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 37 /* 38 * External virtual filesystem routines 39 */ 40 41 #include "opt_ddb.h" 42 #include "opt_watchdog.h" 43 44 #include <sys/param.h> 45 #include <sys/systm.h> 46 #include <sys/asan.h> 47 #include <sys/bio.h> 48 #include <sys/buf.h> 49 #include <sys/capsicum.h> 50 #include <sys/condvar.h> 51 #include <sys/conf.h> 52 #include <sys/counter.h> 53 #include <sys/dirent.h> 54 #include <sys/event.h> 55 #include <sys/eventhandler.h> 56 #include <sys/extattr.h> 57 #include <sys/file.h> 58 #include <sys/fcntl.h> 59 #include <sys/inotify.h> 60 #include <sys/jail.h> 61 #include <sys/kdb.h> 62 #include <sys/kernel.h> 63 #include <sys/kthread.h> 64 #include <sys/ktr.h> 65 #include <sys/limits.h> 66 #include <sys/lockf.h> 67 #include <sys/malloc.h> 68 #include <sys/mount.h> 69 #include <sys/namei.h> 70 #include <sys/pctrie.h> 71 #include <sys/priv.h> 72 #include <sys/reboot.h> 73 #include <sys/refcount.h> 74 #include <sys/rwlock.h> 75 #include <sys/sched.h> 76 #include <sys/sleepqueue.h> 77 #include <sys/smr.h> 78 #include <sys/smp.h> 79 #include <sys/stat.h> 80 #include <sys/stdarg.h> 81 #include <sys/sysctl.h> 82 #include <sys/syslog.h> 83 #include <sys/user.h> 84 #include <sys/vmmeter.h> 85 #include <sys/vnode.h> 86 #include <sys/watchdog.h> 87 88 #include <security/mac/mac_framework.h> 89 90 #include <vm/vm.h> 91 #include <vm/vm_object.h> 92 #include <vm/vm_extern.h> 93 #include <vm/pmap.h> 94 #include <vm/vm_map.h> 95 #include <vm/vm_page.h> 96 #include <vm/vm_kern.h> 97 #include <vm/vnode_pager.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, bool isvnlru); 109 static void v_init_counters(struct vnode *); 110 static void vn_seqc_init(struct vnode *); 111 static void vn_seqc_write_end_free(struct vnode *vp); 112 static void vgonel(struct vnode *); 113 static bool vhold_recycle_free(struct vnode *); 114 static void vdropl_recycle(struct vnode *vp); 115 static void vdrop_recycle(struct vnode *vp); 116 static void vfs_knllock(void *arg); 117 static void vfs_knlunlock(void *arg); 118 static void vfs_knl_assert_lock(void *arg, int what); 119 static void destroy_vpollinfo(struct vpollinfo *vi); 120 static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, 121 daddr_t startlbn, daddr_t endlbn); 122 static void vnlru_recalc(void); 123 124 static SYSCTL_NODE(_vfs, OID_AUTO, vnode, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 125 "vnode configuration and statistics"); 126 static SYSCTL_NODE(_vfs_vnode, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 127 "vnode configuration"); 128 static SYSCTL_NODE(_vfs_vnode, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 129 "vnode statistics"); 130 static SYSCTL_NODE(_vfs_vnode, OID_AUTO, vnlru, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 131 "vnode recycling"); 132 133 /* 134 * Number of vnodes in existence. Increased whenever getnewvnode() 135 * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode. 136 */ 137 static u_long __exclusive_cache_line numvnodes; 138 139 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, 140 "Number of vnodes in existence (legacy)"); 141 SYSCTL_ULONG(_vfs_vnode_stats, OID_AUTO, count, CTLFLAG_RD, &numvnodes, 0, 142 "Number of vnodes in existence"); 143 144 static counter_u64_t vnodes_created; 145 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created, 146 "Number of vnodes created by getnewvnode (legacy)"); 147 SYSCTL_COUNTER_U64(_vfs_vnode_stats, OID_AUTO, created, CTLFLAG_RD, &vnodes_created, 148 "Number of vnodes created by getnewvnode"); 149 150 /* 151 * Conversion tables for conversion from vnode types to inode formats 152 * and back. 153 */ 154 __enum_uint8(vtype) iftovt_tab[16] = { 155 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 156 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON 157 }; 158 int vttoif_tab[10] = { 159 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 160 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT 161 }; 162 163 /* 164 * List of allocates vnodes in the system. 165 */ 166 static TAILQ_HEAD(freelst, vnode) vnode_list; 167 static struct vnode *vnode_list_free_marker; 168 static struct vnode *vnode_list_reclaim_marker; 169 170 /* 171 * "Free" vnode target. Free vnodes are rarely completely free, but are 172 * just ones that are cheap to recycle. Usually they are for files which 173 * have been stat'd but not read; these usually have inode and namecache 174 * data attached to them. This target is the preferred minimum size of a 175 * sub-cache consisting mostly of such files. The system balances the size 176 * of this sub-cache with its complement to try to prevent either from 177 * thrashing while the other is relatively inactive. The targets express 178 * a preference for the best balance. 179 * 180 * "Above" this target there are 2 further targets (watermarks) related 181 * to recyling of free vnodes. In the best-operating case, the cache is 182 * exactly full, the free list has size between vlowat and vhiwat above the 183 * free target, and recycling from it and normal use maintains this state. 184 * Sometimes the free list is below vlowat or even empty, but this state 185 * is even better for immediate use provided the cache is not full. 186 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free 187 * ones) to reach one of these states. The watermarks are currently hard- 188 * coded as 4% and 9% of the available space higher. These and the default 189 * of 25% for wantfreevnodes are too large if the memory size is large. 190 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim 191 * whenever vnlru_proc() becomes active. 192 */ 193 static long wantfreevnodes; 194 static long __exclusive_cache_line freevnodes; 195 static long freevnodes_old; 196 197 static u_long recycles_count; 198 SYSCTL_ULONG(_vfs, OID_AUTO, recycles, CTLFLAG_RD | CTLFLAG_STATS, &recycles_count, 0, 199 "Number of vnodes recycled to meet vnode cache targets (legacy)"); 200 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, recycles, CTLFLAG_RD | CTLFLAG_STATS, 201 &recycles_count, 0, 202 "Number of vnodes recycled to meet vnode cache targets"); 203 204 static u_long recycles_free_count; 205 SYSCTL_ULONG(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD | CTLFLAG_STATS, 206 &recycles_free_count, 0, 207 "Number of free vnodes recycled to meet vnode cache targets (legacy)"); 208 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, recycles_free, CTLFLAG_RD | CTLFLAG_STATS, 209 &recycles_free_count, 0, 210 "Number of free vnodes recycled to meet vnode cache targets"); 211 212 static counter_u64_t direct_recycles_free_count; 213 SYSCTL_COUNTER_U64(_vfs_vnode_vnlru, OID_AUTO, direct_recycles_free, CTLFLAG_RD, 214 &direct_recycles_free_count, 215 "Number of free vnodes recycled by vn_alloc callers to meet vnode cache targets"); 216 217 static counter_u64_t vnode_skipped_requeues; 218 SYSCTL_COUNTER_U64(_vfs_vnode_stats, OID_AUTO, skipped_requeues, CTLFLAG_RD, &vnode_skipped_requeues, 219 "Number of times LRU requeue was skipped due to lock contention"); 220 221 static __read_mostly bool vnode_can_skip_requeue; 222 SYSCTL_BOOL(_vfs_vnode_param, OID_AUTO, can_skip_requeue, CTLFLAG_RW, 223 &vnode_can_skip_requeue, 0, "Is LRU requeue skippable"); 224 225 static u_long deferred_inact; 226 SYSCTL_ULONG(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, 227 &deferred_inact, 0, "Number of times inactive processing was deferred"); 228 229 /* To keep more than one thread at a time from running vfs_getnewfsid */ 230 static struct mtx mntid_mtx; 231 232 /* 233 * Lock for any access to the following: 234 * vnode_list 235 * numvnodes 236 * freevnodes 237 */ 238 static struct mtx __exclusive_cache_line vnode_list_mtx; 239 240 /* Publicly exported FS */ 241 struct nfs_public nfs_pub; 242 243 static uma_zone_t buf_trie_zone; 244 static smr_t buf_trie_smr; 245 246 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 247 static uma_zone_t vnode_zone; 248 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll"); 249 250 __read_frequently smr_t vfs_smr; 251 252 /* 253 * The workitem queue. 254 * 255 * It is useful to delay writes of file data and filesystem metadata 256 * for tens of seconds so that quickly created and deleted files need 257 * not waste disk bandwidth being created and removed. To realize this, 258 * we append vnodes to a "workitem" queue. When running with a soft 259 * updates implementation, most pending metadata dependencies should 260 * not wait for more than a few seconds. Thus, mounted on block devices 261 * are delayed only about a half the time that file data is delayed. 262 * Similarly, directory updates are more critical, so are only delayed 263 * about a third the time that file data is delayed. Thus, there are 264 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 265 * one each second (driven off the filesystem syncer process). The 266 * syncer_delayno variable indicates the next queue that is to be processed. 267 * Items that need to be processed soon are placed in this queue: 268 * 269 * syncer_workitem_pending[syncer_delayno] 270 * 271 * A delay of fifteen seconds is done by placing the request fifteen 272 * entries later in the queue: 273 * 274 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 275 * 276 */ 277 static int syncer_delayno; 278 static long syncer_mask; 279 LIST_HEAD(synclist, bufobj); 280 static struct synclist *syncer_workitem_pending; 281 /* 282 * The sync_mtx protects: 283 * bo->bo_synclist 284 * sync_vnode_count 285 * syncer_delayno 286 * syncer_state 287 * syncer_workitem_pending 288 * syncer_worklist_len 289 * rushjob 290 */ 291 static struct mtx sync_mtx; 292 static struct cv sync_wakeup; 293 294 #define SYNCER_MAXDELAY 32 295 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 296 static int syncdelay = 30; /* max time to delay syncing data */ 297 static int filedelay = 30; /* time to delay syncing files */ 298 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, 299 "Time to delay syncing files (in seconds)"); 300 static int dirdelay = 29; /* time to delay syncing directories */ 301 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, 302 "Time to delay syncing directories (in seconds)"); 303 static int metadelay = 28; /* time to delay syncing metadata */ 304 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, 305 "Time to delay syncing metadata (in seconds)"); 306 static int rushjob; /* number of slots to run ASAP */ 307 static int stat_rush_requests; /* number of times I/O speeded up */ 308 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, 309 "Number of times I/O speeded up (rush requests)"); 310 311 #define VDBATCH_SIZE 8 312 struct vdbatch { 313 u_int index; 314 struct mtx lock; 315 struct vnode *tab[VDBATCH_SIZE]; 316 }; 317 DPCPU_DEFINE_STATIC(struct vdbatch, vd); 318 319 static void vdbatch_dequeue(struct vnode *vp); 320 321 /* 322 * The syncer will require at least SYNCER_MAXDELAY iterations to shutdown; 323 * we probably don't want to pause for the whole second each time. 324 */ 325 #define SYNCER_SHUTDOWN_SPEEDUP 32 326 static int sync_vnode_count; 327 static int syncer_worklist_len; 328 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 329 syncer_state; 330 331 /* Target for maximum number of vnodes. */ 332 u_long desiredvnodes; 333 static u_long gapvnodes; /* gap between wanted and desired */ 334 static u_long vhiwat; /* enough extras after expansion */ 335 static u_long vlowat; /* minimal extras before expansion */ 336 static bool vstir; /* nonzero to stir non-free vnodes */ 337 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */ 338 339 static u_long vnlru_read_freevnodes(void); 340 341 /* 342 * Note that no attempt is made to sanitize these parameters. 343 */ 344 static int 345 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS) 346 { 347 u_long val; 348 int error; 349 350 val = desiredvnodes; 351 error = sysctl_handle_long(oidp, &val, 0, req); 352 if (error != 0 || req->newptr == NULL) 353 return (error); 354 355 if (val == desiredvnodes) 356 return (0); 357 mtx_lock(&vnode_list_mtx); 358 desiredvnodes = val; 359 wantfreevnodes = desiredvnodes / 4; 360 vnlru_recalc(); 361 mtx_unlock(&vnode_list_mtx); 362 /* 363 * XXX There is no protection against multiple threads changing 364 * desiredvnodes at the same time. Locking above only helps vnlru and 365 * getnewvnode. 366 */ 367 vfs_hash_changesize(desiredvnodes); 368 cache_changesize(desiredvnodes); 369 return (0); 370 } 371 372 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes, 373 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes, 374 "LU", "Target for maximum number of vnodes (legacy)"); 375 SYSCTL_PROC(_vfs_vnode_param, OID_AUTO, limit, 376 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes, 377 "LU", "Target for maximum number of vnodes"); 378 379 static int 380 sysctl_freevnodes(SYSCTL_HANDLER_ARGS) 381 { 382 u_long rfreevnodes; 383 384 rfreevnodes = vnlru_read_freevnodes(); 385 return (sysctl_handle_long(oidp, &rfreevnodes, 0, req)); 386 } 387 388 SYSCTL_PROC(_vfs, OID_AUTO, freevnodes, 389 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_freevnodes, 390 "LU", "Number of \"free\" vnodes (legacy)"); 391 SYSCTL_PROC(_vfs_vnode_stats, OID_AUTO, free, 392 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_freevnodes, 393 "LU", "Number of \"free\" vnodes"); 394 395 static int 396 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS) 397 { 398 u_long val; 399 int error; 400 401 val = wantfreevnodes; 402 error = sysctl_handle_long(oidp, &val, 0, req); 403 if (error != 0 || req->newptr == NULL) 404 return (error); 405 406 if (val == wantfreevnodes) 407 return (0); 408 mtx_lock(&vnode_list_mtx); 409 wantfreevnodes = val; 410 vnlru_recalc(); 411 mtx_unlock(&vnode_list_mtx); 412 return (0); 413 } 414 415 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes, 416 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes, 417 "LU", "Target for minimum number of \"free\" vnodes (legacy)"); 418 SYSCTL_PROC(_vfs_vnode_param, OID_AUTO, wantfree, 419 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes, 420 "LU", "Target for minimum number of \"free\" vnodes"); 421 422 static int vnlru_nowhere; 423 SYSCTL_INT(_vfs_vnode_vnlru, OID_AUTO, failed_runs, CTLFLAG_RD | CTLFLAG_STATS, 424 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 425 426 static int 427 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS) 428 { 429 struct vnode *vp; 430 struct nameidata nd; 431 char *buf; 432 unsigned long ndflags; 433 int error; 434 435 if (req->newptr == NULL) 436 return (EINVAL); 437 if (req->newlen >= PATH_MAX) 438 return (E2BIG); 439 440 buf = malloc(PATH_MAX, M_TEMP, M_WAITOK); 441 error = SYSCTL_IN(req, buf, req->newlen); 442 if (error != 0) 443 goto out; 444 445 buf[req->newlen] = '\0'; 446 447 ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1; 448 NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf); 449 if ((error = namei(&nd)) != 0) 450 goto out; 451 vp = nd.ni_vp; 452 453 if (VN_IS_DOOMED(vp)) { 454 /* 455 * This vnode is being recycled. Return != 0 to let the caller 456 * know that the sysctl had no effect. Return EAGAIN because a 457 * subsequent call will likely succeed (since namei will create 458 * a new vnode if necessary) 459 */ 460 error = EAGAIN; 461 goto putvnode; 462 } 463 464 vgone(vp); 465 putvnode: 466 vput(vp); 467 NDFREE_PNBUF(&nd); 468 out: 469 free(buf, M_TEMP); 470 return (error); 471 } 472 473 static int 474 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS) 475 { 476 struct thread *td = curthread; 477 struct vnode *vp; 478 struct file *fp; 479 int error; 480 int fd; 481 482 if (req->newptr == NULL) 483 return (EBADF); 484 485 error = sysctl_handle_int(oidp, &fd, 0, req); 486 if (error != 0) 487 return (error); 488 error = getvnode(curthread, fd, &cap_fcntl_rights, &fp); 489 if (error != 0) 490 return (error); 491 vp = fp->f_vnode; 492 493 error = vn_lock(vp, LK_EXCLUSIVE); 494 if (error != 0) 495 goto drop; 496 497 vgone(vp); 498 VOP_UNLOCK(vp); 499 drop: 500 fdrop(fp, td); 501 return (error); 502 } 503 504 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode, 505 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, 506 sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname"); 507 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode, 508 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, 509 sysctl_ftry_reclaim_vnode, "I", 510 "Try to reclaim a vnode by its file descriptor"); 511 512 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */ 513 #define vnsz2log 8 514 #ifndef DEBUG_LOCKS 515 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log && 516 sizeof(struct vnode) < 1UL << (vnsz2log + 1), 517 "vnsz2log needs to be updated"); 518 #endif 519 520 /* 521 * Support for the bufobj clean & dirty pctrie. 522 */ 523 static void * 524 buf_trie_alloc(struct pctrie *ptree) 525 { 526 return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT)); 527 } 528 529 static void 530 buf_trie_free(struct pctrie *ptree, void *node) 531 { 532 uma_zfree_smr(buf_trie_zone, node); 533 } 534 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free, 535 buf_trie_smr); 536 537 /* 538 * Lookup the next element greater than or equal to lblkno, accounting for the 539 * fact that, for pctries, negative values are greater than nonnegative ones. 540 */ 541 static struct buf * 542 buf_lookup_ge(struct bufv *bv, daddr_t lblkno) 543 { 544 struct buf *bp; 545 546 bp = BUF_PCTRIE_LOOKUP_GE(&bv->bv_root, lblkno); 547 if (bp == NULL && lblkno < 0) 548 bp = BUF_PCTRIE_LOOKUP_GE(&bv->bv_root, 0); 549 if (bp != NULL && bp->b_lblkno < lblkno) 550 bp = NULL; 551 return (bp); 552 } 553 554 /* 555 * Insert bp, and find the next element smaller than bp, accounting for the fact 556 * that, for pctries, negative values are greater than nonnegative ones. 557 */ 558 static int 559 buf_insert_lookup_le(struct bufv *bv, struct buf *bp, struct buf **n) 560 { 561 int error; 562 563 error = BUF_PCTRIE_INSERT_LOOKUP_LE(&bv->bv_root, bp, n); 564 if (error != EEXIST) { 565 if (*n == NULL && bp->b_lblkno >= 0) 566 *n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, ~0L); 567 if (*n != NULL && (*n)->b_lblkno >= bp->b_lblkno) 568 *n = NULL; 569 } 570 return (error); 571 } 572 573 /* 574 * Initialize the vnode management data structures. 575 * 576 * Reevaluate the following cap on the number of vnodes after the physical 577 * memory size exceeds 512GB. In the limit, as the physical memory size 578 * grows, the ratio of the memory size in KB to vnodes approaches 64:1. 579 */ 580 #ifndef MAXVNODES_MAX 581 #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */ 582 #endif 583 584 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker"); 585 586 static struct vnode * 587 vn_alloc_marker(struct mount *mp) 588 { 589 struct vnode *vp; 590 591 vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 592 vp->v_type = VMARKER; 593 vp->v_mount = mp; 594 595 return (vp); 596 } 597 598 static void 599 vn_free_marker(struct vnode *vp) 600 { 601 602 MPASS(vp->v_type == VMARKER); 603 free(vp, M_VNODE_MARKER); 604 } 605 606 #ifdef KASAN 607 static int 608 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused) 609 { 610 kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0); 611 return (0); 612 } 613 614 static void 615 vnode_dtor(void *mem, int size, void *arg __unused) 616 { 617 size_t end1, end2, off1, off2; 618 619 _Static_assert(offsetof(struct vnode, v_vnodelist) < 620 offsetof(struct vnode, v_dbatchcpu), 621 "KASAN marks require updating"); 622 623 off1 = offsetof(struct vnode, v_vnodelist); 624 off2 = offsetof(struct vnode, v_dbatchcpu); 625 end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist); 626 end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu); 627 628 /* 629 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even 630 * after the vnode has been freed. Try to get some KASAN coverage by 631 * marking everything except those two fields as invalid. Because 632 * KASAN's tracking is not byte-granular, any preceding fields sharing 633 * the same 8-byte aligned word must also be marked valid. 634 */ 635 636 /* Handle the area from the start until v_vnodelist... */ 637 off1 = rounddown2(off1, KASAN_SHADOW_SCALE); 638 kasan_mark(mem, off1, off1, KASAN_UMA_FREED); 639 640 /* ... then the area between v_vnodelist and v_dbatchcpu ... */ 641 off1 = roundup2(end1, KASAN_SHADOW_SCALE); 642 off2 = rounddown2(off2, KASAN_SHADOW_SCALE); 643 if (off2 > off1) 644 kasan_mark((void *)((char *)mem + off1), off2 - off1, 645 off2 - off1, KASAN_UMA_FREED); 646 647 /* ... and finally the area from v_dbatchcpu to the end. */ 648 off2 = roundup2(end2, KASAN_SHADOW_SCALE); 649 kasan_mark((void *)((char *)mem + off2), size - off2, size - off2, 650 KASAN_UMA_FREED); 651 } 652 #endif /* KASAN */ 653 654 /* 655 * Initialize a vnode as it first enters the zone. 656 */ 657 static int 658 vnode_init(void *mem, int size, int flags) 659 { 660 struct vnode *vp; 661 662 vp = mem; 663 bzero(vp, size); 664 /* 665 * Setup locks. 666 */ 667 vp->v_vnlock = &vp->v_lock; 668 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 669 /* 670 * By default, don't allow shared locks unless filesystems opt-in. 671 */ 672 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT, 673 LK_NOSHARE | LK_IS_VNODE); 674 /* 675 * Initialize bufobj. 676 */ 677 bufobj_init(&vp->v_bufobj, vp); 678 /* 679 * Initialize namecache. 680 */ 681 cache_vnode_init(vp); 682 /* 683 * Initialize rangelocks. 684 */ 685 rangelock_init(&vp->v_rl); 686 687 vp->v_dbatchcpu = NOCPU; 688 689 vp->v_state = VSTATE_DEAD; 690 691 /* 692 * Check vhold_recycle_free for an explanation. 693 */ 694 vp->v_holdcnt = VHOLD_NO_SMR; 695 vp->v_type = VNON; 696 mtx_lock(&vnode_list_mtx); 697 TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist); 698 mtx_unlock(&vnode_list_mtx); 699 return (0); 700 } 701 702 /* 703 * Free a vnode when it is cleared from the zone. 704 */ 705 static void 706 vnode_fini(void *mem, int size) 707 { 708 struct vnode *vp; 709 struct bufobj *bo; 710 711 vp = mem; 712 vdbatch_dequeue(vp); 713 mtx_lock(&vnode_list_mtx); 714 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 715 mtx_unlock(&vnode_list_mtx); 716 rangelock_destroy(&vp->v_rl); 717 lockdestroy(vp->v_vnlock); 718 mtx_destroy(&vp->v_interlock); 719 bo = &vp->v_bufobj; 720 rw_destroy(BO_LOCKPTR(bo)); 721 722 kasan_mark(mem, size, size, 0); 723 } 724 725 /* 726 * Provide the size of NFS nclnode and NFS fh for calculation of the 727 * vnode memory consumption. The size is specified directly to 728 * eliminate dependency on NFS-private header. 729 * 730 * Other filesystems may use bigger or smaller (like UFS and ZFS) 731 * private inode data, but the NFS-based estimation is ample enough. 732 * Still, we care about differences in the size between 64- and 32-bit 733 * platforms. 734 * 735 * Namecache structure size is heuristically 736 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1. 737 */ 738 #ifdef _LP64 739 #define NFS_NCLNODE_SZ (528 + 64) 740 #define NC_SZ 148 741 #else 742 #define NFS_NCLNODE_SZ (360 + 32) 743 #define NC_SZ 92 744 #endif 745 746 static void 747 vntblinit(void *dummy __unused) 748 { 749 struct vdbatch *vd; 750 uma_ctor ctor; 751 uma_dtor dtor; 752 int cpu, physvnodes, virtvnodes; 753 754 /* 755 * 'desiredvnodes' is the minimum of a function of the physical memory 756 * size and another of the kernel heap size (UMA limit, a portion of the 757 * KVA). 758 * 759 * Currently, on 64-bit platforms, 'desiredvnodes' is set to 760 * 'virtvnodes' up to a physical memory cutoff of ~1722MB, after which 761 * 'physvnodes' applies instead. With the current automatic tuning for 762 * 'maxfiles' (32 files/MB), 'desiredvnodes' is always greater than it. 763 */ 764 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 32 + 765 min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 32; 766 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) + 767 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ)); 768 desiredvnodes = min(physvnodes, virtvnodes); 769 if (desiredvnodes > MAXVNODES_MAX) { 770 if (bootverbose) 771 printf("Reducing kern.maxvnodes %lu -> %lu\n", 772 desiredvnodes, MAXVNODES_MAX); 773 desiredvnodes = MAXVNODES_MAX; 774 } 775 wantfreevnodes = desiredvnodes / 4; 776 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 777 TAILQ_INIT(&vnode_list); 778 mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF); 779 /* 780 * The lock is taken to appease WITNESS. 781 */ 782 mtx_lock(&vnode_list_mtx); 783 vnlru_recalc(); 784 mtx_unlock(&vnode_list_mtx); 785 vnode_list_free_marker = vn_alloc_marker(NULL); 786 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist); 787 vnode_list_reclaim_marker = vn_alloc_marker(NULL); 788 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist); 789 790 #ifdef KASAN 791 ctor = vnode_ctor; 792 dtor = vnode_dtor; 793 #else 794 ctor = NULL; 795 dtor = NULL; 796 #endif 797 vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor, 798 vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN); 799 uma_zone_set_smr(vnode_zone, vfs_smr); 800 801 /* 802 * Preallocate enough nodes to support one-per buf so that 803 * we can not fail an insert. reassignbuf() callers can not 804 * tolerate the insertion failure. 805 */ 806 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(), 807 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, 808 UMA_ZONE_NOFREE | UMA_ZONE_SMR); 809 buf_trie_smr = uma_zone_get_smr(buf_trie_zone); 810 uma_prealloc(buf_trie_zone, nbuf); 811 812 vnodes_created = counter_u64_alloc(M_WAITOK); 813 direct_recycles_free_count = counter_u64_alloc(M_WAITOK); 814 vnode_skipped_requeues = counter_u64_alloc(M_WAITOK); 815 816 /* 817 * Initialize the filesystem syncer. 818 */ 819 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 820 &syncer_mask); 821 syncer_maxdelay = syncer_mask + 1; 822 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 823 cv_init(&sync_wakeup, "syncer"); 824 825 CPU_FOREACH(cpu) { 826 vd = DPCPU_ID_PTR((cpu), vd); 827 bzero(vd, sizeof(*vd)); 828 mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF); 829 } 830 } 831 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL); 832 833 /* 834 * Mark a mount point as busy. Used to synchronize access and to delay 835 * unmounting. Eventually, mountlist_mtx is not released on failure. 836 * 837 * vfs_busy() is a custom lock, it can block the caller. 838 * vfs_busy() only sleeps if the unmount is active on the mount point. 839 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any 840 * vnode belonging to mp. 841 * 842 * Lookup uses vfs_busy() to traverse mount points. 843 * root fs var fs 844 * / vnode lock A / vnode lock (/var) D 845 * /var vnode lock B /log vnode lock(/var/log) E 846 * vfs_busy lock C vfs_busy lock F 847 * 848 * Within each file system, the lock order is C->A->B and F->D->E. 849 * 850 * When traversing across mounts, the system follows that lock order: 851 * 852 * C->A->B 853 * | 854 * +->F->D->E 855 * 856 * The lookup() process for namei("/var") illustrates the process: 857 * 1. VOP_LOOKUP() obtains B while A is held 858 * 2. vfs_busy() obtains a shared lock on F while A and B are held 859 * 3. vput() releases lock on B 860 * 4. vput() releases lock on A 861 * 5. VFS_ROOT() obtains lock on D while shared lock on F is held 862 * 6. vfs_unbusy() releases shared lock on F 863 * 7. vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A. 864 * Attempt to lock A (instead of vp_crossmp) while D is held would 865 * violate the global order, causing deadlocks. 866 * 867 * dounmount() locks B while F is drained. Note that for stacked 868 * filesystems, D and B in the example above may be the same lock, 869 * which introdues potential lock order reversal deadlock between 870 * dounmount() and step 5 above. These filesystems may avoid the LOR 871 * by setting VV_CROSSLOCK on the covered vnode so that lock B will 872 * remain held until after step 5. 873 */ 874 int 875 vfs_busy(struct mount *mp, int flags) 876 { 877 struct mount_pcpu *mpcpu; 878 879 MPASS((flags & ~MBF_MASK) == 0); 880 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags); 881 882 if (vfs_op_thread_enter(mp, mpcpu)) { 883 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 884 MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0); 885 MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0); 886 vfs_mp_count_add_pcpu(mpcpu, ref, 1); 887 vfs_mp_count_add_pcpu(mpcpu, lockref, 1); 888 vfs_op_thread_exit(mp, mpcpu); 889 if (flags & MBF_MNTLSTLOCK) 890 mtx_unlock(&mountlist_mtx); 891 return (0); 892 } 893 894 MNT_ILOCK(mp); 895 vfs_assert_mount_counters(mp); 896 MNT_REF(mp); 897 /* 898 * If mount point is currently being unmounted, sleep until the 899 * mount point fate is decided. If thread doing the unmounting fails, 900 * it will clear MNTK_UNMOUNT flag before waking us up, indicating 901 * that this mount point has survived the unmount attempt and vfs_busy 902 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE 903 * flag in addition to MNTK_UNMOUNT, indicating that mount point is 904 * about to be really destroyed. vfs_busy needs to release its 905 * reference on the mount point in this case and return with ENOENT, 906 * telling the caller the mount it tried to busy is no longer valid. 907 */ 908 while (mp->mnt_kern_flag & MNTK_UNMOUNT) { 909 KASSERT(TAILQ_EMPTY(&mp->mnt_uppers), 910 ("%s: non-empty upper mount list with pending unmount", 911 __func__)); 912 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) { 913 MNT_REL(mp); 914 MNT_IUNLOCK(mp); 915 CTR1(KTR_VFS, "%s: failed busying before sleeping", 916 __func__); 917 return (ENOENT); 918 } 919 if (flags & MBF_MNTLSTLOCK) 920 mtx_unlock(&mountlist_mtx); 921 mp->mnt_kern_flag |= MNTK_MWAIT; 922 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0); 923 if (flags & MBF_MNTLSTLOCK) 924 mtx_lock(&mountlist_mtx); 925 MNT_ILOCK(mp); 926 } 927 if (flags & MBF_MNTLSTLOCK) 928 mtx_unlock(&mountlist_mtx); 929 mp->mnt_lockref++; 930 MNT_IUNLOCK(mp); 931 return (0); 932 } 933 934 /* 935 * Free a busy filesystem. 936 */ 937 void 938 vfs_unbusy(struct mount *mp) 939 { 940 struct mount_pcpu *mpcpu; 941 int c; 942 943 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 944 945 if (vfs_op_thread_enter(mp, mpcpu)) { 946 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 947 vfs_mp_count_sub_pcpu(mpcpu, lockref, 1); 948 vfs_mp_count_sub_pcpu(mpcpu, ref, 1); 949 vfs_op_thread_exit(mp, mpcpu); 950 return; 951 } 952 953 MNT_ILOCK(mp); 954 vfs_assert_mount_counters(mp); 955 MNT_REL(mp); 956 c = --mp->mnt_lockref; 957 if (mp->mnt_vfs_ops == 0) { 958 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 959 MNT_IUNLOCK(mp); 960 return; 961 } 962 if (c < 0) 963 vfs_dump_mount_counters(mp); 964 if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) { 965 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT); 966 CTR1(KTR_VFS, "%s: waking up waiters", __func__); 967 mp->mnt_kern_flag &= ~MNTK_DRAINING; 968 wakeup(&mp->mnt_lockref); 969 } 970 MNT_IUNLOCK(mp); 971 } 972 973 /* 974 * Lookup a mount point by filesystem identifier. 975 */ 976 struct mount * 977 vfs_getvfs(fsid_t *fsid) 978 { 979 struct mount *mp; 980 981 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 982 mtx_lock(&mountlist_mtx); 983 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 984 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) { 985 vfs_ref(mp); 986 mtx_unlock(&mountlist_mtx); 987 return (mp); 988 } 989 } 990 mtx_unlock(&mountlist_mtx); 991 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 992 return ((struct mount *) 0); 993 } 994 995 /* 996 * Lookup a mount point by filesystem identifier, busying it before 997 * returning. 998 * 999 * To avoid congestion on mountlist_mtx, implement simple direct-mapped 1000 * cache for popular filesystem identifiers. The cache is lockess, using 1001 * the fact that struct mount's are never freed. In worst case we may 1002 * get pointer to unmounted or even different filesystem, so we have to 1003 * check what we got, and go slow way if so. 1004 */ 1005 struct mount * 1006 vfs_busyfs(fsid_t *fsid) 1007 { 1008 #define FSID_CACHE_SIZE 256 1009 typedef struct mount * volatile vmp_t; 1010 static vmp_t cache[FSID_CACHE_SIZE]; 1011 struct mount *mp; 1012 int error; 1013 uint32_t hash; 1014 1015 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 1016 hash = fsid->val[0] ^ fsid->val[1]; 1017 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1); 1018 mp = cache[hash]; 1019 if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0) 1020 goto slow; 1021 if (vfs_busy(mp, 0) != 0) { 1022 cache[hash] = NULL; 1023 goto slow; 1024 } 1025 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) 1026 return (mp); 1027 else 1028 vfs_unbusy(mp); 1029 1030 slow: 1031 mtx_lock(&mountlist_mtx); 1032 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 1033 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) { 1034 error = vfs_busy(mp, MBF_MNTLSTLOCK); 1035 if (error) { 1036 cache[hash] = NULL; 1037 mtx_unlock(&mountlist_mtx); 1038 return (NULL); 1039 } 1040 cache[hash] = mp; 1041 return (mp); 1042 } 1043 } 1044 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 1045 mtx_unlock(&mountlist_mtx); 1046 return ((struct mount *) 0); 1047 } 1048 1049 /* 1050 * Check if a user can access privileged mount options. 1051 */ 1052 int 1053 vfs_suser(struct mount *mp, struct thread *td) 1054 { 1055 int error; 1056 1057 if (jailed(td->td_ucred)) { 1058 /* 1059 * If the jail of the calling thread lacks permission for 1060 * this type of file system, deny immediately. 1061 */ 1062 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag)) 1063 return (EPERM); 1064 1065 /* 1066 * If the file system was mounted outside the jail of the 1067 * calling thread, deny immediately. 1068 */ 1069 if (prison_check(td->td_ucred, mp->mnt_cred) != 0) 1070 return (EPERM); 1071 } 1072 1073 /* 1074 * If file system supports delegated administration, we don't check 1075 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified 1076 * by the file system itself. 1077 * If this is not the user that did original mount, we check for 1078 * the PRIV_VFS_MOUNT_OWNER privilege. 1079 */ 1080 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) && 1081 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 1082 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) 1083 return (error); 1084 } 1085 return (0); 1086 } 1087 1088 /* 1089 * Get a new unique fsid. Try to make its val[0] unique, since this value 1090 * will be used to create fake device numbers for stat(). Also try (but 1091 * not so hard) make its val[0] unique mod 2^16, since some emulators only 1092 * support 16-bit device numbers. We end up with unique val[0]'s for the 1093 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 1094 * 1095 * Keep in mind that several mounts may be running in parallel. Starting 1096 * the search one past where the previous search terminated is both a 1097 * micro-optimization and a defense against returning the same fsid to 1098 * different mounts. 1099 */ 1100 void 1101 vfs_getnewfsid(struct mount *mp) 1102 { 1103 static uint16_t mntid_base; 1104 struct mount *nmp; 1105 fsid_t tfsid; 1106 int mtype; 1107 1108 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 1109 mtx_lock(&mntid_mtx); 1110 mtype = mp->mnt_vfc->vfc_typenum; 1111 tfsid.val[1] = mtype; 1112 mtype = (mtype & 0xFF) << 24; 1113 for (;;) { 1114 tfsid.val[0] = makedev(255, 1115 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 1116 mntid_base++; 1117 if ((nmp = vfs_getvfs(&tfsid)) == NULL) 1118 break; 1119 vfs_rel(nmp); 1120 } 1121 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 1122 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 1123 mtx_unlock(&mntid_mtx); 1124 } 1125 1126 /* 1127 * Knob to control the precision of file timestamps: 1128 * 1129 * 0 = seconds only; nanoseconds zeroed. 1130 * 1 = seconds and nanoseconds, accurate within 1/HZ. 1131 * 2 = seconds and nanoseconds, truncated to microseconds. 1132 * >=3 = seconds and nanoseconds, maximum precision. 1133 */ 1134 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 1135 1136 static int timestamp_precision = TSP_USEC; 1137 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 1138 ×tamp_precision, 0, "File timestamp precision (0: seconds, " 1139 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, " 1140 "3+: sec + ns (max. precision))"); 1141 1142 /* 1143 * Get a current timestamp. 1144 */ 1145 void 1146 vfs_timestamp(struct timespec *tsp) 1147 { 1148 struct timeval tv; 1149 1150 switch (timestamp_precision) { 1151 case TSP_SEC: 1152 tsp->tv_sec = time_second; 1153 tsp->tv_nsec = 0; 1154 break; 1155 case TSP_HZ: 1156 getnanotime(tsp); 1157 break; 1158 case TSP_USEC: 1159 microtime(&tv); 1160 TIMEVAL_TO_TIMESPEC(&tv, tsp); 1161 break; 1162 case TSP_NSEC: 1163 default: 1164 nanotime(tsp); 1165 break; 1166 } 1167 } 1168 1169 /* 1170 * Set vnode attributes to VNOVAL 1171 */ 1172 void 1173 vattr_null(struct vattr *vap) 1174 { 1175 1176 vap->va_type = VNON; 1177 vap->va_size = VNOVAL; 1178 vap->va_bytes = VNOVAL; 1179 vap->va_mode = VNOVAL; 1180 vap->va_nlink = VNOVAL; 1181 vap->va_uid = VNOVAL; 1182 vap->va_gid = VNOVAL; 1183 vap->va_fsid = VNOVAL; 1184 vap->va_fileid = VNOVAL; 1185 vap->va_blocksize = VNOVAL; 1186 vap->va_rdev = VNOVAL; 1187 vap->va_atime.tv_sec = VNOVAL; 1188 vap->va_atime.tv_nsec = VNOVAL; 1189 vap->va_mtime.tv_sec = VNOVAL; 1190 vap->va_mtime.tv_nsec = VNOVAL; 1191 vap->va_ctime.tv_sec = VNOVAL; 1192 vap->va_ctime.tv_nsec = VNOVAL; 1193 vap->va_birthtime.tv_sec = VNOVAL; 1194 vap->va_birthtime.tv_nsec = VNOVAL; 1195 vap->va_flags = VNOVAL; 1196 vap->va_gen = VNOVAL; 1197 vap->va_vaflags = 0; 1198 vap->va_filerev = VNOVAL; 1199 vap->va_bsdflags = 0; 1200 } 1201 1202 /* 1203 * Try to reduce the total number of vnodes. 1204 * 1205 * This routine (and its user) are buggy in at least the following ways: 1206 * - all parameters were picked years ago when RAM sizes were significantly 1207 * smaller 1208 * - it can pick vnodes based on pages used by the vm object, but filesystems 1209 * like ZFS don't use it making the pick broken 1210 * - since ZFS has its own aging policy it gets partially combated by this one 1211 * - a dedicated method should be provided for filesystems to let them decide 1212 * whether the vnode should be recycled 1213 * 1214 * This routine is called when we have too many vnodes. It attempts 1215 * to free <count> vnodes and will potentially free vnodes that still 1216 * have VM backing store (VM backing store is typically the cause 1217 * of a vnode blowout so we want to do this). Therefore, this operation 1218 * is not considered cheap. 1219 * 1220 * A number of conditions may prevent a vnode from being reclaimed. 1221 * the buffer cache may have references on the vnode, a directory 1222 * vnode may still have references due to the namei cache representing 1223 * underlying files, or the vnode may be in active use. It is not 1224 * desirable to reuse such vnodes. These conditions may cause the 1225 * number of vnodes to reach some minimum value regardless of what 1226 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 1227 * 1228 * @param reclaim_nc_src Only reclaim directories with outgoing namecache 1229 * entries if this argument is strue 1230 * @param trigger Only reclaim vnodes with fewer than this many resident 1231 * pages. 1232 * @param target How many vnodes to reclaim. 1233 * @return The number of vnodes that were reclaimed. 1234 */ 1235 static int 1236 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target) 1237 { 1238 struct vnode *vp, *mvp; 1239 struct mount *mp; 1240 struct vm_object *object; 1241 u_long done; 1242 bool retried; 1243 1244 mtx_assert(&vnode_list_mtx, MA_OWNED); 1245 1246 retried = false; 1247 done = 0; 1248 1249 mvp = vnode_list_reclaim_marker; 1250 restart: 1251 vp = mvp; 1252 while (done < target) { 1253 vp = TAILQ_NEXT(vp, v_vnodelist); 1254 if (__predict_false(vp == NULL)) 1255 break; 1256 1257 if (__predict_false(vp->v_type == VMARKER)) 1258 continue; 1259 1260 /* 1261 * If it's been deconstructed already, it's still 1262 * referenced, or it exceeds the trigger, skip it. 1263 * Also skip free vnodes. We are trying to make space 1264 * for more free vnodes, not reduce their count. 1265 */ 1266 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 || 1267 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src))) 1268 goto next_iter; 1269 1270 if (vp->v_type == VBAD || vp->v_type == VNON) 1271 goto next_iter; 1272 1273 object = atomic_load_ptr(&vp->v_object); 1274 if (object == NULL || object->resident_page_count > trigger) { 1275 goto next_iter; 1276 } 1277 1278 /* 1279 * Handle races against vnode allocation. Filesystems lock the 1280 * vnode some time after it gets returned from getnewvnode, 1281 * despite type and hold count being manipulated earlier. 1282 * Resorting to checking v_mount restores guarantees present 1283 * before the global list was reworked to contain all vnodes. 1284 */ 1285 if (!VI_TRYLOCK(vp)) 1286 goto next_iter; 1287 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) { 1288 VI_UNLOCK(vp); 1289 goto next_iter; 1290 } 1291 if (vp->v_mount == NULL) { 1292 VI_UNLOCK(vp); 1293 goto next_iter; 1294 } 1295 vholdl(vp); 1296 VI_UNLOCK(vp); 1297 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1298 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1299 mtx_unlock(&vnode_list_mtx); 1300 1301 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1302 vdrop_recycle(vp); 1303 goto next_iter_unlocked; 1304 } 1305 if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) { 1306 vdrop_recycle(vp); 1307 vn_finished_write(mp); 1308 goto next_iter_unlocked; 1309 } 1310 1311 VI_LOCK(vp); 1312 if (vp->v_usecount > 0 || 1313 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 1314 (vp->v_object != NULL && vp->v_object->handle == vp && 1315 vp->v_object->resident_page_count > trigger)) { 1316 VOP_UNLOCK(vp); 1317 vdropl_recycle(vp); 1318 vn_finished_write(mp); 1319 goto next_iter_unlocked; 1320 } 1321 recycles_count++; 1322 vgonel(vp); 1323 VOP_UNLOCK(vp); 1324 vdropl_recycle(vp); 1325 vn_finished_write(mp); 1326 done++; 1327 next_iter_unlocked: 1328 maybe_yield(); 1329 mtx_lock(&vnode_list_mtx); 1330 goto restart; 1331 next_iter: 1332 MPASS(vp->v_type != VMARKER); 1333 if (!should_yield()) 1334 continue; 1335 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1336 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1337 mtx_unlock(&vnode_list_mtx); 1338 kern_yield(PRI_USER); 1339 mtx_lock(&vnode_list_mtx); 1340 goto restart; 1341 } 1342 if (done == 0 && !retried) { 1343 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1344 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist); 1345 retried = true; 1346 goto restart; 1347 } 1348 return (done); 1349 } 1350 1351 static int max_free_per_call = 10000; 1352 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_free_per_call, 0, 1353 "limit on vnode free requests per call to the vnlru_free routine (legacy)"); 1354 SYSCTL_INT(_vfs_vnode_vnlru, OID_AUTO, max_free_per_call, CTLFLAG_RW, 1355 &max_free_per_call, 0, 1356 "limit on vnode free requests per call to the vnlru_free routine"); 1357 1358 /* 1359 * Attempt to recycle requested amount of free vnodes. 1360 */ 1361 static int 1362 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp, bool isvnlru) 1363 { 1364 struct vnode *vp; 1365 struct mount *mp; 1366 int ocount; 1367 bool retried; 1368 1369 mtx_assert(&vnode_list_mtx, MA_OWNED); 1370 if (count > max_free_per_call) 1371 count = max_free_per_call; 1372 if (count == 0) { 1373 mtx_unlock(&vnode_list_mtx); 1374 return (0); 1375 } 1376 ocount = count; 1377 retried = false; 1378 vp = mvp; 1379 for (;;) { 1380 vp = TAILQ_NEXT(vp, v_vnodelist); 1381 if (__predict_false(vp == NULL)) { 1382 /* 1383 * The free vnode marker can be past eligible vnodes: 1384 * 1. if vdbatch_process trylock failed 1385 * 2. if vtryrecycle failed 1386 * 1387 * If so, start the scan from scratch. 1388 */ 1389 if (!retried && vnlru_read_freevnodes() > 0) { 1390 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1391 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist); 1392 vp = mvp; 1393 retried = true; 1394 continue; 1395 } 1396 1397 /* 1398 * Give up 1399 */ 1400 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1401 TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist); 1402 mtx_unlock(&vnode_list_mtx); 1403 break; 1404 } 1405 if (__predict_false(vp->v_type == VMARKER)) 1406 continue; 1407 if (vp->v_holdcnt > 0) 1408 continue; 1409 /* 1410 * Don't recycle if our vnode is from different type 1411 * of mount point. Note that mp is type-safe, the 1412 * check does not reach unmapped address even if 1413 * vnode is reclaimed. 1414 */ 1415 if (mnt_op != NULL && (mp = vp->v_mount) != NULL && 1416 mp->mnt_op != mnt_op) { 1417 continue; 1418 } 1419 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) { 1420 continue; 1421 } 1422 if (!vhold_recycle_free(vp)) 1423 continue; 1424 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1425 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1426 mtx_unlock(&vnode_list_mtx); 1427 /* 1428 * FIXME: ignores the return value, meaning it may be nothing 1429 * got recycled but it claims otherwise to the caller. 1430 * 1431 * Originally the value started being ignored in 2005 with 1432 * 114a1006a8204aa156e1f9ad6476cdff89cada7f . 1433 * 1434 * Respecting the value can run into significant stalls if most 1435 * vnodes belong to one file system and it has writes 1436 * suspended. In presence of many threads and millions of 1437 * vnodes they keep contending on the vnode_list_mtx lock only 1438 * to find vnodes they can't recycle. 1439 * 1440 * The solution would be to pre-check if the vnode is likely to 1441 * be recycle-able, but it needs to happen with the 1442 * vnode_list_mtx lock held. This runs into a problem where 1443 * VOP_GETWRITEMOUNT (currently needed to find out about if 1444 * writes are frozen) can take locks which LOR against it. 1445 * 1446 * Check nullfs for one example (null_getwritemount). 1447 */ 1448 vtryrecycle(vp, isvnlru); 1449 count--; 1450 if (count == 0) { 1451 break; 1452 } 1453 mtx_lock(&vnode_list_mtx); 1454 vp = mvp; 1455 } 1456 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1457 return (ocount - count); 1458 } 1459 1460 /* 1461 * XXX: returns without vnode_list_mtx locked! 1462 */ 1463 static int 1464 vnlru_free_locked_direct(int count) 1465 { 1466 int ret; 1467 1468 mtx_assert(&vnode_list_mtx, MA_OWNED); 1469 ret = vnlru_free_impl(count, NULL, vnode_list_free_marker, false); 1470 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1471 return (ret); 1472 } 1473 1474 static int 1475 vnlru_free_locked_vnlru(int count) 1476 { 1477 int ret; 1478 1479 mtx_assert(&vnode_list_mtx, MA_OWNED); 1480 ret = vnlru_free_impl(count, NULL, vnode_list_free_marker, true); 1481 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1482 return (ret); 1483 } 1484 1485 static int 1486 vnlru_free_vnlru(int count) 1487 { 1488 1489 mtx_lock(&vnode_list_mtx); 1490 return (vnlru_free_locked_vnlru(count)); 1491 } 1492 1493 void 1494 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp) 1495 { 1496 1497 MPASS(mnt_op != NULL); 1498 MPASS(mvp != NULL); 1499 VNPASS(mvp->v_type == VMARKER, mvp); 1500 mtx_lock(&vnode_list_mtx); 1501 vnlru_free_impl(count, mnt_op, mvp, true); 1502 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1503 } 1504 1505 struct vnode * 1506 vnlru_alloc_marker(void) 1507 { 1508 struct vnode *mvp; 1509 1510 mvp = vn_alloc_marker(NULL); 1511 mtx_lock(&vnode_list_mtx); 1512 TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist); 1513 mtx_unlock(&vnode_list_mtx); 1514 return (mvp); 1515 } 1516 1517 void 1518 vnlru_free_marker(struct vnode *mvp) 1519 { 1520 mtx_lock(&vnode_list_mtx); 1521 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1522 mtx_unlock(&vnode_list_mtx); 1523 vn_free_marker(mvp); 1524 } 1525 1526 static void 1527 vnlru_recalc(void) 1528 { 1529 1530 mtx_assert(&vnode_list_mtx, MA_OWNED); 1531 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100); 1532 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */ 1533 vlowat = vhiwat / 2; 1534 } 1535 1536 /* 1537 * Attempt to recycle vnodes in a context that is always safe to block. 1538 * Calling vlrurecycle() from the bowels of filesystem code has some 1539 * interesting deadlock problems. 1540 */ 1541 static struct proc *vnlruproc; 1542 static int vnlruproc_sig; 1543 static u_long vnlruproc_kicks; 1544 1545 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, kicks, CTLFLAG_RD, &vnlruproc_kicks, 0, 1546 "Number of times vnlru awakened due to vnode shortage"); 1547 1548 #define VNLRU_COUNT_SLOP 100 1549 1550 /* 1551 * The main freevnodes counter is only updated when a counter local to CPU 1552 * diverges from 0 by more than VNLRU_FREEVNODES_SLOP. CPUs are conditionally 1553 * walked to compute a more accurate total. 1554 * 1555 * Note: the actual value at any given moment can still exceed slop, but it 1556 * should not be by significant margin in practice. 1557 */ 1558 #define VNLRU_FREEVNODES_SLOP 126 1559 1560 static void __noinline 1561 vfs_freevnodes_rollup(int8_t *lfreevnodes) 1562 { 1563 1564 atomic_add_long(&freevnodes, *lfreevnodes); 1565 *lfreevnodes = 0; 1566 critical_exit(); 1567 } 1568 1569 static __inline void 1570 vfs_freevnodes_inc(void) 1571 { 1572 int8_t *lfreevnodes; 1573 1574 critical_enter(); 1575 lfreevnodes = PCPU_PTR(vfs_freevnodes); 1576 (*lfreevnodes)++; 1577 if (__predict_false(*lfreevnodes == VNLRU_FREEVNODES_SLOP)) 1578 vfs_freevnodes_rollup(lfreevnodes); 1579 else 1580 critical_exit(); 1581 } 1582 1583 static __inline void 1584 vfs_freevnodes_dec(void) 1585 { 1586 int8_t *lfreevnodes; 1587 1588 critical_enter(); 1589 lfreevnodes = PCPU_PTR(vfs_freevnodes); 1590 (*lfreevnodes)--; 1591 if (__predict_false(*lfreevnodes == -VNLRU_FREEVNODES_SLOP)) 1592 vfs_freevnodes_rollup(lfreevnodes); 1593 else 1594 critical_exit(); 1595 } 1596 1597 static u_long 1598 vnlru_read_freevnodes(void) 1599 { 1600 long slop, rfreevnodes, rfreevnodes_old; 1601 int cpu; 1602 1603 rfreevnodes = atomic_load_long(&freevnodes); 1604 rfreevnodes_old = atomic_load_long(&freevnodes_old); 1605 1606 if (rfreevnodes > rfreevnodes_old) 1607 slop = rfreevnodes - rfreevnodes_old; 1608 else 1609 slop = rfreevnodes_old - rfreevnodes; 1610 if (slop < VNLRU_FREEVNODES_SLOP) 1611 return (rfreevnodes >= 0 ? rfreevnodes : 0); 1612 CPU_FOREACH(cpu) { 1613 rfreevnodes += cpuid_to_pcpu[cpu]->pc_vfs_freevnodes; 1614 } 1615 atomic_store_long(&freevnodes_old, rfreevnodes); 1616 return (freevnodes_old >= 0 ? freevnodes_old : 0); 1617 } 1618 1619 static bool 1620 vnlru_under(u_long rnumvnodes, u_long limit) 1621 { 1622 u_long rfreevnodes, space; 1623 1624 if (__predict_false(rnumvnodes > desiredvnodes)) 1625 return (true); 1626 1627 space = desiredvnodes - rnumvnodes; 1628 if (space < limit) { 1629 rfreevnodes = vnlru_read_freevnodes(); 1630 if (rfreevnodes > wantfreevnodes) 1631 space += rfreevnodes - wantfreevnodes; 1632 } 1633 return (space < limit); 1634 } 1635 1636 static void 1637 vnlru_kick_locked(void) 1638 { 1639 1640 mtx_assert(&vnode_list_mtx, MA_OWNED); 1641 if (vnlruproc_sig == 0) { 1642 vnlruproc_sig = 1; 1643 vnlruproc_kicks++; 1644 wakeup(vnlruproc); 1645 } 1646 } 1647 1648 static void 1649 vnlru_kick_cond(void) 1650 { 1651 1652 if (vnlru_read_freevnodes() > wantfreevnodes) 1653 return; 1654 1655 if (vnlruproc_sig) 1656 return; 1657 mtx_lock(&vnode_list_mtx); 1658 vnlru_kick_locked(); 1659 mtx_unlock(&vnode_list_mtx); 1660 } 1661 1662 static void 1663 vnlru_proc_sleep(void) 1664 { 1665 1666 if (vnlruproc_sig) { 1667 vnlruproc_sig = 0; 1668 wakeup(&vnlruproc_sig); 1669 } 1670 msleep(vnlruproc, &vnode_list_mtx, PVFS|PDROP, "vlruwt", hz); 1671 } 1672 1673 /* 1674 * A lighter version of the machinery below. 1675 * 1676 * Tries to reach goals only by recycling free vnodes and does not invoke 1677 * uma_reclaim(UMA_RECLAIM_DRAIN). 1678 * 1679 * This works around pathological behavior in vnlru in presence of tons of free 1680 * vnodes, but without having to rewrite the machinery at this time. Said 1681 * behavior boils down to continuously trying to reclaim all kinds of vnodes 1682 * (cycling through all levels of "force") when the count is transiently above 1683 * limit. This happens a lot when all vnodes are used up and vn_alloc 1684 * speculatively increments the counter. 1685 * 1686 * Sample testcase: vnode limit 8388608, 20 separate directory trees each with 1687 * 1 million files in total and 20 find(1) processes stating them in parallel 1688 * (one per each tree). 1689 * 1690 * On a kernel with only stock machinery this needs anywhere between 60 and 120 1691 * seconds to execute (time varies *wildly* between runs). With the workaround 1692 * it consistently stays around 20 seconds [it got further down with later 1693 * changes]. 1694 * 1695 * That is to say the entire thing needs a fundamental redesign (most notably 1696 * to accommodate faster recycling), the above only tries to get it ouf the way. 1697 * 1698 * Return values are: 1699 * -1 -- fallback to regular vnlru loop 1700 * 0 -- do nothing, go to sleep 1701 * >0 -- recycle this many vnodes 1702 */ 1703 static long 1704 vnlru_proc_light_pick(void) 1705 { 1706 u_long rnumvnodes, rfreevnodes; 1707 1708 if (vstir || vnlruproc_sig == 1) 1709 return (-1); 1710 1711 rnumvnodes = atomic_load_long(&numvnodes); 1712 rfreevnodes = vnlru_read_freevnodes(); 1713 1714 /* 1715 * vnode limit might have changed and now we may be at a significant 1716 * excess. Bail if we can't sort it out with free vnodes. 1717 * 1718 * Due to atomic updates the count can legitimately go above 1719 * the limit for a short period, don't bother doing anything in 1720 * that case. 1721 */ 1722 if (rnumvnodes > desiredvnodes + VNLRU_COUNT_SLOP + 10) { 1723 if (rnumvnodes - rfreevnodes >= desiredvnodes || 1724 rfreevnodes <= wantfreevnodes) { 1725 return (-1); 1726 } 1727 1728 return (rnumvnodes - desiredvnodes); 1729 } 1730 1731 /* 1732 * Don't try to reach wantfreevnodes target if there are too few vnodes 1733 * to begin with. 1734 */ 1735 if (rnumvnodes < wantfreevnodes) { 1736 return (0); 1737 } 1738 1739 if (rfreevnodes < wantfreevnodes) { 1740 return (-1); 1741 } 1742 1743 return (0); 1744 } 1745 1746 static bool 1747 vnlru_proc_light(void) 1748 { 1749 long freecount; 1750 1751 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1752 1753 freecount = vnlru_proc_light_pick(); 1754 if (freecount == -1) 1755 return (false); 1756 1757 if (freecount != 0) { 1758 vnlru_free_vnlru(freecount); 1759 } 1760 1761 mtx_lock(&vnode_list_mtx); 1762 vnlru_proc_sleep(); 1763 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1764 return (true); 1765 } 1766 1767 static u_long uma_reclaim_calls; 1768 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, uma_reclaim_calls, CTLFLAG_RD | CTLFLAG_STATS, 1769 &uma_reclaim_calls, 0, "Number of calls to uma_reclaim"); 1770 1771 static void 1772 vnlru_proc(void) 1773 { 1774 u_long rnumvnodes, rfreevnodes, target; 1775 unsigned long onumvnodes; 1776 int done, force, trigger, usevnodes; 1777 bool reclaim_nc_src, want_reread; 1778 1779 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc, 1780 SHUTDOWN_PRI_FIRST); 1781 1782 force = 0; 1783 want_reread = false; 1784 for (;;) { 1785 kproc_suspend_check(vnlruproc); 1786 1787 if (force == 0 && vnlru_proc_light()) 1788 continue; 1789 1790 mtx_lock(&vnode_list_mtx); 1791 rnumvnodes = atomic_load_long(&numvnodes); 1792 1793 if (want_reread) { 1794 force = vnlru_under(numvnodes, vhiwat) ? 1 : 0; 1795 want_reread = false; 1796 } 1797 1798 /* 1799 * If numvnodes is too large (due to desiredvnodes being 1800 * adjusted using its sysctl, or emergency growth), first 1801 * try to reduce it by discarding free vnodes. 1802 */ 1803 if (rnumvnodes > desiredvnodes + 10) { 1804 vnlru_free_locked_vnlru(rnumvnodes - desiredvnodes); 1805 mtx_lock(&vnode_list_mtx); 1806 rnumvnodes = atomic_load_long(&numvnodes); 1807 } 1808 /* 1809 * Sleep if the vnode cache is in a good state. This is 1810 * when it is not over-full and has space for about a 4% 1811 * or 9% expansion (by growing its size or inexcessively 1812 * reducing free vnode count). Otherwise, try to reclaim 1813 * space for a 10% expansion. 1814 */ 1815 if (vstir && force == 0) { 1816 force = 1; 1817 vstir = false; 1818 } 1819 if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) { 1820 vnlru_proc_sleep(); 1821 continue; 1822 } 1823 rfreevnodes = vnlru_read_freevnodes(); 1824 1825 onumvnodes = rnumvnodes; 1826 /* 1827 * Calculate parameters for recycling. These are the same 1828 * throughout the loop to give some semblance of fairness. 1829 * The trigger point is to avoid recycling vnodes with lots 1830 * of resident pages. We aren't trying to free memory; we 1831 * are trying to recycle or at least free vnodes. 1832 */ 1833 if (rnumvnodes <= desiredvnodes) 1834 usevnodes = rnumvnodes - rfreevnodes; 1835 else 1836 usevnodes = rnumvnodes; 1837 if (usevnodes <= 0) 1838 usevnodes = 1; 1839 /* 1840 * The trigger value is chosen to give a conservatively 1841 * large value to ensure that it alone doesn't prevent 1842 * making progress. The value can easily be so large that 1843 * it is effectively infinite in some congested and 1844 * misconfigured cases, and this is necessary. Normally 1845 * it is about 8 to 100 (pages), which is quite large. 1846 */ 1847 trigger = vm_cnt.v_page_count * 2 / usevnodes; 1848 if (force < 2) 1849 trigger = vsmalltrigger; 1850 reclaim_nc_src = force >= 3; 1851 target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1); 1852 target = target / 10 + 1; 1853 done = vlrureclaim(reclaim_nc_src, trigger, target); 1854 mtx_unlock(&vnode_list_mtx); 1855 /* 1856 * Total number of vnodes can transiently go slightly above the 1857 * limit (see vn_alloc_hard), no need to call uma_reclaim if 1858 * this happens. 1859 */ 1860 if (onumvnodes + VNLRU_COUNT_SLOP + 1000 > desiredvnodes && 1861 numvnodes <= desiredvnodes) { 1862 uma_reclaim_calls++; 1863 uma_reclaim(UMA_RECLAIM_DRAIN); 1864 } 1865 if (done == 0) { 1866 if (force == 0 || force == 1) { 1867 force = 2; 1868 continue; 1869 } 1870 if (force == 2) { 1871 force = 3; 1872 continue; 1873 } 1874 want_reread = true; 1875 force = 0; 1876 vnlru_nowhere++; 1877 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 1878 } else { 1879 want_reread = true; 1880 kern_yield(PRI_USER); 1881 } 1882 } 1883 } 1884 1885 static struct kproc_desc vnlru_kp = { 1886 "vnlru", 1887 vnlru_proc, 1888 &vnlruproc 1889 }; 1890 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, 1891 &vnlru_kp); 1892 1893 /* 1894 * Routines having to do with the management of the vnode table. 1895 */ 1896 1897 /* 1898 * Try to recycle a freed vnode. 1899 */ 1900 static int 1901 vtryrecycle(struct vnode *vp, bool isvnlru) 1902 { 1903 struct mount *vnmp; 1904 1905 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 1906 VNPASS(vp->v_holdcnt > 0, vp); 1907 /* 1908 * This vnode may found and locked via some other list, if so we 1909 * can't recycle it yet. 1910 */ 1911 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 1912 CTR2(KTR_VFS, 1913 "%s: impossible to recycle, vp %p lock is already held", 1914 __func__, vp); 1915 vdrop_recycle(vp); 1916 return (EWOULDBLOCK); 1917 } 1918 /* 1919 * Don't recycle if its filesystem is being suspended. 1920 */ 1921 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 1922 VOP_UNLOCK(vp); 1923 CTR2(KTR_VFS, 1924 "%s: impossible to recycle, cannot start the write for %p", 1925 __func__, vp); 1926 vdrop_recycle(vp); 1927 return (EBUSY); 1928 } 1929 /* 1930 * If we got this far, we need to acquire the interlock and see if 1931 * anyone picked up this vnode from another list. If not, we will 1932 * mark it with DOOMED via vgonel() so that anyone who does find it 1933 * will skip over it. 1934 */ 1935 VI_LOCK(vp); 1936 if (vp->v_usecount) { 1937 VOP_UNLOCK(vp); 1938 vdropl_recycle(vp); 1939 vn_finished_write(vnmp); 1940 CTR2(KTR_VFS, 1941 "%s: impossible to recycle, %p is already referenced", 1942 __func__, vp); 1943 return (EBUSY); 1944 } 1945 if (!VN_IS_DOOMED(vp)) { 1946 if (isvnlru) 1947 recycles_free_count++; 1948 else 1949 counter_u64_add(direct_recycles_free_count, 1); 1950 vgonel(vp); 1951 } 1952 VOP_UNLOCK(vp); 1953 vdropl_recycle(vp); 1954 vn_finished_write(vnmp); 1955 return (0); 1956 } 1957 1958 /* 1959 * Allocate a new vnode. 1960 * 1961 * The operation never returns an error. Returning an error was disabled 1962 * in r145385 (dated 2005) with the following comment: 1963 * 1964 * XXX Not all VFS_VGET/ffs_vget callers check returns. 1965 * 1966 * Given the age of this commit (almost 15 years at the time of writing this 1967 * comment) restoring the ability to fail requires a significant audit of 1968 * all codepaths. 1969 * 1970 * The routine can try to free a vnode or stall for up to 1 second waiting for 1971 * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation. 1972 */ 1973 static u_long vn_alloc_cyclecount; 1974 static u_long vn_alloc_sleeps; 1975 1976 SYSCTL_ULONG(_vfs_vnode_stats, OID_AUTO, alloc_sleeps, CTLFLAG_RD, &vn_alloc_sleeps, 0, 1977 "Number of times vnode allocation blocked waiting on vnlru"); 1978 1979 static struct vnode * __noinline 1980 vn_alloc_hard(struct mount *mp, u_long rnumvnodes, bool bumped) 1981 { 1982 u_long rfreevnodes; 1983 1984 if (bumped) { 1985 if (rnumvnodes > desiredvnodes + VNLRU_COUNT_SLOP) { 1986 atomic_subtract_long(&numvnodes, 1); 1987 bumped = false; 1988 } 1989 } 1990 1991 mtx_lock(&vnode_list_mtx); 1992 1993 /* 1994 * Reload 'numvnodes', as since we acquired the lock, it may have 1995 * changed significantly if we waited, and 'rnumvnodes' above was only 1996 * actually passed if 'bumped' is true (else it is 0). 1997 */ 1998 rnumvnodes = atomic_load_long(&numvnodes); 1999 if (rnumvnodes + !bumped < desiredvnodes) { 2000 vn_alloc_cyclecount = 0; 2001 mtx_unlock(&vnode_list_mtx); 2002 goto alloc; 2003 } 2004 2005 rfreevnodes = vnlru_read_freevnodes(); 2006 if (vn_alloc_cyclecount++ >= rfreevnodes) { 2007 vn_alloc_cyclecount = 0; 2008 vstir = true; 2009 } 2010 2011 /* 2012 * Grow the vnode cache if it will not be above its target max after 2013 * growing. Otherwise, if there is at least one free vnode, try to 2014 * reclaim 1 item from it before growing the cache (possibly above its 2015 * target max if the reclamation failed or is delayed). 2016 */ 2017 if (vnlru_free_locked_direct(1) > 0) 2018 goto alloc; 2019 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 2020 if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 2021 /* 2022 * Wait for space for a new vnode. 2023 */ 2024 if (bumped) { 2025 atomic_subtract_long(&numvnodes, 1); 2026 bumped = false; 2027 } 2028 mtx_lock(&vnode_list_mtx); 2029 vnlru_kick_locked(); 2030 vn_alloc_sleeps++; 2031 msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz); 2032 if (atomic_load_long(&numvnodes) + 1 > desiredvnodes && 2033 vnlru_read_freevnodes() > 1) 2034 vnlru_free_locked_direct(1); 2035 else 2036 mtx_unlock(&vnode_list_mtx); 2037 } 2038 alloc: 2039 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 2040 if (!bumped) 2041 atomic_add_long(&numvnodes, 1); 2042 vnlru_kick_cond(); 2043 return (uma_zalloc_smr(vnode_zone, M_WAITOK)); 2044 } 2045 2046 static struct vnode * 2047 vn_alloc(struct mount *mp) 2048 { 2049 u_long rnumvnodes; 2050 2051 if (__predict_false(vn_alloc_cyclecount != 0)) 2052 return (vn_alloc_hard(mp, 0, false)); 2053 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; 2054 if (__predict_false(vnlru_under(rnumvnodes, vlowat))) { 2055 return (vn_alloc_hard(mp, rnumvnodes, true)); 2056 } 2057 2058 return (uma_zalloc_smr(vnode_zone, M_WAITOK)); 2059 } 2060 2061 static void 2062 vn_free(struct vnode *vp) 2063 { 2064 2065 atomic_subtract_long(&numvnodes, 1); 2066 uma_zfree_smr(vnode_zone, vp); 2067 } 2068 2069 /* 2070 * Allocate a new vnode. 2071 */ 2072 int 2073 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, 2074 struct vnode **vpp) 2075 { 2076 struct vnode *vp; 2077 struct thread *td; 2078 struct lock_object *lo; 2079 2080 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag); 2081 2082 KASSERT(vops->registered, 2083 ("%s: not registered vector op %p\n", __func__, vops)); 2084 cache_validate_vop_vector(mp, vops); 2085 2086 td = curthread; 2087 if (td->td_vp_reserved != NULL) { 2088 vp = td->td_vp_reserved; 2089 td->td_vp_reserved = NULL; 2090 } else { 2091 vp = vn_alloc(mp); 2092 } 2093 counter_u64_add(vnodes_created, 1); 2094 2095 vn_set_state(vp, VSTATE_UNINITIALIZED); 2096 2097 /* 2098 * Locks are given the generic name "vnode" when created. 2099 * Follow the historic practice of using the filesystem 2100 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc. 2101 * 2102 * Locks live in a witness group keyed on their name. Thus, 2103 * when a lock is renamed, it must also move from the witness 2104 * group of its old name to the witness group of its new name. 2105 * 2106 * The change only needs to be made when the vnode moves 2107 * from one filesystem type to another. We ensure that each 2108 * filesystem use a single static name pointer for its tag so 2109 * that we can compare pointers rather than doing a strcmp(). 2110 */ 2111 lo = &vp->v_vnlock->lock_object; 2112 #ifdef WITNESS 2113 if (lo->lo_name != tag) { 2114 #endif 2115 lo->lo_name = tag; 2116 #ifdef WITNESS 2117 WITNESS_DESTROY(lo); 2118 WITNESS_INIT(lo, tag); 2119 } 2120 #endif 2121 /* 2122 * By default, don't allow shared locks unless filesystems opt-in. 2123 */ 2124 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE; 2125 /* 2126 * Finalize various vnode identity bits. 2127 */ 2128 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp)); 2129 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp)); 2130 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp)); 2131 vp->v_type = VNON; 2132 vp->v_op = vops; 2133 vp->v_irflag = 0; 2134 v_init_counters(vp); 2135 vn_seqc_init(vp); 2136 vp->v_bufobj.bo_ops = &buf_ops_bio; 2137 #ifdef DIAGNOSTIC 2138 if (mp == NULL && vops != &dead_vnodeops) 2139 printf("NULL mp in getnewvnode(9), tag %s\n", tag); 2140 #endif 2141 #ifdef MAC 2142 mac_vnode_init(vp); 2143 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 2144 mac_vnode_associate_singlelabel(mp, vp); 2145 #endif 2146 if (mp != NULL) { 2147 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize; 2148 } 2149 2150 /* 2151 * For the filesystems which do not use vfs_hash_insert(), 2152 * still initialize v_hash to have vfs_hash_index() useful. 2153 * E.g., nullfs uses vfs_hash_index() on the lower vnode for 2154 * its own hashing. 2155 */ 2156 vp->v_hash = (uintptr_t)vp >> vnsz2log; 2157 2158 *vpp = vp; 2159 return (0); 2160 } 2161 2162 void 2163 getnewvnode_reserve(void) 2164 { 2165 struct thread *td; 2166 2167 td = curthread; 2168 MPASS(td->td_vp_reserved == NULL); 2169 td->td_vp_reserved = vn_alloc(NULL); 2170 } 2171 2172 void 2173 getnewvnode_drop_reserve(void) 2174 { 2175 struct thread *td; 2176 2177 td = curthread; 2178 if (td->td_vp_reserved != NULL) { 2179 vn_free(td->td_vp_reserved); 2180 td->td_vp_reserved = NULL; 2181 } 2182 } 2183 2184 static void __noinline 2185 freevnode(struct vnode *vp) 2186 { 2187 struct bufobj *bo; 2188 2189 ASSERT_VOP_UNLOCKED(vp, __func__); 2190 2191 /* 2192 * The vnode has been marked for destruction, so free it. 2193 * 2194 * The vnode will be returned to the zone where it will 2195 * normally remain until it is needed for another vnode. We 2196 * need to cleanup (or verify that the cleanup has already 2197 * been done) any residual data left from its current use 2198 * so as not to contaminate the freshly allocated vnode. 2199 */ 2200 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp); 2201 /* 2202 * Paired with vgone. 2203 */ 2204 vn_seqc_write_end_free(vp); 2205 2206 bo = &vp->v_bufobj; 2207 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 2208 VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp); 2209 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 2210 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 2211 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 2212 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 2213 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp, 2214 ("clean blk trie not empty")); 2215 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 2216 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp, 2217 ("dirty blk trie not empty")); 2218 VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp, 2219 ("Leaked inactivation")); 2220 VI_UNLOCK(vp); 2221 cache_assert_no_entries(vp); 2222 2223 #ifdef MAC 2224 mac_vnode_destroy(vp); 2225 #endif 2226 if (vp->v_pollinfo != NULL) { 2227 int error __diagused; 2228 2229 /* 2230 * Use LK_NOWAIT to shut up witness about the lock. We may get 2231 * here while having another vnode locked when trying to 2232 * satisfy a lookup and needing to recycle. 2233 */ 2234 error = VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT); 2235 VNASSERT(error == 0, vp, 2236 ("freevnode: cannot lock vp %p for pollinfo destroy", vp)); 2237 destroy_vpollinfo(vp->v_pollinfo); 2238 VOP_UNLOCK(vp); 2239 vp->v_pollinfo = NULL; 2240 } 2241 vp->v_mountedhere = NULL; 2242 vp->v_unpcb = NULL; 2243 vp->v_rdev = NULL; 2244 vp->v_fifoinfo = NULL; 2245 vp->v_iflag = 0; 2246 vp->v_vflag = 0; 2247 bo->bo_flag = 0; 2248 vn_free(vp); 2249 } 2250 2251 /* 2252 * Delete from old mount point vnode list, if on one. 2253 */ 2254 static void 2255 delmntque(struct vnode *vp) 2256 { 2257 struct mount *mp; 2258 2259 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); 2260 2261 mp = vp->v_mount; 2262 MNT_ILOCK(mp); 2263 VI_LOCK(vp); 2264 vp->v_mount = NULL; 2265 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 2266 ("bad mount point vnode list size")); 2267 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2268 mp->mnt_nvnodelistsize--; 2269 MNT_REL(mp); 2270 MNT_IUNLOCK(mp); 2271 /* 2272 * The caller expects the interlock to be still held. 2273 */ 2274 ASSERT_VI_LOCKED(vp, __func__); 2275 } 2276 2277 static int 2278 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr) 2279 { 2280 2281 KASSERT(vp->v_mount == NULL, 2282 ("insmntque: vnode already on per mount vnode list")); 2283 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 2284 if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) { 2285 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); 2286 } else { 2287 KASSERT(!dtr, 2288 ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup", 2289 __func__)); 2290 } 2291 2292 /* 2293 * We acquire the vnode interlock early to ensure that the 2294 * vnode cannot be recycled by another process releasing a 2295 * holdcnt on it before we get it on both the vnode list 2296 * and the active vnode list. The mount mutex protects only 2297 * manipulation of the vnode list and the vnode freelist 2298 * mutex protects only manipulation of the active vnode list. 2299 * Hence the need to hold the vnode interlock throughout. 2300 */ 2301 MNT_ILOCK(mp); 2302 VI_LOCK(vp); 2303 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 && 2304 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || 2305 mp->mnt_nvnodelistsize == 0)) && 2306 (vp->v_vflag & VV_FORCEINSMQ) == 0) { 2307 VI_UNLOCK(vp); 2308 MNT_IUNLOCK(mp); 2309 if (dtr) { 2310 vp->v_data = NULL; 2311 vp->v_op = &dead_vnodeops; 2312 vgone(vp); 2313 vput(vp); 2314 } 2315 return (EBUSY); 2316 } 2317 vp->v_mount = mp; 2318 MNT_REF(mp); 2319 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2320 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 2321 ("neg mount point vnode list size")); 2322 mp->mnt_nvnodelistsize++; 2323 VI_UNLOCK(vp); 2324 MNT_IUNLOCK(mp); 2325 return (0); 2326 } 2327 2328 /* 2329 * Insert into list of vnodes for the new mount point, if available. 2330 * insmntque() reclaims the vnode on insertion failure, insmntque1() 2331 * leaves handling of the vnode to the caller. 2332 */ 2333 int 2334 insmntque(struct vnode *vp, struct mount *mp) 2335 { 2336 return (insmntque1_int(vp, mp, true)); 2337 } 2338 2339 int 2340 insmntque1(struct vnode *vp, struct mount *mp) 2341 { 2342 return (insmntque1_int(vp, mp, false)); 2343 } 2344 2345 /* 2346 * Flush out and invalidate all buffers associated with a bufobj 2347 * Called with the underlying object locked. 2348 */ 2349 int 2350 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) 2351 { 2352 int error; 2353 2354 BO_LOCK(bo); 2355 if (flags & V_SAVE) { 2356 error = bufobj_wwait(bo, slpflag, slptimeo); 2357 if (error) { 2358 BO_UNLOCK(bo); 2359 return (error); 2360 } 2361 if (bo->bo_dirty.bv_cnt > 0) { 2362 BO_UNLOCK(bo); 2363 do { 2364 error = BO_SYNC(bo, MNT_WAIT); 2365 } while (error == ERELOOKUP); 2366 if (error != 0) 2367 return (error); 2368 BO_LOCK(bo); 2369 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) { 2370 BO_UNLOCK(bo); 2371 return (EBUSY); 2372 } 2373 } 2374 } 2375 /* 2376 * If you alter this loop please notice that interlock is dropped and 2377 * reacquired in flushbuflist. Special care is needed to ensure that 2378 * no race conditions occur from this. 2379 */ 2380 do { 2381 error = flushbuflist(&bo->bo_clean, 2382 flags, bo, slpflag, slptimeo); 2383 if (error == 0 && !(flags & V_CLEANONLY)) 2384 error = flushbuflist(&bo->bo_dirty, 2385 flags, bo, slpflag, slptimeo); 2386 if (error != 0 && error != EAGAIN) { 2387 BO_UNLOCK(bo); 2388 return (error); 2389 } 2390 } while (error != 0); 2391 2392 /* 2393 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 2394 * have write I/O in-progress but if there is a VM object then the 2395 * VM object can also have read-I/O in-progress. 2396 */ 2397 do { 2398 bufobj_wwait(bo, 0, 0); 2399 if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) { 2400 BO_UNLOCK(bo); 2401 vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx"); 2402 BO_LOCK(bo); 2403 } 2404 } while (bo->bo_numoutput > 0); 2405 BO_UNLOCK(bo); 2406 2407 /* 2408 * Destroy the copy in the VM cache, too. 2409 */ 2410 if (bo->bo_object != NULL && 2411 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) { 2412 VM_OBJECT_WLOCK(bo->bo_object); 2413 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? 2414 OBJPR_CLEANONLY : 0); 2415 VM_OBJECT_WUNLOCK(bo->bo_object); 2416 } 2417 2418 #ifdef INVARIANTS 2419 BO_LOCK(bo); 2420 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO | 2421 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 || 2422 bo->bo_clean.bv_cnt > 0)) 2423 panic("vinvalbuf: flush failed"); 2424 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 && 2425 bo->bo_dirty.bv_cnt > 0) 2426 panic("vinvalbuf: flush dirty failed"); 2427 BO_UNLOCK(bo); 2428 #endif 2429 return (0); 2430 } 2431 2432 /* 2433 * Flush out and invalidate all buffers associated with a vnode. 2434 * Called with the underlying object locked. 2435 */ 2436 int 2437 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 2438 { 2439 2440 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 2441 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 2442 if (vp->v_object != NULL && vp->v_object->handle != vp) 2443 return (0); 2444 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); 2445 } 2446 2447 /* 2448 * Flush out buffers on the specified list. 2449 * 2450 */ 2451 static int 2452 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 2453 int slptimeo) 2454 { 2455 struct buf *bp, *nbp; 2456 int retval, error; 2457 daddr_t lblkno; 2458 b_xflags_t xflags; 2459 2460 ASSERT_BO_WLOCKED(bo); 2461 2462 retval = 0; 2463 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 2464 /* 2465 * If we are flushing both V_NORMAL and V_ALT buffers then 2466 * do not skip any buffers. If we are flushing only V_NORMAL 2467 * buffers then skip buffers marked as BX_ALTDATA. If we are 2468 * flushing only V_ALT buffers then skip buffers not marked 2469 * as BX_ALTDATA. 2470 */ 2471 if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) && 2472 (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) || 2473 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) { 2474 continue; 2475 } 2476 if (nbp != NULL) { 2477 lblkno = nbp->b_lblkno; 2478 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN); 2479 } 2480 retval = EAGAIN; 2481 error = BUF_TIMELOCK(bp, 2482 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), 2483 "flushbuf", slpflag, slptimeo); 2484 if (error) { 2485 BO_LOCK(bo); 2486 return (error != ENOLCK ? error : EAGAIN); 2487 } 2488 KASSERT(bp->b_bufobj == bo, 2489 ("bp %p wrong b_bufobj %p should be %p", 2490 bp, bp->b_bufobj, bo)); 2491 /* 2492 * XXX Since there are no node locks for NFS, I 2493 * believe there is a slight chance that a delayed 2494 * write will occur while sleeping just above, so 2495 * check for it. 2496 */ 2497 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 2498 (flags & V_SAVE)) { 2499 bremfree(bp); 2500 bp->b_flags |= B_ASYNC; 2501 bwrite(bp); 2502 BO_LOCK(bo); 2503 return (EAGAIN); /* XXX: why not loop ? */ 2504 } 2505 bremfree(bp); 2506 bp->b_flags |= (B_INVAL | B_RELBUF); 2507 bp->b_flags &= ~B_ASYNC; 2508 brelse(bp); 2509 BO_LOCK(bo); 2510 if (nbp == NULL) 2511 break; 2512 nbp = gbincore(bo, lblkno); 2513 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2514 != xflags) 2515 break; /* nbp invalid */ 2516 } 2517 return (retval); 2518 } 2519 2520 int 2521 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn) 2522 { 2523 struct buf *bp; 2524 int error; 2525 daddr_t lblkno; 2526 2527 ASSERT_BO_LOCKED(bo); 2528 2529 for (lblkno = startn;;) { 2530 again: 2531 bp = buf_lookup_ge(bufv, lblkno); 2532 if (bp == NULL || bp->b_lblkno >= endn) 2533 break; 2534 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 2535 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0); 2536 if (error != 0) { 2537 BO_RLOCK(bo); 2538 if (error == ENOLCK) 2539 goto again; 2540 return (error); 2541 } 2542 KASSERT(bp->b_bufobj == bo, 2543 ("bp %p wrong b_bufobj %p should be %p", 2544 bp, bp->b_bufobj, bo)); 2545 lblkno = bp->b_lblkno + 1; 2546 if ((bp->b_flags & B_MANAGED) == 0) 2547 bremfree(bp); 2548 bp->b_flags |= B_RELBUF; 2549 /* 2550 * In the VMIO case, use the B_NOREUSE flag to hint that the 2551 * pages backing each buffer in the range are unlikely to be 2552 * reused. Dirty buffers will have the hint applied once 2553 * they've been written. 2554 */ 2555 if ((bp->b_flags & B_VMIO) != 0) 2556 bp->b_flags |= B_NOREUSE; 2557 brelse(bp); 2558 BO_RLOCK(bo); 2559 } 2560 return (0); 2561 } 2562 2563 /* 2564 * Truncate a file's buffer and pages to a specified length. This 2565 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 2566 * sync activity. 2567 */ 2568 int 2569 vtruncbuf(struct vnode *vp, off_t length, int blksize) 2570 { 2571 struct buf *bp, *nbp; 2572 struct bufobj *bo; 2573 daddr_t startlbn; 2574 2575 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__, 2576 vp, blksize, (uintmax_t)length); 2577 2578 /* 2579 * Round up to the *next* lbn. 2580 */ 2581 startlbn = howmany(length, blksize); 2582 2583 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 2584 2585 bo = &vp->v_bufobj; 2586 restart_unlocked: 2587 BO_LOCK(bo); 2588 2589 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN) 2590 ; 2591 2592 if (length > 0) { 2593 /* 2594 * Write out vnode metadata, e.g. indirect blocks. 2595 */ 2596 restartsync: 2597 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2598 if (bp->b_lblkno >= 0) 2599 continue; 2600 /* 2601 * Since we hold the vnode lock this should only 2602 * fail if we're racing with the buf daemon. 2603 */ 2604 if (BUF_LOCK(bp, 2605 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2606 BO_LOCKPTR(bo)) == ENOLCK) 2607 goto restart_unlocked; 2608 2609 VNASSERT((bp->b_flags & B_DELWRI), vp, 2610 ("buf(%p) on dirty queue without DELWRI", bp)); 2611 2612 bremfree(bp); 2613 bawrite(bp); 2614 BO_LOCK(bo); 2615 goto restartsync; 2616 } 2617 } 2618 2619 bufobj_wwait(bo, 0, 0); 2620 BO_UNLOCK(bo); 2621 vnode_pager_setsize(vp, length); 2622 2623 return (0); 2624 } 2625 2626 /* 2627 * Invalidate the cached pages of a file's buffer within the range of block 2628 * numbers [startlbn, endlbn). 2629 */ 2630 void 2631 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn, 2632 int blksize) 2633 { 2634 struct bufobj *bo; 2635 off_t start, end; 2636 2637 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range"); 2638 2639 start = blksize * startlbn; 2640 end = blksize * endlbn; 2641 2642 bo = &vp->v_bufobj; 2643 BO_LOCK(bo); 2644 MPASS(blksize == bo->bo_bsize); 2645 2646 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN) 2647 ; 2648 2649 BO_UNLOCK(bo); 2650 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1)); 2651 } 2652 2653 static int 2654 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, 2655 daddr_t startlbn, daddr_t endlbn) 2656 { 2657 struct bufv *bv; 2658 struct buf *bp, *nbp; 2659 uint8_t anyfreed; 2660 bool clean; 2661 2662 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked"); 2663 ASSERT_BO_LOCKED(bo); 2664 2665 anyfreed = 1; 2666 clean = true; 2667 do { 2668 bv = clean ? &bo->bo_clean : &bo->bo_dirty; 2669 bp = buf_lookup_ge(bv, startlbn); 2670 if (bp == NULL) 2671 continue; 2672 TAILQ_FOREACH_FROM_SAFE(bp, &bv->bv_hd, b_bobufs, nbp) { 2673 if (bp->b_lblkno >= endlbn) 2674 break; 2675 if (BUF_LOCK(bp, 2676 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2677 BO_LOCKPTR(bo)) == ENOLCK) { 2678 BO_LOCK(bo); 2679 return (EAGAIN); 2680 } 2681 2682 bremfree(bp); 2683 bp->b_flags |= B_INVAL | B_RELBUF; 2684 bp->b_flags &= ~B_ASYNC; 2685 brelse(bp); 2686 anyfreed = 2; 2687 2688 BO_LOCK(bo); 2689 if (nbp != NULL && 2690 (((nbp->b_xflags & 2691 (clean ? BX_VNCLEAN : BX_VNDIRTY)) == 0) || 2692 nbp->b_vp != vp || 2693 (nbp->b_flags & B_DELWRI) == (clean? B_DELWRI: 0))) 2694 return (EAGAIN); 2695 } 2696 } while (clean = !clean, anyfreed-- > 0); 2697 return (0); 2698 } 2699 2700 static void 2701 buf_vlist_remove(struct buf *bp) 2702 { 2703 struct bufv *bv; 2704 b_xflags_t flags; 2705 2706 flags = bp->b_xflags; 2707 2708 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2709 ASSERT_BO_WLOCKED(bp->b_bufobj); 2710 KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 && 2711 (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN), 2712 ("%s: buffer %p has invalid queue state", __func__, bp)); 2713 2714 if ((flags & BX_VNDIRTY) != 0) 2715 bv = &bp->b_bufobj->bo_dirty; 2716 else 2717 bv = &bp->b_bufobj->bo_clean; 2718 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno); 2719 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 2720 bv->bv_cnt--; 2721 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 2722 } 2723 2724 /* 2725 * Add the buffer to the sorted clean or dirty block list. Return zero on 2726 * success, EEXIST if a buffer with this identity already exists, or another 2727 * error on allocation failure. 2728 */ 2729 static inline int 2730 buf_vlist_find_or_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 2731 { 2732 struct bufv *bv; 2733 struct buf *n; 2734 int error; 2735 2736 ASSERT_BO_WLOCKED(bo); 2737 KASSERT((bo->bo_flag & BO_NOBUFS) == 0, 2738 ("buf_vlist_add: bo %p does not allow bufs", bo)); 2739 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0, 2740 ("dead bo %p", bo)); 2741 KASSERT((bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) == xflags, 2742 ("buf_vlist_add: b_xflags %#x not set on bp %p", xflags, bp)); 2743 2744 if (xflags & BX_VNDIRTY) 2745 bv = &bo->bo_dirty; 2746 else 2747 bv = &bo->bo_clean; 2748 2749 error = buf_insert_lookup_le(bv, bp, &n); 2750 if (n == NULL) { 2751 KASSERT(error != EEXIST, 2752 ("buf_vlist_add: EEXIST but no existing buf found: bp %p", 2753 bp)); 2754 } else { 2755 KASSERT(n->b_lblkno <= bp->b_lblkno, 2756 ("buf_vlist_add: out of order insert/lookup: bp %p n %p", 2757 bp, n)); 2758 KASSERT((n->b_lblkno == bp->b_lblkno) == (error == EEXIST), 2759 ("buf_vlist_add: inconsistent result for existing buf: " 2760 "error %d bp %p n %p", error, bp, n)); 2761 } 2762 if (error != 0) 2763 return (error); 2764 2765 /* Keep the list ordered. */ 2766 if (n == NULL) { 2767 KASSERT(TAILQ_EMPTY(&bv->bv_hd) || 2768 bp->b_lblkno < TAILQ_FIRST(&bv->bv_hd)->b_lblkno, 2769 ("buf_vlist_add: queue order: " 2770 "%p should be before first %p", 2771 bp, TAILQ_FIRST(&bv->bv_hd))); 2772 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs); 2773 } else { 2774 KASSERT(TAILQ_NEXT(n, b_bobufs) == NULL || 2775 bp->b_lblkno < TAILQ_NEXT(n, b_bobufs)->b_lblkno, 2776 ("buf_vlist_add: queue order: " 2777 "%p should be before next %p", 2778 bp, TAILQ_NEXT(n, b_bobufs))); 2779 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs); 2780 } 2781 2782 bv->bv_cnt++; 2783 return (0); 2784 } 2785 2786 /* 2787 * Add the buffer to the sorted clean or dirty block list. 2788 * 2789 * NOTE: xflags is passed as a constant, optimizing this inline function! 2790 */ 2791 static void 2792 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 2793 { 2794 int error; 2795 2796 KASSERT((bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) == 0, 2797 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 2798 bp->b_xflags |= xflags; 2799 error = buf_vlist_find_or_add(bp, bo, xflags); 2800 if (error) 2801 panic("buf_vlist_add: error=%d", error); 2802 } 2803 2804 /* 2805 * Look up a buffer using the buffer tries. 2806 */ 2807 struct buf * 2808 gbincore(struct bufobj *bo, daddr_t lblkno) 2809 { 2810 struct buf *bp; 2811 2812 ASSERT_BO_LOCKED(bo); 2813 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno); 2814 if (bp != NULL) 2815 return (bp); 2816 return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno)); 2817 } 2818 2819 /* 2820 * Look up a buf using the buffer tries, without the bufobj lock. This relies 2821 * on SMR for safe lookup, and bufs being in a no-free zone to provide type 2822 * stability of the result. Like other lockless lookups, the found buf may 2823 * already be invalid by the time this function returns. 2824 */ 2825 struct buf * 2826 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno) 2827 { 2828 struct buf *bp; 2829 2830 ASSERT_BO_UNLOCKED(bo); 2831 bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno); 2832 if (bp != NULL) 2833 return (bp); 2834 return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno)); 2835 } 2836 2837 /* 2838 * Associate a buffer with a vnode. 2839 */ 2840 int 2841 bgetvp(struct vnode *vp, struct buf *bp) 2842 { 2843 struct bufobj *bo; 2844 int error; 2845 2846 bo = &vp->v_bufobj; 2847 ASSERT_BO_UNLOCKED(bo); 2848 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 2849 2850 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 2851 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 2852 ("bgetvp: bp already attached! %p", bp)); 2853 2854 /* 2855 * Add the buf to the vnode's clean list unless we lost a race and find 2856 * an existing buf in either dirty or clean. 2857 */ 2858 bp->b_vp = vp; 2859 bp->b_bufobj = bo; 2860 bp->b_xflags |= BX_VNCLEAN; 2861 error = EEXIST; 2862 BO_LOCK(bo); 2863 if (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, bp->b_lblkno) == NULL) 2864 error = buf_vlist_find_or_add(bp, bo, BX_VNCLEAN); 2865 BO_UNLOCK(bo); 2866 if (__predict_true(error == 0)) { 2867 vhold(vp); 2868 return (0); 2869 } 2870 if (error != EEXIST) 2871 panic("bgetvp: buf_vlist_add error: %d", error); 2872 bp->b_vp = NULL; 2873 bp->b_bufobj = NULL; 2874 bp->b_xflags &= ~BX_VNCLEAN; 2875 return (error); 2876 } 2877 2878 /* 2879 * Disassociate a buffer from a vnode. 2880 */ 2881 void 2882 brelvp(struct buf *bp) 2883 { 2884 struct bufobj *bo; 2885 struct vnode *vp; 2886 2887 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2888 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 2889 2890 /* 2891 * Delete from old vnode list, if on one. 2892 */ 2893 vp = bp->b_vp; /* XXX */ 2894 bo = bp->b_bufobj; 2895 BO_LOCK(bo); 2896 buf_vlist_remove(bp); 2897 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2898 bo->bo_flag &= ~BO_ONWORKLST; 2899 mtx_lock(&sync_mtx); 2900 LIST_REMOVE(bo, bo_synclist); 2901 syncer_worklist_len--; 2902 mtx_unlock(&sync_mtx); 2903 } 2904 bp->b_vp = NULL; 2905 bp->b_bufobj = NULL; 2906 BO_UNLOCK(bo); 2907 vdrop(vp); 2908 } 2909 2910 /* 2911 * Add an item to the syncer work queue. 2912 */ 2913 static void 2914 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 2915 { 2916 int slot; 2917 2918 ASSERT_BO_WLOCKED(bo); 2919 2920 mtx_lock(&sync_mtx); 2921 if (bo->bo_flag & BO_ONWORKLST) 2922 LIST_REMOVE(bo, bo_synclist); 2923 else { 2924 bo->bo_flag |= BO_ONWORKLST; 2925 syncer_worklist_len++; 2926 } 2927 2928 if (delay > syncer_maxdelay - 2) 2929 delay = syncer_maxdelay - 2; 2930 slot = (syncer_delayno + delay) & syncer_mask; 2931 2932 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 2933 mtx_unlock(&sync_mtx); 2934 } 2935 2936 static int 2937 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 2938 { 2939 int error, len; 2940 2941 mtx_lock(&sync_mtx); 2942 len = syncer_worklist_len - sync_vnode_count; 2943 mtx_unlock(&sync_mtx); 2944 error = SYSCTL_OUT(req, &len, sizeof(len)); 2945 return (error); 2946 } 2947 2948 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, 2949 CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0, 2950 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 2951 2952 static struct proc *updateproc; 2953 static void sched_sync(void); 2954 static struct kproc_desc up_kp = { 2955 "syncer", 2956 sched_sync, 2957 &updateproc 2958 }; 2959 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); 2960 2961 static int 2962 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) 2963 { 2964 struct vnode *vp; 2965 struct mount *mp; 2966 2967 *bo = LIST_FIRST(slp); 2968 if (*bo == NULL) 2969 return (0); 2970 vp = bo2vnode(*bo); 2971 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) 2972 return (1); 2973 /* 2974 * We use vhold in case the vnode does not 2975 * successfully sync. vhold prevents the vnode from 2976 * going away when we unlock the sync_mtx so that 2977 * we can acquire the vnode interlock. 2978 */ 2979 vholdl(vp); 2980 mtx_unlock(&sync_mtx); 2981 VI_UNLOCK(vp); 2982 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2983 vdrop(vp); 2984 mtx_lock(&sync_mtx); 2985 return (*bo == LIST_FIRST(slp)); 2986 } 2987 MPASSERT(mp == NULL || (curthread->td_pflags & TDP_IGNSUSP) != 0 || 2988 (mp->mnt_kern_flag & MNTK_SUSPENDED) == 0, mp, 2989 ("suspended mp syncing vp %p", vp)); 2990 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2991 (void) VOP_FSYNC(vp, MNT_LAZY, td); 2992 VOP_UNLOCK(vp); 2993 vn_finished_write(mp); 2994 BO_LOCK(*bo); 2995 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { 2996 /* 2997 * Put us back on the worklist. The worklist 2998 * routine will remove us from our current 2999 * position and then add us back in at a later 3000 * position. 3001 */ 3002 vn_syncer_add_to_worklist(*bo, syncdelay); 3003 } 3004 BO_UNLOCK(*bo); 3005 vdrop(vp); 3006 mtx_lock(&sync_mtx); 3007 return (0); 3008 } 3009 3010 static int first_printf = 1; 3011 3012 /* 3013 * System filesystem synchronizer daemon. 3014 */ 3015 static void 3016 sched_sync(void) 3017 { 3018 struct synclist *next, *slp; 3019 struct bufobj *bo; 3020 long starttime; 3021 struct thread *td = curthread; 3022 int last_work_seen; 3023 int net_worklist_len; 3024 int syncer_final_iter; 3025 int error; 3026 3027 last_work_seen = 0; 3028 syncer_final_iter = 0; 3029 syncer_state = SYNCER_RUNNING; 3030 starttime = time_uptime; 3031 td->td_pflags |= TDP_NORUNNINGBUF; 3032 3033 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 3034 SHUTDOWN_PRI_LAST); 3035 3036 mtx_lock(&sync_mtx); 3037 for (;;) { 3038 if (syncer_state == SYNCER_FINAL_DELAY && 3039 syncer_final_iter == 0) { 3040 mtx_unlock(&sync_mtx); 3041 kproc_suspend_check(td->td_proc); 3042 mtx_lock(&sync_mtx); 3043 } 3044 net_worklist_len = syncer_worklist_len - sync_vnode_count; 3045 if (syncer_state != SYNCER_RUNNING && 3046 starttime != time_uptime) { 3047 if (first_printf) { 3048 printf("\nSyncing disks, vnodes remaining... "); 3049 first_printf = 0; 3050 } 3051 printf("%d ", net_worklist_len); 3052 } 3053 starttime = time_uptime; 3054 3055 /* 3056 * Push files whose dirty time has expired. Be careful 3057 * of interrupt race on slp queue. 3058 * 3059 * Skip over empty worklist slots when shutting down. 3060 */ 3061 do { 3062 slp = &syncer_workitem_pending[syncer_delayno]; 3063 syncer_delayno += 1; 3064 if (syncer_delayno == syncer_maxdelay) 3065 syncer_delayno = 0; 3066 next = &syncer_workitem_pending[syncer_delayno]; 3067 /* 3068 * If the worklist has wrapped since the 3069 * it was emptied of all but syncer vnodes, 3070 * switch to the FINAL_DELAY state and run 3071 * for one more second. 3072 */ 3073 if (syncer_state == SYNCER_SHUTTING_DOWN && 3074 net_worklist_len == 0 && 3075 last_work_seen == syncer_delayno) { 3076 syncer_state = SYNCER_FINAL_DELAY; 3077 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 3078 } 3079 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 3080 syncer_worklist_len > 0); 3081 3082 /* 3083 * Keep track of the last time there was anything 3084 * on the worklist other than syncer vnodes. 3085 * Return to the SHUTTING_DOWN state if any 3086 * new work appears. 3087 */ 3088 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 3089 last_work_seen = syncer_delayno; 3090 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 3091 syncer_state = SYNCER_SHUTTING_DOWN; 3092 while (!LIST_EMPTY(slp)) { 3093 error = sync_vnode(slp, &bo, td); 3094 if (error == 1) { 3095 LIST_REMOVE(bo, bo_synclist); 3096 LIST_INSERT_HEAD(next, bo, bo_synclist); 3097 continue; 3098 } 3099 3100 if (first_printf == 0) { 3101 /* 3102 * Drop the sync mutex, because some watchdog 3103 * drivers need to sleep while patting 3104 */ 3105 mtx_unlock(&sync_mtx); 3106 wdog_kern_pat(WD_LASTVAL); 3107 mtx_lock(&sync_mtx); 3108 } 3109 } 3110 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 3111 syncer_final_iter--; 3112 /* 3113 * The variable rushjob allows the kernel to speed up the 3114 * processing of the filesystem syncer process. A rushjob 3115 * value of N tells the filesystem syncer to process the next 3116 * N seconds worth of work on its queue ASAP. Currently rushjob 3117 * is used by the soft update code to speed up the filesystem 3118 * syncer process when the incore state is getting so far 3119 * ahead of the disk that the kernel memory pool is being 3120 * threatened with exhaustion. 3121 */ 3122 if (rushjob > 0) { 3123 rushjob -= 1; 3124 continue; 3125 } 3126 /* 3127 * Just sleep for a short period of time between 3128 * iterations when shutting down to allow some I/O 3129 * to happen. 3130 * 3131 * If it has taken us less than a second to process the 3132 * current work, then wait. Otherwise start right over 3133 * again. We can still lose time if any single round 3134 * takes more than two seconds, but it does not really 3135 * matter as we are just trying to generally pace the 3136 * filesystem activity. 3137 */ 3138 if (syncer_state != SYNCER_RUNNING || 3139 time_uptime == starttime) { 3140 thread_lock(td); 3141 sched_prio(td, PPAUSE); 3142 thread_unlock(td); 3143 } 3144 if (syncer_state != SYNCER_RUNNING) 3145 cv_timedwait(&sync_wakeup, &sync_mtx, 3146 hz / SYNCER_SHUTDOWN_SPEEDUP); 3147 else if (time_uptime == starttime) 3148 cv_timedwait(&sync_wakeup, &sync_mtx, hz); 3149 } 3150 } 3151 3152 /* 3153 * Request the syncer daemon to speed up its work. 3154 * We never push it to speed up more than half of its 3155 * normal turn time, otherwise it could take over the cpu. 3156 */ 3157 int 3158 speedup_syncer(void) 3159 { 3160 int ret = 0; 3161 3162 mtx_lock(&sync_mtx); 3163 if (rushjob < syncdelay / 2) { 3164 rushjob += 1; 3165 stat_rush_requests += 1; 3166 ret = 1; 3167 } 3168 mtx_unlock(&sync_mtx); 3169 cv_broadcast(&sync_wakeup); 3170 return (ret); 3171 } 3172 3173 /* 3174 * Tell the syncer to speed up its work and run though its work 3175 * list several times, then tell it to shut down. 3176 */ 3177 static void 3178 syncer_shutdown(void *arg, int howto) 3179 { 3180 3181 if (howto & RB_NOSYNC) 3182 return; 3183 mtx_lock(&sync_mtx); 3184 syncer_state = SYNCER_SHUTTING_DOWN; 3185 rushjob = 0; 3186 mtx_unlock(&sync_mtx); 3187 cv_broadcast(&sync_wakeup); 3188 kproc_shutdown(arg, howto); 3189 } 3190 3191 void 3192 syncer_suspend(void) 3193 { 3194 3195 syncer_shutdown(updateproc, 0); 3196 } 3197 3198 void 3199 syncer_resume(void) 3200 { 3201 3202 mtx_lock(&sync_mtx); 3203 first_printf = 1; 3204 syncer_state = SYNCER_RUNNING; 3205 mtx_unlock(&sync_mtx); 3206 cv_broadcast(&sync_wakeup); 3207 kproc_resume(updateproc); 3208 } 3209 3210 /* 3211 * Move the buffer between the clean and dirty lists of its vnode. 3212 */ 3213 void 3214 reassignbuf(struct buf *bp) 3215 { 3216 struct vnode *vp; 3217 struct bufobj *bo; 3218 int delay; 3219 #ifdef INVARIANTS 3220 struct bufv *bv; 3221 #endif 3222 3223 vp = bp->b_vp; 3224 bo = bp->b_bufobj; 3225 3226 KASSERT((bp->b_flags & B_PAGING) == 0, 3227 ("%s: cannot reassign paging buffer %p", __func__, bp)); 3228 3229 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 3230 bp, bp->b_vp, bp->b_flags); 3231 3232 BO_LOCK(bo); 3233 if ((bo->bo_flag & BO_NONSTERILE) == 0) { 3234 /* 3235 * Coordinate with getblk's unlocked lookup. Make 3236 * BO_NONSTERILE visible before the first reassignbuf produces 3237 * any side effect. This could be outside the bo lock if we 3238 * used a separate atomic flag field. 3239 */ 3240 bo->bo_flag |= BO_NONSTERILE; 3241 atomic_thread_fence_rel(); 3242 } 3243 buf_vlist_remove(bp); 3244 3245 /* 3246 * If dirty, put on list of dirty buffers; otherwise insert onto list 3247 * of clean buffers. 3248 */ 3249 if (bp->b_flags & B_DELWRI) { 3250 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 3251 switch (vp->v_type) { 3252 case VDIR: 3253 delay = dirdelay; 3254 break; 3255 case VCHR: 3256 delay = metadelay; 3257 break; 3258 default: 3259 delay = filedelay; 3260 } 3261 vn_syncer_add_to_worklist(bo, delay); 3262 } 3263 buf_vlist_add(bp, bo, BX_VNDIRTY); 3264 } else { 3265 buf_vlist_add(bp, bo, BX_VNCLEAN); 3266 3267 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 3268 mtx_lock(&sync_mtx); 3269 LIST_REMOVE(bo, bo_synclist); 3270 syncer_worklist_len--; 3271 mtx_unlock(&sync_mtx); 3272 bo->bo_flag &= ~BO_ONWORKLST; 3273 } 3274 } 3275 #ifdef INVARIANTS 3276 bv = &bo->bo_clean; 3277 bp = TAILQ_FIRST(&bv->bv_hd); 3278 KASSERT(bp == NULL || bp->b_bufobj == bo, 3279 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3280 bp = TAILQ_LAST(&bv->bv_hd, buflists); 3281 KASSERT(bp == NULL || bp->b_bufobj == bo, 3282 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3283 bv = &bo->bo_dirty; 3284 bp = TAILQ_FIRST(&bv->bv_hd); 3285 KASSERT(bp == NULL || bp->b_bufobj == bo, 3286 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3287 bp = TAILQ_LAST(&bv->bv_hd, buflists); 3288 KASSERT(bp == NULL || bp->b_bufobj == bo, 3289 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3290 #endif 3291 BO_UNLOCK(bo); 3292 } 3293 3294 static void 3295 v_init_counters(struct vnode *vp) 3296 { 3297 3298 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0, 3299 vp, ("%s called for an initialized vnode", __FUNCTION__)); 3300 ASSERT_VI_UNLOCKED(vp, __FUNCTION__); 3301 3302 refcount_init(&vp->v_holdcnt, 1); 3303 refcount_init(&vp->v_usecount, 1); 3304 } 3305 3306 /* 3307 * Get a usecount on a vnode. 3308 * 3309 * vget and vget_finish may fail to lock the vnode if they lose a race against 3310 * it being doomed. LK_RETRY can be passed in flags to lock it anyway. 3311 * 3312 * Consumers which don't guarantee liveness of the vnode can use SMR to 3313 * try to get a reference. Note this operation can fail since the vnode 3314 * may be awaiting getting freed by the time they get to it. 3315 */ 3316 enum vgetstate 3317 vget_prep_smr(struct vnode *vp) 3318 { 3319 enum vgetstate vs; 3320 3321 VFS_SMR_ASSERT_ENTERED(); 3322 3323 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 3324 vs = VGET_USECOUNT; 3325 } else { 3326 if (vhold_smr(vp)) 3327 vs = VGET_HOLDCNT; 3328 else 3329 vs = VGET_NONE; 3330 } 3331 return (vs); 3332 } 3333 3334 enum vgetstate 3335 vget_prep(struct vnode *vp) 3336 { 3337 enum vgetstate vs; 3338 3339 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 3340 vs = VGET_USECOUNT; 3341 } else { 3342 vhold(vp); 3343 vs = VGET_HOLDCNT; 3344 } 3345 return (vs); 3346 } 3347 3348 void 3349 vget_abort(struct vnode *vp, enum vgetstate vs) 3350 { 3351 3352 switch (vs) { 3353 case VGET_USECOUNT: 3354 vrele(vp); 3355 break; 3356 case VGET_HOLDCNT: 3357 vdrop(vp); 3358 break; 3359 default: 3360 __assert_unreachable(); 3361 } 3362 } 3363 3364 int 3365 vget(struct vnode *vp, int flags) 3366 { 3367 enum vgetstate vs; 3368 3369 vs = vget_prep(vp); 3370 return (vget_finish(vp, flags, vs)); 3371 } 3372 3373 int 3374 vget_finish(struct vnode *vp, int flags, enum vgetstate vs) 3375 { 3376 int error; 3377 3378 if ((flags & LK_INTERLOCK) != 0) 3379 ASSERT_VI_LOCKED(vp, __func__); 3380 else 3381 ASSERT_VI_UNLOCKED(vp, __func__); 3382 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); 3383 VNPASS(vp->v_holdcnt > 0, vp); 3384 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); 3385 3386 error = vn_lock(vp, flags); 3387 if (__predict_false(error != 0)) { 3388 vget_abort(vp, vs); 3389 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, 3390 vp); 3391 return (error); 3392 } 3393 3394 vget_finish_ref(vp, vs); 3395 return (0); 3396 } 3397 3398 void 3399 vget_finish_ref(struct vnode *vp, enum vgetstate vs) 3400 { 3401 int old; 3402 3403 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); 3404 VNPASS(vp->v_holdcnt > 0, vp); 3405 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); 3406 3407 if (vs == VGET_USECOUNT) 3408 return; 3409 3410 /* 3411 * We hold the vnode. If the usecount is 0 it will be utilized to keep 3412 * the vnode around. Otherwise someone else lended their hold count and 3413 * we have to drop ours. 3414 */ 3415 old = atomic_fetchadd_int(&vp->v_usecount, 1); 3416 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old)); 3417 if (old != 0) { 3418 #ifdef INVARIANTS 3419 old = atomic_fetchadd_int(&vp->v_holdcnt, -1); 3420 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); 3421 #else 3422 refcount_release(&vp->v_holdcnt); 3423 #endif 3424 } 3425 } 3426 3427 void 3428 vref(struct vnode *vp) 3429 { 3430 enum vgetstate vs; 3431 3432 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3433 vs = vget_prep(vp); 3434 vget_finish_ref(vp, vs); 3435 } 3436 3437 void 3438 vrefact(struct vnode *vp) 3439 { 3440 int old __diagused; 3441 3442 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3443 old = refcount_acquire(&vp->v_usecount); 3444 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); 3445 } 3446 3447 void 3448 vlazy(struct vnode *vp) 3449 { 3450 struct mount *mp; 3451 3452 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__)); 3453 3454 if ((vp->v_mflag & VMP_LAZYLIST) != 0) 3455 return; 3456 /* 3457 * We may get here for inactive routines after the vnode got doomed. 3458 */ 3459 if (VN_IS_DOOMED(vp)) 3460 return; 3461 mp = vp->v_mount; 3462 mtx_lock(&mp->mnt_listmtx); 3463 if ((vp->v_mflag & VMP_LAZYLIST) == 0) { 3464 vp->v_mflag |= VMP_LAZYLIST; 3465 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3466 mp->mnt_lazyvnodelistsize++; 3467 } 3468 mtx_unlock(&mp->mnt_listmtx); 3469 } 3470 3471 static void 3472 vunlazy(struct vnode *vp) 3473 { 3474 struct mount *mp; 3475 3476 ASSERT_VI_LOCKED(vp, __func__); 3477 VNPASS(!VN_IS_DOOMED(vp), vp); 3478 3479 mp = vp->v_mount; 3480 mtx_lock(&mp->mnt_listmtx); 3481 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 3482 /* 3483 * Don't remove the vnode from the lazy list if another thread 3484 * has increased the hold count. It may have re-enqueued the 3485 * vnode to the lazy list and is now responsible for its 3486 * removal. 3487 */ 3488 if (vp->v_holdcnt == 0) { 3489 vp->v_mflag &= ~VMP_LAZYLIST; 3490 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3491 mp->mnt_lazyvnodelistsize--; 3492 } 3493 mtx_unlock(&mp->mnt_listmtx); 3494 } 3495 3496 /* 3497 * This routine is only meant to be called from vgonel prior to dooming 3498 * the vnode. 3499 */ 3500 static void 3501 vunlazy_gone(struct vnode *vp) 3502 { 3503 struct mount *mp; 3504 3505 ASSERT_VOP_ELOCKED(vp, __func__); 3506 ASSERT_VI_LOCKED(vp, __func__); 3507 VNPASS(!VN_IS_DOOMED(vp), vp); 3508 3509 if (vp->v_mflag & VMP_LAZYLIST) { 3510 mp = vp->v_mount; 3511 mtx_lock(&mp->mnt_listmtx); 3512 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 3513 vp->v_mflag &= ~VMP_LAZYLIST; 3514 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3515 mp->mnt_lazyvnodelistsize--; 3516 mtx_unlock(&mp->mnt_listmtx); 3517 } 3518 } 3519 3520 static void 3521 vdefer_inactive(struct vnode *vp) 3522 { 3523 3524 ASSERT_VI_LOCKED(vp, __func__); 3525 VNPASS(vp->v_holdcnt > 0, vp); 3526 if (VN_IS_DOOMED(vp)) { 3527 vdropl(vp); 3528 return; 3529 } 3530 if (vp->v_iflag & VI_DEFINACT) { 3531 VNPASS(vp->v_holdcnt > 1, vp); 3532 vdropl(vp); 3533 return; 3534 } 3535 if (vp->v_usecount > 0) { 3536 vp->v_iflag &= ~VI_OWEINACT; 3537 vdropl(vp); 3538 return; 3539 } 3540 vlazy(vp); 3541 vp->v_iflag |= VI_DEFINACT; 3542 VI_UNLOCK(vp); 3543 atomic_add_long(&deferred_inact, 1); 3544 } 3545 3546 static void 3547 vdefer_inactive_unlocked(struct vnode *vp) 3548 { 3549 3550 VI_LOCK(vp); 3551 if ((vp->v_iflag & VI_OWEINACT) == 0) { 3552 vdropl(vp); 3553 return; 3554 } 3555 vdefer_inactive(vp); 3556 } 3557 3558 enum vput_op { VRELE, VPUT, VUNREF }; 3559 3560 /* 3561 * Handle ->v_usecount transitioning to 0. 3562 * 3563 * By releasing the last usecount we take ownership of the hold count which 3564 * provides liveness of the vnode, meaning we have to vdrop. 3565 * 3566 * For all vnodes we may need to perform inactive processing. It requires an 3567 * exclusive lock on the vnode, while it is legal to call here with only a 3568 * shared lock (or no locks). If locking the vnode in an expected manner fails, 3569 * inactive processing gets deferred to the syncer. 3570 * 3571 * XXX Some filesystems pass in an exclusively locked vnode and strongly depend 3572 * on the lock being held all the way until VOP_INACTIVE. This in particular 3573 * happens with UFS which adds half-constructed vnodes to the hash, where they 3574 * can be found by other code. 3575 */ 3576 static void 3577 vput_final(struct vnode *vp, enum vput_op func) 3578 { 3579 int error; 3580 bool want_unlock; 3581 3582 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3583 VNPASS(vp->v_holdcnt > 0, vp); 3584 3585 VI_LOCK(vp); 3586 3587 /* 3588 * By the time we got here someone else might have transitioned 3589 * the count back to > 0. 3590 */ 3591 if (vp->v_usecount > 0) 3592 goto out; 3593 3594 /* 3595 * If the vnode is doomed vgone already performed inactive processing 3596 * (if needed). 3597 */ 3598 if (VN_IS_DOOMED(vp)) 3599 goto out; 3600 3601 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0)) 3602 goto out; 3603 3604 if (vp->v_iflag & VI_DOINGINACT) 3605 goto out; 3606 3607 /* 3608 * Locking operations here will drop the interlock and possibly the 3609 * vnode lock, opening a window where the vnode can get doomed all the 3610 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to 3611 * perform inactive. 3612 */ 3613 vp->v_iflag |= VI_OWEINACT; 3614 want_unlock = false; 3615 error = 0; 3616 switch (func) { 3617 case VRELE: 3618 switch (VOP_ISLOCKED(vp)) { 3619 case LK_EXCLUSIVE: 3620 break; 3621 case LK_EXCLOTHER: 3622 case 0: 3623 want_unlock = true; 3624 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 3625 VI_LOCK(vp); 3626 break; 3627 default: 3628 /* 3629 * The lock has at least one sharer, but we have no way 3630 * to conclude whether this is us. Play it safe and 3631 * defer processing. 3632 */ 3633 error = EAGAIN; 3634 break; 3635 } 3636 break; 3637 case VPUT: 3638 want_unlock = true; 3639 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3640 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | 3641 LK_NOWAIT); 3642 VI_LOCK(vp); 3643 } 3644 break; 3645 case VUNREF: 3646 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3647 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); 3648 VI_LOCK(vp); 3649 } 3650 break; 3651 } 3652 if (error == 0) { 3653 if (func == VUNREF) { 3654 VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp, 3655 ("recursive vunref")); 3656 vp->v_vflag |= VV_UNREF; 3657 } 3658 for (;;) { 3659 error = vinactive(vp); 3660 if (want_unlock) 3661 VOP_UNLOCK(vp); 3662 if (error != ERELOOKUP || !want_unlock) 3663 break; 3664 VOP_LOCK(vp, LK_EXCLUSIVE); 3665 } 3666 if (func == VUNREF) 3667 vp->v_vflag &= ~VV_UNREF; 3668 vdropl(vp); 3669 } else { 3670 vdefer_inactive(vp); 3671 } 3672 return; 3673 out: 3674 if (func == VPUT) 3675 VOP_UNLOCK(vp); 3676 vdropl(vp); 3677 } 3678 3679 /* 3680 * Decrement ->v_usecount for a vnode. 3681 * 3682 * Releasing the last use count requires additional processing, see vput_final 3683 * above for details. 3684 * 3685 * Comment above each variant denotes lock state on entry and exit. 3686 */ 3687 3688 /* 3689 * in: any 3690 * out: same as passed in 3691 */ 3692 void 3693 vrele(struct vnode *vp) 3694 { 3695 3696 ASSERT_VI_UNLOCKED(vp, __func__); 3697 if (!refcount_release(&vp->v_usecount)) 3698 return; 3699 vput_final(vp, VRELE); 3700 } 3701 3702 /* 3703 * in: locked 3704 * out: unlocked 3705 */ 3706 void 3707 vput(struct vnode *vp) 3708 { 3709 3710 ASSERT_VOP_LOCKED(vp, __func__); 3711 ASSERT_VI_UNLOCKED(vp, __func__); 3712 if (!refcount_release(&vp->v_usecount)) { 3713 VOP_UNLOCK(vp); 3714 return; 3715 } 3716 vput_final(vp, VPUT); 3717 } 3718 3719 /* 3720 * in: locked 3721 * out: locked 3722 */ 3723 void 3724 vunref(struct vnode *vp) 3725 { 3726 3727 ASSERT_VOP_LOCKED(vp, __func__); 3728 ASSERT_VI_UNLOCKED(vp, __func__); 3729 if (!refcount_release(&vp->v_usecount)) 3730 return; 3731 vput_final(vp, VUNREF); 3732 } 3733 3734 void 3735 vhold(struct vnode *vp) 3736 { 3737 int old; 3738 3739 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3740 old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3741 VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3742 ("%s: wrong hold count %d", __func__, old)); 3743 if (old == 0) 3744 vfs_freevnodes_dec(); 3745 } 3746 3747 void 3748 vholdnz(struct vnode *vp) 3749 { 3750 3751 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3752 #ifdef INVARIANTS 3753 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3754 VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3755 ("%s: wrong hold count %d", __func__, old)); 3756 #else 3757 atomic_add_int(&vp->v_holdcnt, 1); 3758 #endif 3759 } 3760 3761 /* 3762 * Grab a hold count unless the vnode is freed. 3763 * 3764 * Only use this routine if vfs smr is the only protection you have against 3765 * freeing the vnode. 3766 * 3767 * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag 3768 * is not set. After the flag is set the vnode becomes immutable to anyone but 3769 * the thread which managed to set the flag. 3770 * 3771 * It may be tempting to replace the loop with: 3772 * count = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3773 * if (count & VHOLD_NO_SMR) { 3774 * backpedal and error out; 3775 * } 3776 * 3777 * However, while this is more performant, it hinders debugging by eliminating 3778 * the previously mentioned invariant. 3779 */ 3780 bool 3781 vhold_smr(struct vnode *vp) 3782 { 3783 int count; 3784 3785 VFS_SMR_ASSERT_ENTERED(); 3786 3787 count = atomic_load_int(&vp->v_holdcnt); 3788 for (;;) { 3789 if (count & VHOLD_NO_SMR) { 3790 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, 3791 ("non-zero hold count with flags %d\n", count)); 3792 return (false); 3793 } 3794 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); 3795 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) { 3796 if (count == 0) 3797 vfs_freevnodes_dec(); 3798 return (true); 3799 } 3800 } 3801 } 3802 3803 /* 3804 * Hold a free vnode for recycling. 3805 * 3806 * Note: vnode_init references this comment. 3807 * 3808 * Attempts to recycle only need the global vnode list lock and have no use for 3809 * SMR. 3810 * 3811 * However, vnodes get inserted into the global list before they get fully 3812 * initialized and stay there until UMA decides to free the memory. This in 3813 * particular means the target can be found before it becomes usable and after 3814 * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to 3815 * VHOLD_NO_SMR. 3816 * 3817 * Note: the vnode may gain more references after we transition the count 0->1. 3818 */ 3819 static bool 3820 vhold_recycle_free(struct vnode *vp) 3821 { 3822 int count; 3823 3824 mtx_assert(&vnode_list_mtx, MA_OWNED); 3825 3826 count = atomic_load_int(&vp->v_holdcnt); 3827 for (;;) { 3828 if (count & VHOLD_NO_SMR) { 3829 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, 3830 ("non-zero hold count with flags %d\n", count)); 3831 return (false); 3832 } 3833 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); 3834 if (count > 0) { 3835 return (false); 3836 } 3837 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) { 3838 vfs_freevnodes_dec(); 3839 return (true); 3840 } 3841 } 3842 } 3843 3844 static void __noinline 3845 vdbatch_process(struct vdbatch *vd) 3846 { 3847 struct vnode *vp; 3848 int i; 3849 3850 mtx_assert(&vd->lock, MA_OWNED); 3851 MPASS(curthread->td_pinned > 0); 3852 MPASS(vd->index == VDBATCH_SIZE); 3853 3854 /* 3855 * Attempt to requeue the passed batch, but give up easily. 3856 * 3857 * Despite batching the mechanism is prone to transient *significant* 3858 * lock contention, where vnode_list_mtx becomes the primary bottleneck 3859 * if multiple CPUs get here (one real-world example is highly parallel 3860 * do-nothing make , which will stat *tons* of vnodes). Since it is 3861 * quasi-LRU (read: not that great even if fully honoured) provide an 3862 * option to just dodge the problem. Parties which don't like it are 3863 * welcome to implement something better. 3864 */ 3865 if (vnode_can_skip_requeue) { 3866 if (!mtx_trylock(&vnode_list_mtx)) { 3867 counter_u64_add(vnode_skipped_requeues, 1); 3868 critical_enter(); 3869 for (i = 0; i < VDBATCH_SIZE; i++) { 3870 vp = vd->tab[i]; 3871 vd->tab[i] = NULL; 3872 MPASS(vp->v_dbatchcpu != NOCPU); 3873 vp->v_dbatchcpu = NOCPU; 3874 } 3875 vd->index = 0; 3876 critical_exit(); 3877 return; 3878 3879 } 3880 /* fallthrough to locked processing */ 3881 } else { 3882 mtx_lock(&vnode_list_mtx); 3883 } 3884 3885 mtx_assert(&vnode_list_mtx, MA_OWNED); 3886 critical_enter(); 3887 for (i = 0; i < VDBATCH_SIZE; i++) { 3888 vp = vd->tab[i]; 3889 vd->tab[i] = NULL; 3890 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 3891 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist); 3892 MPASS(vp->v_dbatchcpu != NOCPU); 3893 vp->v_dbatchcpu = NOCPU; 3894 } 3895 mtx_unlock(&vnode_list_mtx); 3896 vd->index = 0; 3897 critical_exit(); 3898 } 3899 3900 static void 3901 vdbatch_enqueue(struct vnode *vp) 3902 { 3903 struct vdbatch *vd; 3904 3905 ASSERT_VI_LOCKED(vp, __func__); 3906 VNPASS(!VN_IS_DOOMED(vp), vp); 3907 3908 if (vp->v_dbatchcpu != NOCPU) { 3909 VI_UNLOCK(vp); 3910 return; 3911 } 3912 3913 sched_pin(); 3914 vd = DPCPU_PTR(vd); 3915 mtx_lock(&vd->lock); 3916 MPASS(vd->index < VDBATCH_SIZE); 3917 MPASS(vd->tab[vd->index] == NULL); 3918 /* 3919 * A hack: we depend on being pinned so that we know what to put in 3920 * ->v_dbatchcpu. 3921 */ 3922 vp->v_dbatchcpu = curcpu; 3923 vd->tab[vd->index] = vp; 3924 vd->index++; 3925 VI_UNLOCK(vp); 3926 if (vd->index == VDBATCH_SIZE) 3927 vdbatch_process(vd); 3928 mtx_unlock(&vd->lock); 3929 sched_unpin(); 3930 } 3931 3932 /* 3933 * This routine must only be called for vnodes which are about to be 3934 * deallocated. Supporting dequeue for arbitrary vndoes would require 3935 * validating that the locked batch matches. 3936 */ 3937 static void 3938 vdbatch_dequeue(struct vnode *vp) 3939 { 3940 struct vdbatch *vd; 3941 int i; 3942 short cpu; 3943 3944 VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp); 3945 3946 cpu = vp->v_dbatchcpu; 3947 if (cpu == NOCPU) 3948 return; 3949 3950 vd = DPCPU_ID_PTR(cpu, vd); 3951 mtx_lock(&vd->lock); 3952 for (i = 0; i < vd->index; i++) { 3953 if (vd->tab[i] != vp) 3954 continue; 3955 vp->v_dbatchcpu = NOCPU; 3956 vd->index--; 3957 vd->tab[i] = vd->tab[vd->index]; 3958 vd->tab[vd->index] = NULL; 3959 break; 3960 } 3961 mtx_unlock(&vd->lock); 3962 /* 3963 * Either we dequeued the vnode above or the target CPU beat us to it. 3964 */ 3965 MPASS(vp->v_dbatchcpu == NOCPU); 3966 } 3967 3968 /* 3969 * Drop the hold count of the vnode. 3970 * 3971 * It will only get freed if this is the last hold *and* it has been vgone'd. 3972 * 3973 * Because the vnode vm object keeps a hold reference on the vnode if 3974 * there is at least one resident non-cached page, the vnode cannot 3975 * leave the active list without the page cleanup done. 3976 */ 3977 static void __noinline 3978 vdropl_final(struct vnode *vp) 3979 { 3980 3981 ASSERT_VI_LOCKED(vp, __func__); 3982 VNPASS(VN_IS_DOOMED(vp), vp); 3983 /* 3984 * Set the VHOLD_NO_SMR flag. 3985 * 3986 * We may be racing against vhold_smr. If they win we can just pretend 3987 * we never got this far, they will vdrop later. 3988 */ 3989 if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) { 3990 vfs_freevnodes_inc(); 3991 VI_UNLOCK(vp); 3992 /* 3993 * We lost the aforementioned race. Any subsequent access is 3994 * invalid as they might have managed to vdropl on their own. 3995 */ 3996 return; 3997 } 3998 /* 3999 * Don't bump freevnodes as this one is going away. 4000 */ 4001 freevnode(vp); 4002 } 4003 4004 void 4005 vdrop(struct vnode *vp) 4006 { 4007 4008 ASSERT_VI_UNLOCKED(vp, __func__); 4009 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4010 if (refcount_release_if_not_last(&vp->v_holdcnt)) 4011 return; 4012 VI_LOCK(vp); 4013 vdropl(vp); 4014 } 4015 4016 static __always_inline void 4017 vdropl_impl(struct vnode *vp, bool enqueue) 4018 { 4019 4020 ASSERT_VI_LOCKED(vp, __func__); 4021 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4022 if (!refcount_release(&vp->v_holdcnt)) { 4023 VI_UNLOCK(vp); 4024 return; 4025 } 4026 VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp); 4027 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp); 4028 if (VN_IS_DOOMED(vp)) { 4029 vdropl_final(vp); 4030 return; 4031 } 4032 4033 vfs_freevnodes_inc(); 4034 if (vp->v_mflag & VMP_LAZYLIST) { 4035 vunlazy(vp); 4036 } 4037 4038 if (!enqueue) { 4039 VI_UNLOCK(vp); 4040 return; 4041 } 4042 4043 /* 4044 * Also unlocks the interlock. We can't assert on it as we 4045 * released our hold and by now the vnode might have been 4046 * freed. 4047 */ 4048 vdbatch_enqueue(vp); 4049 } 4050 4051 void 4052 vdropl(struct vnode *vp) 4053 { 4054 4055 vdropl_impl(vp, true); 4056 } 4057 4058 /* 4059 * vdrop a vnode when recycling 4060 * 4061 * This is a special case routine only to be used when recycling, differs from 4062 * regular vdrop by not requeieing the vnode on LRU. 4063 * 4064 * Consider a case where vtryrecycle continuously fails with all vnodes (due to 4065 * e.g., frozen writes on the filesystem), filling the batch and causing it to 4066 * be requeued. Then vnlru will end up revisiting the same vnodes. This is a 4067 * loop which can last for as long as writes are frozen. 4068 */ 4069 static void 4070 vdropl_recycle(struct vnode *vp) 4071 { 4072 4073 vdropl_impl(vp, false); 4074 } 4075 4076 static void 4077 vdrop_recycle(struct vnode *vp) 4078 { 4079 4080 VI_LOCK(vp); 4081 vdropl_recycle(vp); 4082 } 4083 4084 /* 4085 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 4086 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 4087 */ 4088 static int 4089 vinactivef(struct vnode *vp) 4090 { 4091 int error; 4092 4093 ASSERT_VOP_ELOCKED(vp, "vinactive"); 4094 ASSERT_VI_LOCKED(vp, "vinactive"); 4095 VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp); 4096 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4097 vp->v_iflag |= VI_DOINGINACT; 4098 vp->v_iflag &= ~VI_OWEINACT; 4099 VI_UNLOCK(vp); 4100 4101 /* 4102 * Before moving off the active list, we must be sure that any 4103 * modified pages are converted into the vnode's dirty 4104 * buffers, since these will no longer be checked once the 4105 * vnode is on the inactive list. 4106 * 4107 * The write-out of the dirty pages is asynchronous. At the 4108 * point that VOP_INACTIVE() is called, there could still be 4109 * pending I/O and dirty pages in the object. 4110 */ 4111 if ((vp->v_vflag & VV_NOSYNC) == 0) 4112 vnode_pager_clean_async(vp); 4113 4114 error = VOP_INACTIVE(vp); 4115 VI_LOCK(vp); 4116 VNPASS(vp->v_iflag & VI_DOINGINACT, vp); 4117 vp->v_iflag &= ~VI_DOINGINACT; 4118 return (error); 4119 } 4120 4121 int 4122 vinactive(struct vnode *vp) 4123 { 4124 4125 ASSERT_VOP_ELOCKED(vp, "vinactive"); 4126 ASSERT_VI_LOCKED(vp, "vinactive"); 4127 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4128 4129 if ((vp->v_iflag & VI_OWEINACT) == 0) 4130 return (0); 4131 if (vp->v_iflag & VI_DOINGINACT) 4132 return (0); 4133 if (vp->v_usecount > 0) { 4134 vp->v_iflag &= ~VI_OWEINACT; 4135 return (0); 4136 } 4137 return (vinactivef(vp)); 4138 } 4139 4140 /* 4141 * Remove any vnodes in the vnode table belonging to mount point mp. 4142 * 4143 * If FORCECLOSE is not specified, there should not be any active ones, 4144 * return error if any are found (nb: this is a user error, not a 4145 * system error). If FORCECLOSE is specified, detach any active vnodes 4146 * that are found. 4147 * 4148 * If WRITECLOSE is set, only flush out regular file vnodes open for 4149 * writing. 4150 * 4151 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 4152 * 4153 * `rootrefs' specifies the base reference count for the root vnode 4154 * of this filesystem. The root vnode is considered busy if its 4155 * v_usecount exceeds this value. On a successful return, vflush(, td) 4156 * will call vrele() on the root vnode exactly rootrefs times. 4157 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 4158 * be zero. 4159 */ 4160 #ifdef DIAGNOSTIC 4161 static int busyprt = 0; /* print out busy vnodes */ 4162 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 4163 #endif 4164 4165 int 4166 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 4167 { 4168 struct vnode *vp, *mvp, *rootvp = NULL; 4169 struct vattr vattr; 4170 int busy = 0, error; 4171 4172 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 4173 rootrefs, flags); 4174 if (rootrefs > 0) { 4175 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 4176 ("vflush: bad args")); 4177 /* 4178 * Get the filesystem root vnode. We can vput() it 4179 * immediately, since with rootrefs > 0, it won't go away. 4180 */ 4181 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 4182 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 4183 __func__, error); 4184 return (error); 4185 } 4186 vput(rootvp); 4187 } 4188 loop: 4189 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 4190 vholdl(vp); 4191 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 4192 if (error) { 4193 vdrop(vp); 4194 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 4195 goto loop; 4196 } 4197 /* 4198 * Skip over a vnodes marked VV_SYSTEM. 4199 */ 4200 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 4201 VOP_UNLOCK(vp); 4202 vdrop(vp); 4203 continue; 4204 } 4205 /* 4206 * If WRITECLOSE is set, flush out unlinked but still open 4207 * files (even if open only for reading) and regular file 4208 * vnodes open for writing. 4209 */ 4210 if (flags & WRITECLOSE) { 4211 vnode_pager_clean_async(vp); 4212 do { 4213 error = VOP_FSYNC(vp, MNT_WAIT, td); 4214 } while (error == ERELOOKUP); 4215 if (error != 0) { 4216 VOP_UNLOCK(vp); 4217 vdrop(vp); 4218 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 4219 return (error); 4220 } 4221 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 4222 VI_LOCK(vp); 4223 4224 if ((vp->v_type == VNON || 4225 (error == 0 && vattr.va_nlink > 0)) && 4226 (vp->v_writecount <= 0 || vp->v_type != VREG)) { 4227 VOP_UNLOCK(vp); 4228 vdropl(vp); 4229 continue; 4230 } 4231 } else 4232 VI_LOCK(vp); 4233 /* 4234 * With v_usecount == 0, all we need to do is clear out the 4235 * vnode data structures and we are done. 4236 * 4237 * If FORCECLOSE is set, forcibly close the vnode. 4238 */ 4239 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 4240 vgonel(vp); 4241 } else { 4242 busy++; 4243 #ifdef DIAGNOSTIC 4244 if (busyprt) 4245 vn_printf(vp, "vflush: busy vnode "); 4246 #endif 4247 } 4248 VOP_UNLOCK(vp); 4249 vdropl(vp); 4250 } 4251 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 4252 /* 4253 * If just the root vnode is busy, and if its refcount 4254 * is equal to `rootrefs', then go ahead and kill it. 4255 */ 4256 VI_LOCK(rootvp); 4257 KASSERT(busy > 0, ("vflush: not busy")); 4258 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 4259 ("vflush: usecount %d < rootrefs %d", 4260 rootvp->v_usecount, rootrefs)); 4261 if (busy == 1 && rootvp->v_usecount == rootrefs) { 4262 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 4263 vgone(rootvp); 4264 VOP_UNLOCK(rootvp); 4265 busy = 0; 4266 } else 4267 VI_UNLOCK(rootvp); 4268 } 4269 if (busy) { 4270 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 4271 busy); 4272 return (EBUSY); 4273 } 4274 for (; rootrefs > 0; rootrefs--) 4275 vrele(rootvp); 4276 return (0); 4277 } 4278 4279 /* 4280 * Recycle an unused vnode. 4281 */ 4282 int 4283 vrecycle(struct vnode *vp) 4284 { 4285 int recycled; 4286 4287 VI_LOCK(vp); 4288 recycled = vrecyclel(vp); 4289 VI_UNLOCK(vp); 4290 return (recycled); 4291 } 4292 4293 /* 4294 * vrecycle, with the vp interlock held. 4295 */ 4296 int 4297 vrecyclel(struct vnode *vp) 4298 { 4299 int recycled; 4300 4301 ASSERT_VOP_ELOCKED(vp, __func__); 4302 ASSERT_VI_LOCKED(vp, __func__); 4303 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4304 recycled = 0; 4305 if (vp->v_usecount == 0) { 4306 recycled = 1; 4307 vgonel(vp); 4308 } 4309 return (recycled); 4310 } 4311 4312 /* 4313 * Eliminate all activity associated with a vnode 4314 * in preparation for reuse. 4315 */ 4316 void 4317 vgone(struct vnode *vp) 4318 { 4319 VI_LOCK(vp); 4320 vgonel(vp); 4321 VI_UNLOCK(vp); 4322 } 4323 4324 /* 4325 * Notify upper mounts about reclaimed or unlinked vnode. 4326 */ 4327 void 4328 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event) 4329 { 4330 struct mount *mp; 4331 struct mount_upper_node *ump; 4332 4333 mp = atomic_load_ptr(&vp->v_mount); 4334 if (mp == NULL) 4335 return; 4336 if (TAILQ_EMPTY(&mp->mnt_notify)) 4337 return; 4338 4339 MNT_ILOCK(mp); 4340 mp->mnt_upper_pending++; 4341 KASSERT(mp->mnt_upper_pending > 0, 4342 ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending)); 4343 TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) { 4344 MNT_IUNLOCK(mp); 4345 switch (event) { 4346 case VFS_NOTIFY_UPPER_RECLAIM: 4347 VFS_RECLAIM_LOWERVP(ump->mp, vp); 4348 break; 4349 case VFS_NOTIFY_UPPER_UNLINK: 4350 VFS_UNLINK_LOWERVP(ump->mp, vp); 4351 break; 4352 } 4353 MNT_ILOCK(mp); 4354 } 4355 mp->mnt_upper_pending--; 4356 if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 && 4357 mp->mnt_upper_pending == 0) { 4358 mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER; 4359 wakeup(&mp->mnt_uppers); 4360 } 4361 MNT_IUNLOCK(mp); 4362 } 4363 4364 /* 4365 * vgone, with the vp interlock held. 4366 */ 4367 static void 4368 vgonel(struct vnode *vp) 4369 { 4370 struct thread *td; 4371 struct mount *mp; 4372 vm_object_t object; 4373 bool active, doinginact, oweinact; 4374 4375 ASSERT_VOP_ELOCKED(vp, "vgonel"); 4376 ASSERT_VI_LOCKED(vp, "vgonel"); 4377 VNASSERT(vp->v_holdcnt, vp, 4378 ("vgonel: vp %p has no reference.", vp)); 4379 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4380 td = curthread; 4381 4382 /* 4383 * Don't vgonel if we're already doomed. 4384 */ 4385 if (VN_IS_DOOMED(vp)) { 4386 VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \ 4387 vn_get_state(vp) == VSTATE_DEAD, vp); 4388 return; 4389 } 4390 /* 4391 * Paired with freevnode. 4392 */ 4393 vn_seqc_write_begin_locked(vp); 4394 vunlazy_gone(vp); 4395 vn_irflag_set_locked(vp, VIRF_DOOMED); 4396 vn_set_state(vp, VSTATE_DESTROYING); 4397 4398 /* 4399 * Check to see if the vnode is in use. If so, we have to 4400 * call VOP_CLOSE() and VOP_INACTIVE(). 4401 * 4402 * It could be that VOP_INACTIVE() requested reclamation, in 4403 * which case we should avoid recursion, so check 4404 * VI_DOINGINACT. This is not precise but good enough. 4405 */ 4406 active = vp->v_usecount > 0; 4407 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 4408 doinginact = (vp->v_iflag & VI_DOINGINACT) != 0; 4409 4410 /* 4411 * If we need to do inactive VI_OWEINACT will be set. 4412 */ 4413 if (vp->v_iflag & VI_DEFINACT) { 4414 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 4415 vp->v_iflag &= ~VI_DEFINACT; 4416 vdropl(vp); 4417 } else { 4418 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); 4419 VI_UNLOCK(vp); 4420 } 4421 cache_purge_vgone(vp); 4422 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); 4423 4424 /* 4425 * If purging an active vnode, it must be closed and 4426 * deactivated before being reclaimed. 4427 */ 4428 if (active) 4429 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 4430 if (!doinginact) { 4431 do { 4432 if (oweinact || active) { 4433 VI_LOCK(vp); 4434 vinactivef(vp); 4435 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 4436 VI_UNLOCK(vp); 4437 } 4438 } while (oweinact); 4439 } 4440 if (vp->v_type == VSOCK) 4441 vfs_unp_reclaim(vp); 4442 4443 /* 4444 * Clean out any buffers associated with the vnode. 4445 * If the flush fails, just toss the buffers. 4446 */ 4447 mp = NULL; 4448 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 4449 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 4450 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { 4451 while (vinvalbuf(vp, 0, 0, 0) != 0) 4452 ; 4453 } 4454 4455 BO_LOCK(&vp->v_bufobj); 4456 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && 4457 vp->v_bufobj.bo_dirty.bv_cnt == 0 && 4458 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && 4459 vp->v_bufobj.bo_clean.bv_cnt == 0, 4460 ("vp %p bufobj not invalidated", vp)); 4461 4462 /* 4463 * For VMIO bufobj, BO_DEAD is set later, or in 4464 * vm_object_terminate() after the object's page queue is 4465 * flushed. 4466 */ 4467 object = vp->v_bufobj.bo_object; 4468 if (object == NULL) 4469 vp->v_bufobj.bo_flag |= BO_DEAD; 4470 BO_UNLOCK(&vp->v_bufobj); 4471 4472 /* 4473 * Handle the VM part. Tmpfs handles v_object on its own (the 4474 * OBJT_VNODE check). Nullfs or other bypassing filesystems 4475 * should not touch the object borrowed from the lower vnode 4476 * (the handle check). 4477 */ 4478 if (object != NULL && object->type == OBJT_VNODE && 4479 object->handle == vp) 4480 vnode_destroy_vobject(vp); 4481 4482 /* 4483 * Reclaim the vnode. 4484 */ 4485 if (VOP_RECLAIM(vp)) 4486 panic("vgone: cannot reclaim"); 4487 if (mp != NULL) 4488 vn_finished_secondary_write(mp); 4489 VNASSERT(vp->v_object == NULL, vp, 4490 ("vop_reclaim left v_object vp=%p", vp)); 4491 /* 4492 * Clear the advisory locks and wake up waiting threads. 4493 */ 4494 if (vp->v_lockf != NULL) { 4495 (void)VOP_ADVLOCKPURGE(vp); 4496 vp->v_lockf = NULL; 4497 } 4498 /* 4499 * Delete from old mount point vnode list. 4500 */ 4501 if (vp->v_mount == NULL) { 4502 VI_LOCK(vp); 4503 } else { 4504 delmntque(vp); 4505 ASSERT_VI_LOCKED(vp, "vgonel 2"); 4506 } 4507 /* 4508 * Done with purge, reset to the standard lock and invalidate 4509 * the vnode. 4510 */ 4511 vp->v_vnlock = &vp->v_lock; 4512 vp->v_op = &dead_vnodeops; 4513 vp->v_type = VBAD; 4514 vn_set_state(vp, VSTATE_DEAD); 4515 } 4516 4517 /* 4518 * Print out a description of a vnode. 4519 */ 4520 static const char *const vtypename[] = { 4521 [VNON] = "VNON", 4522 [VREG] = "VREG", 4523 [VDIR] = "VDIR", 4524 [VBLK] = "VBLK", 4525 [VCHR] = "VCHR", 4526 [VLNK] = "VLNK", 4527 [VSOCK] = "VSOCK", 4528 [VFIFO] = "VFIFO", 4529 [VBAD] = "VBAD", 4530 [VMARKER] = "VMARKER", 4531 }; 4532 _Static_assert(nitems(vtypename) == VLASTTYPE + 1, 4533 "vnode type name not added to vtypename"); 4534 4535 static const char *const vstatename[] = { 4536 [VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED", 4537 [VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED", 4538 [VSTATE_DESTROYING] = "VSTATE_DESTROYING", 4539 [VSTATE_DEAD] = "VSTATE_DEAD", 4540 }; 4541 _Static_assert(nitems(vstatename) == VLASTSTATE + 1, 4542 "vnode state name not added to vstatename"); 4543 4544 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0, 4545 "new hold count flag not added to vn_printf"); 4546 4547 void 4548 vn_printf(struct vnode *vp, const char *fmt, ...) 4549 { 4550 va_list ap; 4551 char buf[256], buf2[16]; 4552 u_long flags; 4553 u_int holdcnt; 4554 short irflag; 4555 4556 va_start(ap, fmt); 4557 vprintf(fmt, ap); 4558 va_end(ap); 4559 printf("%p: ", (void *)vp); 4560 printf("type %s state %s op %p\n", vtypename[vp->v_type], 4561 vstatename[vp->v_state], vp->v_op); 4562 holdcnt = atomic_load_int(&vp->v_holdcnt); 4563 printf(" usecount %d, writecount %d, refcount %d seqc users %d", 4564 vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS, 4565 vp->v_seqc_users); 4566 switch (vp->v_type) { 4567 case VDIR: 4568 printf(" mountedhere %p\n", vp->v_mountedhere); 4569 break; 4570 case VCHR: 4571 printf(" rdev %p\n", vp->v_rdev); 4572 break; 4573 case VSOCK: 4574 printf(" socket %p\n", vp->v_unpcb); 4575 break; 4576 case VFIFO: 4577 printf(" fifoinfo %p\n", vp->v_fifoinfo); 4578 break; 4579 default: 4580 printf("\n"); 4581 break; 4582 } 4583 buf[0] = '\0'; 4584 buf[1] = '\0'; 4585 if (holdcnt & VHOLD_NO_SMR) 4586 strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf)); 4587 printf(" hold count flags (%s)\n", buf + 1); 4588 4589 buf[0] = '\0'; 4590 buf[1] = '\0'; 4591 irflag = vn_irflag_read(vp); 4592 if (irflag & VIRF_DOOMED) 4593 strlcat(buf, "|VIRF_DOOMED", sizeof(buf)); 4594 if (irflag & VIRF_PGREAD) 4595 strlcat(buf, "|VIRF_PGREAD", sizeof(buf)); 4596 if (irflag & VIRF_MOUNTPOINT) 4597 strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf)); 4598 if (irflag & VIRF_TEXT_REF) 4599 strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf)); 4600 flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF); 4601 if (flags != 0) { 4602 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags); 4603 strlcat(buf, buf2, sizeof(buf)); 4604 } 4605 if (vp->v_vflag & VV_ROOT) 4606 strlcat(buf, "|VV_ROOT", sizeof(buf)); 4607 if (vp->v_vflag & VV_ISTTY) 4608 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 4609 if (vp->v_vflag & VV_NOSYNC) 4610 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 4611 if (vp->v_vflag & VV_ETERNALDEV) 4612 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 4613 if (vp->v_vflag & VV_CACHEDLABEL) 4614 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 4615 if (vp->v_vflag & VV_VMSIZEVNLOCK) 4616 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf)); 4617 if (vp->v_vflag & VV_COPYONWRITE) 4618 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 4619 if (vp->v_vflag & VV_SYSTEM) 4620 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 4621 if (vp->v_vflag & VV_PROCDEP) 4622 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 4623 if (vp->v_vflag & VV_DELETED) 4624 strlcat(buf, "|VV_DELETED", sizeof(buf)); 4625 if (vp->v_vflag & VV_MD) 4626 strlcat(buf, "|VV_MD", sizeof(buf)); 4627 if (vp->v_vflag & VV_FORCEINSMQ) 4628 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 4629 if (vp->v_vflag & VV_READLINK) 4630 strlcat(buf, "|VV_READLINK", sizeof(buf)); 4631 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 4632 VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM | 4633 VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK); 4634 if (flags != 0) { 4635 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 4636 strlcat(buf, buf2, sizeof(buf)); 4637 } 4638 if (vp->v_iflag & VI_MOUNT) 4639 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 4640 if (vp->v_iflag & VI_DOINGINACT) 4641 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 4642 if (vp->v_iflag & VI_OWEINACT) 4643 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 4644 if (vp->v_iflag & VI_DEFINACT) 4645 strlcat(buf, "|VI_DEFINACT", sizeof(buf)); 4646 if (vp->v_iflag & VI_FOPENING) 4647 strlcat(buf, "|VI_FOPENING", sizeof(buf)); 4648 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT | 4649 VI_OWEINACT | VI_DEFINACT | VI_FOPENING); 4650 if (flags != 0) { 4651 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 4652 strlcat(buf, buf2, sizeof(buf)); 4653 } 4654 if (vp->v_mflag & VMP_LAZYLIST) 4655 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf)); 4656 flags = vp->v_mflag & ~(VMP_LAZYLIST); 4657 if (flags != 0) { 4658 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags); 4659 strlcat(buf, buf2, sizeof(buf)); 4660 } 4661 printf(" flags (%s)", buf + 1); 4662 if (mtx_owned(VI_MTX(vp))) 4663 printf(" VI_LOCKed"); 4664 printf("\n"); 4665 if (vp->v_object != NULL) 4666 printf(" v_object %p ref %d pages %d " 4667 "cleanbuf %d dirtybuf %d\n", 4668 vp->v_object, vp->v_object->ref_count, 4669 vp->v_object->resident_page_count, 4670 vp->v_bufobj.bo_clean.bv_cnt, 4671 vp->v_bufobj.bo_dirty.bv_cnt); 4672 printf(" "); 4673 lockmgr_printinfo(vp->v_vnlock); 4674 if (vp->v_data != NULL) 4675 VOP_PRINT(vp); 4676 } 4677 4678 #ifdef DDB 4679 /* 4680 * List all of the locked vnodes in the system. 4681 * Called when debugging the kernel. 4682 */ 4683 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE) 4684 { 4685 struct mount *mp; 4686 struct vnode *vp; 4687 4688 /* 4689 * Note: because this is DDB, we can't obey the locking semantics 4690 * for these structures, which means we could catch an inconsistent 4691 * state and dereference a nasty pointer. Not much to be done 4692 * about that. 4693 */ 4694 db_printf("Locked vnodes\n"); 4695 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4696 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4697 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) 4698 vn_printf(vp, "vnode "); 4699 } 4700 } 4701 } 4702 4703 /* 4704 * Show details about the given vnode. 4705 */ 4706 DB_SHOW_COMMAND(vnode, db_show_vnode) 4707 { 4708 struct vnode *vp; 4709 4710 if (!have_addr) 4711 return; 4712 vp = (struct vnode *)addr; 4713 vn_printf(vp, "vnode "); 4714 } 4715 4716 /* 4717 * Show details about the given mount point. 4718 */ 4719 DB_SHOW_COMMAND(mount, db_show_mount) 4720 { 4721 struct mount *mp; 4722 struct vfsopt *opt; 4723 struct statfs *sp; 4724 struct vnode *vp; 4725 char buf[512]; 4726 uint64_t mflags; 4727 u_int flags; 4728 4729 if (!have_addr) { 4730 /* No address given, print short info about all mount points. */ 4731 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4732 db_printf("%p %s on %s (%s)\n", mp, 4733 mp->mnt_stat.f_mntfromname, 4734 mp->mnt_stat.f_mntonname, 4735 mp->mnt_stat.f_fstypename); 4736 if (db_pager_quit) 4737 break; 4738 } 4739 db_printf("\nMore info: show mount <addr>\n"); 4740 return; 4741 } 4742 4743 mp = (struct mount *)addr; 4744 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 4745 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 4746 4747 buf[0] = '\0'; 4748 mflags = mp->mnt_flag; 4749 #define MNT_FLAG(flag) do { \ 4750 if (mflags & (flag)) { \ 4751 if (buf[0] != '\0') \ 4752 strlcat(buf, ", ", sizeof(buf)); \ 4753 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 4754 mflags &= ~(flag); \ 4755 } \ 4756 } while (0) 4757 MNT_FLAG(MNT_RDONLY); 4758 MNT_FLAG(MNT_SYNCHRONOUS); 4759 MNT_FLAG(MNT_NOEXEC); 4760 MNT_FLAG(MNT_NOSUID); 4761 MNT_FLAG(MNT_NFS4ACLS); 4762 MNT_FLAG(MNT_UNION); 4763 MNT_FLAG(MNT_ASYNC); 4764 MNT_FLAG(MNT_SUIDDIR); 4765 MNT_FLAG(MNT_SOFTDEP); 4766 MNT_FLAG(MNT_NOSYMFOLLOW); 4767 MNT_FLAG(MNT_GJOURNAL); 4768 MNT_FLAG(MNT_MULTILABEL); 4769 MNT_FLAG(MNT_ACLS); 4770 MNT_FLAG(MNT_NOATIME); 4771 MNT_FLAG(MNT_NOCLUSTERR); 4772 MNT_FLAG(MNT_NOCLUSTERW); 4773 MNT_FLAG(MNT_SUJ); 4774 MNT_FLAG(MNT_EXRDONLY); 4775 MNT_FLAG(MNT_EXPORTED); 4776 MNT_FLAG(MNT_DEFEXPORTED); 4777 MNT_FLAG(MNT_EXPORTANON); 4778 MNT_FLAG(MNT_EXKERB); 4779 MNT_FLAG(MNT_EXPUBLIC); 4780 MNT_FLAG(MNT_LOCAL); 4781 MNT_FLAG(MNT_QUOTA); 4782 MNT_FLAG(MNT_ROOTFS); 4783 MNT_FLAG(MNT_USER); 4784 MNT_FLAG(MNT_IGNORE); 4785 MNT_FLAG(MNT_UPDATE); 4786 MNT_FLAG(MNT_DELEXPORT); 4787 MNT_FLAG(MNT_RELOAD); 4788 MNT_FLAG(MNT_FORCE); 4789 MNT_FLAG(MNT_SNAPSHOT); 4790 MNT_FLAG(MNT_BYFSID); 4791 MNT_FLAG(MNT_NAMEDATTR); 4792 #undef MNT_FLAG 4793 if (mflags != 0) { 4794 if (buf[0] != '\0') 4795 strlcat(buf, ", ", sizeof(buf)); 4796 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4797 "0x%016jx", mflags); 4798 } 4799 db_printf(" mnt_flag = %s\n", buf); 4800 4801 buf[0] = '\0'; 4802 flags = mp->mnt_kern_flag; 4803 #define MNT_KERN_FLAG(flag) do { \ 4804 if (flags & (flag)) { \ 4805 if (buf[0] != '\0') \ 4806 strlcat(buf, ", ", sizeof(buf)); \ 4807 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 4808 flags &= ~(flag); \ 4809 } \ 4810 } while (0) 4811 MNT_KERN_FLAG(MNTK_UNMOUNTF); 4812 MNT_KERN_FLAG(MNTK_ASYNC); 4813 MNT_KERN_FLAG(MNTK_SOFTDEP); 4814 MNT_KERN_FLAG(MNTK_NOMSYNC); 4815 MNT_KERN_FLAG(MNTK_DRAINING); 4816 MNT_KERN_FLAG(MNTK_REFEXPIRE); 4817 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 4818 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 4819 MNT_KERN_FLAG(MNTK_NO_IOPF); 4820 MNT_KERN_FLAG(MNTK_RECURSE); 4821 MNT_KERN_FLAG(MNTK_UPPER_WAITER); 4822 MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE); 4823 MNT_KERN_FLAG(MNTK_USES_BCACHE); 4824 MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG); 4825 MNT_KERN_FLAG(MNTK_FPLOOKUP); 4826 MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER); 4827 MNT_KERN_FLAG(MNTK_NOASYNC); 4828 MNT_KERN_FLAG(MNTK_UNMOUNT); 4829 MNT_KERN_FLAG(MNTK_MWAIT); 4830 MNT_KERN_FLAG(MNTK_SUSPEND); 4831 MNT_KERN_FLAG(MNTK_SUSPEND2); 4832 MNT_KERN_FLAG(MNTK_SUSPENDED); 4833 MNT_KERN_FLAG(MNTK_NULL_NOCACHE); 4834 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 4835 #undef MNT_KERN_FLAG 4836 if (flags != 0) { 4837 if (buf[0] != '\0') 4838 strlcat(buf, ", ", sizeof(buf)); 4839 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4840 "0x%08x", flags); 4841 } 4842 db_printf(" mnt_kern_flag = %s\n", buf); 4843 4844 db_printf(" mnt_opt = "); 4845 opt = TAILQ_FIRST(mp->mnt_opt); 4846 if (opt != NULL) { 4847 db_printf("%s", opt->name); 4848 opt = TAILQ_NEXT(opt, link); 4849 while (opt != NULL) { 4850 db_printf(", %s", opt->name); 4851 opt = TAILQ_NEXT(opt, link); 4852 } 4853 } 4854 db_printf("\n"); 4855 4856 sp = &mp->mnt_stat; 4857 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 4858 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 4859 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 4860 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 4861 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 4862 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 4863 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 4864 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 4865 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 4866 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 4867 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 4868 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 4869 4870 db_printf(" mnt_cred = { uid=%u ruid=%u", 4871 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 4872 if (jailed(mp->mnt_cred)) 4873 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 4874 db_printf(" }\n"); 4875 db_printf(" mnt_ref = %d (with %d in the struct)\n", 4876 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref); 4877 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 4878 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 4879 db_printf(" mnt_lazyvnodelistsize = %d\n", 4880 mp->mnt_lazyvnodelistsize); 4881 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n", 4882 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount); 4883 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 4884 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 4885 db_printf(" mnt_lockref = %d (with %d in the struct)\n", 4886 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref); 4887 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 4888 db_printf(" mnt_secondary_accwrites = %d\n", 4889 mp->mnt_secondary_accwrites); 4890 db_printf(" mnt_gjprovider = %s\n", 4891 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 4892 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops); 4893 4894 db_printf("\n\nList of active vnodes\n"); 4895 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4896 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) { 4897 vn_printf(vp, "vnode "); 4898 if (db_pager_quit) 4899 break; 4900 } 4901 } 4902 db_printf("\n\nList of inactive vnodes\n"); 4903 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4904 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) { 4905 vn_printf(vp, "vnode "); 4906 if (db_pager_quit) 4907 break; 4908 } 4909 } 4910 } 4911 #endif /* DDB */ 4912 4913 /* 4914 * Fill in a struct xvfsconf based on a struct vfsconf. 4915 */ 4916 static int 4917 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 4918 { 4919 struct xvfsconf xvfsp; 4920 4921 bzero(&xvfsp, sizeof(xvfsp)); 4922 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4923 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4924 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4925 xvfsp.vfc_flags = vfsp->vfc_flags; 4926 /* 4927 * These are unused in userland, we keep them 4928 * to not break binary compatibility. 4929 */ 4930 xvfsp.vfc_vfsops = NULL; 4931 xvfsp.vfc_next = NULL; 4932 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4933 } 4934 4935 #ifdef COMPAT_FREEBSD32 4936 struct xvfsconf32 { 4937 uint32_t vfc_vfsops; 4938 char vfc_name[MFSNAMELEN]; 4939 int32_t vfc_typenum; 4940 int32_t vfc_refcount; 4941 int32_t vfc_flags; 4942 uint32_t vfc_next; 4943 }; 4944 4945 static int 4946 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 4947 { 4948 struct xvfsconf32 xvfsp; 4949 4950 bzero(&xvfsp, sizeof(xvfsp)); 4951 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4952 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4953 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4954 xvfsp.vfc_flags = vfsp->vfc_flags; 4955 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4956 } 4957 #endif 4958 4959 /* 4960 * Top level filesystem related information gathering. 4961 */ 4962 static int 4963 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 4964 { 4965 struct vfsconf *vfsp; 4966 int error; 4967 4968 error = 0; 4969 vfsconf_slock(); 4970 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4971 #ifdef COMPAT_FREEBSD32 4972 if (req->flags & SCTL_MASK32) 4973 error = vfsconf2x32(req, vfsp); 4974 else 4975 #endif 4976 error = vfsconf2x(req, vfsp); 4977 if (error) 4978 break; 4979 } 4980 vfsconf_sunlock(); 4981 return (error); 4982 } 4983 4984 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | 4985 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, 4986 "S,xvfsconf", "List of all configured filesystems"); 4987 4988 #ifndef BURN_BRIDGES 4989 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 4990 4991 static int 4992 vfs_sysctl(SYSCTL_HANDLER_ARGS) 4993 { 4994 int *name = (int *)arg1 - 1; /* XXX */ 4995 u_int namelen = arg2 + 1; /* XXX */ 4996 struct vfsconf *vfsp; 4997 4998 log(LOG_WARNING, "userland calling deprecated sysctl, " 4999 "please rebuild world\n"); 5000 5001 #if 1 || defined(COMPAT_PRELITE2) 5002 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 5003 if (namelen == 1) 5004 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 5005 #endif 5006 5007 switch (name[1]) { 5008 case VFS_MAXTYPENUM: 5009 if (namelen != 2) 5010 return (ENOTDIR); 5011 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 5012 case VFS_CONF: 5013 if (namelen != 3) 5014 return (ENOTDIR); /* overloaded */ 5015 vfsconf_slock(); 5016 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 5017 if (vfsp->vfc_typenum == name[2]) 5018 break; 5019 } 5020 vfsconf_sunlock(); 5021 if (vfsp == NULL) 5022 return (EOPNOTSUPP); 5023 #ifdef COMPAT_FREEBSD32 5024 if (req->flags & SCTL_MASK32) 5025 return (vfsconf2x32(req, vfsp)); 5026 else 5027 #endif 5028 return (vfsconf2x(req, vfsp)); 5029 } 5030 return (EOPNOTSUPP); 5031 } 5032 5033 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | 5034 CTLFLAG_MPSAFE, vfs_sysctl, 5035 "Generic filesystem"); 5036 5037 #if 1 || defined(COMPAT_PRELITE2) 5038 5039 static int 5040 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 5041 { 5042 int error; 5043 struct vfsconf *vfsp; 5044 struct ovfsconf ovfs; 5045 5046 vfsconf_slock(); 5047 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 5048 bzero(&ovfs, sizeof(ovfs)); 5049 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 5050 strcpy(ovfs.vfc_name, vfsp->vfc_name); 5051 ovfs.vfc_index = vfsp->vfc_typenum; 5052 ovfs.vfc_refcount = vfsp->vfc_refcount; 5053 ovfs.vfc_flags = vfsp->vfc_flags; 5054 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 5055 if (error != 0) { 5056 vfsconf_sunlock(); 5057 return (error); 5058 } 5059 } 5060 vfsconf_sunlock(); 5061 return (0); 5062 } 5063 5064 #endif /* 1 || COMPAT_PRELITE2 */ 5065 #endif /* !BURN_BRIDGES */ 5066 5067 static void 5068 unmount_or_warn(struct mount *mp) 5069 { 5070 int error; 5071 5072 error = dounmount(mp, MNT_FORCE, curthread); 5073 if (error != 0) { 5074 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); 5075 if (error == EBUSY) 5076 printf("BUSY)\n"); 5077 else 5078 printf("%d)\n", error); 5079 } 5080 } 5081 5082 /* 5083 * Unmount all filesystems. The list is traversed in reverse order 5084 * of mounting to avoid dependencies. 5085 */ 5086 void 5087 vfs_unmountall(void) 5088 { 5089 struct mount *mp, *tmp; 5090 5091 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 5092 5093 /* 5094 * Since this only runs when rebooting, it is not interlocked. 5095 */ 5096 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { 5097 vfs_ref(mp); 5098 5099 /* 5100 * Forcibly unmounting "/dev" before "/" would prevent clean 5101 * unmount of the latter. 5102 */ 5103 if (mp == rootdevmp) 5104 continue; 5105 5106 unmount_or_warn(mp); 5107 } 5108 5109 if (rootdevmp != NULL) 5110 unmount_or_warn(rootdevmp); 5111 } 5112 5113 static void 5114 vfs_deferred_inactive(struct vnode *vp, int lkflags) 5115 { 5116 5117 ASSERT_VI_LOCKED(vp, __func__); 5118 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp); 5119 if ((vp->v_iflag & VI_OWEINACT) == 0) { 5120 vdropl(vp); 5121 return; 5122 } 5123 if (vn_lock(vp, lkflags) == 0) { 5124 VI_LOCK(vp); 5125 vinactive(vp); 5126 VOP_UNLOCK(vp); 5127 vdropl(vp); 5128 return; 5129 } 5130 vdefer_inactive_unlocked(vp); 5131 } 5132 5133 static int 5134 vfs_periodic_inactive_filter(struct vnode *vp, void *arg) 5135 { 5136 5137 return (vp->v_iflag & VI_DEFINACT); 5138 } 5139 5140 static void __noinline 5141 vfs_periodic_inactive(struct mount *mp, int flags) 5142 { 5143 struct vnode *vp, *mvp; 5144 int lkflags; 5145 5146 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 5147 if (flags != MNT_WAIT) 5148 lkflags |= LK_NOWAIT; 5149 5150 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) { 5151 if ((vp->v_iflag & VI_DEFINACT) == 0) { 5152 VI_UNLOCK(vp); 5153 continue; 5154 } 5155 vp->v_iflag &= ~VI_DEFINACT; 5156 vfs_deferred_inactive(vp, lkflags); 5157 } 5158 } 5159 5160 static inline bool 5161 vfs_want_msync(struct vnode *vp) 5162 { 5163 struct vm_object *obj; 5164 5165 /* 5166 * This test may be performed without any locks held. 5167 * We rely on vm_object's type stability. 5168 */ 5169 if (vp->v_vflag & VV_NOSYNC) 5170 return (false); 5171 obj = vp->v_object; 5172 return (obj != NULL && vm_object_mightbedirty(obj)); 5173 } 5174 5175 static int 5176 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused) 5177 { 5178 5179 if (vp->v_vflag & VV_NOSYNC) 5180 return (false); 5181 if (vp->v_iflag & VI_DEFINACT) 5182 return (true); 5183 return (vfs_want_msync(vp)); 5184 } 5185 5186 static void __noinline 5187 vfs_periodic_msync_inactive(struct mount *mp, int flags) 5188 { 5189 struct vnode *vp, *mvp; 5190 int lkflags; 5191 bool seen_defer; 5192 5193 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 5194 if (flags != MNT_WAIT) 5195 lkflags |= LK_NOWAIT; 5196 5197 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) { 5198 seen_defer = false; 5199 if (vp->v_iflag & VI_DEFINACT) { 5200 vp->v_iflag &= ~VI_DEFINACT; 5201 seen_defer = true; 5202 } 5203 if (!vfs_want_msync(vp)) { 5204 if (seen_defer) 5205 vfs_deferred_inactive(vp, lkflags); 5206 else 5207 VI_UNLOCK(vp); 5208 continue; 5209 } 5210 if (vget(vp, lkflags) == 0) { 5211 if ((vp->v_vflag & VV_NOSYNC) == 0) { 5212 if (flags == MNT_WAIT) 5213 vnode_pager_clean_sync(vp); 5214 else 5215 vnode_pager_clean_async(vp); 5216 } 5217 vput(vp); 5218 if (seen_defer) 5219 vdrop(vp); 5220 } else { 5221 if (seen_defer) 5222 vdefer_inactive_unlocked(vp); 5223 } 5224 } 5225 } 5226 5227 void 5228 vfs_periodic(struct mount *mp, int flags) 5229 { 5230 5231 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 5232 5233 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0) 5234 vfs_periodic_inactive(mp, flags); 5235 else 5236 vfs_periodic_msync_inactive(mp, flags); 5237 } 5238 5239 static void 5240 destroy_vpollinfo_free(struct vpollinfo *vi) 5241 { 5242 5243 knlist_destroy(&vi->vpi_selinfo.si_note); 5244 mtx_destroy(&vi->vpi_lock); 5245 free(vi, M_VNODEPOLL); 5246 } 5247 5248 static void 5249 destroy_vpollinfo(struct vpollinfo *vi) 5250 { 5251 KASSERT(TAILQ_EMPTY(&vi->vpi_inotify), 5252 ("%s: pollinfo %p has lingering watches", __func__, vi)); 5253 knlist_clear(&vi->vpi_selinfo.si_note, 1); 5254 seldrain(&vi->vpi_selinfo); 5255 destroy_vpollinfo_free(vi); 5256 } 5257 5258 /* 5259 * Initialize per-vnode helper structure to hold poll-related state. 5260 */ 5261 void 5262 v_addpollinfo(struct vnode *vp) 5263 { 5264 struct vpollinfo *vi; 5265 5266 if (atomic_load_ptr(&vp->v_pollinfo) != NULL) 5267 return; 5268 vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO); 5269 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 5270 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 5271 vfs_knlunlock, vfs_knl_assert_lock); 5272 TAILQ_INIT(&vi->vpi_inotify); 5273 VI_LOCK(vp); 5274 if (vp->v_pollinfo != NULL) { 5275 VI_UNLOCK(vp); 5276 destroy_vpollinfo_free(vi); 5277 return; 5278 } 5279 vp->v_pollinfo = vi; 5280 VI_UNLOCK(vp); 5281 } 5282 5283 /* 5284 * Record a process's interest in events which might happen to 5285 * a vnode. Because poll uses the historic select-style interface 5286 * internally, this routine serves as both the ``check for any 5287 * pending events'' and the ``record my interest in future events'' 5288 * functions. (These are done together, while the lock is held, 5289 * to avoid race conditions.) 5290 */ 5291 int 5292 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 5293 { 5294 5295 v_addpollinfo(vp); 5296 mtx_lock(&vp->v_pollinfo->vpi_lock); 5297 if (vp->v_pollinfo->vpi_revents & events) { 5298 /* 5299 * This leaves events we are not interested 5300 * in available for the other process which 5301 * which presumably had requested them 5302 * (otherwise they would never have been 5303 * recorded). 5304 */ 5305 events &= vp->v_pollinfo->vpi_revents; 5306 vp->v_pollinfo->vpi_revents &= ~events; 5307 5308 mtx_unlock(&vp->v_pollinfo->vpi_lock); 5309 return (events); 5310 } 5311 vp->v_pollinfo->vpi_events |= events; 5312 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 5313 mtx_unlock(&vp->v_pollinfo->vpi_lock); 5314 return (0); 5315 } 5316 5317 /* 5318 * Routine to create and manage a filesystem syncer vnode. 5319 */ 5320 #define sync_close ((int (*)(struct vop_close_args *))nullop) 5321 static int sync_fsync(struct vop_fsync_args *); 5322 static int sync_inactive(struct vop_inactive_args *); 5323 static int sync_reclaim(struct vop_reclaim_args *); 5324 5325 static struct vop_vector sync_vnodeops = { 5326 .vop_bypass = VOP_EOPNOTSUPP, 5327 .vop_close = sync_close, 5328 .vop_fsync = sync_fsync, 5329 .vop_getwritemount = vop_stdgetwritemount, 5330 .vop_inactive = sync_inactive, 5331 .vop_need_inactive = vop_stdneed_inactive, 5332 .vop_reclaim = sync_reclaim, 5333 .vop_lock1 = vop_stdlock, 5334 .vop_unlock = vop_stdunlock, 5335 .vop_islocked = vop_stdislocked, 5336 .vop_fplookup_vexec = VOP_EAGAIN, 5337 .vop_fplookup_symlink = VOP_EAGAIN, 5338 }; 5339 VFS_VOP_VECTOR_REGISTER(sync_vnodeops); 5340 5341 /* 5342 * Create a new filesystem syncer vnode for the specified mount point. 5343 */ 5344 void 5345 vfs_allocate_syncvnode(struct mount *mp) 5346 { 5347 struct vnode *vp; 5348 struct bufobj *bo; 5349 static long start, incr, next; 5350 int error; 5351 5352 /* Allocate a new vnode */ 5353 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 5354 if (error != 0) 5355 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 5356 vp->v_type = VNON; 5357 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5358 vp->v_vflag |= VV_FORCEINSMQ; 5359 error = insmntque1(vp, mp); 5360 if (error != 0) 5361 panic("vfs_allocate_syncvnode: insmntque() failed"); 5362 vp->v_vflag &= ~VV_FORCEINSMQ; 5363 vn_set_state(vp, VSTATE_CONSTRUCTED); 5364 VOP_UNLOCK(vp); 5365 /* 5366 * Place the vnode onto the syncer worklist. We attempt to 5367 * scatter them about on the list so that they will go off 5368 * at evenly distributed times even if all the filesystems 5369 * are mounted at once. 5370 */ 5371 next += incr; 5372 if (next == 0 || next > syncer_maxdelay) { 5373 start /= 2; 5374 incr /= 2; 5375 if (start == 0) { 5376 start = syncer_maxdelay / 2; 5377 incr = syncer_maxdelay; 5378 } 5379 next = start; 5380 } 5381 bo = &vp->v_bufobj; 5382 BO_LOCK(bo); 5383 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 5384 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 5385 mtx_lock(&sync_mtx); 5386 sync_vnode_count++; 5387 if (mp->mnt_syncer == NULL) { 5388 mp->mnt_syncer = vp; 5389 vp = NULL; 5390 } 5391 mtx_unlock(&sync_mtx); 5392 BO_UNLOCK(bo); 5393 if (vp != NULL) { 5394 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5395 vgone(vp); 5396 vput(vp); 5397 } 5398 } 5399 5400 void 5401 vfs_deallocate_syncvnode(struct mount *mp) 5402 { 5403 struct vnode *vp; 5404 5405 mtx_lock(&sync_mtx); 5406 vp = mp->mnt_syncer; 5407 if (vp != NULL) 5408 mp->mnt_syncer = NULL; 5409 mtx_unlock(&sync_mtx); 5410 if (vp != NULL) 5411 vrele(vp); 5412 } 5413 5414 /* 5415 * Do a lazy sync of the filesystem. 5416 */ 5417 static int 5418 sync_fsync(struct vop_fsync_args *ap) 5419 { 5420 struct vnode *syncvp = ap->a_vp; 5421 struct mount *mp = syncvp->v_mount; 5422 int error, save; 5423 struct bufobj *bo; 5424 5425 /* 5426 * We only need to do something if this is a lazy evaluation. 5427 */ 5428 if (ap->a_waitfor != MNT_LAZY) 5429 return (0); 5430 5431 /* 5432 * Move ourselves to the back of the sync list. 5433 */ 5434 bo = &syncvp->v_bufobj; 5435 BO_LOCK(bo); 5436 vn_syncer_add_to_worklist(bo, syncdelay); 5437 BO_UNLOCK(bo); 5438 5439 /* 5440 * Walk the list of vnodes pushing all that are dirty and 5441 * not already on the sync list. 5442 */ 5443 if (vfs_busy(mp, MBF_NOWAIT) != 0) 5444 return (0); 5445 VOP_UNLOCK(syncvp); 5446 save = curthread_pflags_set(TDP_SYNCIO); 5447 /* 5448 * The filesystem at hand may be idle with free vnodes stored in the 5449 * batch. Return them instead of letting them stay there indefinitely. 5450 */ 5451 vfs_periodic(mp, MNT_NOWAIT); 5452 error = VFS_SYNC(mp, MNT_LAZY); 5453 curthread_pflags_restore(save); 5454 vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY); 5455 vfs_unbusy(mp); 5456 return (error); 5457 } 5458 5459 /* 5460 * The syncer vnode is no referenced. 5461 */ 5462 static int 5463 sync_inactive(struct vop_inactive_args *ap) 5464 { 5465 5466 vgone(ap->a_vp); 5467 return (0); 5468 } 5469 5470 /* 5471 * The syncer vnode is no longer needed and is being decommissioned. 5472 * 5473 * Modifications to the worklist must be protected by sync_mtx. 5474 */ 5475 static int 5476 sync_reclaim(struct vop_reclaim_args *ap) 5477 { 5478 struct vnode *vp = ap->a_vp; 5479 struct bufobj *bo; 5480 5481 bo = &vp->v_bufobj; 5482 BO_LOCK(bo); 5483 mtx_lock(&sync_mtx); 5484 if (vp->v_mount->mnt_syncer == vp) 5485 vp->v_mount->mnt_syncer = NULL; 5486 if (bo->bo_flag & BO_ONWORKLST) { 5487 LIST_REMOVE(bo, bo_synclist); 5488 syncer_worklist_len--; 5489 sync_vnode_count--; 5490 bo->bo_flag &= ~BO_ONWORKLST; 5491 } 5492 mtx_unlock(&sync_mtx); 5493 BO_UNLOCK(bo); 5494 5495 return (0); 5496 } 5497 5498 int 5499 vn_need_pageq_flush(struct vnode *vp) 5500 { 5501 struct vm_object *obj; 5502 5503 obj = vp->v_object; 5504 return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 5505 vm_object_mightbedirty(obj)); 5506 } 5507 5508 /* 5509 * Check if vnode represents a disk device 5510 */ 5511 bool 5512 vn_isdisk_error(struct vnode *vp, int *errp) 5513 { 5514 int error; 5515 5516 if (vp->v_type != VCHR) { 5517 error = ENOTBLK; 5518 goto out; 5519 } 5520 error = 0; 5521 dev_lock(); 5522 if (vp->v_rdev == NULL) 5523 error = ENXIO; 5524 else if (vp->v_rdev->si_devsw == NULL) 5525 error = ENXIO; 5526 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 5527 error = ENOTBLK; 5528 dev_unlock(); 5529 out: 5530 *errp = error; 5531 return (error == 0); 5532 } 5533 5534 bool 5535 vn_isdisk(struct vnode *vp) 5536 { 5537 int error; 5538 5539 return (vn_isdisk_error(vp, &error)); 5540 } 5541 5542 /* 5543 * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see 5544 * the comment above cache_fplookup for details. 5545 */ 5546 int 5547 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred) 5548 { 5549 int error; 5550 5551 VFS_SMR_ASSERT_ENTERED(); 5552 5553 /* Check the owner. */ 5554 if (cred->cr_uid == file_uid) { 5555 if (file_mode & S_IXUSR) 5556 return (0); 5557 goto out_error; 5558 } 5559 5560 /* Otherwise, check the groups (first match) */ 5561 if (groupmember(file_gid, cred)) { 5562 if (file_mode & S_IXGRP) 5563 return (0); 5564 goto out_error; 5565 } 5566 5567 /* Otherwise, check everyone else. */ 5568 if (file_mode & S_IXOTH) 5569 return (0); 5570 out_error: 5571 /* 5572 * Permission check failed, but it is possible denial will get overwritten 5573 * (e.g., when root is traversing through a 700 directory owned by someone 5574 * else). 5575 * 5576 * vaccess() calls priv_check_cred which in turn can descent into MAC 5577 * modules overriding this result. It's quite unclear what semantics 5578 * are allowed for them to operate, thus for safety we don't call them 5579 * from within the SMR section. This also means if any such modules 5580 * are present, we have to let the regular lookup decide. 5581 */ 5582 error = priv_check_cred_vfs_lookup_nomac(cred); 5583 switch (error) { 5584 case 0: 5585 return (0); 5586 case EAGAIN: 5587 /* 5588 * MAC modules present. 5589 */ 5590 return (EAGAIN); 5591 case EPERM: 5592 return (EACCES); 5593 default: 5594 return (error); 5595 } 5596 } 5597 5598 /* 5599 * Common filesystem object access control check routine. Accepts a 5600 * vnode's type, "mode", uid and gid, requested access mode, and credentials. 5601 * Returns 0 on success, or an errno on failure. 5602 */ 5603 int 5604 vaccess(__enum_uint8(vtype) type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 5605 accmode_t accmode, struct ucred *cred) 5606 { 5607 accmode_t dac_granted; 5608 accmode_t priv_granted; 5609 5610 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 5611 ("invalid bit in accmode")); 5612 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 5613 ("VAPPEND without VWRITE")); 5614 5615 /* 5616 * Look for a normal, non-privileged way to access the file/directory 5617 * as requested. If it exists, go with that. 5618 */ 5619 5620 dac_granted = 0; 5621 5622 /* Check the owner. */ 5623 if (cred->cr_uid == file_uid) { 5624 dac_granted |= VADMIN; 5625 if (file_mode & S_IXUSR) 5626 dac_granted |= VEXEC; 5627 if (file_mode & S_IRUSR) 5628 dac_granted |= VREAD; 5629 if (file_mode & S_IWUSR) 5630 dac_granted |= (VWRITE | VAPPEND); 5631 5632 if ((accmode & dac_granted) == accmode) 5633 return (0); 5634 5635 goto privcheck; 5636 } 5637 5638 /* Otherwise, check the groups (first match) */ 5639 if (groupmember(file_gid, cred)) { 5640 if (file_mode & S_IXGRP) 5641 dac_granted |= VEXEC; 5642 if (file_mode & S_IRGRP) 5643 dac_granted |= VREAD; 5644 if (file_mode & S_IWGRP) 5645 dac_granted |= (VWRITE | VAPPEND); 5646 5647 if ((accmode & dac_granted) == accmode) 5648 return (0); 5649 5650 goto privcheck; 5651 } 5652 5653 /* Otherwise, check everyone else. */ 5654 if (file_mode & S_IXOTH) 5655 dac_granted |= VEXEC; 5656 if (file_mode & S_IROTH) 5657 dac_granted |= VREAD; 5658 if (file_mode & S_IWOTH) 5659 dac_granted |= (VWRITE | VAPPEND); 5660 if ((accmode & dac_granted) == accmode) 5661 return (0); 5662 5663 privcheck: 5664 /* 5665 * Build a privilege mask to determine if the set of privileges 5666 * satisfies the requirements when combined with the granted mask 5667 * from above. For each privilege, if the privilege is required, 5668 * bitwise or the request type onto the priv_granted mask. 5669 */ 5670 priv_granted = 0; 5671 5672 if (type == VDIR) { 5673 /* 5674 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 5675 * requests, instead of PRIV_VFS_EXEC. 5676 */ 5677 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5678 !priv_check_cred(cred, PRIV_VFS_LOOKUP)) 5679 priv_granted |= VEXEC; 5680 } else { 5681 /* 5682 * Ensure that at least one execute bit is on. Otherwise, 5683 * a privileged user will always succeed, and we don't want 5684 * this to happen unless the file really is executable. 5685 */ 5686 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5687 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 5688 !priv_check_cred(cred, PRIV_VFS_EXEC)) 5689 priv_granted |= VEXEC; 5690 } 5691 5692 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 5693 !priv_check_cred(cred, PRIV_VFS_READ)) 5694 priv_granted |= VREAD; 5695 5696 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 5697 !priv_check_cred(cred, PRIV_VFS_WRITE)) 5698 priv_granted |= (VWRITE | VAPPEND); 5699 5700 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 5701 !priv_check_cred(cred, PRIV_VFS_ADMIN)) 5702 priv_granted |= VADMIN; 5703 5704 if ((accmode & (priv_granted | dac_granted)) == accmode) { 5705 return (0); 5706 } 5707 5708 return ((accmode & VADMIN) ? EPERM : EACCES); 5709 } 5710 5711 /* 5712 * Credential check based on process requesting service, and per-attribute 5713 * permissions. 5714 */ 5715 int 5716 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 5717 struct thread *td, accmode_t accmode) 5718 { 5719 5720 /* 5721 * Kernel-invoked always succeeds. 5722 */ 5723 if (cred == NOCRED) 5724 return (0); 5725 5726 /* 5727 * Do not allow privileged processes in jail to directly manipulate 5728 * system attributes. 5729 */ 5730 switch (attrnamespace) { 5731 case EXTATTR_NAMESPACE_SYSTEM: 5732 /* Potentially should be: return (EPERM); */ 5733 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM)); 5734 case EXTATTR_NAMESPACE_USER: 5735 return (VOP_ACCESS(vp, accmode, cred, td)); 5736 default: 5737 return (EPERM); 5738 } 5739 } 5740 5741 #ifdef INVARIANTS 5742 void 5743 assert_vi_locked(struct vnode *vp, const char *str) 5744 { 5745 VNASSERT(mtx_owned(VI_MTX(vp)), vp, 5746 ("%s: vnode interlock is not locked but should be", str)); 5747 } 5748 5749 void 5750 assert_vi_unlocked(struct vnode *vp, const char *str) 5751 { 5752 VNASSERT(!mtx_owned(VI_MTX(vp)), vp, 5753 ("%s: vnode interlock is locked but should not be", str)); 5754 } 5755 5756 void 5757 assert_vop_locked(struct vnode *vp, const char *str) 5758 { 5759 bool locked; 5760 5761 if (KERNEL_PANICKED() || vp == NULL) 5762 return; 5763 5764 #ifdef WITNESS 5765 locked = !((vp->v_irflag & VIRF_CROSSMP) == 0 && 5766 witness_is_owned(&vp->v_vnlock->lock_object) == -1); 5767 #else 5768 int state = VOP_ISLOCKED(vp); 5769 locked = state != 0 && state != LK_EXCLOTHER; 5770 #endif 5771 VNASSERT(locked, vp, ("%s: vnode is not locked but should be", str)); 5772 } 5773 5774 void 5775 assert_vop_unlocked(struct vnode *vp, const char *str) 5776 { 5777 bool locked; 5778 5779 if (KERNEL_PANICKED() || vp == NULL) 5780 return; 5781 5782 #ifdef WITNESS 5783 locked = (vp->v_irflag & VIRF_CROSSMP) == 0 && 5784 witness_is_owned(&vp->v_vnlock->lock_object) == 1; 5785 #else 5786 locked = VOP_ISLOCKED(vp) == LK_EXCLUSIVE; 5787 #endif 5788 VNASSERT(!locked, vp, ("%s: vnode is locked but should not be", str)); 5789 } 5790 5791 void 5792 assert_vop_elocked(struct vnode *vp, const char *str) 5793 { 5794 bool locked; 5795 5796 if (KERNEL_PANICKED() || vp == NULL) 5797 return; 5798 5799 locked = VOP_ISLOCKED(vp) == LK_EXCLUSIVE; 5800 VNASSERT(locked, vp, 5801 ("%s: vnode is not exclusive locked but should be", str)); 5802 } 5803 #endif /* INVARIANTS */ 5804 5805 void 5806 vop_rename_fail(struct vop_rename_args *ap) 5807 { 5808 5809 if (ap->a_tvp != NULL) 5810 vput(ap->a_tvp); 5811 if (ap->a_tdvp == ap->a_tvp) 5812 vrele(ap->a_tdvp); 5813 else 5814 vput(ap->a_tdvp); 5815 vrele(ap->a_fdvp); 5816 vrele(ap->a_fvp); 5817 } 5818 5819 void 5820 vop_rename_pre(void *ap) 5821 { 5822 struct vop_rename_args *a = ap; 5823 5824 #ifdef INVARIANTS 5825 struct mount *tmp; 5826 5827 if (a->a_tvp) 5828 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 5829 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 5830 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 5831 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 5832 5833 /* Check the source (from). */ 5834 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 5835 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 5836 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 5837 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 5838 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 5839 5840 /* Check the target. */ 5841 if (a->a_tvp) 5842 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 5843 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 5844 5845 tmp = NULL; 5846 VOP_GETWRITEMOUNT(a->a_tdvp, &tmp); 5847 lockmgr_assert(&tmp->mnt_renamelock, KA_XLOCKED); 5848 vfs_rel(tmp); 5849 #endif 5850 /* 5851 * It may be tempting to add vn_seqc_write_begin/end calls here and 5852 * in vop_rename_post but that's not going to work out since some 5853 * filesystems relookup vnodes mid-rename. This is probably a bug. 5854 * 5855 * For now filesystems are expected to do the relevant calls after they 5856 * decide what vnodes to operate on. 5857 */ 5858 if (a->a_tdvp != a->a_fdvp) 5859 vhold(a->a_fdvp); 5860 if (a->a_tvp != a->a_fvp) 5861 vhold(a->a_fvp); 5862 vhold(a->a_tdvp); 5863 if (a->a_tvp) 5864 vhold(a->a_tvp); 5865 } 5866 5867 #ifdef INVARIANTS 5868 void 5869 vop_fplookup_vexec_debugpre(void *ap __unused) 5870 { 5871 5872 VFS_SMR_ASSERT_ENTERED(); 5873 } 5874 5875 void 5876 vop_fplookup_vexec_debugpost(void *ap, int rc) 5877 { 5878 struct vop_fplookup_vexec_args *a; 5879 struct vnode *vp; 5880 5881 a = ap; 5882 vp = a->a_vp; 5883 5884 VFS_SMR_ASSERT_ENTERED(); 5885 if (rc == EOPNOTSUPP) 5886 VNPASS(VN_IS_DOOMED(vp), vp); 5887 } 5888 5889 void 5890 vop_fplookup_symlink_debugpre(void *ap __unused) 5891 { 5892 5893 VFS_SMR_ASSERT_ENTERED(); 5894 } 5895 5896 void 5897 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused) 5898 { 5899 5900 VFS_SMR_ASSERT_ENTERED(); 5901 } 5902 5903 static void 5904 vop_fsync_debugprepost(struct vnode *vp, const char *name) 5905 { 5906 struct mount *mp; 5907 5908 if (vp->v_type == VCHR) 5909 ; 5910 /* 5911 * The shared vs. exclusive locking policy for fsync() 5912 * is actually determined by vp's write mount as indicated 5913 * by VOP_GETWRITEMOUNT(), which for stacked filesystems 5914 * may not be the same as vp->v_mount. However, if the 5915 * underlying filesystem which really handles the fsync() 5916 * supports shared locking, the stacked filesystem must also 5917 * be prepared for its VOP_FSYNC() operation to be called 5918 * with only a shared lock. On the other hand, if the 5919 * stacked filesystem claims support for shared write 5920 * locking but the underlying filesystem does not, and the 5921 * caller incorrectly uses a shared lock, this condition 5922 * should still be caught when the stacked filesystem 5923 * invokes VOP_FSYNC() on the underlying filesystem. 5924 */ 5925 else { 5926 mp = NULL; 5927 VOP_GETWRITEMOUNT(vp, &mp); 5928 if (vn_lktype_write(mp, vp) == LK_SHARED) 5929 ASSERT_VOP_LOCKED(vp, name); 5930 else 5931 ASSERT_VOP_ELOCKED(vp, name); 5932 if (mp != NULL) 5933 vfs_rel(mp); 5934 } 5935 } 5936 5937 void 5938 vop_fsync_debugpre(void *a) 5939 { 5940 struct vop_fsync_args *ap; 5941 5942 ap = a; 5943 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5944 } 5945 5946 void 5947 vop_fsync_debugpost(void *a, int rc __unused) 5948 { 5949 struct vop_fsync_args *ap; 5950 5951 ap = a; 5952 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5953 } 5954 5955 void 5956 vop_fdatasync_debugpre(void *a) 5957 { 5958 struct vop_fdatasync_args *ap; 5959 5960 ap = a; 5961 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5962 } 5963 5964 void 5965 vop_fdatasync_debugpost(void *a, int rc __unused) 5966 { 5967 struct vop_fdatasync_args *ap; 5968 5969 ap = a; 5970 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5971 } 5972 5973 void 5974 vop_strategy_debugpre(void *ap) 5975 { 5976 struct vop_strategy_args *a; 5977 struct buf *bp; 5978 5979 a = ap; 5980 bp = a->a_bp; 5981 5982 /* 5983 * Cluster ops lock their component buffers but not the IO container. 5984 */ 5985 if ((bp->b_flags & B_CLUSTER) != 0) 5986 return; 5987 5988 BUF_ASSERT_LOCKED(bp); 5989 } 5990 5991 void 5992 vop_lock_debugpre(void *ap) 5993 { 5994 struct vop_lock1_args *a = ap; 5995 5996 if ((a->a_flags & LK_INTERLOCK) == 0) 5997 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5998 else 5999 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 6000 } 6001 6002 void 6003 vop_lock_debugpost(void *ap, int rc) 6004 { 6005 struct vop_lock1_args *a = ap; 6006 6007 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 6008 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) 6009 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 6010 } 6011 6012 void 6013 vop_unlock_debugpre(void *ap) 6014 { 6015 struct vop_unlock_args *a = ap; 6016 struct vnode *vp = a->a_vp; 6017 6018 VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp); 6019 ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK"); 6020 } 6021 6022 void 6023 vop_need_inactive_debugpre(void *ap) 6024 { 6025 struct vop_need_inactive_args *a = ap; 6026 6027 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 6028 } 6029 6030 void 6031 vop_need_inactive_debugpost(void *ap, int rc) 6032 { 6033 struct vop_need_inactive_args *a = ap; 6034 6035 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 6036 } 6037 #endif /* INVARIANTS */ 6038 6039 void 6040 vop_allocate_post(void *ap, int rc) 6041 { 6042 struct vop_allocate_args *a; 6043 6044 a = ap; 6045 if (rc == 0) 6046 INOTIFY(a->a_vp, IN_MODIFY); 6047 } 6048 6049 void 6050 vop_copy_file_range_post(void *ap, int rc) 6051 { 6052 struct vop_copy_file_range_args *a; 6053 6054 a = ap; 6055 if (rc == 0) { 6056 INOTIFY(a->a_invp, IN_ACCESS); 6057 INOTIFY(a->a_outvp, IN_MODIFY); 6058 } 6059 } 6060 6061 void 6062 vop_create_pre(void *ap) 6063 { 6064 struct vop_create_args *a; 6065 struct vnode *dvp; 6066 6067 a = ap; 6068 dvp = a->a_dvp; 6069 vn_seqc_write_begin(dvp); 6070 } 6071 6072 void 6073 vop_create_post(void *ap, int rc) 6074 { 6075 struct vop_create_args *a; 6076 struct vnode *dvp; 6077 6078 a = ap; 6079 dvp = a->a_dvp; 6080 vn_seqc_write_end(dvp); 6081 if (!rc) { 6082 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6083 INOTIFY_NAME(*a->a_vpp, dvp, a->a_cnp, IN_CREATE); 6084 } 6085 } 6086 6087 void 6088 vop_deallocate_post(void *ap, int rc) 6089 { 6090 struct vop_deallocate_args *a; 6091 6092 a = ap; 6093 if (rc == 0) 6094 INOTIFY(a->a_vp, IN_MODIFY); 6095 } 6096 6097 void 6098 vop_whiteout_pre(void *ap) 6099 { 6100 struct vop_whiteout_args *a; 6101 struct vnode *dvp; 6102 6103 a = ap; 6104 dvp = a->a_dvp; 6105 vn_seqc_write_begin(dvp); 6106 } 6107 6108 void 6109 vop_whiteout_post(void *ap, int rc) 6110 { 6111 struct vop_whiteout_args *a; 6112 struct vnode *dvp; 6113 6114 a = ap; 6115 dvp = a->a_dvp; 6116 vn_seqc_write_end(dvp); 6117 } 6118 6119 void 6120 vop_deleteextattr_pre(void *ap) 6121 { 6122 struct vop_deleteextattr_args *a; 6123 struct vnode *vp; 6124 6125 a = ap; 6126 vp = a->a_vp; 6127 vn_seqc_write_begin(vp); 6128 } 6129 6130 void 6131 vop_deleteextattr_post(void *ap, int rc) 6132 { 6133 struct vop_deleteextattr_args *a; 6134 struct vnode *vp; 6135 6136 a = ap; 6137 vp = a->a_vp; 6138 vn_seqc_write_end(vp); 6139 if (!rc) { 6140 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 6141 INOTIFY(vp, IN_ATTRIB); 6142 } 6143 } 6144 6145 void 6146 vop_link_pre(void *ap) 6147 { 6148 struct vop_link_args *a; 6149 struct vnode *vp, *tdvp; 6150 6151 a = ap; 6152 vp = a->a_vp; 6153 tdvp = a->a_tdvp; 6154 vn_seqc_write_begin(vp); 6155 vn_seqc_write_begin(tdvp); 6156 } 6157 6158 void 6159 vop_link_post(void *ap, int rc) 6160 { 6161 struct vop_link_args *a; 6162 struct vnode *vp, *tdvp; 6163 6164 a = ap; 6165 vp = a->a_vp; 6166 tdvp = a->a_tdvp; 6167 vn_seqc_write_end(vp); 6168 vn_seqc_write_end(tdvp); 6169 if (!rc) { 6170 VFS_KNOTE_LOCKED(vp, NOTE_LINK); 6171 VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE); 6172 INOTIFY_NAME(vp, tdvp, a->a_cnp, _IN_ATTRIB_LINKCOUNT); 6173 INOTIFY_NAME(vp, tdvp, a->a_cnp, IN_CREATE); 6174 } 6175 } 6176 6177 void 6178 vop_mkdir_pre(void *ap) 6179 { 6180 struct vop_mkdir_args *a; 6181 struct vnode *dvp; 6182 6183 a = ap; 6184 dvp = a->a_dvp; 6185 vn_seqc_write_begin(dvp); 6186 } 6187 6188 void 6189 vop_mkdir_post(void *ap, int rc) 6190 { 6191 struct vop_mkdir_args *a; 6192 struct vnode *dvp; 6193 6194 a = ap; 6195 dvp = a->a_dvp; 6196 vn_seqc_write_end(dvp); 6197 if (!rc) { 6198 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 6199 INOTIFY_NAME(*a->a_vpp, dvp, a->a_cnp, IN_CREATE); 6200 } 6201 } 6202 6203 #ifdef INVARIANTS 6204 void 6205 vop_mkdir_debugpost(void *ap, int rc) 6206 { 6207 struct vop_mkdir_args *a; 6208 6209 a = ap; 6210 if (!rc) 6211 cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp); 6212 } 6213 #endif 6214 6215 void 6216 vop_mknod_pre(void *ap) 6217 { 6218 struct vop_mknod_args *a; 6219 struct vnode *dvp; 6220 6221 a = ap; 6222 dvp = a->a_dvp; 6223 vn_seqc_write_begin(dvp); 6224 } 6225 6226 void 6227 vop_mknod_post(void *ap, int rc) 6228 { 6229 struct vop_mknod_args *a; 6230 struct vnode *dvp; 6231 6232 a = ap; 6233 dvp = a->a_dvp; 6234 vn_seqc_write_end(dvp); 6235 if (!rc) { 6236 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6237 INOTIFY_NAME(*a->a_vpp, dvp, a->a_cnp, IN_CREATE); 6238 } 6239 } 6240 6241 void 6242 vop_reclaim_post(void *ap, int rc) 6243 { 6244 struct vop_reclaim_args *a; 6245 struct vnode *vp; 6246 6247 a = ap; 6248 vp = a->a_vp; 6249 ASSERT_VOP_IN_SEQC(vp); 6250 if (!rc) { 6251 VFS_KNOTE_LOCKED(vp, NOTE_REVOKE); 6252 INOTIFY_REVOKE(vp); 6253 } 6254 } 6255 6256 void 6257 vop_remove_pre(void *ap) 6258 { 6259 struct vop_remove_args *a; 6260 struct vnode *dvp, *vp; 6261 6262 a = ap; 6263 dvp = a->a_dvp; 6264 vp = a->a_vp; 6265 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_UNLINK); 6266 vn_seqc_write_begin(dvp); 6267 vn_seqc_write_begin(vp); 6268 } 6269 6270 void 6271 vop_remove_post(void *ap, int rc) 6272 { 6273 struct vop_remove_args *a; 6274 struct vnode *dvp, *vp; 6275 6276 a = ap; 6277 dvp = a->a_dvp; 6278 vp = a->a_vp; 6279 vn_seqc_write_end(dvp); 6280 vn_seqc_write_end(vp); 6281 if (!rc) { 6282 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6283 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 6284 INOTIFY_NAME(vp, dvp, a->a_cnp, _IN_ATTRIB_LINKCOUNT); 6285 INOTIFY_NAME(vp, dvp, a->a_cnp, IN_DELETE); 6286 } 6287 } 6288 6289 void 6290 vop_rename_post(void *ap, int rc) 6291 { 6292 struct vop_rename_args *a = ap; 6293 long hint; 6294 6295 if (!rc) { 6296 hint = NOTE_WRITE; 6297 if (a->a_fdvp == a->a_tdvp) { 6298 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) 6299 hint |= NOTE_LINK; 6300 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 6301 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 6302 } else { 6303 hint |= NOTE_EXTEND; 6304 if (a->a_fvp->v_type == VDIR) 6305 hint |= NOTE_LINK; 6306 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 6307 6308 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && 6309 a->a_tvp->v_type == VDIR) 6310 hint &= ~NOTE_LINK; 6311 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 6312 } 6313 6314 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 6315 if (a->a_tvp) 6316 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 6317 INOTIFY_MOVE(a->a_fvp, a->a_fdvp, a->a_fcnp, a->a_tvp, 6318 a->a_tdvp, a->a_tcnp); 6319 } 6320 if (a->a_tdvp != a->a_fdvp) 6321 vdrop(a->a_fdvp); 6322 if (a->a_tvp != a->a_fvp) 6323 vdrop(a->a_fvp); 6324 vdrop(a->a_tdvp); 6325 if (a->a_tvp) 6326 vdrop(a->a_tvp); 6327 } 6328 6329 void 6330 vop_rmdir_pre(void *ap) 6331 { 6332 struct vop_rmdir_args *a; 6333 struct vnode *dvp, *vp; 6334 6335 a = ap; 6336 dvp = a->a_dvp; 6337 vp = a->a_vp; 6338 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_UNLINK); 6339 vn_seqc_write_begin(dvp); 6340 vn_seqc_write_begin(vp); 6341 } 6342 6343 void 6344 vop_rmdir_post(void *ap, int rc) 6345 { 6346 struct vop_rmdir_args *a; 6347 struct vnode *dvp, *vp; 6348 6349 a = ap; 6350 dvp = a->a_dvp; 6351 vp = a->a_vp; 6352 vn_seqc_write_end(dvp); 6353 vn_seqc_write_end(vp); 6354 if (!rc) { 6355 vp->v_vflag |= VV_UNLINKED; 6356 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 6357 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 6358 INOTIFY_NAME(vp, dvp, a->a_cnp, IN_DELETE); 6359 } 6360 } 6361 6362 void 6363 vop_setattr_pre(void *ap) 6364 { 6365 struct vop_setattr_args *a; 6366 struct vnode *vp; 6367 6368 a = ap; 6369 vp = a->a_vp; 6370 vn_seqc_write_begin(vp); 6371 } 6372 6373 void 6374 vop_setattr_post(void *ap, int rc) 6375 { 6376 struct vop_setattr_args *a; 6377 struct vnode *vp; 6378 6379 a = ap; 6380 vp = a->a_vp; 6381 vn_seqc_write_end(vp); 6382 if (!rc) { 6383 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 6384 INOTIFY(vp, IN_ATTRIB); 6385 } 6386 } 6387 6388 void 6389 vop_setacl_pre(void *ap) 6390 { 6391 struct vop_setacl_args *a; 6392 struct vnode *vp; 6393 6394 a = ap; 6395 vp = a->a_vp; 6396 vn_seqc_write_begin(vp); 6397 } 6398 6399 void 6400 vop_setacl_post(void *ap, int rc __unused) 6401 { 6402 struct vop_setacl_args *a; 6403 struct vnode *vp; 6404 6405 a = ap; 6406 vp = a->a_vp; 6407 vn_seqc_write_end(vp); 6408 } 6409 6410 void 6411 vop_setextattr_pre(void *ap) 6412 { 6413 struct vop_setextattr_args *a; 6414 struct vnode *vp; 6415 6416 a = ap; 6417 vp = a->a_vp; 6418 vn_seqc_write_begin(vp); 6419 } 6420 6421 void 6422 vop_setextattr_post(void *ap, int rc) 6423 { 6424 struct vop_setextattr_args *a; 6425 struct vnode *vp; 6426 6427 a = ap; 6428 vp = a->a_vp; 6429 vn_seqc_write_end(vp); 6430 if (!rc) { 6431 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 6432 INOTIFY(vp, IN_ATTRIB); 6433 } 6434 } 6435 6436 void 6437 vop_symlink_pre(void *ap) 6438 { 6439 struct vop_symlink_args *a; 6440 struct vnode *dvp; 6441 6442 a = ap; 6443 dvp = a->a_dvp; 6444 vn_seqc_write_begin(dvp); 6445 } 6446 6447 void 6448 vop_symlink_post(void *ap, int rc) 6449 { 6450 struct vop_symlink_args *a; 6451 struct vnode *dvp; 6452 6453 a = ap; 6454 dvp = a->a_dvp; 6455 vn_seqc_write_end(dvp); 6456 if (!rc) { 6457 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6458 INOTIFY_NAME(*a->a_vpp, dvp, a->a_cnp, IN_CREATE); 6459 } 6460 } 6461 6462 void 6463 vop_open_post(void *ap, int rc) 6464 { 6465 struct vop_open_args *a = ap; 6466 6467 if (!rc) { 6468 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); 6469 INOTIFY(a->a_vp, IN_OPEN); 6470 } 6471 } 6472 6473 void 6474 vop_close_post(void *ap, int rc) 6475 { 6476 struct vop_close_args *a = ap; 6477 6478 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ 6479 !VN_IS_DOOMED(a->a_vp))) { 6480 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 6481 NOTE_CLOSE_WRITE : NOTE_CLOSE); 6482 INOTIFY(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 6483 IN_CLOSE_WRITE : IN_CLOSE_NOWRITE); 6484 } 6485 } 6486 6487 void 6488 vop_read_post(void *ap, int rc) 6489 { 6490 struct vop_read_args *a = ap; 6491 6492 if (!rc) { 6493 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 6494 INOTIFY(a->a_vp, IN_ACCESS); 6495 } 6496 } 6497 6498 void 6499 vop_read_pgcache_post(void *ap, int rc) 6500 { 6501 struct vop_read_pgcache_args *a = ap; 6502 6503 if (!rc) 6504 VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ); 6505 } 6506 6507 static struct knlist fs_knlist; 6508 6509 static void 6510 vfs_event_init(void *arg) 6511 { 6512 knlist_init_mtx(&fs_knlist, NULL); 6513 } 6514 /* XXX - correct order? */ 6515 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 6516 6517 void 6518 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 6519 { 6520 6521 KNOTE_UNLOCKED(&fs_knlist, event); 6522 } 6523 6524 static int filt_fsattach(struct knote *kn); 6525 static void filt_fsdetach(struct knote *kn); 6526 static int filt_fsevent(struct knote *kn, long hint); 6527 6528 const struct filterops fs_filtops = { 6529 .f_isfd = 0, 6530 .f_attach = filt_fsattach, 6531 .f_detach = filt_fsdetach, 6532 .f_event = filt_fsevent, 6533 }; 6534 6535 static int 6536 filt_fsattach(struct knote *kn) 6537 { 6538 6539 kn->kn_flags |= EV_CLEAR; 6540 knlist_add(&fs_knlist, kn, 0); 6541 return (0); 6542 } 6543 6544 static void 6545 filt_fsdetach(struct knote *kn) 6546 { 6547 6548 knlist_remove(&fs_knlist, kn, 0); 6549 } 6550 6551 static int 6552 filt_fsevent(struct knote *kn, long hint) 6553 { 6554 6555 kn->kn_fflags |= kn->kn_sfflags & hint; 6556 6557 return (kn->kn_fflags != 0); 6558 } 6559 6560 static int 6561 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 6562 { 6563 struct vfsidctl vc; 6564 int error; 6565 struct mount *mp; 6566 6567 if (req->newptr == NULL) 6568 return (EINVAL); 6569 error = SYSCTL_IN(req, &vc, sizeof(vc)); 6570 if (error) 6571 return (error); 6572 if (vc.vc_vers != VFS_CTL_VERS1) 6573 return (EINVAL); 6574 mp = vfs_getvfs(&vc.vc_fsid); 6575 if (mp == NULL) 6576 return (ENOENT); 6577 /* ensure that a specific sysctl goes to the right filesystem. */ 6578 if (strcmp(vc.vc_fstypename, "*") != 0 && 6579 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 6580 vfs_rel(mp); 6581 return (EINVAL); 6582 } 6583 VCTLTOREQ(&vc, req); 6584 error = VFS_SYSCTL(mp, vc.vc_op, req); 6585 vfs_rel(mp); 6586 return (error); 6587 } 6588 6589 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR, 6590 NULL, 0, sysctl_vfs_ctl, "", 6591 "Sysctl by fsid"); 6592 6593 /* 6594 * Function to initialize a va_filerev field sensibly. 6595 * XXX: Wouldn't a random number make a lot more sense ?? 6596 */ 6597 u_quad_t 6598 init_va_filerev(void) 6599 { 6600 struct bintime bt; 6601 6602 getbinuptime(&bt); 6603 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 6604 } 6605 6606 static int filt_vfsread(struct knote *kn, long hint); 6607 static int filt_vfswrite(struct knote *kn, long hint); 6608 static int filt_vfsvnode(struct knote *kn, long hint); 6609 static void filt_vfsdetach(struct knote *kn); 6610 static int filt_vfsdump(struct proc *p, struct knote *kn, 6611 struct kinfo_knote *kin); 6612 6613 static const struct filterops vfsread_filtops = { 6614 .f_isfd = 1, 6615 .f_detach = filt_vfsdetach, 6616 .f_event = filt_vfsread, 6617 .f_userdump = filt_vfsdump, 6618 }; 6619 static const struct filterops vfswrite_filtops = { 6620 .f_isfd = 1, 6621 .f_detach = filt_vfsdetach, 6622 .f_event = filt_vfswrite, 6623 .f_userdump = filt_vfsdump, 6624 }; 6625 static const struct filterops vfsvnode_filtops = { 6626 .f_isfd = 1, 6627 .f_detach = filt_vfsdetach, 6628 .f_event = filt_vfsvnode, 6629 .f_userdump = filt_vfsdump, 6630 }; 6631 6632 static void 6633 vfs_knllock(void *arg) 6634 { 6635 struct vnode *vp = arg; 6636 6637 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 6638 } 6639 6640 static void 6641 vfs_knlunlock(void *arg) 6642 { 6643 struct vnode *vp = arg; 6644 6645 VOP_UNLOCK(vp); 6646 } 6647 6648 static void 6649 vfs_knl_assert_lock(void *arg, int what) 6650 { 6651 #ifdef INVARIANTS 6652 struct vnode *vp = arg; 6653 6654 if (what == LA_LOCKED) 6655 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 6656 else 6657 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 6658 #endif 6659 } 6660 6661 int 6662 vfs_kqfilter(struct vop_kqfilter_args *ap) 6663 { 6664 struct vnode *vp = ap->a_vp; 6665 struct knote *kn = ap->a_kn; 6666 struct knlist *knl; 6667 6668 KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ && 6669 kn->kn_filter != EVFILT_WRITE), 6670 ("READ/WRITE filter on a FIFO leaked through")); 6671 switch (kn->kn_filter) { 6672 case EVFILT_READ: 6673 kn->kn_fop = &vfsread_filtops; 6674 break; 6675 case EVFILT_WRITE: 6676 kn->kn_fop = &vfswrite_filtops; 6677 break; 6678 case EVFILT_VNODE: 6679 kn->kn_fop = &vfsvnode_filtops; 6680 break; 6681 default: 6682 return (EINVAL); 6683 } 6684 6685 kn->kn_hook = (caddr_t)vp; 6686 6687 v_addpollinfo(vp); 6688 if (vp->v_pollinfo == NULL) 6689 return (ENOMEM); 6690 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 6691 vhold(vp); 6692 knlist_add(knl, kn, 0); 6693 6694 return (0); 6695 } 6696 6697 /* 6698 * Detach knote from vnode 6699 */ 6700 static void 6701 filt_vfsdetach(struct knote *kn) 6702 { 6703 struct vnode *vp = (struct vnode *)kn->kn_hook; 6704 6705 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 6706 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 6707 vdrop(vp); 6708 } 6709 6710 /*ARGSUSED*/ 6711 static int 6712 filt_vfsread(struct knote *kn, long hint) 6713 { 6714 struct vnode *vp = (struct vnode *)kn->kn_hook; 6715 off_t size; 6716 int res; 6717 6718 /* 6719 * filesystem is gone, so set the EOF flag and schedule 6720 * the knote for deletion. 6721 */ 6722 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6723 VI_LOCK(vp); 6724 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6725 VI_UNLOCK(vp); 6726 return (1); 6727 } 6728 6729 if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0) 6730 return (0); 6731 6732 VI_LOCK(vp); 6733 kn->kn_data = size - kn->kn_fp->f_offset; 6734 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; 6735 VI_UNLOCK(vp); 6736 return (res); 6737 } 6738 6739 /*ARGSUSED*/ 6740 static int 6741 filt_vfswrite(struct knote *kn, long hint) 6742 { 6743 struct vnode *vp = (struct vnode *)kn->kn_hook; 6744 6745 VI_LOCK(vp); 6746 6747 /* 6748 * filesystem is gone, so set the EOF flag and schedule 6749 * the knote for deletion. 6750 */ 6751 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) 6752 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6753 6754 kn->kn_data = 0; 6755 VI_UNLOCK(vp); 6756 return (1); 6757 } 6758 6759 static int 6760 filt_vfsvnode(struct knote *kn, long hint) 6761 { 6762 struct vnode *vp = (struct vnode *)kn->kn_hook; 6763 int res; 6764 6765 VI_LOCK(vp); 6766 if (kn->kn_sfflags & hint) 6767 kn->kn_fflags |= hint; 6768 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6769 kn->kn_flags |= EV_EOF; 6770 VI_UNLOCK(vp); 6771 return (1); 6772 } 6773 res = (kn->kn_fflags != 0); 6774 VI_UNLOCK(vp); 6775 return (res); 6776 } 6777 6778 static int 6779 filt_vfsdump(struct proc *p, struct knote *kn, struct kinfo_knote *kin) 6780 { 6781 struct vattr va; 6782 struct vnode *vp; 6783 char *fullpath, *freepath; 6784 int error; 6785 6786 kin->knt_extdata = KNOTE_EXTDATA_VNODE; 6787 6788 vp = kn->kn_fp->f_vnode; 6789 kin->knt_vnode.knt_vnode_type = vntype_to_kinfo(vp->v_type); 6790 6791 va.va_fsid = VNOVAL; 6792 vn_lock(vp, LK_SHARED | LK_RETRY); 6793 error = VOP_GETATTR(vp, &va, curthread->td_ucred); 6794 VOP_UNLOCK(vp); 6795 if (error != 0) 6796 return (error); 6797 kin->knt_vnode.knt_vnode_fsid = va.va_fsid; 6798 kin->knt_vnode.knt_vnode_fileid = va.va_fileid; 6799 6800 freepath = NULL; 6801 fullpath = "-"; 6802 error = vn_fullpath(vp, &fullpath, &freepath); 6803 if (error == 0) { 6804 strlcpy(kin->knt_vnode.knt_vnode_fullpath, fullpath, 6805 sizeof(kin->knt_vnode.knt_vnode_fullpath)); 6806 } 6807 if (freepath != NULL) 6808 free(freepath, M_TEMP); 6809 6810 return (0); 6811 } 6812 6813 int 6814 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 6815 { 6816 int error; 6817 6818 if (dp->d_reclen > ap->a_uio->uio_resid) 6819 return (ENAMETOOLONG); 6820 error = uiomove(dp, dp->d_reclen, ap->a_uio); 6821 if (error) { 6822 if (ap->a_ncookies != NULL) { 6823 if (ap->a_cookies != NULL) 6824 free(ap->a_cookies, M_TEMP); 6825 ap->a_cookies = NULL; 6826 *ap->a_ncookies = 0; 6827 } 6828 return (error); 6829 } 6830 if (ap->a_ncookies == NULL) 6831 return (0); 6832 6833 KASSERT(ap->a_cookies, 6834 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 6835 6836 *ap->a_cookies = realloc(*ap->a_cookies, 6837 (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO); 6838 (*ap->a_cookies)[*ap->a_ncookies] = off; 6839 *ap->a_ncookies += 1; 6840 return (0); 6841 } 6842 6843 /* 6844 * The purpose of this routine is to remove granularity from accmode_t, 6845 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 6846 * VADMIN and VAPPEND. 6847 * 6848 * If it returns 0, the caller is supposed to continue with the usual 6849 * access checks using 'accmode' as modified by this routine. If it 6850 * returns nonzero value, the caller is supposed to return that value 6851 * as errno. 6852 * 6853 * Note that after this routine runs, accmode may be zero. 6854 */ 6855 int 6856 vfs_unixify_accmode(accmode_t *accmode) 6857 { 6858 /* 6859 * There is no way to specify explicit "deny" rule using 6860 * file mode or POSIX.1e ACLs. 6861 */ 6862 if (*accmode & VEXPLICIT_DENY) { 6863 *accmode = 0; 6864 return (0); 6865 } 6866 6867 /* 6868 * None of these can be translated into usual access bits. 6869 * Also, the common case for NFSv4 ACLs is to not contain 6870 * either of these bits. Caller should check for VWRITE 6871 * on the containing directory instead. 6872 */ 6873 if (*accmode & (VDELETE_CHILD | VDELETE)) 6874 return (EPERM); 6875 6876 if (*accmode & VADMIN_PERMS) { 6877 *accmode &= ~VADMIN_PERMS; 6878 *accmode |= VADMIN; 6879 } 6880 6881 /* 6882 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 6883 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 6884 */ 6885 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 6886 6887 return (0); 6888 } 6889 6890 /* 6891 * Clear out a doomed vnode (if any) and replace it with a new one as long 6892 * as the fs is not being unmounted. Return the root vnode to the caller. 6893 */ 6894 static int __noinline 6895 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp) 6896 { 6897 struct vnode *vp; 6898 int error; 6899 6900 restart: 6901 if (mp->mnt_rootvnode != NULL) { 6902 MNT_ILOCK(mp); 6903 vp = mp->mnt_rootvnode; 6904 if (vp != NULL) { 6905 if (!VN_IS_DOOMED(vp)) { 6906 vrefact(vp); 6907 MNT_IUNLOCK(mp); 6908 error = vn_lock(vp, flags); 6909 if (error == 0) { 6910 *vpp = vp; 6911 return (0); 6912 } 6913 vrele(vp); 6914 goto restart; 6915 } 6916 /* 6917 * Clear the old one. 6918 */ 6919 mp->mnt_rootvnode = NULL; 6920 } 6921 MNT_IUNLOCK(mp); 6922 if (vp != NULL) { 6923 vfs_op_barrier_wait(mp); 6924 vrele(vp); 6925 } 6926 } 6927 error = VFS_CACHEDROOT(mp, flags, vpp); 6928 if (error != 0) 6929 return (error); 6930 if (mp->mnt_vfs_ops == 0) { 6931 MNT_ILOCK(mp); 6932 if (mp->mnt_vfs_ops != 0) { 6933 MNT_IUNLOCK(mp); 6934 return (0); 6935 } 6936 if (mp->mnt_rootvnode == NULL) { 6937 vrefact(*vpp); 6938 mp->mnt_rootvnode = *vpp; 6939 } else { 6940 if (mp->mnt_rootvnode != *vpp) { 6941 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) { 6942 panic("%s: mismatch between vnode returned " 6943 " by VFS_CACHEDROOT and the one cached " 6944 " (%p != %p)", 6945 __func__, *vpp, mp->mnt_rootvnode); 6946 } 6947 } 6948 } 6949 MNT_IUNLOCK(mp); 6950 } 6951 return (0); 6952 } 6953 6954 int 6955 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp) 6956 { 6957 struct mount_pcpu *mpcpu; 6958 struct vnode *vp; 6959 int error; 6960 6961 if (!vfs_op_thread_enter(mp, mpcpu)) 6962 return (vfs_cache_root_fallback(mp, flags, vpp)); 6963 vp = atomic_load_ptr(&mp->mnt_rootvnode); 6964 if (vp == NULL || VN_IS_DOOMED(vp)) { 6965 vfs_op_thread_exit(mp, mpcpu); 6966 return (vfs_cache_root_fallback(mp, flags, vpp)); 6967 } 6968 vrefact(vp); 6969 vfs_op_thread_exit(mp, mpcpu); 6970 error = vn_lock(vp, flags); 6971 if (error != 0) { 6972 vrele(vp); 6973 return (vfs_cache_root_fallback(mp, flags, vpp)); 6974 } 6975 *vpp = vp; 6976 return (0); 6977 } 6978 6979 struct vnode * 6980 vfs_cache_root_clear(struct mount *mp) 6981 { 6982 struct vnode *vp; 6983 6984 /* 6985 * ops > 0 guarantees there is nobody who can see this vnode 6986 */ 6987 MPASS(mp->mnt_vfs_ops > 0); 6988 vp = mp->mnt_rootvnode; 6989 if (vp != NULL) 6990 vn_seqc_write_begin(vp); 6991 mp->mnt_rootvnode = NULL; 6992 return (vp); 6993 } 6994 6995 void 6996 vfs_cache_root_set(struct mount *mp, struct vnode *vp) 6997 { 6998 6999 MPASS(mp->mnt_vfs_ops > 0); 7000 vrefact(vp); 7001 mp->mnt_rootvnode = vp; 7002 } 7003 7004 /* 7005 * These are helper functions for filesystems to traverse all 7006 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 7007 * 7008 * This interface replaces MNT_VNODE_FOREACH. 7009 */ 7010 7011 struct vnode * 7012 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 7013 { 7014 struct vnode *vp; 7015 7016 maybe_yield(); 7017 MNT_ILOCK(mp); 7018 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7019 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL; 7020 vp = TAILQ_NEXT(vp, v_nmntvnodes)) { 7021 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 7022 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 7023 continue; 7024 VI_LOCK(vp); 7025 if (VN_IS_DOOMED(vp)) { 7026 VI_UNLOCK(vp); 7027 continue; 7028 } 7029 break; 7030 } 7031 if (vp == NULL) { 7032 __mnt_vnode_markerfree_all(mvp, mp); 7033 /* MNT_IUNLOCK(mp); -- done in above function */ 7034 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 7035 return (NULL); 7036 } 7037 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 7038 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 7039 MNT_IUNLOCK(mp); 7040 return (vp); 7041 } 7042 7043 struct vnode * 7044 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 7045 { 7046 struct vnode *vp; 7047 7048 *mvp = vn_alloc_marker(mp); 7049 MNT_ILOCK(mp); 7050 MNT_REF(mp); 7051 7052 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 7053 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 7054 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 7055 continue; 7056 VI_LOCK(vp); 7057 if (VN_IS_DOOMED(vp)) { 7058 VI_UNLOCK(vp); 7059 continue; 7060 } 7061 break; 7062 } 7063 if (vp == NULL) { 7064 MNT_REL(mp); 7065 MNT_IUNLOCK(mp); 7066 vn_free_marker(*mvp); 7067 *mvp = NULL; 7068 return (NULL); 7069 } 7070 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 7071 MNT_IUNLOCK(mp); 7072 return (vp); 7073 } 7074 7075 void 7076 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 7077 { 7078 7079 if (*mvp == NULL) { 7080 MNT_IUNLOCK(mp); 7081 return; 7082 } 7083 7084 mtx_assert(MNT_MTX(mp), MA_OWNED); 7085 7086 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7087 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 7088 MNT_REL(mp); 7089 MNT_IUNLOCK(mp); 7090 vn_free_marker(*mvp); 7091 *mvp = NULL; 7092 } 7093 7094 /* 7095 * These are helper functions for filesystems to traverse their 7096 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h 7097 */ 7098 static void 7099 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 7100 { 7101 7102 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7103 7104 MNT_ILOCK(mp); 7105 MNT_REL(mp); 7106 MNT_IUNLOCK(mp); 7107 vn_free_marker(*mvp); 7108 *mvp = NULL; 7109 } 7110 7111 /* 7112 * Relock the mp mount vnode list lock with the vp vnode interlock in the 7113 * conventional lock order during mnt_vnode_next_lazy iteration. 7114 * 7115 * On entry, the mount vnode list lock is held and the vnode interlock is not. 7116 * The list lock is dropped and reacquired. On success, both locks are held. 7117 * On failure, the mount vnode list lock is held but the vnode interlock is 7118 * not, and the procedure may have yielded. 7119 */ 7120 static bool 7121 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp, 7122 struct vnode *vp) 7123 { 7124 7125 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER && 7126 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp, 7127 ("%s: bad marker", __func__)); 7128 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp, 7129 ("%s: inappropriate vnode", __func__)); 7130 ASSERT_VI_UNLOCKED(vp, __func__); 7131 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 7132 7133 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist); 7134 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist); 7135 7136 /* 7137 * Note we may be racing against vdrop which transitioned the hold 7138 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine, 7139 * if we are the only user after we get the interlock we will just 7140 * vdrop. 7141 */ 7142 vhold(vp); 7143 mtx_unlock(&mp->mnt_listmtx); 7144 VI_LOCK(vp); 7145 if (VN_IS_DOOMED(vp)) { 7146 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); 7147 goto out_lost; 7148 } 7149 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 7150 /* 7151 * There is nothing to do if we are the last user. 7152 */ 7153 if (!refcount_release_if_not_last(&vp->v_holdcnt)) 7154 goto out_lost; 7155 mtx_lock(&mp->mnt_listmtx); 7156 return (true); 7157 out_lost: 7158 vdropl(vp); 7159 maybe_yield(); 7160 mtx_lock(&mp->mnt_listmtx); 7161 return (false); 7162 } 7163 7164 static struct vnode * 7165 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 7166 void *cbarg) 7167 { 7168 struct vnode *vp; 7169 7170 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 7171 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7172 restart: 7173 vp = TAILQ_NEXT(*mvp, v_lazylist); 7174 while (vp != NULL) { 7175 if (vp->v_type == VMARKER) { 7176 vp = TAILQ_NEXT(vp, v_lazylist); 7177 continue; 7178 } 7179 /* 7180 * See if we want to process the vnode. Note we may encounter a 7181 * long string of vnodes we don't care about and hog the list 7182 * as a result. Check for it and requeue the marker. 7183 */ 7184 VNPASS(!VN_IS_DOOMED(vp), vp); 7185 if (!cb(vp, cbarg)) { 7186 if (!should_yield()) { 7187 vp = TAILQ_NEXT(vp, v_lazylist); 7188 continue; 7189 } 7190 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, 7191 v_lazylist); 7192 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, 7193 v_lazylist); 7194 mtx_unlock(&mp->mnt_listmtx); 7195 kern_yield(PRI_USER); 7196 mtx_lock(&mp->mnt_listmtx); 7197 goto restart; 7198 } 7199 /* 7200 * Try-lock because this is the wrong lock order. 7201 */ 7202 if (!VI_TRYLOCK(vp) && 7203 !mnt_vnode_next_lazy_relock(*mvp, mp, vp)) 7204 goto restart; 7205 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 7206 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 7207 ("alien vnode on the lazy list %p %p", vp, mp)); 7208 VNPASS(vp->v_mount == mp, vp); 7209 VNPASS(!VN_IS_DOOMED(vp), vp); 7210 break; 7211 } 7212 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 7213 7214 /* Check if we are done */ 7215 if (vp == NULL) { 7216 mtx_unlock(&mp->mnt_listmtx); 7217 mnt_vnode_markerfree_lazy(mvp, mp); 7218 return (NULL); 7219 } 7220 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist); 7221 mtx_unlock(&mp->mnt_listmtx); 7222 ASSERT_VI_LOCKED(vp, "lazy iter"); 7223 return (vp); 7224 } 7225 7226 struct vnode * 7227 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 7228 void *cbarg) 7229 { 7230 7231 maybe_yield(); 7232 mtx_lock(&mp->mnt_listmtx); 7233 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 7234 } 7235 7236 struct vnode * 7237 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 7238 void *cbarg) 7239 { 7240 struct vnode *vp; 7241 7242 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist)) 7243 return (NULL); 7244 7245 *mvp = vn_alloc_marker(mp); 7246 MNT_ILOCK(mp); 7247 MNT_REF(mp); 7248 MNT_IUNLOCK(mp); 7249 7250 mtx_lock(&mp->mnt_listmtx); 7251 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist); 7252 if (vp == NULL) { 7253 mtx_unlock(&mp->mnt_listmtx); 7254 mnt_vnode_markerfree_lazy(mvp, mp); 7255 return (NULL); 7256 } 7257 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist); 7258 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 7259 } 7260 7261 void 7262 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 7263 { 7264 7265 if (*mvp == NULL) 7266 return; 7267 7268 mtx_lock(&mp->mnt_listmtx); 7269 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 7270 mtx_unlock(&mp->mnt_listmtx); 7271 mnt_vnode_markerfree_lazy(mvp, mp); 7272 } 7273 7274 int 7275 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp) 7276 { 7277 7278 if ((cnp->cn_flags & NOEXECCHECK) != 0) { 7279 cnp->cn_flags &= ~NOEXECCHECK; 7280 return (0); 7281 } 7282 7283 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread)); 7284 } 7285 7286 /* 7287 * Do not use this variant unless you have means other than the hold count 7288 * to prevent the vnode from getting freed. 7289 */ 7290 void 7291 vn_seqc_write_begin_locked(struct vnode *vp) 7292 { 7293 7294 ASSERT_VI_LOCKED(vp, __func__); 7295 VNPASS(vp->v_holdcnt > 0, vp); 7296 VNPASS(vp->v_seqc_users >= 0, vp); 7297 vp->v_seqc_users++; 7298 if (vp->v_seqc_users == 1) 7299 seqc_sleepable_write_begin(&vp->v_seqc); 7300 } 7301 7302 void 7303 vn_seqc_write_begin(struct vnode *vp) 7304 { 7305 7306 VI_LOCK(vp); 7307 vn_seqc_write_begin_locked(vp); 7308 VI_UNLOCK(vp); 7309 } 7310 7311 void 7312 vn_seqc_write_end_locked(struct vnode *vp) 7313 { 7314 7315 ASSERT_VI_LOCKED(vp, __func__); 7316 VNPASS(vp->v_seqc_users > 0, vp); 7317 vp->v_seqc_users--; 7318 if (vp->v_seqc_users == 0) 7319 seqc_sleepable_write_end(&vp->v_seqc); 7320 } 7321 7322 void 7323 vn_seqc_write_end(struct vnode *vp) 7324 { 7325 7326 VI_LOCK(vp); 7327 vn_seqc_write_end_locked(vp); 7328 VI_UNLOCK(vp); 7329 } 7330 7331 /* 7332 * Special case handling for allocating and freeing vnodes. 7333 * 7334 * The counter remains unchanged on free so that a doomed vnode will 7335 * keep testing as in modify as long as it is accessible with SMR. 7336 */ 7337 static void 7338 vn_seqc_init(struct vnode *vp) 7339 { 7340 7341 vp->v_seqc = 0; 7342 vp->v_seqc_users = 0; 7343 } 7344 7345 static void 7346 vn_seqc_write_end_free(struct vnode *vp) 7347 { 7348 7349 VNPASS(seqc_in_modify(vp->v_seqc), vp); 7350 VNPASS(vp->v_seqc_users == 1, vp); 7351 } 7352 7353 void 7354 vn_irflag_set_locked(struct vnode *vp, short toset) 7355 { 7356 short flags; 7357 7358 ASSERT_VI_LOCKED(vp, __func__); 7359 flags = vn_irflag_read(vp); 7360 VNASSERT((flags & toset) == 0, vp, 7361 ("%s: some of the passed flags already set (have %d, passed %d)\n", 7362 __func__, flags, toset)); 7363 atomic_store_short(&vp->v_irflag, flags | toset); 7364 } 7365 7366 void 7367 vn_irflag_set(struct vnode *vp, short toset) 7368 { 7369 7370 VI_LOCK(vp); 7371 vn_irflag_set_locked(vp, toset); 7372 VI_UNLOCK(vp); 7373 } 7374 7375 void 7376 vn_irflag_set_cond_locked(struct vnode *vp, short toset) 7377 { 7378 short flags; 7379 7380 ASSERT_VI_LOCKED(vp, __func__); 7381 flags = vn_irflag_read(vp); 7382 atomic_store_short(&vp->v_irflag, flags | toset); 7383 } 7384 7385 void 7386 vn_irflag_set_cond(struct vnode *vp, short toset) 7387 { 7388 7389 VI_LOCK(vp); 7390 vn_irflag_set_cond_locked(vp, toset); 7391 VI_UNLOCK(vp); 7392 } 7393 7394 void 7395 vn_irflag_unset_locked(struct vnode *vp, short tounset) 7396 { 7397 short flags; 7398 7399 ASSERT_VI_LOCKED(vp, __func__); 7400 flags = vn_irflag_read(vp); 7401 VNASSERT((flags & tounset) == tounset, vp, 7402 ("%s: some of the passed flags not set (have %d, passed %d)\n", 7403 __func__, flags, tounset)); 7404 atomic_store_short(&vp->v_irflag, flags & ~tounset); 7405 } 7406 7407 void 7408 vn_irflag_unset(struct vnode *vp, short tounset) 7409 { 7410 7411 VI_LOCK(vp); 7412 vn_irflag_unset_locked(vp, tounset); 7413 VI_UNLOCK(vp); 7414 } 7415 7416 int 7417 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred) 7418 { 7419 struct vattr vattr; 7420 int error; 7421 7422 ASSERT_VOP_LOCKED(vp, __func__); 7423 error = VOP_GETATTR(vp, &vattr, cred); 7424 if (__predict_true(error == 0)) { 7425 if (vattr.va_size <= OFF_MAX) 7426 *size = vattr.va_size; 7427 else 7428 error = EFBIG; 7429 } 7430 return (error); 7431 } 7432 7433 int 7434 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred) 7435 { 7436 int error; 7437 7438 VOP_LOCK(vp, LK_SHARED); 7439 error = vn_getsize_locked(vp, size, cred); 7440 VOP_UNLOCK(vp); 7441 return (error); 7442 } 7443 7444 #ifdef INVARIANTS 7445 void 7446 vn_set_state_validate(struct vnode *vp, __enum_uint8(vstate) state) 7447 { 7448 7449 switch (vp->v_state) { 7450 case VSTATE_UNINITIALIZED: 7451 switch (state) { 7452 case VSTATE_CONSTRUCTED: 7453 case VSTATE_DESTROYING: 7454 return; 7455 default: 7456 break; 7457 } 7458 break; 7459 case VSTATE_CONSTRUCTED: 7460 ASSERT_VOP_ELOCKED(vp, __func__); 7461 switch (state) { 7462 case VSTATE_DESTROYING: 7463 return; 7464 default: 7465 break; 7466 } 7467 break; 7468 case VSTATE_DESTROYING: 7469 ASSERT_VOP_ELOCKED(vp, __func__); 7470 switch (state) { 7471 case VSTATE_DEAD: 7472 return; 7473 default: 7474 break; 7475 } 7476 break; 7477 case VSTATE_DEAD: 7478 switch (state) { 7479 case VSTATE_UNINITIALIZED: 7480 return; 7481 default: 7482 break; 7483 } 7484 break; 7485 } 7486 7487 vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state); 7488 panic("invalid state transition %d -> %d\n", vp->v_state, state); 7489 } 7490 #endif 7491