struct proc *vnlruproc; 1291 static int vnlruproc_sig; 1292 1293 /* 1294 * The main freevnodes counter is only updated when threads requeue their vnode 1295 * batches. CPUs are conditionally walked to compute a more accurate total. 1296 * 1297 * Limit how much of a slop are we willing to tolerate. Note: the actual value 1298 * at any given moment can still exceed slop, but it should not be by significant 1299 * margin in practice. 1300 */ 1301 #define VNLRU_FREEVNODES_SLOP 128 1302 1303 static u_long 1304 vnlru_read_freevnodes(void) 1305 { 1306 struct vdbatch *vd; 1307 long slop; 1308 int cpu; 1309 1310 mtx_assert(&vnode_list_mtx, MA_OWNED); 1311 if (freevnodes > freevnodes_old) 1312 slop = freevnodes - freevnodes_old; 1313 else 1314 slop = freevnodes_old - freevnodes; 1315 if (slop < VNLRU_FREEVNODES_SLOP) 1316 return (freevnodes >= 0 ? freevnodes : 0); 1317 freevnodes_old = freevnodes; 1318 CPU_FOREACH(cpu) { 1319 vd = DPCPU_ID_PTR((cpu), vd); 1320 freevnodes_old += vd->freevnodes; 1321 } 1322 return (freevnodes_old >= 0 ? freevnodes_old : 0); 1323 } 1324 1325 static bool 1326 vnlru_under(u_long rnumvnodes, u_long limit) 1327 { 1328 u_long rfreevnodes, space; 1329 1330 if (__predict_false(rnumvnodes > desiredvnodes)) 1331 return (true); 1332 1333 space = desiredvnodes - rnumvnodes; 1334 if (space < limit) { 1335 rfreevnodes = vnlru_read_freevnodes(); 1336 if (rfreevnodes > wantfreevnodes) 1337 space += rfreevnodes - wantfreevnodes; 1338 } 1339 return (space < limit); 1340 } 1341 1342 static bool 1343 vnlru_under_unlocked(u_long rnumvnodes, u_long limit) 1344 { 1345 long rfreevnodes, space; 1346 1347 if (__predict_false(rnumvnodes > desiredvnodes)) 1348 return (true); 1349 1350 space = desiredvnodes - rnumvnodes; 1351 if (space < limit) { 1352 rfreevnodes = atomic_load_long(&freevnodes); 1353 if (rfreevnodes > wantfreevnodes) 1354 space += rfreevnodes - wantfreevnodes; 1355 } 1356 return (space < limit); 1357 } 1358 1359 static void 1360 vnlru_kick(void) 1361 { 1362 1363 mtx_assert(&vnode_list_mtx, MA_OWNED); 1364 if (vnlruproc_sig == 0) { 1365 vnlruproc_sig = 1; 1366 wakeup(vnlruproc); 1367 } 1368 } 1369 1370 static void 1371 vnlru_proc(void) 1372 { 1373 u_long rnumvnodes, rfreevnodes, target; 1374 unsigned long onumvnodes; 1375 int done, force, trigger, usevnodes; 1376 bool reclaim_nc_src, want_reread; 1377 1378 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc, 1379 SHUTDOWN_PRI_FIRST); 1380 1381 force = 0; 1382 want_reread = false; 1383 for (;;) { 1384 kproc_suspend_check(vnlruproc); 1385 mtx_lock(&vnode_list_mtx); 1386 rnumvnodes = atomic_load_long(&numvnodes); 1387 1388 if (want_reread) { 1389 force = vnlru_under(numvnodes, vhiwat) ? 1 : 0; 1390 want_reread = false; 1391 } 1392 1393 /* 1394 * If numvnodes is too large (due to desiredvnodes being 1395 * adjusted using its sysctl, or emergency growth), first 1396 * try to reduce it by discarding from the free list. 1397 */ 1398 if (rnumvnodes > desiredvnodes) { 1399 vnlru_free_locked(rnumvnodes - desiredvnodes, NULL); 1400 rnumvnodes = atomic_load_long(&numvnodes); 1401 } 1402 /* 1403 * Sleep if the vnode cache is in a good state. This is 1404 * when it is not over-full and has space for about a 4% 1405 * or 9% expansion (by growing its size or inexcessively 1406 * reducing its free list). Otherwise, try to reclaim 1407 * space for a 10% expansion. 1408 */ 1409 if (vstir && force == 0) { 1410 force = 1; 1411 vstir = 0; 1412 } 1413 if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) { 1414 vnlruproc_sig = 0; 1415 wakeup(&vnlruproc_sig); 1416 msleep(vnlruproc, &vnode_list_mtx, 1417 PVFS|PDROP, "vlruwt", hz); 1418 continue; 1419 } 1420 rfreevnodes = vnlru_read_freevnodes(); 1421 1422 onumvnodes = rnumvnodes; 1423 /* 1424 * Calculate parameters for recycling. These are the same 1425 * throughout the loop to give some semblance of fairness. 1426 * The trigger point is to avoid recycling vnodes with lots 1427 * of resident pages. We aren't trying to free memory; we 1428 * are trying to recycle or at least free vnodes. 1429 */ 1430 if (rnumvnodes <= desiredvnodes) 1431 usevnodes = rnumvnodes - rfreevnodes; 1432 else 1433 usevnodes = rnumvnodes; 1434 if (usevnodes <= 0) 1435 usevnodes = 1; 1436 /* 1437 * The trigger value is is chosen to give a conservatively 1438 * large value to ensure that it alone doesn't prevent 1439 * making progress. The value can easily be so large that 1440 * it is effectively infinite in some congested and 1441 * misconfigured cases, and this is necessary. Normally 1442 * it is about 8 to 100 (pages), which is quite large. 1443 */ 1444 trigger = vm_cnt.v_page_count * 2 / usevnodes; 1445 if (force < 2) 1446 trigger = vsmalltrigger; 1447 reclaim_nc_src = force >= 3; 1448 target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1); 1449 target = target / 10 + 1; 1450 done = vlrureclaim(reclaim_nc_src, trigger, target); 1451 mtx_unlock(&vnode_list_mtx); 1452 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes) 1453 uma_reclaim(UMA_RECLAIM_DRAIN); 1454 if (done == 0) { 1455 if (force == 0 || force == 1) { 1456 force = 2; 1457 continue; 1458 } 1459 if (force == 2) { 1460 force = 3; 1461 continue; 1462 } 1463 want_reread = true; 1464 force = 0; 1465 vnlru_nowhere++; 1466 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 1467 } else { 1468 want_reread = true; 1469 kern_yield(PRI_USER); 1470 } 1471 } 1472 } 1473 1474 static struct kproc_desc vnlru_kp = { 1475 "vnlru", 1476 vnlru_proc, 1477 &vnlruproc 1478 }; 1479 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, 1480 &vnlru_kp); 1481 1482 /* 1483 * Routines having to do with the management of the vnode table. 1484 */ 1485 1486 /* 1487 * Try to recycle a freed vnode. We abort if anyone picks up a reference 1488 * before we actually vgone(). This function must be called with the vnode 1489 * held to prevent the vnode from being returned to the free list midway 1490 * through vgone(). 1491 */ 1492 static int 1493 vtryrecycle(struct vnode *vp) 1494 { 1495 struct mount *vnmp; 1496 1497 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 1498 VNASSERT(vp->v_holdcnt, vp, 1499 ("vtryrecycle: Recycling vp %p without a reference.", vp)); 1500 /* 1501 * This vnode may found and locked via some other list, if so we 1502 * can't recycle it yet. 1503 */ 1504 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 1505 CTR2(KTR_VFS, 1506 "%s: impossible to recycle, vp %p lock is already held", 1507 __func__, vp); 1508 return (EWOULDBLOCK); 1509 } 1510 /* 1511 * Don't recycle if its filesystem is being suspended. 1512 */ 1513 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 1514 VOP_UNLOCK(vp); 1515 CTR2(KTR_VFS, 1516 "%s: impossible to recycle, cannot start the write for %p", 1517 __func__, vp); 1518 return (EBUSY); 1519 } 1520 /* 1521 * If we got this far, we need to acquire the interlock and see if 1522 * anyone picked up this vnode from another list. If not, we will 1523 * mark it with DOOMED via vgonel() so that anyone who does find it 1524 * will skip over it. 1525 */ 1526 VI_LOCK(vp); 1527 if (vp->v_usecount) { 1528 VOP_UNLOCK(vp); 1529 VI_UNLOCK(vp); 1530 vn_finished_write(vnmp); 1531 CTR2(KTR_VFS, 1532 "%s: impossible to recycle, %p is already referenced", 1533 __func__, vp); 1534 return (EBUSY); 1535 } 1536 if (!VN_IS_DOOMED(vp)) { 1537 counter_u64_add(recycles_free_count, 1); 1538 vgonel(vp); 1539 } 1540 VOP_UNLOCK(vp); 1541 VI_UNLOCK(vp); 1542 vn_finished_write(vnmp); 1543 return (0); 1544 } 1545 1546 /* 1547 * Allocate a new vnode. 1548 * 1549 * The operation never returns an error. Returning an error was disabled 1550 * in r145385 (dated 2005) with the following comment: 1551 * 1552 * XXX Not all VFS_VGET/ffs_vget callers check returns. 1553 * 1554 * Given the age of this commit (almost 15 years at the time of writing this 1555 * comment) restoring the ability to fail requires a significant audit of 1556 * all codepaths. 1557 * 1558 * The routine can try to free a vnode or stall for up to 1 second waiting for 1559 * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation. 1560 */ 1561 static u_long vn_alloc_cyclecount; 1562 1563 static struct vnode * __noinline 1564 vn_alloc_hard(struct mount *mp) 1565 { 1566 u_long rnumvnodes, rfreevnodes; 1567 1568 mtx_lock(&vnode_list_mtx); 1569 rnumvnodes = atomic_load_long(&numvnodes); 1570 if (rnumvnodes + 1 < desiredvnodes) { 1571 vn_alloc_cyclecount = 0; 1572 goto alloc; 1573 } 1574 rfreevnodes = vnlru_read_freevnodes(); 1575 if (vn_alloc_cyclecount++ >= rfreevnodes) { 1576 vn_alloc_cyclecount = 0; 1577 vstir = 1; 1578 } 1579 /* 1580 * Grow the vnode cache if it will not be above its target max 1581 * after growing. Otherwise, if the free list is nonempty, try 1582 * to reclaim 1 item from it before growing the cache (possibly 1583 * above its target max if the reclamation failed or is delayed). 1584 * Otherwise, wait for some space. In all cases, schedule 1585 * vnlru_proc() if we are getting short of space. The watermarks 1586 * should be chosen so that we never wait or even reclaim from 1587 * the free list to below its target minimum. 1588 */ 1589 if (vnlru_free_locked(1, NULL) > 0) 1590 goto alloc; 1591 if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 1592 /* 1593 * Wait for space for a new vnode. 1594 */ 1595 vnlru_kick(); 1596 msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz); 1597 if (atomic_load_long(&numvnodes) + 1 > desiredvnodes && 1598 vnlru_read_freevnodes() > 1) 1599 vnlru_free_locked(1, NULL); 1600 } 1601 alloc: 1602 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; 1603 if (vnlru_under(rnumvnodes, vlowat)) 1604 vnlru_kick(); 1605 mtx_unlock(&vnode_list_mtx); 1606 return (uma_zalloc(vnode_zone, M_WAITOK)); 1607 } 1608 1609 static struct vnode * 1610 vn_alloc(struct mount *mp) 1611 { 1612 u_long rnumvnodes; 1613 1614 if (__predict_false(vn_alloc_cyclecount != 0)) 1615 return (vn_alloc_hard(mp)); 1616 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; 1617 if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) { 1618 atomic_subtract_long(&numvnodes, 1); 1619 return (vn_alloc_hard(mp)); 1620 } 1621 1622 return (uma_zalloc(vnode_zone, M_WAITOK)); 1623 } 1624 1625 static void 1626 vn_free(struct vnode *vp) 1627 { 1628 1629 atomic_subtract_long(&numvnodes, 1); 1630 uma_zfree(vnode_zone, vp); 1631 } 1632 1633 /* 1634 * Return the next vnode from the free list. 1635 */ 1636 int 1637 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, 1638 struct vnode **vpp) 1639 { 1640 struct vnode *vp; 1641 struct thread *td; 1642 struct lock_object *lo; 1643 1644 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag); 1645 1646 KASSERT(vops->registered, 1647 ("%s: not registered vector op %p\n", __func__, vops)); 1648 1649 td = curthread; 1650 if (td->td_vp_reserved != NULL) { 1651 vp = td->td_vp_reserved; 1652 td->td_vp_reserved = NULL; 1653 } else { 1654 vp = vn_alloc(mp); 1655 } 1656 counter_u64_add(vnodes_created, 1); 1657 /* 1658 * Locks are given the generic name "vnode" when created. 1659 * Follow the historic practice of using the filesystem 1660 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc. 1661 * 1662 * Locks live in a witness group keyed on their name. Thus, 1663 * when a lock is renamed, it must also move from the witness 1664 * group of its old name to the witness group of its new name. 1665 * 1666 * The change only needs to be made when the vnode moves 1667 * from one filesystem type to another. We ensure that each 1668 * filesystem use a single static name pointer for its tag so 1669 * that we can compare pointers rather than doing a strcmp(). 1670 */ 1671 lo = &vp->v_vnlock->lock_object; 1672 #ifdef WITNESS 1673 if (lo->lo_name != tag) { 1674 #endif 1675 lo->lo_name = tag; 1676 #ifdef WITNESS 1677 WITNESS_DESTROY(lo); 1678 WITNESS_INIT(lo, tag); 1679 } 1680 #endif 1681 /* 1682 * By default, don't allow shared locks unless filesystems opt-in. 1683 */ 1684 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE; 1685 /* 1686 * Finalize various vnode identity bits. 1687 */ 1688 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp)); 1689 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp)); 1690 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp)); 1691 vp->v_type = VNON; 1692 vp->v_op = vops; 1693 v_init_counters(vp); 1694 vp->v_bufobj.bo_ops = &buf_ops_bio; 1695 #ifdef DIAGNOSTIC 1696 if (mp == NULL && vops != &dead_vnodeops) 1697 printf("NULL mp in getnewvnode(9), tag %s\n", tag); 1698 #endif 1699 #ifdef MAC 1700 mac_vnode_init(vp); 1701 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 1702 mac_vnode_associate_singlelabel(mp, vp); 1703 #endif 1704 if (mp != NULL) { 1705 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize; 1706 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) 1707 vp->v_vflag |= VV_NOKNOTE; 1708 } 1709 1710 /* 1711 * For the filesystems which do not use vfs_hash_insert(), 1712 * still initialize v_hash to have vfs_hash_index() useful. 1713 * E.g., nullfs uses vfs_hash_index() on the lower vnode for 1714 * its own hashing. 1715 */ 1716 vp->v_hash = (uintptr_t)vp >> vnsz2log; 1717 1718 *vpp = vp; 1719 return (0); 1720 } 1721 1722 void 1723 getnewvnode_reserve(void) 1724 { 1725 struct thread *td; 1726 1727 td = curthread; 1728 MPASS(td->td_vp_reserved == NULL); 1729 td->td_vp_reserved = vn_alloc(NULL); 1730 } 1731 1732 void 1733 getnewvnode_drop_reserve(void) 1734 { 1735 struct thread *td; 1736 1737 td = curthread; 1738 if (td->td_vp_reserved != NULL) { 1739 vn_free(td->td_vp_reserved); 1740 td->td_vp_reserved = NULL; 1741 } 1742 } 1743 1744 static void 1745 freevnode(struct vnode *vp) 1746 { 1747 struct bufobj *bo; 1748 1749 /* 1750 * The vnode has been marked for destruction, so free it. 1751 * 1752 * The vnode will be returned to the zone where it will 1753 * normally remain until it is needed for another vnode. We 1754 * need to cleanup (or verify that the cleanup has already 1755 * been done) any residual data left from its current use 1756 * so as not to contaminate the freshly allocated vnode. 1757 */ 1758 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp); 1759 bo = &vp->v_bufobj; 1760 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 1761 VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count")); 1762 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 1763 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 1764 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 1765 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 1766 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp, 1767 ("clean blk trie not empty")); 1768 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 1769 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp, 1770 ("dirty blk trie not empty")); 1771 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); 1772 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); 1773 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for ..")); 1774 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp, 1775 ("Dangling rangelock waiters")); 1776 VI_UNLOCK(vp); 1777 #ifdef MAC 1778 mac_vnode_destroy(vp); 1779 #endif 1780 if (vp->v_pollinfo != NULL) { 1781 destroy_vpollinfo(vp->v_pollinfo); 1782 vp->v_pollinfo = NULL; 1783 } 1784 #ifdef INVARIANTS 1785 /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */ 1786 vp->v_op = NULL; 1787 #endif 1788 vp->v_mountedhere = NULL; 1789 vp->v_unpcb = NULL; 1790 vp->v_rdev = NULL; 1791 vp->v_fifoinfo = NULL; 1792 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 1793 vp->v_irflag = 0; 1794 vp->v_iflag = 0; 1795 vp->v_vflag = 0; 1796 bo->bo_flag = 0; 1797 vn_free(vp); 1798 } 1799 1800 /* 1801 * Delete from old mount point vnode list, if on one. 1802 */ 1803 static void 1804 delmntque(struct vnode *vp) 1805 { 1806 struct mount *mp; 1807 1808 mp = vp->v_mount; 1809 if (mp == NULL) 1810 return; 1811 MNT_ILOCK(mp); 1812 VI_LOCK(vp); 1813 if (vp->v_mflag & VMP_LAZYLIST) { 1814 mtx_lock(&mp->mnt_listmtx); 1815 if (vp->v_mflag & VMP_LAZYLIST) { 1816 vp->v_mflag &= ~VMP_LAZYLIST; 1817 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 1818 mp->mnt_lazyvnodelistsize--; 1819 } 1820 mtx_unlock(&mp->mnt_listmtx); 1821 } 1822 vp->v_mount = NULL; 1823 VI_UNLOCK(vp); 1824 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 1825 ("bad mount point vnode list size")); 1826 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1827 mp->mnt_nvnodelistsize--; 1828 MNT_REL(mp); 1829 MNT_IUNLOCK(mp); 1830 } 1831 1832 static void 1833 insmntque_stddtr(struct vnode *vp, void *dtr_arg) 1834 { 1835 1836 vp->v_data = NULL; 1837 vp->v_op = &dead_vnodeops; 1838 vgone(vp); 1839 vput(vp); 1840 } 1841 1842 /* 1843 * Insert into list of vnodes for the new mount point, if available. 1844 */ 1845 int 1846 insmntque1(struct vnode *vp, struct mount *mp, 1847 void (*dtr)(struct vnode *, void *), void *dtr_arg) 1848 { 1849 1850 KASSERT(vp->v_mount == NULL, 1851 ("insmntque: vnode already on per mount vnode list")); 1852 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 1853 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); 1854 1855 /* 1856 * We acquire the vnode interlock early to ensure that the 1857 * vnode cannot be recycled by another process releasing a 1858 * holdcnt on it before we get it on both the vnode list 1859 * and the active vnode list. The mount mutex protects only 1860 * manipulation of the vnode list and the vnode freelist 1861 * mutex protects only manipulation of the active vnode list. 1862 * Hence the need to hold the vnode interlock throughout. 1863 */ 1864 MNT_ILOCK(mp); 1865 VI_LOCK(vp); 1866 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 && 1867 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || 1868 mp->mnt_nvnodelistsize == 0)) && 1869 (vp->v_vflag & VV_FORCEINSMQ) == 0) { 1870 VI_UNLOCK(vp); 1871 MNT_IUNLOCK(mp); 1872 if (dtr != NULL) 1873 dtr(vp, dtr_arg); 1874 return (EBUSY); 1875 } 1876 vp->v_mount = mp; 1877 MNT_REF(mp); 1878 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1879 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 1880 ("neg mount point vnode list size")); 1881 mp->mnt_nvnodelistsize++; 1882 VI_UNLOCK(vp); 1883 MNT_IUNLOCK(mp); 1884 return (0); 1885 } 1886 1887 int 1888 insmntque(struct vnode *vp, struct mount *mp) 1889 { 1890 1891 return (insmntque1(vp, mp, insmntque_stddtr, NULL)); 1892 } 1893 1894 /* 1895 * Flush out and invalidate all buffers associated with a bufobj 1896 * Called with the underlying object locked. 1897 */ 1898 int 1899 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) 1900 { 1901 int error; 1902 1903 BO_LOCK(bo); 1904 if (flags & V_SAVE) { 1905 error = bufobj_wwait(bo, slpflag, slptimeo); 1906 if (error) { 1907 BO_UNLOCK(bo); 1908 return (error); 1909 } 1910 if (bo->bo_dirty.bv_cnt > 0) { 1911 BO_UNLOCK(bo); 1912 if ((error = BO_SYNC(bo, MNT_WAIT)) != 0) 1913 return (error); 1914 /* 1915 * XXX We could save a lock/unlock if this was only 1916 * enabled under INVARIANTS 1917 */ 1918 BO_LOCK(bo); 1919 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 1920 panic("vinvalbuf: dirty bufs"); 1921 } 1922 } 1923 /* 1924 * If you alter this loop please notice that interlock is dropped and 1925 * reacquired in flushbuflist. Special care is needed to ensure that 1926 * no race conditions occur from this. 1927 */ 1928 do { 1929 error = flushbuflist(&bo->bo_clean, 1930 flags, bo, slpflag, slptimeo); 1931 if (error == 0 && !(flags & V_CLEANONLY)) 1932 error = flushbuflist(&bo->bo_dirty, 1933 flags, bo, slpflag, slptimeo); 1934 if (error != 0 && error != EAGAIN) { 1935 BO_UNLOCK(bo); 1936 return (error); 1937 } 1938 } while (error != 0); 1939 1940 /* 1941 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1942 * have write I/O in-progress but if there is a VM object then the 1943 * VM object can also have read-I/O in-progress. 1944 */ 1945 do { 1946 bufobj_wwait(bo, 0, 0); 1947 if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) { 1948 BO_UNLOCK(bo); 1949 vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx"); 1950 BO_LOCK(bo); 1951 } 1952 } while (bo->bo_numoutput > 0); 1953 BO_UNLOCK(bo); 1954 1955 /* 1956 * Destroy the copy in the VM cache, too. 1957 */ 1958 if (bo->bo_object != NULL && 1959 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) { 1960 VM_OBJECT_WLOCK(bo->bo_object); 1961 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? 1962 OBJPR_CLEANONLY : 0); 1963 VM_OBJECT_WUNLOCK(bo->bo_object); 1964 } 1965 1966 #ifdef INVARIANTS 1967 BO_LOCK(bo); 1968 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO | 1969 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 || 1970 bo->bo_clean.bv_cnt > 0)) 1971 panic("vinvalbuf: flush failed"); 1972 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 && 1973 bo->bo_dirty.bv_cnt > 0) 1974 panic("vinvalbuf: flush dirty failed"); 1975 BO_UNLOCK(bo); 1976 #endif 1977 return (0); 1978 } 1979 1980 /* 1981 * Flush out and invalidate all buffers associated with a vnode. 1982 * Called with the underlying object locked. 1983 */ 1984 int 1985 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 1986 { 1987 1988 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 1989 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1990 if (vp->v_object != NULL && vp->v_object->handle != vp) 1991 return (0); 1992 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); 1993 } 1994 1995 /* 1996 * Flush out buffers on the specified list. 1997 * 1998 */ 1999 static int 2000 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 2001 int slptimeo) 2002 { 2003 struct buf *bp, *nbp; 2004 int retval, error; 2005 daddr_t lblkno; 2006 b_xflags_t xflags; 2007 2008 ASSERT_BO_WLOCKED(bo); 2009 2010 retval = 0; 2011 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 2012 /* 2013 * If we are flushing both V_NORMAL and V_ALT buffers then 2014 * do not skip any buffers. If we are flushing only V_NORMAL 2015 * buffers then skip buffers marked as BX_ALTDATA. If we are 2016 * flushing only V_ALT buffers then skip buffers not marked 2017 * as BX_ALTDATA. 2018 */ 2019 if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) && 2020 (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) || 2021 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) { 2022 continue; 2023 } 2024 if (nbp != NULL) { 2025 lblkno = nbp->b_lblkno; 2026 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN); 2027 } 2028 retval = EAGAIN; 2029 error = BUF_TIMELOCK(bp, 2030 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), 2031 "flushbuf", slpflag, slptimeo); 2032 if (error) { 2033 BO_LOCK(bo); 2034 return (error != ENOLCK ? error : EAGAIN); 2035 } 2036 KASSERT(bp->b_bufobj == bo, 2037 ("bp %p wrong b_bufobj %p should be %p", 2038 bp, bp->b_bufobj, bo)); 2039 /* 2040 * XXX Since there are no node locks for NFS, I 2041 * believe there is a slight chance that a delayed 2042 * write will occur while sleeping just above, so 2043 * check for it. 2044 */ 2045 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 2046 (flags & V_SAVE)) { 2047 bremfree(bp); 2048 bp->b_flags |= B_ASYNC; 2049 bwrite(bp); 2050 BO_LOCK(bo); 2051 return (EAGAIN); /* XXX: why not loop ? */ 2052 } 2053 bremfree(bp); 2054 bp->b_flags |= (B_INVAL | B_RELBUF); 2055 bp->b_flags &= ~B_ASYNC; 2056 brelse(bp); 2057 BO_LOCK(bo); 2058 if (nbp == NULL) 2059 break; 2060 nbp = gbincore(bo, lblkno); 2061 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2062 != xflags) 2063 break; /* nbp invalid */ 2064 } 2065 return (retval); 2066 } 2067 2068 int 2069 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn) 2070 { 2071 struct buf *bp; 2072 int error; 2073 daddr_t lblkno; 2074 2075 ASSERT_BO_LOCKED(bo); 2076 2077 for (lblkno = startn;;) { 2078 again: 2079 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno); 2080 if (bp == NULL || bp->b_lblkno >= endn || 2081 bp->b_lblkno < startn) 2082 break; 2083 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 2084 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0); 2085 if (error != 0) { 2086 BO_RLOCK(bo); 2087 if (error == ENOLCK) 2088 goto again; 2089 return (error); 2090 } 2091 KASSERT(bp->b_bufobj == bo, 2092 ("bp %p wrong b_bufobj %p should be %p", 2093 bp, bp->b_bufobj, bo)); 2094 lblkno = bp->b_lblkno + 1; 2095 if ((bp->b_flags & B_MANAGED) == 0) 2096 bremfree(bp); 2097 bp->b_flags |= B_RELBUF; 2098 /* 2099 * In the VMIO case, use the B_NOREUSE flag to hint that the 2100 * pages backing each buffer in the range are unlikely to be 2101 * reused. Dirty buffers will have the hint applied once 2102 * they've been written. 2103 */ 2104 if ((bp->b_flags & B_VMIO) != 0) 2105 bp->b_flags |= B_NOREUSE; 2106 brelse(bp); 2107 BO_RLOCK(bo); 2108 } 2109 return (0); 2110 } 2111 2112 /* 2113 * Truncate a file's buffer and pages to a specified length. This 2114 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 2115 * sync activity. 2116 */ 2117 int 2118 vtruncbuf(struct vnode *vp, off_t length, int blksize) 2119 { 2120 struct buf *bp, *nbp; 2121 struct bufobj *bo; 2122 daddr_t startlbn; 2123 2124 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__, 2125 vp, blksize, (uintmax_t)length); 2126 2127 /* 2128 * Round up to the *next* lbn. 2129 */ 2130 startlbn = howmany(length, blksize); 2131 2132 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 2133 2134 bo = &vp->v_bufobj; 2135 restart_unlocked: 2136 BO_LOCK(bo); 2137 2138 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN) 2139 ; 2140 2141 if (length > 0) { 2142 restartsync: 2143 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2144 if (bp->b_lblkno > 0) 2145 continue; 2146 /* 2147 * Since we hold the vnode lock this should only 2148 * fail if we're racing with the buf daemon. 2149 */ 2150 if (BUF_LOCK(bp, 2151 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2152 BO_LOCKPTR(bo)) == ENOLCK) 2153 goto restart_unlocked; 2154 2155 VNASSERT((bp->b_flags & B_DELWRI), vp, 2156 ("buf(%p) on dirty queue without DELWRI", bp)); 2157 2158 bremfree(bp); 2159 bawrite(bp); 2160 BO_LOCK(bo); 2161 goto restartsync; 2162 } 2163 } 2164 2165 bufobj_wwait(bo, 0, 0); 2166 BO_UNLOCK(bo); 2167 vnode_pager_setsize(vp, length); 2168 2169 return (0); 2170 } 2171 2172 /* 2173 * Invalidate the cached pages of a file's buffer within the range of block 2174 * numbers [startlbn, endlbn). 2175 */ 2176 void 2177 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn, 2178 int blksize) 2179 { 2180 struct bufobj *bo; 2181 off_t start, end; 2182 2183 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range"); 2184 2185 start = blksize * startlbn; 2186 end = blksize * endlbn; 2187 2188 bo = &vp->v_bufobj; 2189 BO_LOCK(bo); 2190 MPASS(blksize == bo->bo_bsize); 2191 2192 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN) 2193 ; 2194 2195 BO_UNLOCK(bo); 2196 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1)); 2197 } 2198 2199 static int 2200 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, 2201 daddr_t startlbn, daddr_t endlbn) 2202 { 2203 struct buf *bp, *nbp; 2204 bool anyfreed; 2205 2206 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked"); 2207 ASSERT_BO_LOCKED(bo); 2208 2209 do { 2210 anyfreed = false; 2211 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 2212 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) 2213 continue; 2214 if (BUF_LOCK(bp, 2215 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2216 BO_LOCKPTR(bo)) == ENOLCK) { 2217 BO_LOCK(bo); 2218 return (EAGAIN); 2219 } 2220 2221 bremfree(bp); 2222 bp->b_flags |= B_INVAL | B_RELBUF; 2223 bp->b_flags &= ~B_ASYNC; 2224 brelse(bp); 2225 anyfreed = true; 2226 2227 BO_LOCK(bo); 2228 if (nbp != NULL && 2229 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 2230 nbp->b_vp != vp || 2231 (nbp->b_flags & B_DELWRI) != 0)) 2232 return (EAGAIN); 2233 } 2234 2235 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2236 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) 2237 continue; 2238 if (BUF_LOCK(bp, 2239 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2240 BO_LOCKPTR(bo)) == ENOLCK) { 2241 BO_LOCK(bo); 2242 return (EAGAIN); 2243 } 2244 bremfree(bp); 2245 bp->b_flags |= B_INVAL | B_RELBUF; 2246 bp->b_flags &= ~B_ASYNC; 2247 brelse(bp); 2248 anyfreed = true; 2249 2250 BO_LOCK(bo); 2251 if (nbp != NULL && 2252 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 2253 (nbp->b_vp != vp) || 2254 (nbp->b_flags & B_DELWRI) == 0)) 2255 return (EAGAIN); 2256 } 2257 } while (anyfreed); 2258 return (0); 2259 } 2260 2261 static void 2262 buf_vlist_remove(struct buf *bp) 2263 { 2264 struct bufv *bv; 2265 2266 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2267 ASSERT_BO_WLOCKED(bp->b_bufobj); 2268 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) != 2269 (BX_VNDIRTY|BX_VNCLEAN), 2270 ("buf_vlist_remove: Buf %p is on two lists", bp)); 2271 if (bp->b_xflags & BX_VNDIRTY) 2272 bv = &bp->b_bufobj->bo_dirty; 2273 else 2274 bv = &bp->b_bufobj->bo_clean; 2275 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno); 2276 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 2277 bv->bv_cnt--; 2278 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 2279 } 2280 2281 /* 2282 * Add the buffer to the sorted clean or dirty block list. 2283 * 2284 * NOTE: xflags is passed as a constant, optimizing this inline function! 2285 */ 2286 static void 2287 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 2288 { 2289 struct bufv *bv; 2290 struct buf *n; 2291 int error; 2292 2293 ASSERT_BO_WLOCKED(bo); 2294 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0, 2295 ("dead bo %p", bo)); 2296 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 2297 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 2298 bp->b_xflags |= xflags; 2299 if (xflags & BX_VNDIRTY) 2300 bv = &bo->bo_dirty; 2301 else 2302 bv = &bo->bo_clean; 2303 2304 /* 2305 * Keep the list ordered. Optimize empty list insertion. Assume 2306 * we tend to grow at the tail so lookup_le should usually be cheaper 2307 * than _ge. 2308 */ 2309 if (bv->bv_cnt == 0 || 2310 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno) 2311 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 2312 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL) 2313 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs); 2314 else 2315 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs); 2316 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp); 2317 if (error) 2318 panic("buf_vlist_add: Preallocated nodes insufficient."); 2319 bv->bv_cnt++; 2320 } 2321 2322 /* 2323 * Look up a buffer using the buffer tries. 2324 */ 2325 struct buf * 2326 gbincore(struct bufobj *bo, daddr_t lblkno) 2327 { 2328 struct buf *bp; 2329 2330 ASSERT_BO_LOCKED(bo); 2331 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno); 2332 if (bp != NULL) 2333 return (bp); 2334 return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno); 2335 } 2336 2337 /* 2338 * Associate a buffer with a vnode. 2339 */ 2340 void 2341 bgetvp(struct vnode *vp, struct buf *bp) 2342 { 2343 struct bufobj *bo; 2344 2345 bo = &vp->v_bufobj; 2346 ASSERT_BO_WLOCKED(bo); 2347 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 2348 2349 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 2350 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 2351 ("bgetvp: bp already attached! %p", bp)); 2352 2353 vhold(vp); 2354 bp->b_vp = vp; 2355 bp->b_bufobj = bo; 2356 /* 2357 * Insert onto list for new vnode. 2358 */ 2359 buf_vlist_add(bp, bo, BX_VNCLEAN); 2360 } 2361 2362 /* 2363 * Disassociate a buffer from a vnode. 2364 */ 2365 void 2366 brelvp(struct buf *bp) 2367 { 2368 struct bufobj *bo; 2369 struct vnode *vp; 2370 2371 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2372 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 2373 2374 /* 2375 * Delete from old vnode list, if on one. 2376 */ 2377 vp = bp->b_vp; /* XXX */ 2378 bo = bp->b_bufobj; 2379 BO_LOCK(bo); 2380 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2381 buf_vlist_remove(bp); 2382 else 2383 panic("brelvp: Buffer %p not on queue.", bp); 2384 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2385 bo->bo_flag &= ~BO_ONWORKLST; 2386 mtx_lock(&sync_mtx); 2387 LIST_REMOVE(bo, bo_synclist); 2388 syncer_worklist_len--; 2389 mtx_unlock(&sync_mtx); 2390 } 2391 bp->b_vp = NULL; 2392 bp->b_bufobj = NULL; 2393 BO_UNLOCK(bo); 2394 vdrop(vp); 2395 } 2396 2397 /* 2398 * Add an item to the syncer work queue. 2399 */ 2400 static void 2401 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 2402 { 2403 int slot; 2404 2405 ASSERT_BO_WLOCKED(bo); 2406 2407 mtx_lock(&sync_mtx); 2408 if (bo->bo_flag & BO_ONWORKLST) 2409 LIST_REMOVE(bo, bo_synclist); 2410 else { 2411 bo->bo_flag |= BO_ONWORKLST; 2412 syncer_worklist_len++; 2413 } 2414 2415 if (delay > syncer_maxdelay - 2) 2416 delay = syncer_maxdelay - 2; 2417 slot = (syncer_delayno + delay) & syncer_mask; 2418 2419 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 2420 mtx_unlock(&sync_mtx); 2421 } 2422 2423 static int 2424 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 2425 { 2426 int error, len; 2427 2428 mtx_lock(&sync_mtx); 2429 len = syncer_worklist_len - sync_vnode_count; 2430 mtx_unlock(&sync_mtx); 2431 error = SYSCTL_OUT(req, &len, sizeof(len)); 2432 return (error); 2433 } 2434 2435 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, 2436 CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0, 2437 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 2438 2439 static struct proc *updateproc; 2440 static void sched_sync(void); 2441 static struct kproc_desc up_kp = { 2442 "syncer", 2443 sched_sync, 2444 &updateproc 2445 }; 2446 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); 2447 2448 static int 2449 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) 2450 { 2451 struct vnode *vp; 2452 struct mount *mp; 2453 2454 *bo = LIST_FIRST(slp); 2455 if (*bo == NULL) 2456 return (0); 2457 vp = bo2vnode(*bo); 2458 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) 2459 return (1); 2460 /* 2461 * We use vhold in case the vnode does not 2462 * successfully sync. vhold prevents the vnode from 2463 * going away when we unlock the sync_mtx so that 2464 * we can acquire the vnode interlock. 2465 */ 2466 vholdl(vp); 2467 mtx_unlock(&sync_mtx); 2468 VI_UNLOCK(vp); 2469 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2470 vdrop(vp); 2471 mtx_lock(&sync_mtx); 2472 return (*bo == LIST_FIRST(slp)); 2473 } 2474 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2475 (void) VOP_FSYNC(vp, MNT_LAZY, td); 2476 VOP_UNLOCK(vp); 2477 vn_finished_write(mp); 2478 BO_LOCK(*bo); 2479 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { 2480 /* 2481 * Put us back on the worklist. The worklist 2482 * routine will remove us from our current 2483 * position and then add us back in at a later 2484 * position. 2485 */ 2486 vn_syncer_add_to_worklist(*bo, syncdelay); 2487 } 2488 BO_UNLOCK(*bo); 2489 vdrop(vp); 2490 mtx_lock(&sync_mtx); 2491 return (0); 2492 } 2493 2494 static int first_printf = 1; 2495 2496 /* 2497 * System filesystem synchronizer daemon. 2498 */ 2499 static void 2500 sched_sync(void) 2501 { 2502 struct synclist *next, *slp; 2503 struct bufobj *bo; 2504 long starttime; 2505 struct thread *td = curthread; 2506 int last_work_seen; 2507 int net_worklist_len; 2508 int syncer_final_iter; 2509 int error; 2510 2511 last_work_seen = 0; 2512 syncer_final_iter = 0; 2513 syncer_state = SYNCER_RUNNING; 2514 starttime = time_uptime; 2515 td->td_pflags |= TDP_NORUNNINGBUF; 2516 2517 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 2518 SHUTDOWN_PRI_LAST); 2519 2520 mtx_lock(&sync_mtx); 2521 for (;;) { 2522 if (syncer_state == SYNCER_FINAL_DELAY && 2523 syncer_final_iter == 0) { 2524 mtx_unlock(&sync_mtx); 2525 kproc_suspend_check(td->td_proc); 2526 mtx_lock(&sync_mtx); 2527 } 2528 net_worklist_len = syncer_worklist_len - sync_vnode_count; 2529 if (syncer_state != SYNCER_RUNNING && 2530 starttime != time_uptime) { 2531 if (first_printf) { 2532 printf("\nSyncing disks, vnodes remaining... "); 2533 first_printf = 0; 2534 } 2535 printf("%d ", net_worklist_len); 2536 } 2537 starttime = time_uptime; 2538 2539 /* 2540 * Push files whose dirty time has expired. Be careful 2541 * of interrupt race on slp queue. 2542 * 2543 * Skip over empty worklist slots when shutting down. 2544 */ 2545 do { 2546 slp = &syncer_workitem_pending[syncer_delayno]; 2547 syncer_delayno += 1; 2548 if (syncer_delayno == syncer_maxdelay) 2549 syncer_delayno = 0; 2550 next = &syncer_workitem_pending[syncer_delayno]; 2551 /* 2552 * If the worklist has wrapped since the 2553 * it was emptied of all but syncer vnodes, 2554 * switch to the FINAL_DELAY state and run 2555 * for one more second. 2556 */ 2557 if (syncer_state == SYNCER_SHUTTING_DOWN && 2558 net_worklist_len == 0 && 2559 last_work_seen == syncer_delayno) { 2560 syncer_state = SYNCER_FINAL_DELAY; 2561 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 2562 } 2563 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 2564 syncer_worklist_len > 0); 2565 2566 /* 2567 * Keep track of the last time there was anything 2568 * on the worklist other than syncer vnodes. 2569 * Return to the SHUTTING_DOWN state if any 2570 * new work appears. 2571 */ 2572 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 2573 last_work_seen = syncer_delayno; 2574 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 2575 syncer_state = SYNCER_SHUTTING_DOWN; 2576 while (!LIST_EMPTY(slp)) { 2577 error = sync_vnode(slp, &bo, td); 2578 if (error == 1) { 2579 LIST_REMOVE(bo, bo_synclist); 2580 LIST_INSERT_HEAD(next, bo, bo_synclist); 2581 continue; 2582 } 2583 2584 if (first_printf == 0) { 2585 /* 2586 * Drop the sync mutex, because some watchdog 2587 * drivers need to sleep while patting 2588 */ 2589 mtx_unlock(&sync_mtx); 2590 wdog_kern_pat(WD_LASTVAL); 2591 mtx_lock(&sync_mtx); 2592 } 2593 2594 } 2595 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 2596 syncer_final_iter--; 2597 /* 2598 * The variable rushjob allows the kernel to speed up the 2599 * processing of the filesystem syncer process. A rushjob 2600 * value of N tells the filesystem syncer to process the next 2601 * N seconds worth of work on its queue ASAP. Currently rushjob 2602 * is used by the soft update code to speed up the filesystem 2603 * syncer process when the incore state is getting so far 2604 * ahead of the disk that the kernel memory pool is being 2605 * threatened with exhaustion. 2606 */ 2607 if (rushjob > 0) { 2608 rushjob -= 1; 2609 continue; 2610 } 2611 /* 2612 * Just sleep for a short period of time between 2613 * iterations when shutting down to allow some I/O 2614 * to happen. 2615 * 2616 * If it has taken us less than a second to process the 2617 * current work, then wait. Otherwise start right over 2618 * again. We can still lose time if any single round 2619 * takes more than two seconds, but it does not really 2620 * matter as we are just trying to generally pace the 2621 * filesystem activity. 2622 */ 2623 if (syncer_state != SYNCER_RUNNING || 2624 time_uptime == starttime) { 2625 thread_lock(td); 2626 sched_prio(td, PPAUSE); 2627 thread_unlock(td); 2628 } 2629 if (syncer_state != SYNCER_RUNNING) 2630 cv_timedwait(&sync_wakeup, &sync_mtx, 2631 hz / SYNCER_SHUTDOWN_SPEEDUP); 2632 else if (time_uptime == starttime) 2633 cv_timedwait(&sync_wakeup, &sync_mtx, hz); 2634 } 2635 } 2636 2637 /* 2638 * Request the syncer daemon to speed up its work. 2639 * We never push it to speed up more than half of its 2640 * normal turn time, otherwise it could take over the cpu. 2641 */ 2642 int 2643 speedup_syncer(void) 2644 { 2645 int ret = 0; 2646 2647 mtx_lock(&sync_mtx); 2648 if (rushjob < syncdelay / 2) { 2649 rushjob += 1; 2650 stat_rush_requests += 1; 2651 ret = 1; 2652 } 2653 mtx_unlock(&sync_mtx); 2654 cv_broadcast(&sync_wakeup); 2655 return (ret); 2656 } 2657 2658 /* 2659 * Tell the syncer to speed up its work and run though its work 2660 * list several times, then tell it to shut down. 2661 */ 2662 static void 2663 syncer_shutdown(void *arg, int howto) 2664 { 2665 2666 if (howto & RB_NOSYNC) 2667 return; 2668 mtx_lock(&sync_mtx); 2669 syncer_state = SYNCER_SHUTTING_DOWN; 2670 rushjob = 0; 2671 mtx_unlock(&sync_mtx); 2672 cv_broadcast(&sync_wakeup); 2673 kproc_shutdown(arg, howto); 2674 } 2675 2676 void 2677 syncer_suspend(void) 2678 { 2679 2680 syncer_shutdown(updateproc, 0); 2681 } 2682 2683 void 2684 syncer_resume(void) 2685 { 2686 2687 mtx_lock(&sync_mtx); 2688 first_printf = 1; 2689 syncer_state = SYNCER_RUNNING; 2690 mtx_unlock(&sync_mtx); 2691 cv_broadcast(&sync_wakeup); 2692 kproc_resume(updateproc); 2693 } 2694 2695 /* 2696 * Reassign a buffer from one vnode to another. 2697 * Used to assign file specific control information 2698 * (indirect blocks) to the vnode to which they belong. 2699 */ 2700 void 2701 reassignbuf(struct buf *bp) 2702 { 2703 struct vnode *vp; 2704 struct bufobj *bo; 2705 int delay; 2706 #ifdef INVARIANTS 2707 struct bufv *bv; 2708 #endif 2709 2710 vp = bp->b_vp; 2711 bo = bp->b_bufobj; 2712 ++reassignbufcalls; 2713 2714 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 2715 bp, bp->b_vp, bp->b_flags); 2716 /* 2717 * B_PAGING flagged buffers cannot be reassigned because their vp 2718 * is not fully linked in. 2719 */ 2720 if (bp->b_flags & B_PAGING) 2721 panic("cannot reassign paging buffer"); 2722 2723 /* 2724 * Delete from old vnode list, if on one. 2725 */ 2726 BO_LOCK(bo); 2727 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2728 buf_vlist_remove(bp); 2729 else 2730 panic("reassignbuf: Buffer %p not on queue.", bp); 2731 /* 2732 * If dirty, put on list of dirty buffers; otherwise insert onto list 2733 * of clean buffers. 2734 */ 2735 if (bp->b_flags & B_DELWRI) { 2736 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 2737 switch (vp->v_type) { 2738 case VDIR: 2739 delay = dirdelay; 2740 break; 2741 case VCHR: 2742 delay = metadelay; 2743 break; 2744 default: 2745 delay = filedelay; 2746 } 2747 vn_syncer_add_to_worklist(bo, delay); 2748 } 2749 buf_vlist_add(bp, bo, BX_VNDIRTY); 2750 } else { 2751 buf_vlist_add(bp, bo, BX_VNCLEAN); 2752 2753 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2754 mtx_lock(&sync_mtx); 2755 LIST_REMOVE(bo, bo_synclist); 2756 syncer_worklist_len--; 2757 mtx_unlock(&sync_mtx); 2758 bo->bo_flag &= ~BO_ONWORKLST; 2759 } 2760 } 2761 #ifdef INVARIANTS 2762 bv = &bo->bo_clean; 2763 bp = TAILQ_FIRST(&bv->bv_hd); 2764 KASSERT(bp == NULL || bp->b_bufobj == bo, 2765 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2766 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2767 KASSERT(bp == NULL || bp->b_bufobj == bo, 2768 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2769 bv = &bo->bo_dirty; 2770 bp = TAILQ_FIRST(&bv->bv_hd); 2771 KASSERT(bp == NULL || bp->b_bufobj == bo, 2772 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2773 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2774 KASSERT(bp == NULL || bp->b_bufobj == bo, 2775 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2776 #endif 2777 BO_UNLOCK(bo); 2778 } 2779 2780 static void 2781 v_init_counters(struct vnode *vp) 2782 { 2783 2784 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0, 2785 vp, ("%s called for an initialized vnode", __FUNCTION__)); 2786 ASSERT_VI_UNLOCKED(vp, __FUNCTION__); 2787 2788 refcount_init(&vp->v_holdcnt, 1); 2789 refcount_init(&vp->v_usecount, 1); 2790 } 2791 2792 /* 2793 * Increment si_usecount of the associated device, if any. 2794 */ 2795 static void 2796 v_incr_devcount(struct vnode *vp) 2797 { 2798 2799 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2800 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2801 dev_lock(); 2802 vp->v_rdev->si_usecount++; 2803 dev_unlock(); 2804 } 2805 } 2806 2807 /* 2808 * Decrement si_usecount of the associated device, if any. 2809 */ 2810 static void 2811 v_decr_devcount(struct vnode *vp) 2812 { 2813 2814 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2815 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2816 dev_lock(); 2817 vp->v_rdev->si_usecount--; 2818 dev_unlock(); 2819 } 2820 } 2821 2822 /* 2823 * Grab a particular vnode from the free list, increment its 2824 * reference count and lock it. VIRF_DOOMED is set if the vnode 2825 * is being destroyed. Only callers who specify LK_RETRY will 2826 * see doomed vnodes. If inactive processing was delayed in 2827 * vput try to do it here. 2828 * 2829 * usecount is manipulated using atomics without holding any locks, 2830 * except when transitioning 0->1 in which case the interlock is held. 2831 2832 * holdcnt is manipulated using atomics without holding any locks, 2833 * except when transitioning 1->0 in which case the interlock is held. 2834 */ 2835 enum vgetstate 2836 vget_prep(struct vnode *vp) 2837 { 2838 enum vgetstate vs; 2839 2840 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2841 vs = VGET_USECOUNT; 2842 } else { 2843 vhold(vp); 2844 vs = VGET_HOLDCNT; 2845 } 2846 return (vs); 2847 } 2848 2849 int 2850 vget(struct vnode *vp, int flags, struct thread *td) 2851 { 2852 enum vgetstate vs; 2853 2854 MPASS(td == curthread); 2855 2856 vs = vget_prep(vp); 2857 return (vget_finish(vp, flags, vs)); 2858 } 2859 2860 int 2861 vget_finish(struct vnode *vp, int flags, enum vgetstate vs) 2862 { 2863 int error, oweinact; 2864 2865 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 2866 ("%s: invalid lock operation", __func__)); 2867 2868 if ((flags & LK_INTERLOCK) != 0) 2869 ASSERT_VI_LOCKED(vp, __func__); 2870 else 2871 ASSERT_VI_UNLOCKED(vp, __func__); 2872 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__)); 2873 if (vs == VGET_USECOUNT) { 2874 VNASSERT(vp->v_usecount > 0, vp, 2875 ("%s: vnode without usecount when VGET_USECOUNT was passed", 2876 __func__)); 2877 } 2878 2879 if ((error = vn_lock(vp, flags)) != 0) { 2880 if (vs == VGET_USECOUNT) 2881 vrele(vp); 2882 else 2883 vdrop(vp); 2884 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, 2885 vp); 2886 return (error); 2887 } 2888 2889 if (vs == VGET_USECOUNT) { 2890 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2891 ("%s: vnode with usecount and VI_OWEINACT set", __func__)); 2892 return (0); 2893 } 2894 2895 /* 2896 * We hold the vnode. If the usecount is 0 it will be utilized to keep 2897 * the vnode around. Otherwise someone else lended their hold count and 2898 * we have to drop ours. 2899 */ 2900 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2901 #ifdef INVARIANTS 2902 int old = atomic_fetchadd_int(&vp->v_holdcnt, -1); 2903 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); 2904 #else 2905 refcount_release(&vp->v_holdcnt); 2906 #endif 2907 VNODE_REFCOUNT_FENCE_ACQ(); 2908 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2909 ("%s: vnode with usecount and VI_OWEINACT set", __func__)); 2910 return (0); 2911 } 2912 2913 /* 2914 * We don't guarantee that any particular close will 2915 * trigger inactive processing so just make a best effort 2916 * here at preventing a reference to a removed file. If 2917 * we don't succeed no harm is done. 2918 * 2919 * Upgrade our holdcnt to a usecount. 2920 */ 2921 VI_LOCK(vp); 2922 /* 2923 * See the previous section. By the time we get here we may find 2924 * ourselves in the same spot. 2925 */ 2926 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2927 #ifdef INVARIANTS 2928 int old = atomic_fetchadd_int(&vp->v_holdcnt, -1); 2929 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); 2930 #else 2931 refcount_release(&vp->v_holdcnt); 2932 #endif 2933 VNODE_REFCOUNT_FENCE_ACQ(); 2934 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2935 ("%s: vnode with usecount and VI_OWEINACT set", 2936 __func__)); 2937 VI_UNLOCK(vp); 2938 return (0); 2939 } 2940 if ((vp->v_iflag & VI_OWEINACT) == 0) { 2941 oweinact = 0; 2942 } else { 2943 oweinact = 1; 2944 vp->v_iflag &= ~VI_OWEINACT; 2945 VNODE_REFCOUNT_FENCE_REL(); 2946 } 2947 v_incr_devcount(vp); 2948 refcount_acquire(&vp->v_usecount); 2949 if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE && 2950 (flags & LK_NOWAIT) == 0) 2951 vinactive(vp); 2952 VI_UNLOCK(vp); 2953 return (0); 2954 } 2955 2956 /* 2957 * Increase the reference (use) and hold count of a vnode. 2958 * This will also remove the vnode from the free list if it is presently free. 2959 */ 2960 void 2961 vref(struct vnode *vp) 2962 { 2963 2964 ASSERT_VI_UNLOCKED(vp, __func__); 2965 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2966 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2967 VNODE_REFCOUNT_FENCE_ACQ(); 2968 VNASSERT(vp->v_holdcnt > 0, vp, 2969 ("%s: active vnode not held", __func__)); 2970 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2971 ("%s: vnode with usecount and VI_OWEINACT set", __func__)); 2972 return; 2973 } 2974 VI_LOCK(vp); 2975 vrefl(vp); 2976 VI_UNLOCK(vp); 2977 } 2978 2979 void 2980 vrefl(struct vnode *vp) 2981 { 2982 2983 ASSERT_VI_LOCKED(vp, __func__); 2984 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2985 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2986 VNODE_REFCOUNT_FENCE_ACQ(); 2987 VNASSERT(vp->v_holdcnt > 0, vp, 2988 ("%s: active vnode not held", __func__)); 2989 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2990 ("%s: vnode with usecount and VI_OWEINACT set", __func__)); 2991 return; 2992 } 2993 vholdl(vp); 2994 if ((vp->v_iflag & VI_OWEINACT) != 0) { 2995 vp->v_iflag &= ~VI_OWEINACT; 2996 VNODE_REFCOUNT_FENCE_REL(); 2997 } 2998 v_incr_devcount(vp); 2999 refcount_acquire(&vp->v_usecount); 3000 } 3001 3002 void 3003 vrefact(struct vnode *vp) 3004 { 3005 3006 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3007 #ifdef INVARIANTS 3008 int old = atomic_fetchadd_int(&vp->v_usecount, 1); 3009 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); 3010 #else 3011 refcount_acquire(&vp->v_usecount); 3012 #endif 3013 } 3014 3015 /* 3016 * Return reference count of a vnode. 3017 * 3018 * The results of this call are only guaranteed when some mechanism is used to 3019 * stop other processes from gaining references to the vnode. This may be the 3020 * case if the caller holds the only reference. This is also useful when stale 3021 * data is acceptable as race conditions may be accounted for by some other 3022 * means. 3023 */ 3024 int 3025 vrefcnt(struct vnode *vp) 3026 { 3027 3028 return (vp->v_usecount); 3029 } 3030 3031 void 3032 vlazy(struct vnode *vp) 3033 { 3034 struct mount *mp; 3035 3036 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__)); 3037 3038 if ((vp->v_mflag & VMP_LAZYLIST) != 0) 3039 return; 3040 mp = vp->v_mount; 3041 mtx_lock(&mp->mnt_listmtx); 3042 if ((vp->v_mflag & VMP_LAZYLIST) == 0) { 3043 vp->v_mflag |= VMP_LAZYLIST; 3044 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3045 mp->mnt_lazyvnodelistsize++; 3046 } 3047 mtx_unlock(&mp->mnt_listmtx); 3048 } 3049 3050 static void 3051 vdefer_inactive(struct vnode *vp) 3052 { 3053 3054 ASSERT_VI_LOCKED(vp, __func__); 3055 VNASSERT(vp->v_iflag & VI_OWEINACT, vp, 3056 ("%s: vnode without VI_OWEINACT", __func__)); 3057 if (VN_IS_DOOMED(vp)) { 3058 vdropl(vp); 3059 return; 3060 } 3061 if (vp->v_iflag & VI_DEFINACT) { 3062 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 3063 vdropl(vp); 3064 return; 3065 } 3066 vlazy(vp); 3067 vp->v_iflag |= VI_DEFINACT; 3068 VI_UNLOCK(vp); 3069 counter_u64_add(deferred_inact, 1); 3070 } 3071 3072 static void 3073 vdefer_inactive_cond(struct vnode *vp) 3074 { 3075 3076 VI_LOCK(vp); 3077 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); 3078 if ((vp->v_iflag & VI_OWEINACT) == 0) { 3079 vdropl(vp); 3080 return; 3081 } 3082 vdefer_inactive(vp); 3083 } 3084 3085 enum vputx_op { VPUTX_VRELE, VPUTX_VPUT, VPUTX_VUNREF }; 3086 3087 /* 3088 * Decrement the use and hold counts for a vnode. 3089 * 3090 * See an explanation near vget() as to why atomic operation is safe. 3091 */ 3092 static void 3093 vputx(struct vnode *vp, enum vputx_op func) 3094 { 3095 int error; 3096 3097 KASSERT(vp != NULL, ("vputx: null vp")); 3098 if (func == VPUTX_VUNREF) 3099 ASSERT_VOP_LOCKED(vp, "vunref"); 3100 ASSERT_VI_UNLOCKED(vp, __func__); 3101 VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp, 3102 ("%s: wrong ref counts", __func__)); 3103 3104 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3105 3106 /* 3107 * We want to hold the vnode until the inactive finishes to 3108 * prevent vgone() races. We drop the use count here and the 3109 * hold count below when we're done. 3110 * 3111 * If we release the last usecount we take ownership of the hold 3112 * count which provides liveness of the vnode, in which case we 3113 * have to vdrop. 3114 */ 3115 if (!refcount_release(&vp->v_usecount)) 3116 return; 3117 VI_LOCK(vp); 3118 v_decr_devcount(vp); 3119 /* 3120 * By the time we got here someone else might have transitioned 3121 * the count back to > 0. 3122 */ 3123 if (vp->v_usecount > 0) { 3124 vdropl(vp); 3125 return; 3126 } 3127 if (vp->v_iflag & VI_DOINGINACT) { 3128 vdropl(vp); 3129 return; 3130 } 3131 3132 /* 3133 * Check if the fs wants to perform inactive processing. Note we 3134 * may be only holding the interlock, in which case it is possible 3135 * someone else called vgone on the vnode and ->v_data is now NULL. 3136 * Since vgone performs inactive on its own there is nothing to do 3137 * here but to drop our hold count. 3138 */ 3139 if (__predict_false(VN_IS_DOOMED(vp)) || 3140 VOP_NEED_INACTIVE(vp) == 0) { 3141 vdropl(vp); 3142 return; 3143 } 3144 3145 /* 3146 * We must call VOP_INACTIVE with the node locked. Mark 3147 * as VI_DOINGINACT to avoid recursion. 3148 */ 3149 vp->v_iflag |= VI_OWEINACT; 3150 switch (func) { 3151 case VPUTX_VRELE: 3152 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 3153 VI_LOCK(vp); 3154 break; 3155 case VPUTX_VPUT: 3156 error = VOP_LOCK(vp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT); 3157 VI_LOCK(vp); 3158 break; 3159 case VPUTX_VUNREF: 3160 error = 0; 3161 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3162 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); 3163 VI_LOCK(vp); 3164 } 3165 break; 3166 } 3167 VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp, 3168 ("vnode with usecount and VI_OWEINACT set")); 3169 if (error == 0) { 3170 if (vp->v_iflag & VI_OWEINACT) 3171 vinactive(vp); 3172 if (func != VPUTX_VUNREF) 3173 VOP_UNLOCK(vp); 3174 vdropl(vp); 3175 } else if (vp->v_iflag & VI_OWEINACT) { 3176 vdefer_inactive(vp); 3177 } else { 3178 vdropl(vp); 3179 } 3180 } 3181 3182 /* 3183 * Vnode put/release. 3184 * If count drops to zero, call inactive routine and return to freelist. 3185 */ 3186 void 3187 vrele(struct vnode *vp) 3188 { 3189 3190 vputx(vp, VPUTX_VRELE); 3191 } 3192 3193 /* 3194 * Release an already locked vnode. This give the same effects as 3195 * unlock+vrele(), but takes less time and avoids releasing and 3196 * re-aquiring the lock (as vrele() acquires the lock internally.) 3197 * 3198 * It is an invariant that all VOP_* calls operate on a held vnode. 3199 * We may be only having an implicit hold stemming from our usecount, 3200 * which we are about to release. If we unlock the vnode afterwards we 3201 * open a time window where someone else dropped the last usecount and 3202 * proceeded to free the vnode before our unlock finished. For this 3203 * reason we unlock the vnode early. This is a little bit wasteful as 3204 * it may be the vnode is exclusively locked and inactive processing is 3205 * needed, in which case we are adding work. 3206 */ 3207 void 3208 vput(struct vnode *vp) 3209 { 3210 3211 VOP_UNLOCK(vp); 3212 vputx(vp, VPUTX_VPUT); 3213 } 3214 3215 /* 3216 * Release an exclusively locked vnode. Do not unlock the vnode lock. 3217 */ 3218 void 3219 vunref(struct vnode *vp) 3220 { 3221 3222 vputx(vp, VPUTX_VUNREF); 3223 } 3224 3225 void 3226 vhold(struct vnode *vp) 3227 { 3228 struct vdbatch *vd; 3229 int old; 3230 3231 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3232 old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3233 VNASSERT(old >= 0, vp, ("%s: wrong hold count %d", __func__, old)); 3234 if (old != 0) 3235 return; 3236 critical_enter(); 3237 vd = DPCPU_PTR(vd); 3238 vd->freevnodes--; 3239 critical_exit(); 3240 } 3241 3242 void 3243 vholdl(struct vnode *vp) 3244 { 3245 3246 ASSERT_VI_LOCKED(vp, __func__); 3247 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3248 vhold(vp); 3249 } 3250 3251 void 3252 vholdnz(struct vnode *vp) 3253 { 3254 3255 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3256 #ifdef INVARIANTS 3257 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3258 VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old)); 3259 #else 3260 atomic_add_int(&vp->v_holdcnt, 1); 3261 #endif 3262 } 3263 3264 static void __noinline 3265 vdbatch_process(struct vdbatch *vd) 3266 { 3267 struct vnode *vp; 3268 int i; 3269 3270 mtx_assert(&vd->lock, MA_OWNED); 3271 MPASS(curthread->td_pinned > 0); 3272 MPASS(vd->index == VDBATCH_SIZE); 3273 3274 mtx_lock(&vnode_list_mtx); 3275 critical_enter(); 3276 freevnodes += vd->freevnodes; 3277 for (i = 0; i < VDBATCH_SIZE; i++) { 3278 vp = vd->tab[i]; 3279 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 3280 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist); 3281 MPASS(vp->v_dbatchcpu != NOCPU); 3282 vp->v_dbatchcpu = NOCPU; 3283 } 3284 mtx_unlock(&vnode_list_mtx); 3285 critical_exit(); 3286 vd->freevnodes = 0; 3287 bzero(vd->tab, sizeof(vd->tab)); 3288 vd->index = 0; 3289 } 3290 3291 static void 3292 vdbatch_enqueue(struct vnode *vp) 3293 { 3294 struct vdbatch *vd; 3295 3296 ASSERT_VI_LOCKED(vp, __func__); 3297 VNASSERT(!VN_IS_DOOMED(vp), vp, 3298 ("%s: deferring requeue of a doomed vnode", __func__)); 3299 3300 critical_enter(); 3301 vd = DPCPU_PTR(vd); 3302 vd->freevnodes++; 3303 if (vp->v_dbatchcpu != NOCPU) { 3304 VI_UNLOCK(vp); 3305 critical_exit(); 3306 return; 3307 } 3308 3309 sched_pin(); 3310 critical_exit(); 3311 mtx_lock(&vd->lock); 3312 MPASS(vd->index < VDBATCH_SIZE); 3313 MPASS(vd->tab[vd->index] == NULL); 3314 /* 3315 * A hack: we depend on being pinned so that we know what to put in 3316 * ->v_dbatchcpu. 3317 */ 3318 vp->v_dbatchcpu = curcpu; 3319 vd->tab[vd->index] = vp; 3320 vd->index++; 3321 VI_UNLOCK(vp); 3322 if (vd->index == VDBATCH_SIZE) 3323 vdbatch_process(vd); 3324 mtx_unlock(&vd->lock); 3325 sched_unpin(); 3326 } 3327 3328 /* 3329 * This routine must only be called for vnodes which are about to be 3330 * deallocated. Supporting dequeue for arbitrary vndoes would require 3331 * validating that the locked batch matches. 3332 */ 3333 static void 3334 vdbatch_dequeue(struct vnode *vp) 3335 { 3336 struct vdbatch *vd; 3337 int i; 3338 short cpu; 3339 3340 VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp, 3341 ("%s: called for a used vnode\n", __func__)); 3342 3343 cpu = vp->v_dbatchcpu; 3344 if (cpu == NOCPU) 3345 return; 3346 3347 vd = DPCPU_ID_PTR(cpu, vd); 3348 mtx_lock(&vd->lock); 3349 for (i = 0; i < vd->index; i++) { 3350 if (vd->tab[i] != vp) 3351 continue; 3352 vp->v_dbatchcpu = NOCPU; 3353 vd->index--; 3354 vd->tab[i] = vd->tab[vd->index]; 3355 vd->tab[vd->index] = NULL; 3356 break; 3357 } 3358 mtx_unlock(&vd->lock); 3359 /* 3360 * Either we dequeued the vnode above or the target CPU beat us to it. 3361 */ 3362 MPASS(vp->v_dbatchcpu == NOCPU); 3363 } 3364 3365 /* 3366 * Drop the hold count of the vnode. If this is the last reference to 3367 * the vnode we place it on the free list unless it has been vgone'd 3368 * (marked VIRF_DOOMED) in which case we will free it. 3369 * 3370 * Because the vnode vm object keeps a hold reference on the vnode if 3371 * there is at least one resident non-cached page, the vnode cannot 3372 * leave the active list without the page cleanup done. 3373 */ 3374 static void 3375 vdrop_deactivate(struct vnode *vp) 3376 { 3377 struct mount *mp; 3378 3379 ASSERT_VI_LOCKED(vp, __func__); 3380 /* 3381 * Mark a vnode as free: remove it from its active list 3382 * and put it up for recycling on the freelist. 3383 */ 3384 VNASSERT(!VN_IS_DOOMED(vp), vp, 3385 ("vdrop: returning doomed vnode")); 3386 VNASSERT(vp->v_op != NULL, vp, 3387 ("vdrop: vnode already reclaimed.")); 3388 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 3389 ("vnode with VI_OWEINACT set")); 3390 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, 3391 ("vnode with VI_DEFINACT set")); 3392 if (vp->v_mflag & VMP_LAZYLIST) { 3393 mp = vp->v_mount; 3394 mtx_lock(&mp->mnt_listmtx); 3395 VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST")); 3396 /* 3397 * Don't remove the vnode from the lazy list if another thread 3398 * has increased the hold count. It may have re-enqueued the 3399 * vnode to the lazy list and is now responsible for its 3400 * removal. 3401 */ 3402 if (vp->v_holdcnt == 0) { 3403 vp->v_mflag &= ~VMP_LAZYLIST; 3404 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3405 mp->mnt_lazyvnodelistsize--; 3406 } 3407 mtx_unlock(&mp->mnt_listmtx); 3408 } 3409 vdbatch_enqueue(vp); 3410 } 3411 3412 void 3413 vdrop(struct vnode *vp) 3414 { 3415 3416 ASSERT_VI_UNLOCKED(vp, __func__); 3417 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3418 if (refcount_release_if_not_last(&vp->v_holdcnt)) 3419 return; 3420 VI_LOCK(vp); 3421 vdropl(vp); 3422 } 3423 3424 void 3425 vdropl(struct vnode *vp) 3426 { 3427 3428 ASSERT_VI_LOCKED(vp, __func__); 3429 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3430 if (!refcount_release(&vp->v_holdcnt)) { 3431 VI_UNLOCK(vp); 3432 return; 3433 } 3434 if (VN_IS_DOOMED(vp)) { 3435 freevnode(vp); 3436 return; 3437 } 3438 vdrop_deactivate(vp); 3439 } 3440 3441 /* 3442 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 3443 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 3444 * OWEINACT tracks whether a vnode missed a call to inactive due to a 3445 * failed lock upgrade. 3446 */ 3447 void 3448 vinactive(struct vnode *vp) 3449 { 3450 struct vm_object *obj; 3451 3452 ASSERT_VOP_ELOCKED(vp, "vinactive"); 3453 ASSERT_VI_LOCKED(vp, "vinactive"); 3454 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, 3455 ("vinactive: recursed on VI_DOINGINACT")); 3456 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3457 vp->v_iflag |= VI_DOINGINACT; 3458 vp->v_iflag &= ~VI_OWEINACT; 3459 VI_UNLOCK(vp); 3460 /* 3461 * Before moving off the active list, we must be sure that any 3462 * modified pages are converted into the vnode's dirty 3463 * buffers, since these will no longer be checked once the 3464 * vnode is on the inactive list. 3465 * 3466 * The write-out of the dirty pages is asynchronous. At the 3467 * point that VOP_INACTIVE() is called, there could still be 3468 * pending I/O and dirty pages in the object. 3469 */ 3470 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 3471 vm_object_mightbedirty(obj)) { 3472 VM_OBJECT_WLOCK(obj); 3473 vm_object_page_clean(obj, 0, 0, 0); 3474 VM_OBJECT_WUNLOCK(obj); 3475 } 3476 VOP_INACTIVE(vp, curthread); 3477 VI_LOCK(vp); 3478 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, 3479 ("vinactive: lost VI_DOINGINACT")); 3480 vp->v_iflag &= ~VI_DOINGINACT; 3481 } 3482 3483 /* 3484 * Remove any vnodes in the vnode table belonging to mount point mp. 3485 * 3486 * If FORCECLOSE is not specified, there should not be any active ones, 3487 * return error if any are found (nb: this is a user error, not a 3488 * system error). If FORCECLOSE is specified, detach any active vnodes 3489 * that are found. 3490 * 3491 * If WRITECLOSE is set, only flush out regular file vnodes open for 3492 * writing. 3493 * 3494 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 3495 * 3496 * `rootrefs' specifies the base reference count for the root vnode 3497 * of this filesystem. The root vnode is considered busy if its 3498 * v_usecount exceeds this value. On a successful return, vflush(, td) 3499 * will call vrele() on the root vnode exactly rootrefs times. 3500 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 3501 * be zero. 3502 */ 3503 #ifdef DIAGNOSTIC 3504 static int busyprt = 0; /* print out busy vnodes */ 3505 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 3506 #endif 3507 3508 int 3509 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 3510 { 3511 struct vnode *vp, *mvp, *rootvp = NULL; 3512 struct vattr vattr; 3513 int busy = 0, error; 3514 3515 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 3516 rootrefs, flags); 3517 if (rootrefs > 0) { 3518 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 3519 ("vflush: bad args")); 3520 /* 3521 * Get the filesystem root vnode. We can vput() it 3522 * immediately, since with rootrefs > 0, it won't go away. 3523 */ 3524 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 3525 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 3526 __func__, error); 3527 return (error); 3528 } 3529 vput(rootvp); 3530 } 3531 loop: 3532 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 3533 vholdl(vp); 3534 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 3535 if (error) { 3536 vdrop(vp); 3537 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3538 goto loop; 3539 } 3540 /* 3541 * Skip over a vnodes marked VV_SYSTEM. 3542 */ 3543 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 3544 VOP_UNLOCK(vp); 3545 vdrop(vp); 3546 continue; 3547 } 3548 /* 3549 * If WRITECLOSE is set, flush out unlinked but still open 3550 * files (even if open only for reading) and regular file 3551 * vnodes open for writing. 3552 */ 3553 if (flags & WRITECLOSE) { 3554 if (vp->v_object != NULL) { 3555 VM_OBJECT_WLOCK(vp->v_object); 3556 vm_object_page_clean(vp->v_object, 0, 0, 0); 3557 VM_OBJECT_WUNLOCK(vp->v_object); 3558 } 3559 error = VOP_FSYNC(vp, MNT_WAIT, td); 3560 if (error != 0) { 3561 VOP_UNLOCK(vp); 3562 vdrop(vp); 3563 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3564 return (error); 3565 } 3566 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 3567 VI_LOCK(vp); 3568 3569 if ((vp->v_type == VNON || 3570 (error == 0 && vattr.va_nlink > 0)) && 3571 (vp->v_writecount <= 0 || vp->v_type != VREG)) { 3572 VOP_UNLOCK(vp); 3573 vdropl(vp); 3574 continue; 3575 } 3576 } else 3577 VI_LOCK(vp); 3578 /* 3579 * With v_usecount == 0, all we need to do is clear out the 3580 * vnode data structures and we are done. 3581 * 3582 * If FORCECLOSE is set, forcibly close the vnode. 3583 */ 3584 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 3585 vgonel(vp); 3586 } else { 3587 busy++; 3588 #ifdef DIAGNOSTIC 3589 if (busyprt) 3590 vn_printf(vp, "vflush: busy vnode "); 3591 #endif 3592 } 3593 VOP_UNLOCK(vp); 3594 vdropl(vp); 3595 } 3596 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 3597 /* 3598 * If just the root vnode is busy, and if its refcount 3599 * is equal to `rootrefs', then go ahead and kill it. 3600 */ 3601 VI_LOCK(rootvp); 3602 KASSERT(busy > 0, ("vflush: not busy")); 3603 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 3604 ("vflush: usecount %d < rootrefs %d", 3605 rootvp->v_usecount, rootrefs)); 3606 if (busy == 1 && rootvp->v_usecount == rootrefs) { 3607 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 3608 vgone(rootvp); 3609 VOP_UNLOCK(rootvp); 3610 busy = 0; 3611 } else 3612 VI_UNLOCK(rootvp); 3613 } 3614 if (busy) { 3615 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 3616 busy); 3617 return (EBUSY); 3618 } 3619 for (; rootrefs > 0; rootrefs--) 3620 vrele(rootvp); 3621 return (0); 3622 } 3623 3624 /* 3625 * Recycle an unused vnode to the front of the free list. 3626 */ 3627 int 3628 vrecycle(struct vnode *vp) 3629 { 3630 int recycled; 3631 3632 VI_LOCK(vp); 3633 recycled = vrecyclel(vp); 3634 VI_UNLOCK(vp); 3635 return (recycled); 3636 } 3637 3638 /* 3639 * vrecycle, with the vp interlock held. 3640 */ 3641 int 3642 vrecyclel(struct vnode *vp) 3643 { 3644 int recycled; 3645 3646 ASSERT_VOP_ELOCKED(vp, __func__); 3647 ASSERT_VI_LOCKED(vp, __func__); 3648 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3649 recycled = 0; 3650 if (vp->v_usecount == 0) { 3651 recycled = 1; 3652 vgonel(vp); 3653 } 3654 return (recycled); 3655 } 3656 3657 /* 3658 * Eliminate all activity associated with a vnode 3659 * in preparation for reuse. 3660 */ 3661 void 3662 vgone(struct vnode *vp) 3663 { 3664 VI_LOCK(vp); 3665 vgonel(vp); 3666 VI_UNLOCK(vp); 3667 } 3668 3669 static void 3670 notify_lowervp_vfs_dummy(struct mount *mp __unused, 3671 struct vnode *lowervp __unused) 3672 { 3673 } 3674 3675 /* 3676 * Notify upper mounts about reclaimed or unlinked vnode. 3677 */ 3678 void 3679 vfs_notify_upper(struct vnode *vp, int event) 3680 { 3681 static struct vfsops vgonel_vfsops = { 3682 .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy, 3683 .vfs_unlink_lowervp = notify_lowervp_vfs_dummy, 3684 }; 3685 struct mount *mp, *ump, *mmp; 3686 3687 mp = vp->v_mount; 3688 if (mp == NULL) 3689 return; 3690 if (TAILQ_EMPTY(&mp->mnt_uppers)) 3691 return; 3692 3693 mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO); 3694 mmp->mnt_op = &vgonel_vfsops; 3695 mmp->mnt_kern_flag |= MNTK_MARKER; 3696 MNT_ILOCK(mp); 3697 mp->mnt_kern_flag |= MNTK_VGONE_UPPER; 3698 for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) { 3699 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) { 3700 ump = TAILQ_NEXT(ump, mnt_upper_link); 3701 continue; 3702 } 3703 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link); 3704 MNT_IUNLOCK(mp); 3705 switch (event) { 3706 case VFS_NOTIFY_UPPER_RECLAIM: 3707 VFS_RECLAIM_LOWERVP(ump, vp); 3708 break; 3709 case VFS_NOTIFY_UPPER_UNLINK: 3710 VFS_UNLINK_LOWERVP(ump, vp); 3711 break; 3712 default: 3713 KASSERT(0, ("invalid event %d", event)); 3714 break; 3715 } 3716 MNT_ILOCK(mp); 3717 ump = TAILQ_NEXT(mmp, mnt_upper_link); 3718 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link); 3719 } 3720 free(mmp, M_TEMP); 3721 mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER; 3722 if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) { 3723 mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER; 3724 wakeup(&mp->mnt_uppers); 3725 } 3726 MNT_IUNLOCK(mp); 3727 } 3728 3729 /* 3730 * vgone, with the vp interlock held. 3731 */ 3732 static void 3733 vgonel(struct vnode *vp) 3734 { 3735 struct thread *td; 3736 struct mount *mp; 3737 vm_object_t object; 3738 bool active, oweinact; 3739 3740 ASSERT_VOP_ELOCKED(vp, "vgonel"); 3741 ASSERT_VI_LOCKED(vp, "vgonel"); 3742 VNASSERT(vp->v_holdcnt, vp, 3743 ("vgonel: vp %p has no reference.", vp)); 3744 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3745 td = curthread; 3746 3747 /* 3748 * Don't vgonel if we're already doomed. 3749 */ 3750 if (vp->v_irflag & VIRF_DOOMED) 3751 return; 3752 vp->v_irflag |= VIRF_DOOMED; 3753 3754 /* 3755 * Check to see if the vnode is in use. If so, we have to call 3756 * VOP_CLOSE() and VOP_INACTIVE(). 3757 */ 3758 active = vp->v_usecount > 0; 3759 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 3760 /* 3761 * If we need to do inactive VI_OWEINACT will be set. 3762 */ 3763 if (vp->v_iflag & VI_DEFINACT) { 3764 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 3765 vp->v_iflag &= ~VI_DEFINACT; 3766 vdropl(vp); 3767 } else { 3768 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); 3769 VI_UNLOCK(vp); 3770 } 3771 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); 3772 3773 /* 3774 * If purging an active vnode, it must be closed and 3775 * deactivated before being reclaimed. 3776 */ 3777 if (active) 3778 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 3779 if (oweinact || active) { 3780 VI_LOCK(vp); 3781 if ((vp->v_iflag & VI_DOINGINACT) == 0) 3782 vinactive(vp); 3783 VI_UNLOCK(vp); 3784 } 3785 if (vp->v_type == VSOCK) 3786 vfs_unp_reclaim(vp); 3787 3788 /* 3789 * Clean out any buffers associated with the vnode. 3790 * If the flush fails, just toss the buffers. 3791 */ 3792 mp = NULL; 3793 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 3794 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 3795 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { 3796 while (vinvalbuf(vp, 0, 0, 0) != 0) 3797 ; 3798 } 3799 3800 BO_LOCK(&vp->v_bufobj); 3801 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && 3802 vp->v_bufobj.bo_dirty.bv_cnt == 0 && 3803 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && 3804 vp->v_bufobj.bo_clean.bv_cnt == 0, 3805 ("vp %p bufobj not invalidated", vp)); 3806 3807 /* 3808 * For VMIO bufobj, BO_DEAD is set later, or in 3809 * vm_object_terminate() after the object's page queue is 3810 * flushed. 3811 */ 3812 object = vp->v_bufobj.bo_object; 3813 if (object == NULL) 3814 vp->v_bufobj.bo_flag |= BO_DEAD; 3815 BO_UNLOCK(&vp->v_bufobj); 3816 3817 /* 3818 * Handle the VM part. Tmpfs handles v_object on its own (the 3819 * OBJT_VNODE check). Nullfs or other bypassing filesystems 3820 * should not touch the object borrowed from the lower vnode 3821 * (the handle check). 3822 */ 3823 if (object != NULL && object->type == OBJT_VNODE && 3824 object->handle == vp) 3825 vnode_destroy_vobject(vp); 3826 3827 /* 3828 * Reclaim the vnode. 3829 */ 3830 if (VOP_RECLAIM(vp, td)) 3831 panic("vgone: cannot reclaim"); 3832 if (mp != NULL) 3833 vn_finished_secondary_write(mp); 3834 VNASSERT(vp->v_object == NULL, vp, 3835 ("vop_reclaim left v_object vp=%p", vp)); 3836 /* 3837 * Clear the advisory locks and wake up waiting threads. 3838 */ 3839 (void)VOP_ADVLOCKPURGE(vp); 3840 vp->v_lockf = NULL; 3841 /* 3842 * Delete from old mount point vnode list. 3843 */ 3844 delmntque(vp); 3845 cache_purge(vp); 3846 /* 3847 * Done with purge, reset to the standard lock and invalidate 3848 * the vnode. 3849 */ 3850 VI_LOCK(vp); 3851 vp->v_vnlock = &vp->v_lock; 3852 vp->v_op = &dead_vnodeops; 3853 vp->v_type = VBAD; 3854 } 3855 3856 /* 3857 * Calculate the total number of references to a special device. 3858 */ 3859 int 3860 vcount(struct vnode *vp) 3861 { 3862 int count; 3863 3864 dev_lock(); 3865 count = vp->v_rdev->si_usecount; 3866 dev_unlock(); 3867 return (count); 3868 } 3869 3870 /* 3871 * Print out a description of a vnode. 3872 */ 3873 static char *typename[] = 3874 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", 3875 "VMARKER"}; 3876 3877 void 3878 vn_printf(struct vnode *vp, const char *fmt, ...) 3879 { 3880 va_list ap; 3881 char buf[256], buf2[16]; 3882 u_long flags; 3883 3884 va_start(ap, fmt); 3885 vprintf(fmt, ap); 3886 va_end(ap); 3887 printf("%p: ", (void *)vp); 3888 printf("type %s\n", typename[vp->v_type]); 3889 printf(" usecount %d, writecount %d, refcount %d", 3890 vp->v_usecount, vp->v_writecount, vp->v_holdcnt); 3891 switch (vp->v_type) { 3892 case VDIR: 3893 printf(" mountedhere %p\n", vp->v_mountedhere); 3894 break; 3895 case VCHR: 3896 printf(" rdev %p\n", vp->v_rdev); 3897 break; 3898 case VSOCK: 3899 printf(" socket %p\n", vp->v_unpcb); 3900 break; 3901 case VFIFO: 3902 printf(" fifoinfo %p\n", vp->v_fifoinfo); 3903 break; 3904 default: 3905 printf("\n"); 3906 break; 3907 } 3908 buf[0] = '\0'; 3909 buf[1] = '\0'; 3910 if (vp->v_irflag & VIRF_DOOMED) 3911 strlcat(buf, "|VIRF_DOOMED", sizeof(buf)); 3912 flags = vp->v_irflag & ~(VIRF_DOOMED); 3913 if (flags != 0) { 3914 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags); 3915 strlcat(buf, buf2, sizeof(buf)); 3916 } 3917 if (vp->v_vflag & VV_ROOT) 3918 strlcat(buf, "|VV_ROOT", sizeof(buf)); 3919 if (vp->v_vflag & VV_ISTTY) 3920 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 3921 if (vp->v_vflag & VV_NOSYNC) 3922 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 3923 if (vp->v_vflag & VV_ETERNALDEV) 3924 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 3925 if (vp->v_vflag & VV_CACHEDLABEL) 3926 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 3927 if (vp->v_vflag & VV_VMSIZEVNLOCK) 3928 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf)); 3929 if (vp->v_vflag & VV_COPYONWRITE) 3930 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 3931 if (vp->v_vflag & VV_SYSTEM) 3932 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 3933 if (vp->v_vflag & VV_PROCDEP) 3934 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 3935 if (vp->v_vflag & VV_NOKNOTE) 3936 strlcat(buf, "|VV_NOKNOTE", sizeof(buf)); 3937 if (vp->v_vflag & VV_DELETED) 3938 strlcat(buf, "|VV_DELETED", sizeof(buf)); 3939 if (vp->v_vflag & VV_MD) 3940 strlcat(buf, "|VV_MD", sizeof(buf)); 3941 if (vp->v_vflag & VV_FORCEINSMQ) 3942 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 3943 if (vp->v_vflag & VV_READLINK) 3944 strlcat(buf, "|VV_READLINK", sizeof(buf)); 3945 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 3946 VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | 3947 VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ); 3948 if (flags != 0) { 3949 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 3950 strlcat(buf, buf2, sizeof(buf)); 3951 } 3952 if (vp->v_iflag & VI_TEXT_REF) 3953 strlcat(buf, "|VI_TEXT_REF", sizeof(buf)); 3954 if (vp->v_iflag & VI_MOUNT) 3955 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 3956 if (vp->v_iflag & VI_DOINGINACT) 3957 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 3958 if (vp->v_iflag & VI_OWEINACT) 3959 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 3960 if (vp->v_iflag & VI_DEFINACT) 3961 strlcat(buf, "|VI_DEFINACT", sizeof(buf)); 3962 flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT | 3963 VI_OWEINACT | VI_DEFINACT); 3964 if (flags != 0) { 3965 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 3966 strlcat(buf, buf2, sizeof(buf)); 3967 } 3968 if (vp->v_mflag & VMP_LAZYLIST) 3969 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf)); 3970 flags = vp->v_mflag & ~(VMP_LAZYLIST); 3971 if (flags != 0) { 3972 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags); 3973 strlcat(buf, buf2, sizeof(buf)); 3974 } 3975 printf(" flags (%s)\n", buf + 1); 3976 if (mtx_owned(VI_MTX(vp))) 3977 printf(" VI_LOCKed"); 3978 if (vp->v_object != NULL) 3979 printf(" v_object %p ref %d pages %d " 3980 "cleanbuf %d dirtybuf %d\n", 3981 vp->v_object, vp->v_object->ref_count, 3982 vp->v_object->resident_page_count, 3983 vp->v_bufobj.bo_clean.bv_cnt, 3984 vp->v_bufobj.bo_dirty.bv_cnt); 3985 printf(" "); 3986 lockmgr_printinfo(vp->v_vnlock); 3987 if (vp->v_data != NULL) 3988 VOP_PRINT(vp); 3989 } 3990 3991 #ifdef DDB 3992 /* 3993 * List all of the locked vnodes in the system. 3994 * Called when debugging the kernel. 3995 */ 3996 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 3997 { 3998 struct mount *mp; 3999 struct vnode *vp; 4000 4001 /* 4002 * Note: because this is DDB, we can't obey the locking semantics 4003 * for these structures, which means we could catch an inconsistent 4004 * state and dereference a nasty pointer. Not much to be done 4005 * about that. 4006 */ 4007 db_printf("Locked vnodes\n"); 4008 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4009 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4010 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) 4011 vn_printf(vp, "vnode "); 4012 } 4013 } 4014 } 4015 4016 /* 4017 * Show details about the given vnode. 4018 */ 4019 DB_SHOW_COMMAND(vnode, db_show_vnode) 4020 { 4021 struct vnode *vp; 4022 4023 if (!have_addr) 4024 return; 4025 vp = (struct vnode *)addr; 4026 vn_printf(vp, "vnode "); 4027 } 4028 4029 /* 4030 * Show details about the given mount point. 4031 */ 4032 DB_SHOW_COMMAND(mount, db_show_mount) 4033 { 4034 struct mount *mp; 4035 struct vfsopt *opt; 4036 struct statfs *sp; 4037 struct vnode *vp; 4038 char buf[512]; 4039 uint64_t mflags; 4040 u_int flags; 4041 4042 if (!have_addr) { 4043 /* No address given, print short info about all mount points. */ 4044 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4045 db_printf("%p %s on %s (%s)\n", mp, 4046 mp->mnt_stat.f_mntfromname, 4047 mp->mnt_stat.f_mntonname, 4048 mp->mnt_stat.f_fstypename); 4049 if (db_pager_quit) 4050 break; 4051 } 4052 db_printf("\nMore info: show mount <addr>\n"); 4053 return; 4054 } 4055 4056 mp = (struct mount *)addr; 4057 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 4058 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 4059 4060 buf[0] = '\0'; 4061 mflags = mp->mnt_flag; 4062 #define MNT_FLAG(flag) do { \ 4063 if (mflags & (flag)) { \ 4064 if (buf[0] != '\0') \ 4065 strlcat(buf, ", ", sizeof(buf)); \ 4066 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 4067 mflags &= ~(flag); \ 4068 } \ 4069 } while (0) 4070 MNT_FLAG(MNT_RDONLY); 4071 MNT_FLAG(MNT_SYNCHRONOUS); 4072 MNT_FLAG(MNT_NOEXEC); 4073 MNT_FLAG(MNT_NOSUID); 4074 MNT_FLAG(MNT_NFS4ACLS); 4075 MNT_FLAG(MNT_UNION); 4076 MNT_FLAG(MNT_ASYNC); 4077 MNT_FLAG(MNT_SUIDDIR); 4078 MNT_FLAG(MNT_SOFTDEP); 4079 MNT_FLAG(MNT_NOSYMFOLLOW); 4080 MNT_FLAG(MNT_GJOURNAL); 4081 MNT_FLAG(MNT_MULTILABEL); 4082 MNT_FLAG(MNT_ACLS); 4083 MNT_FLAG(MNT_NOATIME); 4084 MNT_FLAG(MNT_NOCLUSTERR); 4085 MNT_FLAG(MNT_NOCLUSTERW); 4086 MNT_FLAG(MNT_SUJ); 4087 MNT_FLAG(MNT_EXRDONLY); 4088 MNT_FLAG(MNT_EXPORTED); 4089 MNT_FLAG(MNT_DEFEXPORTED); 4090 MNT_FLAG(MNT_EXPORTANON); 4091 MNT_FLAG(MNT_EXKERB); 4092 MNT_FLAG(MNT_EXPUBLIC); 4093 MNT_FLAG(MNT_LOCAL); 4094 MNT_FLAG(MNT_QUOTA); 4095 MNT_FLAG(MNT_ROOTFS); 4096 MNT_FLAG(MNT_USER); 4097 MNT_FLAG(MNT_IGNORE); 4098 MNT_FLAG(MNT_UPDATE); 4099 MNT_FLAG(MNT_DELEXPORT); 4100 MNT_FLAG(MNT_RELOAD); 4101 MNT_FLAG(MNT_FORCE); 4102 MNT_FLAG(MNT_SNAPSHOT); 4103 MNT_FLAG(MNT_BYFSID); 4104 #undef MNT_FLAG 4105 if (mflags != 0) { 4106 if (buf[0] != '\0') 4107 strlcat(buf, ", ", sizeof(buf)); 4108 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4109 "0x%016jx", mflags); 4110 } 4111 db_printf(" mnt_flag = %s\n", buf); 4112 4113 buf[0] = '\0'; 4114 flags = mp->mnt_kern_flag; 4115 #define MNT_KERN_FLAG(flag) do { \ 4116 if (flags & (flag)) { \ 4117 if (buf[0] != '\0') \ 4118 strlcat(buf, ", ", sizeof(buf)); \ 4119 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 4120 flags &= ~(flag); \ 4121 } \ 4122 } while (0) 4123 MNT_KERN_FLAG(MNTK_UNMOUNTF); 4124 MNT_KERN_FLAG(MNTK_ASYNC); 4125 MNT_KERN_FLAG(MNTK_SOFTDEP); 4126 MNT_KERN_FLAG(MNTK_DRAINING); 4127 MNT_KERN_FLAG(MNTK_REFEXPIRE); 4128 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 4129 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 4130 MNT_KERN_FLAG(MNTK_NO_IOPF); 4131 MNT_KERN_FLAG(MNTK_VGONE_UPPER); 4132 MNT_KERN_FLAG(MNTK_VGONE_WAITER); 4133 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT); 4134 MNT_KERN_FLAG(MNTK_MARKER); 4135 MNT_KERN_FLAG(MNTK_USES_BCACHE); 4136 MNT_KERN_FLAG(MNTK_NOASYNC); 4137 MNT_KERN_FLAG(MNTK_UNMOUNT); 4138 MNT_KERN_FLAG(MNTK_MWAIT); 4139 MNT_KERN_FLAG(MNTK_SUSPEND); 4140 MNT_KERN_FLAG(MNTK_SUSPEND2); 4141 MNT_KERN_FLAG(MNTK_SUSPENDED); 4142 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 4143 MNT_KERN_FLAG(MNTK_NOKNOTE); 4144 #undef MNT_KERN_FLAG 4145 if (flags != 0) { 4146 if (buf[0] != '\0') 4147 strlcat(buf, ", ", sizeof(buf)); 4148 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4149 "0x%08x", flags); 4150 } 4151 db_printf(" mnt_kern_flag = %s\n", buf); 4152 4153 db_printf(" mnt_opt = "); 4154 opt = TAILQ_FIRST(mp->mnt_opt); 4155 if (opt != NULL) { 4156 db_printf("%s", opt->name); 4157 opt = TAILQ_NEXT(opt, link); 4158 while (opt != NULL) { 4159 db_printf(", %s", opt->name); 4160 opt = TAILQ_NEXT(opt, link); 4161 } 4162 } 4163 db_printf("\n"); 4164 4165 sp = &mp->mnt_stat; 4166 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 4167 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 4168 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 4169 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 4170 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 4171 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 4172 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 4173 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 4174 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 4175 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 4176 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 4177 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 4178 4179 db_printf(" mnt_cred = { uid=%u ruid=%u", 4180 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 4181 if (jailed(mp->mnt_cred)) 4182 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 4183 db_printf(" }\n"); 4184 db_printf(" mnt_ref = %d (with %d in the struct)\n", 4185 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref); 4186 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 4187 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 4188 db_printf(" mnt_lazyvnodelistsize = %d\n", 4189 mp->mnt_lazyvnodelistsize); 4190 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n", 4191 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount); 4192 db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen); 4193 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 4194 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 4195 db_printf(" mnt_lockref = %d (with %d in the struct)\n", 4196 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref); 4197 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 4198 db_printf(" mnt_secondary_accwrites = %d\n", 4199 mp->mnt_secondary_accwrites); 4200 db_printf(" mnt_gjprovider = %s\n", 4201 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 4202 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops); 4203 4204 db_printf("\n\nList of active vnodes\n"); 4205 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4206 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) { 4207 vn_printf(vp, "vnode "); 4208 if (db_pager_quit) 4209 break; 4210 } 4211 } 4212 db_printf("\n\nList of inactive vnodes\n"); 4213 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4214 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) { 4215 vn_printf(vp, "vnode "); 4216 if (db_pager_quit) 4217 break; 4218 } 4219 } 4220 } 4221 #endif /* DDB */ 4222 4223 /* 4224 * Fill in a struct xvfsconf based on a struct vfsconf. 4225 */ 4226 static int 4227 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 4228 { 4229 struct xvfsconf xvfsp; 4230 4231 bzero(&xvfsp, sizeof(xvfsp)); 4232 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4233 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4234 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4235 xvfsp.vfc_flags = vfsp->vfc_flags; 4236 /* 4237 * These are unused in userland, we keep them 4238 * to not break binary compatibility. 4239 */ 4240 xvfsp.vfc_vfsops = NULL; 4241 xvfsp.vfc_next = NULL; 4242 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4243 } 4244 4245 #ifdef COMPAT_FREEBSD32 4246 struct xvfsconf32 { 4247 uint32_t vfc_vfsops; 4248 char vfc_name[MFSNAMELEN]; 4249 int32_t vfc_typenum; 4250 int32_t vfc_refcount; 4251 int32_t vfc_flags; 4252 uint32_t vfc_next; 4253 }; 4254 4255 static int 4256 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 4257 { 4258 struct xvfsconf32 xvfsp; 4259 4260 bzero(&xvfsp, sizeof(xvfsp)); 4261 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4262 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4263 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4264 xvfsp.vfc_flags = vfsp->vfc_flags; 4265 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4266 } 4267 #endif 4268 4269 /* 4270 * Top level filesystem related information gathering. 4271 */ 4272 static int 4273 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 4274 { 4275 struct vfsconf *vfsp; 4276 int error; 4277 4278 error = 0; 4279 vfsconf_slock(); 4280 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4281 #ifdef COMPAT_FREEBSD32 4282 if (req->flags & SCTL_MASK32) 4283 error = vfsconf2x32(req, vfsp); 4284 else 4285 #endif 4286 error = vfsconf2x(req, vfsp); 4287 if (error) 4288 break; 4289 } 4290 vfsconf_sunlock(); 4291 return (error); 4292 } 4293 4294 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | 4295 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, 4296 "S,xvfsconf", "List of all configured filesystems"); 4297 4298 #ifndef BURN_BRIDGES 4299 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 4300 4301 static int 4302 vfs_sysctl(SYSCTL_HANDLER_ARGS) 4303 { 4304 int *name = (int *)arg1 - 1; /* XXX */ 4305 u_int namelen = arg2 + 1; /* XXX */ 4306 struct vfsconf *vfsp; 4307 4308 log(LOG_WARNING, "userland calling deprecated sysctl, " 4309 "please rebuild world\n"); 4310 4311 #if 1 || defined(COMPAT_PRELITE2) 4312 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 4313 if (namelen == 1) 4314 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 4315 #endif 4316 4317 switch (name[1]) { 4318 case VFS_MAXTYPENUM: 4319 if (namelen != 2) 4320 return (ENOTDIR); 4321 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 4322 case VFS_CONF: 4323 if (namelen != 3) 4324 return (ENOTDIR); /* overloaded */ 4325 vfsconf_slock(); 4326 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4327 if (vfsp->vfc_typenum == name[2]) 4328 break; 4329 } 4330 vfsconf_sunlock(); 4331 if (vfsp == NULL) 4332 return (EOPNOTSUPP); 4333 #ifdef COMPAT_FREEBSD32 4334 if (req->flags & SCTL_MASK32) 4335 return (vfsconf2x32(req, vfsp)); 4336 else 4337 #endif 4338 return (vfsconf2x(req, vfsp)); 4339 } 4340 return (EOPNOTSUPP); 4341 } 4342 4343 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | 4344 CTLFLAG_MPSAFE, vfs_sysctl, 4345 "Generic filesystem"); 4346 4347 #if 1 || defined(COMPAT_PRELITE2) 4348 4349 static int 4350 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 4351 { 4352 int error; 4353 struct vfsconf *vfsp; 4354 struct ovfsconf ovfs; 4355 4356 vfsconf_slock(); 4357 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4358 bzero(&ovfs, sizeof(ovfs)); 4359 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 4360 strcpy(ovfs.vfc_name, vfsp->vfc_name); 4361 ovfs.vfc_index = vfsp->vfc_typenum; 4362 ovfs.vfc_refcount = vfsp->vfc_refcount; 4363 ovfs.vfc_flags = vfsp->vfc_flags; 4364 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 4365 if (error != 0) { 4366 vfsconf_sunlock(); 4367 return (error); 4368 } 4369 } 4370 vfsconf_sunlock(); 4371 return (0); 4372 } 4373 4374 #endif /* 1 || COMPAT_PRELITE2 */ 4375 #endif /* !BURN_BRIDGES */ 4376 4377 #define KINFO_VNODESLOP 10 4378 #ifdef notyet 4379 /* 4380 * Dump vnode list (via sysctl). 4381 */ 4382 /* ARGSUSED */ 4383 static int 4384 sysctl_vnode(SYSCTL_HANDLER_ARGS) 4385 { 4386 struct xvnode *xvn; 4387 struct mount *mp; 4388 struct vnode *vp; 4389 int error, len, n; 4390 4391 /* 4392 * Stale numvnodes access is not fatal here. 4393 */ 4394 req->lock = 0; 4395 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 4396 if (!req->oldptr) 4397 /* Make an estimate */ 4398 return (SYSCTL_OUT(req, 0, len)); 4399 4400 error = sysctl_wire_old_buffer(req, 0); 4401 if (error != 0) 4402 return (error); 4403 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 4404 n = 0; 4405 mtx_lock(&mountlist_mtx); 4406 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4407 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) 4408 continue; 4409 MNT_ILOCK(mp); 4410 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4411 if (n == len) 4412 break; 4413 vref(vp); 4414 xvn[n].xv_size = sizeof *xvn; 4415 xvn[n].xv_vnode = vp; 4416 xvn[n].xv_id = 0; /* XXX compat */ 4417 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 4418 XV_COPY(usecount); 4419 XV_COPY(writecount); 4420 XV_COPY(holdcnt); 4421 XV_COPY(mount); 4422 XV_COPY(numoutput); 4423 XV_COPY(type); 4424 #undef XV_COPY 4425 xvn[n].xv_flag = vp->v_vflag; 4426 4427 switch (vp->v_type) { 4428 case VREG: 4429 case VDIR: 4430 case VLNK: 4431 break; 4432 case VBLK: 4433 case VCHR: 4434 if (vp->v_rdev == NULL) { 4435 vrele(vp); 4436 continue; 4437 } 4438 xvn[n].xv_dev = dev2udev(vp->v_rdev); 4439 break; 4440 case VSOCK: 4441 xvn[n].xv_socket = vp->v_socket; 4442 break; 4443 case VFIFO: 4444 xvn[n].xv_fifo = vp->v_fifoinfo; 4445 break; 4446 case VNON: 4447 case VBAD: 4448 default: 4449 /* shouldn't happen? */ 4450 vrele(vp); 4451 continue; 4452 } 4453 vrele(vp); 4454 ++n; 4455 } 4456 MNT_IUNLOCK(mp); 4457 mtx_lock(&mountlist_mtx); 4458 vfs_unbusy(mp); 4459 if (n == len) 4460 break; 4461 } 4462 mtx_unlock(&mountlist_mtx); 4463 4464 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 4465 free(xvn, M_TEMP); 4466 return (error); 4467 } 4468 4469 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD | 4470 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode", 4471 ""); 4472 #endif 4473 4474 static void 4475 unmount_or_warn(struct mount *mp) 4476 { 4477 int error; 4478 4479 error = dounmount(mp, MNT_FORCE, curthread); 4480 if (error != 0) { 4481 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); 4482 if (error == EBUSY) 4483 printf("BUSY)\n"); 4484 else 4485 printf("%d)\n", error); 4486 } 4487 } 4488 4489 /* 4490 * Unmount all filesystems. The list is traversed in reverse order 4491 * of mounting to avoid dependencies. 4492 */ 4493 void 4494 vfs_unmountall(void) 4495 { 4496 struct mount *mp, *tmp; 4497 4498 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 4499 4500 /* 4501 * Since this only runs when rebooting, it is not interlocked. 4502 */ 4503 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { 4504 vfs_ref(mp); 4505 4506 /* 4507 * Forcibly unmounting "/dev" before "/" would prevent clean 4508 * unmount of the latter. 4509 */ 4510 if (mp == rootdevmp) 4511 continue; 4512 4513 unmount_or_warn(mp); 4514 } 4515 4516 if (rootdevmp != NULL) 4517 unmount_or_warn(rootdevmp); 4518 } 4519 4520 static void 4521 vfs_deferred_inactive(struct vnode *vp, int lkflags) 4522 { 4523 4524 ASSERT_VI_LOCKED(vp, __func__); 4525 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set")); 4526 if ((vp->v_iflag & VI_OWEINACT) == 0) { 4527 vdropl(vp); 4528 return; 4529 } 4530 if (vn_lock(vp, lkflags) == 0) { 4531 VI_LOCK(vp); 4532 if ((vp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == VI_OWEINACT) 4533 vinactive(vp); 4534 VOP_UNLOCK(vp); 4535 vdropl(vp); 4536 return; 4537 } 4538 vdefer_inactive_cond(vp); 4539 } 4540 4541 static int 4542 vfs_periodic_inactive_filter(struct vnode *vp, void *arg) 4543 { 4544 4545 return (vp->v_iflag & VI_DEFINACT); 4546 } 4547 4548 static void __noinline 4549 vfs_periodic_inactive(struct mount *mp, int flags) 4550 { 4551 struct vnode *vp, *mvp; 4552 int lkflags; 4553 4554 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 4555 if (flags != MNT_WAIT) 4556 lkflags |= LK_NOWAIT; 4557 4558 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) { 4559 if ((vp->v_iflag & VI_DEFINACT) == 0) { 4560 VI_UNLOCK(vp); 4561 continue; 4562 } 4563 vp->v_iflag &= ~VI_DEFINACT; 4564 vfs_deferred_inactive(vp, lkflags); 4565 } 4566 } 4567 4568 static inline bool 4569 vfs_want_msync(struct vnode *vp) 4570 { 4571 struct vm_object *obj; 4572 4573 /* 4574 * This test may be performed without any locks held. 4575 * We rely on vm_object's type stability. 4576 */ 4577 if (vp->v_vflag & VV_NOSYNC) 4578 return (false); 4579 obj = vp->v_object; 4580 return (obj != NULL && vm_object_mightbedirty(obj)); 4581 } 4582 4583 static int 4584 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused) 4585 { 4586 4587 if (vp->v_vflag & VV_NOSYNC) 4588 return (false); 4589 if (vp->v_iflag & VI_DEFINACT) 4590 return (true); 4591 return (vfs_want_msync(vp)); 4592 } 4593 4594 static void __noinline 4595 vfs_periodic_msync_inactive(struct mount *mp, int flags) 4596 { 4597 struct vnode *vp, *mvp; 4598 struct vm_object *obj; 4599 struct thread *td; 4600 int lkflags, objflags; 4601 bool seen_defer; 4602 4603 td = curthread; 4604 4605 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 4606 if (flags != MNT_WAIT) { 4607 lkflags |= LK_NOWAIT; 4608 objflags = OBJPC_NOSYNC; 4609 } else { 4610 objflags = OBJPC_SYNC; 4611 } 4612 4613 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) { 4614 seen_defer = false; 4615 if (vp->v_iflag & VI_DEFINACT) { 4616 vp->v_iflag &= ~VI_DEFINACT; 4617 seen_defer = true; 4618 } 4619 if (!vfs_want_msync(vp)) { 4620 if (seen_defer) 4621 vfs_deferred_inactive(vp, lkflags); 4622 else 4623 VI_UNLOCK(vp); 4624 continue; 4625 } 4626 if (vget(vp, lkflags, td) == 0) { 4627 obj = vp->v_object; 4628 if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) { 4629 VM_OBJECT_WLOCK(obj); 4630 vm_object_page_clean(obj, 0, 0, objflags); 4631 VM_OBJECT_WUNLOCK(obj); 4632 } 4633 vput(vp); 4634 if (seen_defer) 4635 vdrop(vp); 4636 } else { 4637 if (seen_defer) 4638 vdefer_inactive_cond(vp); 4639 } 4640 } 4641 } 4642 4643 void 4644 vfs_periodic(struct mount *mp, int flags) 4645 { 4646 4647 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 4648 4649 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0) 4650 vfs_periodic_inactive(mp, flags); 4651 else 4652 vfs_periodic_msync_inactive(mp, flags); 4653 } 4654 4655 static void 4656 destroy_vpollinfo_free(struct vpollinfo *vi) 4657 { 4658 4659 knlist_destroy(&vi->vpi_selinfo.si_note); 4660 mtx_destroy(&vi->vpi_lock); 4661 uma_zfree(vnodepoll_zone, vi); 4662 } 4663 4664 static void 4665 destroy_vpollinfo(struct vpollinfo *vi) 4666 { 4667 4668 knlist_clear(&vi->vpi_selinfo.si_note, 1); 4669 seldrain(&vi->vpi_selinfo); 4670 destroy_vpollinfo_free(vi); 4671 } 4672 4673 /* 4674 * Initialize per-vnode helper structure to hold poll-related state. 4675 */ 4676 void 4677 v_addpollinfo(struct vnode *vp) 4678 { 4679 struct vpollinfo *vi; 4680 4681 if (vp->v_pollinfo != NULL) 4682 return; 4683 vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO); 4684 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 4685 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 4686 vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked); 4687 VI_LOCK(vp); 4688 if (vp->v_pollinfo != NULL) { 4689 VI_UNLOCK(vp); 4690 destroy_vpollinfo_free(vi); 4691 return; 4692 } 4693 vp->v_pollinfo = vi; 4694 VI_UNLOCK(vp); 4695 } 4696 4697 /* 4698 * Record a process's interest in events which might happen to 4699 * a vnode. Because poll uses the historic select-style interface 4700 * internally, this routine serves as both the ``check for any 4701 * pending events'' and the ``record my interest in future events'' 4702 * functions. (These are done together, while the lock is held, 4703 * to avoid race conditions.) 4704 */ 4705 int 4706 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 4707 { 4708 4709 v_addpollinfo(vp); 4710 mtx_lock(&vp->v_pollinfo->vpi_lock); 4711 if (vp->v_pollinfo->vpi_revents & events) { 4712 /* 4713 * This leaves events we are not interested 4714 * in available for the other process which 4715 * which presumably had requested them 4716 * (otherwise they would never have been 4717 * recorded). 4718 */ 4719 events &= vp->v_pollinfo->vpi_revents; 4720 vp->v_pollinfo->vpi_revents &= ~events; 4721 4722 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4723 return (events); 4724 } 4725 vp->v_pollinfo->vpi_events |= events; 4726 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 4727 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4728 return (0); 4729 } 4730 4731 /* 4732 * Routine to create and manage a filesystem syncer vnode. 4733 */ 4734 #define sync_close ((int (*)(struct vop_close_args *))nullop) 4735 static int sync_fsync(struct vop_fsync_args *); 4736 static int sync_inactive(struct vop_inactive_args *); 4737 static int sync_reclaim(struct vop_reclaim_args *); 4738 4739 static struct vop_vector sync_vnodeops = { 4740 .vop_bypass = VOP_EOPNOTSUPP, 4741 .vop_close = sync_close, /* close */ 4742 .vop_fsync = sync_fsync, /* fsync */ 4743 .vop_inactive = sync_inactive, /* inactive */ 4744 .vop_need_inactive = vop_stdneed_inactive, /* need_inactive */ 4745 .vop_reclaim = sync_reclaim, /* reclaim */ 4746 .vop_lock1 = vop_stdlock, /* lock */ 4747 .vop_unlock = vop_stdunlock, /* unlock */ 4748 .vop_islocked = vop_stdislocked, /* islocked */ 4749 }; 4750 VFS_VOP_VECTOR_REGISTER(sync_vnodeops); 4751 4752 /* 4753 * Create a new filesystem syncer vnode for the specified mount point. 4754 */ 4755 void 4756 vfs_allocate_syncvnode(struct mount *mp) 4757 { 4758 struct vnode *vp; 4759 struct bufobj *bo; 4760 static long start, incr, next; 4761 int error; 4762 4763 /* Allocate a new vnode */ 4764 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 4765 if (error != 0) 4766 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 4767 vp->v_type = VNON; 4768 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4769 vp->v_vflag |= VV_FORCEINSMQ; 4770 error = insmntque(vp, mp); 4771 if (error != 0) 4772 panic("vfs_allocate_syncvnode: insmntque() failed"); 4773 vp->v_vflag &= ~VV_FORCEINSMQ; 4774 VOP_UNLOCK(vp); 4775 /* 4776 * Place the vnode onto the syncer worklist. We attempt to 4777 * scatter them about on the list so that they will go off 4778 * at evenly distributed times even if all the filesystems 4779 * are mounted at once. 4780 */ 4781 next += incr; 4782 if (next == 0 || next > syncer_maxdelay) { 4783 start /= 2; 4784 incr /= 2; 4785 if (start == 0) { 4786 start = syncer_maxdelay / 2; 4787 incr = syncer_maxdelay; 4788 } 4789 next = start; 4790 } 4791 bo = &vp->v_bufobj; 4792 BO_LOCK(bo); 4793 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 4794 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 4795 mtx_lock(&sync_mtx); 4796 sync_vnode_count++; 4797 if (mp->mnt_syncer == NULL) { 4798 mp->mnt_syncer = vp; 4799 vp = NULL; 4800 } 4801 mtx_unlock(&sync_mtx); 4802 BO_UNLOCK(bo); 4803 if (vp != NULL) { 4804 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4805 vgone(vp); 4806 vput(vp); 4807 } 4808 } 4809 4810 void 4811 vfs_deallocate_syncvnode(struct mount *mp) 4812 { 4813 struct vnode *vp; 4814 4815 mtx_lock(&sync_mtx); 4816 vp = mp->mnt_syncer; 4817 if (vp != NULL) 4818 mp->mnt_syncer = NULL; 4819 mtx_unlock(&sync_mtx); 4820 if (vp != NULL) 4821 vrele(vp); 4822 } 4823 4824 /* 4825 * Do a lazy sync of the filesystem. 4826 */ 4827 static int 4828 sync_fsync(struct vop_fsync_args *ap) 4829 { 4830 struct vnode *syncvp = ap->a_vp; 4831 struct mount *mp = syncvp->v_mount; 4832 int error, save; 4833 struct bufobj *bo; 4834 4835 /* 4836 * We only need to do something if this is a lazy evaluation. 4837 */ 4838 if (ap->a_waitfor != MNT_LAZY) 4839 return (0); 4840 4841 /* 4842 * Move ourselves to the back of the sync list. 4843 */ 4844 bo = &syncvp->v_bufobj; 4845 BO_LOCK(bo); 4846 vn_syncer_add_to_worklist(bo, syncdelay); 4847 BO_UNLOCK(bo); 4848 4849 /* 4850 * Walk the list of vnodes pushing all that are dirty and 4851 * not already on the sync list. 4852 */ 4853 if (vfs_busy(mp, MBF_NOWAIT) != 0) 4854 return (0); 4855 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 4856 vfs_unbusy(mp); 4857 return (0); 4858 } 4859 save = curthread_pflags_set(TDP_SYNCIO); 4860 /* 4861 * The filesystem at hand may be idle with free vnodes stored in the 4862 * batch. Return them instead of letting them stay there indefinitely. 4863 */ 4864 vfs_periodic(mp, MNT_NOWAIT); 4865 error = VFS_SYNC(mp, MNT_LAZY); 4866 curthread_pflags_restore(save); 4867 vn_finished_write(mp); 4868 vfs_unbusy(mp); 4869 return (error); 4870 } 4871 4872 /* 4873 * The syncer vnode is no referenced. 4874 */ 4875 static int 4876 sync_inactive(struct vop_inactive_args *ap) 4877 { 4878 4879 vgone(ap->a_vp); 4880 return (0); 4881 } 4882 4883 /* 4884 * The syncer vnode is no longer needed and is being decommissioned. 4885 * 4886 * Modifications to the worklist must be protected by sync_mtx. 4887 */ 4888 static int 4889 sync_reclaim(struct vop_reclaim_args *ap) 4890 { 4891 struct vnode *vp = ap->a_vp; 4892 struct bufobj *bo; 4893 4894 bo = &vp->v_bufobj; 4895 BO_LOCK(bo); 4896 mtx_lock(&sync_mtx); 4897 if (vp->v_mount->mnt_syncer == vp) 4898 vp->v_mount->mnt_syncer = NULL; 4899 if (bo->bo_flag & BO_ONWORKLST) { 4900 LIST_REMOVE(bo, bo_synclist); 4901 syncer_worklist_len--; 4902 sync_vnode_count--; 4903 bo->bo_flag &= ~BO_ONWORKLST; 4904 } 4905 mtx_unlock(&sync_mtx); 4906 BO_UNLOCK(bo); 4907 4908 return (0); 4909 } 4910 4911 int 4912 vn_need_pageq_flush(struct vnode *vp) 4913 { 4914 struct vm_object *obj; 4915 int need; 4916 4917 MPASS(mtx_owned(VI_MTX(vp))); 4918 need = 0; 4919 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 4920 vm_object_mightbedirty(obj)) 4921 need = 1; 4922 return (need); 4923 } 4924 4925 /* 4926 * Check if vnode represents a disk device 4927 */ 4928 int 4929 vn_isdisk(struct vnode *vp, int *errp) 4930 { 4931 int error; 4932 4933 if (vp->v_type != VCHR) { 4934 error = ENOTBLK; 4935 goto out; 4936 } 4937 error = 0; 4938 dev_lock(); 4939 if (vp->v_rdev == NULL) 4940 error = ENXIO; 4941 else if (vp->v_rdev->si_devsw == NULL) 4942 error = ENXIO; 4943 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 4944 error = ENOTBLK; 4945 dev_unlock(); 4946 out: 4947 if (errp != NULL) 4948 *errp = error; 4949 return (error == 0); 4950 } 4951 4952 /* 4953 * Common filesystem object access control check routine. Accepts a 4954 * vnode's type, "mode", uid and gid, requested access mode, credentials, 4955 * and optional call-by-reference privused argument allowing vaccess() 4956 * to indicate to the caller whether privilege was used to satisfy the 4957 * request (obsoleted). Returns 0 on success, or an errno on failure. 4958 */ 4959 int 4960 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 4961 accmode_t accmode, struct ucred *cred, int *privused) 4962 { 4963 accmode_t dac_granted; 4964 accmode_t priv_granted; 4965 4966 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 4967 ("invalid bit in accmode")); 4968 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 4969 ("VAPPEND without VWRITE")); 4970 4971 /* 4972 * Look for a normal, non-privileged way to access the file/directory 4973 * as requested. If it exists, go with that. 4974 */ 4975 4976 if (privused != NULL) 4977 *privused = 0; 4978 4979 dac_granted = 0; 4980 4981 /* Check the owner. */ 4982 if (cred->cr_uid == file_uid) { 4983 dac_granted |= VADMIN; 4984 if (file_mode & S_IXUSR) 4985 dac_granted |= VEXEC; 4986 if (file_mode & S_IRUSR) 4987 dac_granted |= VREAD; 4988 if (file_mode & S_IWUSR) 4989 dac_granted |= (VWRITE | VAPPEND); 4990 4991 if ((accmode & dac_granted) == accmode) 4992 return (0); 4993 4994 goto privcheck; 4995 } 4996 4997 /* Otherwise, check the groups (first match) */ 4998 if (groupmember(file_gid, cred)) { 4999 if (file_mode & S_IXGRP) 5000 dac_granted |= VEXEC; 5001 if (file_mode & S_IRGRP) 5002 dac_granted |= VREAD; 5003 if (file_mode & S_IWGRP) 5004 dac_granted |= (VWRITE | VAPPEND); 5005 5006 if ((accmode & dac_granted) == accmode) 5007 return (0); 5008 5009 goto privcheck; 5010 } 5011 5012 /* Otherwise, check everyone else. */ 5013 if (file_mode & S_IXOTH) 5014 dac_granted |= VEXEC; 5015 if (file_mode & S_IROTH) 5016 dac_granted |= VREAD; 5017 if (file_mode & S_IWOTH) 5018 dac_granted |= (VWRITE | VAPPEND); 5019 if ((accmode & dac_granted) == accmode) 5020 return (0); 5021 5022 privcheck: 5023 /* 5024 * Build a privilege mask to determine if the set of privileges 5025 * satisfies the requirements when combined with the granted mask 5026 * from above. For each privilege, if the privilege is required, 5027 * bitwise or the request type onto the priv_granted mask. 5028 */ 5029 priv_granted = 0; 5030 5031 if (type == VDIR) { 5032 /* 5033 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 5034 * requests, instead of PRIV_VFS_EXEC. 5035 */ 5036 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5037 !priv_check_cred(cred, PRIV_VFS_LOOKUP)) 5038 priv_granted |= VEXEC; 5039 } else { 5040 /* 5041 * Ensure that at least one execute bit is on. Otherwise, 5042 * a privileged user will always succeed, and we don't want 5043 * this to happen unless the file really is executable. 5044 */ 5045 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5046 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 5047 !priv_check_cred(cred, PRIV_VFS_EXEC)) 5048 priv_granted |= VEXEC; 5049 } 5050 5051 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 5052 !priv_check_cred(cred, PRIV_VFS_READ)) 5053 priv_granted |= VREAD; 5054 5055 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 5056 !priv_check_cred(cred, PRIV_VFS_WRITE)) 5057 priv_granted |= (VWRITE | VAPPEND); 5058 5059 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 5060 !priv_check_cred(cred, PRIV_VFS_ADMIN)) 5061 priv_granted |= VADMIN; 5062 5063 if ((accmode & (priv_granted | dac_granted)) == accmode) { 5064 /* XXX audit: privilege used */ 5065 if (privused != NULL) 5066 *privused = 1; 5067 return (0); 5068 } 5069 5070 return ((accmode & VADMIN) ? EPERM : EACCES); 5071 } 5072 5073 /* 5074 * Credential check based on process requesting service, and per-attribute 5075 * permissions. 5076 */ 5077 int 5078 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 5079 struct thread *td, accmode_t accmode) 5080 { 5081 5082 /* 5083 * Kernel-invoked always succeeds. 5084 */ 5085 if (cred == NOCRED) 5086 return (0); 5087 5088 /* 5089 * Do not allow privileged processes in jail to directly manipulate 5090 * system attributes. 5091 */ 5092 switch (attrnamespace) { 5093 case EXTATTR_NAMESPACE_SYSTEM: 5094 /* Potentially should be: return (EPERM); */ 5095 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM)); 5096 case EXTATTR_NAMESPACE_USER: 5097 return (VOP_ACCESS(vp, accmode, cred, td)); 5098 default: 5099 return (EPERM); 5100 } 5101 } 5102 5103 #ifdef DEBUG_VFS_LOCKS 5104 /* 5105 * This only exists to suppress warnings from unlocked specfs accesses. It is 5106 * no longer ok to have an unlocked VFS. 5107 */ 5108 #define IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL || \ 5109 (vp)->v_type == VCHR || (vp)->v_type == VBAD) 5110 5111 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 5112 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, 5113 "Drop into debugger on lock violation"); 5114 5115 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 5116 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 5117 0, "Check for interlock across VOPs"); 5118 5119 int vfs_badlock_print = 1; /* Print lock violations. */ 5120 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 5121 0, "Print lock violations"); 5122 5123 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */ 5124 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode, 5125 0, "Print vnode details on lock violations"); 5126 5127 #ifdef KDB 5128 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 5129 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, 5130 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations"); 5131 #endif 5132 5133 static void 5134 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 5135 { 5136 5137 #ifdef KDB 5138 if (vfs_badlock_backtrace) 5139 kdb_backtrace(); 5140 #endif 5141 if (vfs_badlock_vnode) 5142 vn_printf(vp, "vnode "); 5143 if (vfs_badlock_print) 5144 printf("%s: %p %s\n", str, (void *)vp, msg); 5145 if (vfs_badlock_ddb) 5146 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 5147 } 5148 5149 void 5150 assert_vi_locked(struct vnode *vp, const char *str) 5151 { 5152 5153 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 5154 vfs_badlock("interlock is not locked but should be", str, vp); 5155 } 5156 5157 void 5158 assert_vi_unlocked(struct vnode *vp, const char *str) 5159 { 5160 5161 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 5162 vfs_badlock("interlock is locked but should not be", str, vp); 5163 } 5164 5165 void 5166 assert_vop_locked(struct vnode *vp, const char *str) 5167 { 5168 int locked; 5169 5170 if (!IGNORE_LOCK(vp)) { 5171 locked = VOP_ISLOCKED(vp); 5172 if (locked == 0 || locked == LK_EXCLOTHER) 5173 vfs_badlock("is not locked but should be", str, vp); 5174 } 5175 } 5176 5177 void 5178 assert_vop_unlocked(struct vnode *vp, const char *str) 5179 { 5180 5181 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE) 5182 vfs_badlock("is locked but should not be", str, vp); 5183 } 5184 5185 void 5186 assert_vop_elocked(struct vnode *vp, const char *str) 5187 { 5188 5189 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 5190 vfs_badlock("is not exclusive locked but should be", str, vp); 5191 } 5192 #endif /* DEBUG_VFS_LOCKS */ 5193 5194 void 5195 vop_rename_fail(struct vop_rename_args *ap) 5196 { 5197 5198 if (ap->a_tvp != NULL) 5199 vput(ap->a_tvp); 5200 if (ap->a_tdvp == ap->a_tvp) 5201 vrele(ap->a_tdvp); 5202 else 5203 vput(ap->a_tdvp); 5204 vrele(ap->a_fdvp); 5205 vrele(ap->a_fvp); 5206 } 5207 5208 void 5209 vop_rename_pre(void *ap) 5210 { 5211 struct vop_rename_args *a = ap; 5212 5213 #ifdef DEBUG_VFS_LOCKS 5214 if (a->a_tvp) 5215 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 5216 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 5217 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 5218 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 5219 5220 /* Check the source (from). */ 5221 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 5222 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 5223 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 5224 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 5225 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 5226 5227 /* Check the target. */ 5228 if (a->a_tvp) 5229 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 5230 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 5231 #endif 5232 if (a->a_tdvp != a->a_fdvp) 5233 vhold(a->a_fdvp); 5234 if (a->a_tvp != a->a_fvp) 5235 vhold(a->a_fvp); 5236 vhold(a->a_tdvp); 5237 if (a->a_tvp) 5238 vhold(a->a_tvp); 5239 } 5240 5241 #ifdef DEBUG_VFS_LOCKS 5242 void 5243 vop_strategy_pre(void *ap) 5244 { 5245 struct vop_strategy_args *a; 5246 struct buf *bp; 5247 5248 a = ap; 5249 bp = a->a_bp; 5250 5251 /* 5252 * Cluster ops lock their component buffers but not the IO container. 5253 */ 5254 if ((bp->b_flags & B_CLUSTER) != 0) 5255 return; 5256 5257 if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) { 5258 if (vfs_badlock_print) 5259 printf( 5260 "VOP_STRATEGY: bp is not locked but should be\n"); 5261 if (vfs_badlock_ddb) 5262 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 5263 } 5264 } 5265 5266 void 5267 vop_lock_pre(void *ap) 5268 { 5269 struct vop_lock1_args *a = ap; 5270 5271 if ((a->a_flags & LK_INTERLOCK) == 0) 5272 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5273 else 5274 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 5275 } 5276 5277 void 5278 vop_lock_post(void *ap, int rc) 5279 { 5280 struct vop_lock1_args *a = ap; 5281 5282 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5283 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) 5284 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 5285 } 5286 5287 void 5288 vop_unlock_pre(void *ap) 5289 { 5290 struct vop_unlock_args *a = ap; 5291 5292 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 5293 } 5294 5295 void 5296 vop_unlock_post(void *ap, int rc) 5297 { 5298 return; 5299 } 5300 5301 void 5302 vop_need_inactive_pre(void *ap) 5303 { 5304 struct vop_need_inactive_args *a = ap; 5305 5306 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 5307 } 5308 5309 void 5310 vop_need_inactive_post(void *ap, int rc) 5311 { 5312 struct vop_need_inactive_args *a = ap; 5313 5314 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 5315 } 5316 #endif 5317 5318 void 5319 vop_create_post(void *ap, int rc) 5320 { 5321 struct vop_create_args *a = ap; 5322 5323 if (!rc) 5324 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 5325 } 5326 5327 void 5328 vop_deleteextattr_post(void *ap, int rc) 5329 { 5330 struct vop_deleteextattr_args *a = ap; 5331 5332 if (!rc) 5333 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 5334 } 5335 5336 void 5337 vop_link_post(void *ap, int rc) 5338 { 5339 struct vop_link_args *a = ap; 5340 5341 if (!rc) { 5342 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK); 5343 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE); 5344 } 5345 } 5346 5347 void 5348 vop_mkdir_post(void *ap, int rc) 5349 { 5350 struct vop_mkdir_args *a = ap; 5351 5352 if (!rc) 5353 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 5354 } 5355 5356 void 5357 vop_mknod_post(void *ap, int rc) 5358 { 5359 struct vop_mknod_args *a = ap; 5360 5361 if (!rc) 5362 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 5363 } 5364 5365 void 5366 vop_reclaim_post(void *ap, int rc) 5367 { 5368 struct vop_reclaim_args *a = ap; 5369 5370 if (!rc) 5371 VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE); 5372 } 5373 5374 void 5375 vop_remove_post(void *ap, int rc) 5376 { 5377 struct vop_remove_args *a = ap; 5378 5379 if (!rc) { 5380 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 5381 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 5382 } 5383 } 5384 5385 void 5386 vop_rename_post(void *ap, int rc) 5387 { 5388 struct vop_rename_args *a = ap; 5389 long hint; 5390 5391 if (!rc) { 5392 hint = NOTE_WRITE; 5393 if (a->a_fdvp == a->a_tdvp) { 5394 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) 5395 hint |= NOTE_LINK; 5396 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 5397 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 5398 } else { 5399 hint |= NOTE_EXTEND; 5400 if (a->a_fvp->v_type == VDIR) 5401 hint |= NOTE_LINK; 5402 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 5403 5404 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && 5405 a->a_tvp->v_type == VDIR) 5406 hint &= ~NOTE_LINK; 5407 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 5408 } 5409 5410 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 5411 if (a->a_tvp) 5412 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 5413 } 5414 if (a->a_tdvp != a->a_fdvp) 5415 vdrop(a->a_fdvp); 5416 if (a->a_tvp != a->a_fvp) 5417 vdrop(a->a_fvp); 5418 vdrop(a->a_tdvp); 5419 if (a->a_tvp) 5420 vdrop(a->a_tvp); 5421 } 5422 5423 void 5424 vop_rmdir_post(void *ap, int rc) 5425 { 5426 struct vop_rmdir_args *a = ap; 5427 5428 if (!rc) { 5429 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 5430 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 5431 } 5432 } 5433 5434 void 5435 vop_setattr_post(void *ap, int rc) 5436 { 5437 struct vop_setattr_args *a = ap; 5438 5439 if (!rc) 5440 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 5441 } 5442 5443 void 5444 vop_setextattr_post(void *ap, int rc) 5445 { 5446 struct vop_setextattr_args *a = ap; 5447 5448 if (!rc) 5449 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 5450 } 5451 5452 void 5453 vop_symlink_post(void *ap, int rc) 5454 { 5455 struct vop_symlink_args *a = ap; 5456 5457 if (!rc) 5458 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 5459 } 5460 5461 void 5462 vop_open_post(void *ap, int rc) 5463 { 5464 struct vop_open_args *a = ap; 5465 5466 if (!rc) 5467 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); 5468 } 5469 5470 void 5471 vop_close_post(void *ap, int rc) 5472 { 5473 struct vop_close_args *a = ap; 5474 5475 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ 5476 !VN_IS_DOOMED(a->a_vp))) { 5477 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 5478 NOTE_CLOSE_WRITE : NOTE_CLOSE); 5479 } 5480 } 5481 5482 void 5483 vop_read_post(void *ap, int rc) 5484 { 5485 struct vop_read_args *a = ap; 5486 5487 if (!rc) 5488 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 5489 } 5490 5491 void 5492 vop_readdir_post(void *ap, int rc) 5493 { 5494 struct vop_readdir_args *a = ap; 5495 5496 if (!rc) 5497 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 5498 } 5499 5500 static struct knlist fs_knlist; 5501 5502 static void 5503 vfs_event_init(void *arg) 5504 { 5505 knlist_init_mtx(&fs_knlist, NULL); 5506 } 5507 /* XXX - correct order? */ 5508 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 5509 5510 void 5511 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 5512 { 5513 5514 KNOTE_UNLOCKED(&fs_knlist, event); 5515 } 5516 5517 static int filt_fsattach(struct knote *kn); 5518 static void filt_fsdetach(struct knote *kn); 5519 static int filt_fsevent(struct knote *kn, long hint); 5520 5521 struct filterops fs_filtops = { 5522 .f_isfd = 0, 5523 .f_attach = filt_fsattach, 5524 .f_detach = filt_fsdetach, 5525 .f_event = filt_fsevent 5526 }; 5527 5528 static int 5529 filt_fsattach(struct knote *kn) 5530 { 5531 5532 kn->kn_flags |= EV_CLEAR; 5533 knlist_add(&fs_knlist, kn, 0); 5534 return (0); 5535 } 5536 5537 static void 5538 filt_fsdetach(struct knote *kn) 5539 { 5540 5541 knlist_remove(&fs_knlist, kn, 0); 5542 } 5543 5544 static int 5545 filt_fsevent(struct knote *kn, long hint) 5546 { 5547 5548 kn->kn_fflags |= hint; 5549 return (kn->kn_fflags != 0); 5550 } 5551 5552 static int 5553 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 5554 { 5555 struct vfsidctl vc; 5556 int error; 5557 struct mount *mp; 5558 5559 error = SYSCTL_IN(req, &vc, sizeof(vc)); 5560 if (error) 5561 return (error); 5562 if (vc.vc_vers != VFS_CTL_VERS1) 5563 return (EINVAL); 5564 mp = vfs_getvfs(&vc.vc_fsid); 5565 if (mp == NULL) 5566 return (ENOENT); 5567 /* ensure that a specific sysctl goes to the right filesystem. */ 5568 if (strcmp(vc.vc_fstypename, "*") != 0 && 5569 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 5570 vfs_rel(mp); 5571 return (EINVAL); 5572 } 5573 VCTLTOREQ(&vc, req); 5574 error = VFS_SYSCTL(mp, vc.vc_op, req); 5575 vfs_rel(mp); 5576 return (error); 5577 } 5578 5579 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR, 5580 NULL, 0, sysctl_vfs_ctl, "", 5581 "Sysctl by fsid"); 5582 5583 /* 5584 * Function to initialize a va_filerev field sensibly. 5585 * XXX: Wouldn't a random number make a lot more sense ?? 5586 */ 5587 u_quad_t 5588 init_va_filerev(void) 5589 { 5590 struct bintime bt; 5591 5592 getbinuptime(&bt); 5593 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 5594 } 5595 5596 static int filt_vfsread(struct knote *kn, long hint); 5597 static int filt_vfswrite(struct knote *kn, long hint); 5598 static int filt_vfsvnode(struct knote *kn, long hint); 5599 static void filt_vfsdetach(struct knote *kn); 5600 static struct filterops vfsread_filtops = { 5601 .f_isfd = 1, 5602 .f_detach = filt_vfsdetach, 5603 .f_event = filt_vfsread 5604 }; 5605 static struct filterops vfswrite_filtops = { 5606 .f_isfd = 1, 5607 .f_detach = filt_vfsdetach, 5608 .f_event = filt_vfswrite 5609 }; 5610 static struct filterops vfsvnode_filtops = { 5611 .f_isfd = 1, 5612 .f_detach = filt_vfsdetach, 5613 .f_event = filt_vfsvnode 5614 }; 5615 5616 static void 5617 vfs_knllock(void *arg) 5618 { 5619 struct vnode *vp = arg; 5620 5621 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5622 } 5623 5624 static void 5625 vfs_knlunlock(void *arg) 5626 { 5627 struct vnode *vp = arg; 5628 5629 VOP_UNLOCK(vp); 5630 } 5631 5632 static void 5633 vfs_knl_assert_locked(void *arg) 5634 { 5635 #ifdef DEBUG_VFS_LOCKS 5636 struct vnode *vp = arg; 5637 5638 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 5639 #endif 5640 } 5641 5642 static void 5643 vfs_knl_assert_unlocked(void *arg) 5644 { 5645 #ifdef DEBUG_VFS_LOCKS 5646 struct vnode *vp = arg; 5647 5648 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 5649 #endif 5650 } 5651 5652 int 5653 vfs_kqfilter(struct vop_kqfilter_args *ap) 5654 { 5655 struct vnode *vp = ap->a_vp; 5656 struct knote *kn = ap->a_kn; 5657 struct knlist *knl; 5658 5659 switch (kn->kn_filter) { 5660 case EVFILT_READ: 5661 kn->kn_fop = &vfsread_filtops; 5662 break; 5663 case EVFILT_WRITE: 5664 kn->kn_fop = &vfswrite_filtops; 5665 break; 5666 case EVFILT_VNODE: 5667 kn->kn_fop = &vfsvnode_filtops; 5668 break; 5669 default: 5670 return (EINVAL); 5671 } 5672 5673 kn->kn_hook = (caddr_t)vp; 5674 5675 v_addpollinfo(vp); 5676 if (vp->v_pollinfo == NULL) 5677 return (ENOMEM); 5678 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 5679 vhold(vp); 5680 knlist_add(knl, kn, 0); 5681 5682 return (0); 5683 } 5684 5685 /* 5686 * Detach knote from vnode 5687 */ 5688 static void 5689 filt_vfsdetach(struct knote *kn) 5690 { 5691 struct vnode *vp = (struct vnode *)kn->kn_hook; 5692 5693 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 5694 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 5695 vdrop(vp); 5696 } 5697 5698 /*ARGSUSED*/ 5699 static int 5700 filt_vfsread(struct knote *kn, long hint) 5701 { 5702 struct vnode *vp = (struct vnode *)kn->kn_hook; 5703 struct vattr va; 5704 int res; 5705 5706 /* 5707 * filesystem is gone, so set the EOF flag and schedule 5708 * the knote for deletion. 5709 */ 5710 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 5711 VI_LOCK(vp); 5712 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 5713 VI_UNLOCK(vp); 5714 return (1); 5715 } 5716 5717 if (VOP_GETATTR(vp, &va, curthread->td_ucred)) 5718 return (0); 5719 5720 VI_LOCK(vp); 5721 kn->kn_data = va.va_size - kn->kn_fp->f_offset; 5722 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; 5723 VI_UNLOCK(vp); 5724 return (res); 5725 } 5726 5727 /*ARGSUSED*/ 5728 static int 5729 filt_vfswrite(struct knote *kn, long hint) 5730 { 5731 struct vnode *vp = (struct vnode *)kn->kn_hook; 5732 5733 VI_LOCK(vp); 5734 5735 /* 5736 * filesystem is gone, so set the EOF flag and schedule 5737 * the knote for deletion. 5738 */ 5739 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) 5740 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 5741 5742 kn->kn_data = 0; 5743 VI_UNLOCK(vp); 5744 return (1); 5745 } 5746 5747 static int 5748 filt_vfsvnode(struct knote *kn, long hint) 5749 { 5750 struct vnode *vp = (struct vnode *)kn->kn_hook; 5751 int res; 5752 5753 VI_LOCK(vp); 5754 if (kn->kn_sfflags & hint) 5755 kn->kn_fflags |= hint; 5756 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 5757 kn->kn_flags |= EV_EOF; 5758 VI_UNLOCK(vp); 5759 return (1); 5760 } 5761 res = (kn->kn_fflags != 0); 5762 VI_UNLOCK(vp); 5763 return (res); 5764 } 5765 5766 /* 5767 * Returns whether the directory is empty or not. 5768 * If it is empty, the return value is 0; otherwise 5769 * the return value is an error value (which may 5770 * be ENOTEMPTY). 5771 */ 5772 int 5773 vfs_emptydir(struct vnode *vp) 5774 { 5775 struct uio uio; 5776 struct iovec iov; 5777 struct dirent *dirent, *dp, *endp; 5778 int error, eof; 5779 5780 error = 0; 5781 eof = 0; 5782 5783 ASSERT_VOP_LOCKED(vp, "vfs_emptydir"); 5784 5785 dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK); 5786 iov.iov_base = dirent; 5787 iov.iov_len = sizeof(struct dirent); 5788 5789 uio.uio_iov = &iov; 5790 uio.uio_iovcnt = 1; 5791 uio.uio_offset = 0; 5792 uio.uio_resid = sizeof(struct dirent); 5793 uio.uio_segflg = UIO_SYSSPACE; 5794 uio.uio_rw = UIO_READ; 5795 uio.uio_td = curthread; 5796 5797 while (eof == 0 && error == 0) { 5798 error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof, 5799 NULL, NULL); 5800 if (error != 0) 5801 break; 5802 endp = (void *)((uint8_t *)dirent + 5803 sizeof(struct dirent) - uio.uio_resid); 5804 for (dp = dirent; dp < endp; 5805 dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) { 5806 if (dp->d_type == DT_WHT) 5807 continue; 5808 if (dp->d_namlen == 0) 5809 continue; 5810 if (dp->d_type != DT_DIR && 5811 dp->d_type != DT_UNKNOWN) { 5812 error = ENOTEMPTY; 5813 break; 5814 } 5815 if (dp->d_namlen > 2) { 5816 error = ENOTEMPTY; 5817 break; 5818 } 5819 if (dp->d_namlen == 1 && 5820 dp->d_name[0] != '.') { 5821 error = ENOTEMPTY; 5822 break; 5823 } 5824 if (dp->d_namlen == 2 && 5825 dp->d_name[1] != '.') { 5826 error = ENOTEMPTY; 5827 break; 5828 } 5829 uio.uio_resid = sizeof(struct dirent); 5830 } 5831 } 5832 free(dirent, M_TEMP); 5833 return (error); 5834 } 5835 5836 int 5837 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 5838 { 5839 int error; 5840 5841 if (dp->d_reclen > ap->a_uio->uio_resid) 5842 return (ENAMETOOLONG); 5843 error = uiomove(dp, dp->d_reclen, ap->a_uio); 5844 if (error) { 5845 if (ap->a_ncookies != NULL) { 5846 if (ap->a_cookies != NULL) 5847 free(ap->a_cookies, M_TEMP); 5848 ap->a_cookies = NULL; 5849 *ap->a_ncookies = 0; 5850 } 5851 return (error); 5852 } 5853 if (ap->a_ncookies == NULL) 5854 return (0); 5855 5856 KASSERT(ap->a_cookies, 5857 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 5858 5859 *ap->a_cookies = realloc(*ap->a_cookies, 5860 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); 5861 (*ap->a_cookies)[*ap->a_ncookies] = off; 5862 *ap->a_ncookies += 1; 5863 return (0); 5864 } 5865 5866 /* 5867 * Mark for update the access time of the file if the filesystem 5868 * supports VOP_MARKATIME. This functionality is used by execve and 5869 * mmap, so we want to avoid the I/O implied by directly setting 5870 * va_atime for the sake of efficiency. 5871 */ 5872 void 5873 vfs_mark_atime(struct vnode *vp, struct ucred *cred) 5874 { 5875 struct mount *mp; 5876 5877 mp = vp->v_mount; 5878 ASSERT_VOP_LOCKED(vp, "vfs_mark_atime"); 5879 if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) 5880 (void)VOP_MARKATIME(vp); 5881 } 5882 5883 /* 5884 * The purpose of this routine is to remove granularity from accmode_t, 5885 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 5886 * VADMIN and VAPPEND. 5887 * 5888 * If it returns 0, the caller is supposed to continue with the usual 5889 * access checks using 'accmode' as modified by this routine. If it 5890 * returns nonzero value, the caller is supposed to return that value 5891 * as errno. 5892 * 5893 * Note that after this routine runs, accmode may be zero. 5894 */ 5895 int 5896 vfs_unixify_accmode(accmode_t *accmode) 5897 { 5898 /* 5899 * There is no way to specify explicit "deny" rule using 5900 * file mode or POSIX.1e ACLs. 5901 */ 5902 if (*accmode & VEXPLICIT_DENY) { 5903 *accmode = 0; 5904 return (0); 5905 } 5906 5907 /* 5908 * None of these can be translated into usual access bits. 5909 * Also, the common case for NFSv4 ACLs is to not contain 5910 * either of these bits. Caller should check for VWRITE 5911 * on the containing directory instead. 5912 */ 5913 if (*accmode & (VDELETE_CHILD | VDELETE)) 5914 return (EPERM); 5915 5916 if (*accmode & VADMIN_PERMS) { 5917 *accmode &= ~VADMIN_PERMS; 5918 *accmode |= VADMIN; 5919 } 5920 5921 /* 5922 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 5923 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 5924 */ 5925 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 5926 5927 return (0); 5928 } 5929 5930 /* 5931 * Clear out a doomed vnode (if any) and replace it with a new one as long 5932 * as the fs is not being unmounted. Return the root vnode to the caller. 5933 */ 5934 static int __noinline 5935 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp) 5936 { 5937 struct vnode *vp; 5938 int error; 5939 5940 restart: 5941 if (mp->mnt_rootvnode != NULL) { 5942 MNT_ILOCK(mp); 5943 vp = mp->mnt_rootvnode; 5944 if (vp != NULL) { 5945 if (!VN_IS_DOOMED(vp)) { 5946 vrefact(vp); 5947 MNT_IUNLOCK(mp); 5948 error = vn_lock(vp, flags); 5949 if (error == 0) { 5950 *vpp = vp; 5951 return (0); 5952 } 5953 vrele(vp); 5954 goto restart; 5955 } 5956 /* 5957 * Clear the old one. 5958 */ 5959 mp->mnt_rootvnode = NULL; 5960 } 5961 MNT_IUNLOCK(mp); 5962 if (vp != NULL) { 5963 /* 5964 * Paired with a fence in vfs_op_thread_exit(). 5965 */ 5966 atomic_thread_fence_acq(); 5967 vfs_op_barrier_wait(mp); 5968 vrele(vp); 5969 } 5970 } 5971 error = VFS_CACHEDROOT(mp, flags, vpp); 5972 if (error != 0) 5973 return (error); 5974 if (mp->mnt_vfs_ops == 0) { 5975 MNT_ILOCK(mp); 5976 if (mp->mnt_vfs_ops != 0) { 5977 MNT_IUNLOCK(mp); 5978 return (0); 5979 } 5980 if (mp->mnt_rootvnode == NULL) { 5981 vrefact(*vpp); 5982 mp->mnt_rootvnode = *vpp; 5983 } else { 5984 if (mp->mnt_rootvnode != *vpp) { 5985 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) { 5986 panic("%s: mismatch between vnode returned " 5987 " by VFS_CACHEDROOT and the one cached " 5988 " (%p != %p)", 5989 __func__, *vpp, mp->mnt_rootvnode); 5990 } 5991 } 5992 } 5993 MNT_IUNLOCK(mp); 5994 } 5995 return (0); 5996 } 5997 5998 int 5999 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp) 6000 { 6001 struct vnode *vp; 6002 int error; 6003 6004 if (!vfs_op_thread_enter(mp)) 6005 return (vfs_cache_root_fallback(mp, flags, vpp)); 6006 vp = (struct vnode *)atomic_load_ptr(&mp->mnt_rootvnode); 6007 if (vp == NULL || VN_IS_DOOMED(vp)) { 6008 vfs_op_thread_exit(mp); 6009 return (vfs_cache_root_fallback(mp, flags, vpp)); 6010 } 6011 vrefact(vp); 6012 vfs_op_thread_exit(mp); 6013 error = vn_lock(vp, flags); 6014 if (error != 0) { 6015 vrele(vp); 6016 return (vfs_cache_root_fallback(mp, flags, vpp)); 6017 } 6018 *vpp = vp; 6019 return (0); 6020 } 6021 6022 struct vnode * 6023 vfs_cache_root_clear(struct mount *mp) 6024 { 6025 struct vnode *vp; 6026 6027 /* 6028 * ops > 0 guarantees there is nobody who can see this vnode 6029 */ 6030 MPASS(mp->mnt_vfs_ops > 0); 6031 vp = mp->mnt_rootvnode; 6032 mp->mnt_rootvnode = NULL; 6033 return (vp); 6034 } 6035 6036 void 6037 vfs_cache_root_set(struct mount *mp, struct vnode *vp) 6038 { 6039 6040 MPASS(mp->mnt_vfs_ops > 0); 6041 vrefact(vp); 6042 mp->mnt_rootvnode = vp; 6043 } 6044 6045 /* 6046 * These are helper functions for filesystems to traverse all 6047 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 6048 * 6049 * This interface replaces MNT_VNODE_FOREACH. 6050 */ 6051 6052 6053 struct vnode * 6054 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 6055 { 6056 struct vnode *vp; 6057 6058 if (should_yield()) 6059 kern_yield(PRI_USER); 6060 MNT_ILOCK(mp); 6061 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6062 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL; 6063 vp = TAILQ_NEXT(vp, v_nmntvnodes)) { 6064 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 6065 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 6066 continue; 6067 VI_LOCK(vp); 6068 if (VN_IS_DOOMED(vp)) { 6069 VI_UNLOCK(vp); 6070 continue; 6071 } 6072 break; 6073 } 6074 if (vp == NULL) { 6075 __mnt_vnode_markerfree_all(mvp, mp); 6076 /* MNT_IUNLOCK(mp); -- done in above function */ 6077 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 6078 return (NULL); 6079 } 6080 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 6081 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 6082 MNT_IUNLOCK(mp); 6083 return (vp); 6084 } 6085 6086 struct vnode * 6087 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 6088 { 6089 struct vnode *vp; 6090 6091 *mvp = vn_alloc_marker(mp); 6092 MNT_ILOCK(mp); 6093 MNT_REF(mp); 6094 6095 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 6096 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 6097 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 6098 continue; 6099 VI_LOCK(vp); 6100 if (VN_IS_DOOMED(vp)) { 6101 VI_UNLOCK(vp); 6102 continue; 6103 } 6104 break; 6105 } 6106 if (vp == NULL) { 6107 MNT_REL(mp); 6108 MNT_IUNLOCK(mp); 6109 vn_free_marker(*mvp); 6110 *mvp = NULL; 6111 return (NULL); 6112 } 6113 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 6114 MNT_IUNLOCK(mp); 6115 return (vp); 6116 } 6117 6118 void 6119 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 6120 { 6121 6122 if (*mvp == NULL) { 6123 MNT_IUNLOCK(mp); 6124 return; 6125 } 6126 6127 mtx_assert(MNT_MTX(mp), MA_OWNED); 6128 6129 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6130 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 6131 MNT_REL(mp); 6132 MNT_IUNLOCK(mp); 6133 vn_free_marker(*mvp); 6134 *mvp = NULL; 6135 } 6136 6137 /* 6138 * These are helper functions for filesystems to traverse their 6139 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h 6140 */ 6141 static void 6142 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 6143 { 6144 6145 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6146 6147 MNT_ILOCK(mp); 6148 MNT_REL(mp); 6149 MNT_IUNLOCK(mp); 6150 vn_free_marker(*mvp); 6151 *mvp = NULL; 6152 } 6153 6154 /* 6155 * Relock the mp mount vnode list lock with the vp vnode interlock in the 6156 * conventional lock order during mnt_vnode_next_lazy iteration. 6157 * 6158 * On entry, the mount vnode list lock is held and the vnode interlock is not. 6159 * The list lock is dropped and reacquired. On success, both locks are held. 6160 * On failure, the mount vnode list lock is held but the vnode interlock is 6161 * not, and the procedure may have yielded. 6162 */ 6163 static bool 6164 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp, 6165 struct vnode *vp) 6166 { 6167 const struct vnode *tmp; 6168 bool held, ret; 6169 6170 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER && 6171 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp, 6172 ("%s: bad marker", __func__)); 6173 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp, 6174 ("%s: inappropriate vnode", __func__)); 6175 ASSERT_VI_UNLOCKED(vp, __func__); 6176 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 6177 6178 ret = false; 6179 6180 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist); 6181 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist); 6182 6183 /* 6184 * Use a hold to prevent vp from disappearing while the mount vnode 6185 * list lock is dropped and reacquired. Normally a hold would be 6186 * acquired with vhold(), but that might try to acquire the vnode 6187 * interlock, which would be a LOR with the mount vnode list lock. 6188 */ 6189 held = refcount_acquire_if_not_zero(&vp->v_holdcnt); 6190 mtx_unlock(&mp->mnt_listmtx); 6191 if (!held) 6192 goto abort; 6193 VI_LOCK(vp); 6194 if (!refcount_release_if_not_last(&vp->v_holdcnt)) { 6195 vdropl(vp); 6196 goto abort; 6197 } 6198 mtx_lock(&mp->mnt_listmtx); 6199 6200 /* 6201 * Determine whether the vnode is still the next one after the marker, 6202 * excepting any other markers. If the vnode has not been doomed by 6203 * vgone() then the hold should have ensured that it remained on the 6204 * lazy list. If it has been doomed but is still on the lazy list, 6205 * don't abort, but rather skip over it (avoid spinning on doomed 6206 * vnodes). 6207 */ 6208 tmp = mvp; 6209 do { 6210 tmp = TAILQ_NEXT(tmp, v_lazylist); 6211 } while (tmp != NULL && tmp->v_type == VMARKER); 6212 if (tmp != vp) { 6213 mtx_unlock(&mp->mnt_listmtx); 6214 VI_UNLOCK(vp); 6215 goto abort; 6216 } 6217 6218 ret = true; 6219 goto out; 6220 abort: 6221 maybe_yield(); 6222 mtx_lock(&mp->mnt_listmtx); 6223 out: 6224 if (ret) 6225 ASSERT_VI_LOCKED(vp, __func__); 6226 else 6227 ASSERT_VI_UNLOCKED(vp, __func__); 6228 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 6229 return (ret); 6230 } 6231 6232 static struct vnode * 6233 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 6234 void *cbarg) 6235 { 6236 struct vnode *vp, *nvp; 6237 6238 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 6239 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6240 restart: 6241 vp = TAILQ_NEXT(*mvp, v_lazylist); 6242 while (vp != NULL) { 6243 if (vp->v_type == VMARKER) { 6244 vp = TAILQ_NEXT(vp, v_lazylist); 6245 continue; 6246 } 6247 /* 6248 * See if we want to process the vnode. Note we may encounter a 6249 * long string of vnodes we don't care about and hog the list 6250 * as a result. Check for it and requeue the marker. 6251 */ 6252 if (VN_IS_DOOMED(vp) || !cb(vp, cbarg)) { 6253 if (!should_yield()) { 6254 vp = TAILQ_NEXT(vp, v_lazylist); 6255 continue; 6256 } 6257 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, 6258 v_lazylist); 6259 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, 6260 v_lazylist); 6261 mtx_unlock(&mp->mnt_listmtx); 6262 kern_yield(PRI_USER); 6263 mtx_lock(&mp->mnt_listmtx); 6264 goto restart; 6265 } 6266 /* 6267 * Try-lock because this is the wrong lock order. If that does 6268 * not succeed, drop the mount vnode list lock and try to 6269 * reacquire it and the vnode interlock in the right order. 6270 */ 6271 if (!VI_TRYLOCK(vp) && 6272 !mnt_vnode_next_lazy_relock(*mvp, mp, vp)) 6273 goto restart; 6274 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 6275 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 6276 ("alien vnode on the lazy list %p %p", vp, mp)); 6277 if (vp->v_mount == mp && !VN_IS_DOOMED(vp)) 6278 break; 6279 nvp = TAILQ_NEXT(vp, v_lazylist); 6280 VI_UNLOCK(vp); 6281 vp = nvp; 6282 } 6283 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 6284 6285 /* Check if we are done */ 6286 if (vp == NULL) { 6287 mtx_unlock(&mp->mnt_listmtx); 6288 mnt_vnode_markerfree_lazy(mvp, mp); 6289 return (NULL); 6290 } 6291 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist); 6292 mtx_unlock(&mp->mnt_listmtx); 6293 ASSERT_VI_LOCKED(vp, "lazy iter"); 6294 return (vp); 6295 } 6296 6297 struct vnode * 6298 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 6299 void *cbarg) 6300 { 6301 6302 if (should_yield()) 6303 kern_yield(PRI_USER); 6304 mtx_lock(&mp->mnt_listmtx); 6305 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 6306 } 6307 6308 struct vnode * 6309 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 6310 void *cbarg) 6311 { 6312 struct vnode *vp; 6313 6314 *mvp = vn_alloc_marker(mp); 6315 MNT_ILOCK(mp); 6316 MNT_REF(mp); 6317 MNT_IUNLOCK(mp); 6318 6319 mtx_lock(&mp->mnt_listmtx); 6320 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist); 6321 if (vp == NULL) { 6322 mtx_unlock(&mp->mnt_listmtx); 6323 mnt_vnode_markerfree_lazy(mvp, mp); 6324 return (NULL); 6325 } 6326 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist); 6327 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 6328 } 6329 6330 void 6331 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 6332 { 6333 6334 if (*mvp == NULL) 6335 return; 6336 6337 mtx_lock(&mp->mnt_listmtx); 6338 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 6339 mtx_unlock(&mp->mnt_listmtx); 6340 mnt_vnode_markerfree_lazy(mvp, mp); 6341 } 6342