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 static void 3572 vput_final(struct vnode *vp, enum vput_op func) 3573 { 3574 int error; 3575 bool want_unlock; 3576 3577 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3578 VNPASS(vp->v_holdcnt > 0, vp); 3579 3580 VI_LOCK(vp); 3581 3582 /* 3583 * By the time we got here someone else might have transitioned 3584 * the count back to > 0. 3585 */ 3586 if (vp->v_usecount > 0) 3587 goto out; 3588 3589 /* 3590 * If the vnode is doomed vgone already performed inactive processing 3591 * (if needed). 3592 */ 3593 if (VN_IS_DOOMED(vp)) 3594 goto out; 3595 3596 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0)) 3597 goto out; 3598 3599 if (vp->v_iflag & VI_DOINGINACT) 3600 goto out; 3601 3602 /* 3603 * Locking operations here will drop the interlock and possibly the 3604 * vnode lock, opening a window where the vnode can get doomed all the 3605 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to 3606 * perform inactive. 3607 */ 3608 vp->v_iflag |= VI_OWEINACT; 3609 want_unlock = false; 3610 error = 0; 3611 switch (func) { 3612 case VRELE: 3613 switch (VOP_ISLOCKED(vp)) { 3614 case LK_EXCLUSIVE: 3615 break; 3616 case LK_EXCLOTHER: 3617 case 0: 3618 want_unlock = true; 3619 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 3620 VI_LOCK(vp); 3621 break; 3622 default: 3623 /* 3624 * The lock has at least one sharer, but we have no way 3625 * to conclude whether this is us. Play it safe and 3626 * defer processing. 3627 */ 3628 error = EAGAIN; 3629 break; 3630 } 3631 break; 3632 case VPUT: 3633 want_unlock = true; 3634 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3635 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | 3636 LK_NOWAIT); 3637 VI_LOCK(vp); 3638 } 3639 break; 3640 case VUNREF: 3641 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3642 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); 3643 VI_LOCK(vp); 3644 } 3645 break; 3646 } 3647 if (error == 0) { 3648 if (func == VUNREF) { 3649 VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp, 3650 ("recursive vunref")); 3651 vp->v_vflag |= VV_UNREF; 3652 } 3653 for (;;) { 3654 error = vinactive(vp); 3655 if (want_unlock) 3656 VOP_UNLOCK(vp); 3657 if (error != ERELOOKUP || !want_unlock) 3658 break; 3659 VOP_LOCK(vp, LK_EXCLUSIVE); 3660 } 3661 if (func == VUNREF) 3662 vp->v_vflag &= ~VV_UNREF; 3663 vdropl(vp); 3664 } else { 3665 vdefer_inactive(vp); 3666 } 3667 return; 3668 out: 3669 if (func == VPUT) 3670 VOP_UNLOCK(vp); 3671 vdropl(vp); 3672 } 3673 3674 /* 3675 * Decrement ->v_usecount for a vnode. 3676 * 3677 * Releasing the last use count requires additional processing, see vput_final 3678 * above for details. 3679 * 3680 * Comment above each variant denotes lock state on entry and exit. 3681 */ 3682 3683 /* 3684 * in: any 3685 * out: same as passed in 3686 */ 3687 void 3688 vrele(struct vnode *vp) 3689 { 3690 3691 ASSERT_VI_UNLOCKED(vp, __func__); 3692 if (!refcount_release(&vp->v_usecount)) 3693 return; 3694 vput_final(vp, VRELE); 3695 } 3696 3697 /* 3698 * in: locked 3699 * out: unlocked 3700 */ 3701 void 3702 vput(struct vnode *vp) 3703 { 3704 3705 ASSERT_VOP_LOCKED(vp, __func__); 3706 ASSERT_VI_UNLOCKED(vp, __func__); 3707 if (!refcount_release(&vp->v_usecount)) { 3708 VOP_UNLOCK(vp); 3709 return; 3710 } 3711 vput_final(vp, VPUT); 3712 } 3713 3714 /* 3715 * in: locked 3716 * out: locked 3717 */ 3718 void 3719 vunref(struct vnode *vp) 3720 { 3721 3722 ASSERT_VOP_LOCKED(vp, __func__); 3723 ASSERT_VI_UNLOCKED(vp, __func__); 3724 if (!refcount_release(&vp->v_usecount)) 3725 return; 3726 vput_final(vp, VUNREF); 3727 } 3728 3729 void 3730 vhold(struct vnode *vp) 3731 { 3732 int old; 3733 3734 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3735 old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3736 VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3737 ("%s: wrong hold count %d", __func__, old)); 3738 if (old == 0) 3739 vfs_freevnodes_dec(); 3740 } 3741 3742 void 3743 vholdnz(struct vnode *vp) 3744 { 3745 3746 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3747 #ifdef INVARIANTS 3748 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3749 VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3750 ("%s: wrong hold count %d", __func__, old)); 3751 #else 3752 atomic_add_int(&vp->v_holdcnt, 1); 3753 #endif 3754 } 3755 3756 /* 3757 * Grab a hold count unless the vnode is freed. 3758 * 3759 * Only use this routine if vfs smr is the only protection you have against 3760 * freeing the vnode. 3761 * 3762 * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag 3763 * is not set. After the flag is set the vnode becomes immutable to anyone but 3764 * the thread which managed to set the flag. 3765 * 3766 * It may be tempting to replace the loop with: 3767 * count = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3768 * if (count & VHOLD_NO_SMR) { 3769 * backpedal and error out; 3770 * } 3771 * 3772 * However, while this is more performant, it hinders debugging by eliminating 3773 * the previously mentioned invariant. 3774 */ 3775 bool 3776 vhold_smr(struct vnode *vp) 3777 { 3778 int count; 3779 3780 VFS_SMR_ASSERT_ENTERED(); 3781 3782 count = atomic_load_int(&vp->v_holdcnt); 3783 for (;;) { 3784 if (count & VHOLD_NO_SMR) { 3785 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, 3786 ("non-zero hold count with flags %d\n", count)); 3787 return (false); 3788 } 3789 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); 3790 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) { 3791 if (count == 0) 3792 vfs_freevnodes_dec(); 3793 return (true); 3794 } 3795 } 3796 } 3797 3798 /* 3799 * Hold a free vnode for recycling. 3800 * 3801 * Note: vnode_init references this comment. 3802 * 3803 * Attempts to recycle only need the global vnode list lock and have no use for 3804 * SMR. 3805 * 3806 * However, vnodes get inserted into the global list before they get fully 3807 * initialized and stay there until UMA decides to free the memory. This in 3808 * particular means the target can be found before it becomes usable and after 3809 * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to 3810 * VHOLD_NO_SMR. 3811 * 3812 * Note: the vnode may gain more references after we transition the count 0->1. 3813 */ 3814 static bool 3815 vhold_recycle_free(struct vnode *vp) 3816 { 3817 int count; 3818 3819 mtx_assert(&vnode_list_mtx, MA_OWNED); 3820 3821 count = atomic_load_int(&vp->v_holdcnt); 3822 for (;;) { 3823 if (count & VHOLD_NO_SMR) { 3824 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, 3825 ("non-zero hold count with flags %d\n", count)); 3826 return (false); 3827 } 3828 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); 3829 if (count > 0) { 3830 return (false); 3831 } 3832 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) { 3833 vfs_freevnodes_dec(); 3834 return (true); 3835 } 3836 } 3837 } 3838 3839 static void __noinline 3840 vdbatch_process(struct vdbatch *vd) 3841 { 3842 struct vnode *vp; 3843 int i; 3844 3845 mtx_assert(&vd->lock, MA_OWNED); 3846 MPASS(curthread->td_pinned > 0); 3847 MPASS(vd->index == VDBATCH_SIZE); 3848 3849 /* 3850 * Attempt to requeue the passed batch, but give up easily. 3851 * 3852 * Despite batching the mechanism is prone to transient *significant* 3853 * lock contention, where vnode_list_mtx becomes the primary bottleneck 3854 * if multiple CPUs get here (one real-world example is highly parallel 3855 * do-nothing make , which will stat *tons* of vnodes). Since it is 3856 * quasi-LRU (read: not that great even if fully honoured) provide an 3857 * option to just dodge the problem. Parties which don't like it are 3858 * welcome to implement something better. 3859 */ 3860 if (vnode_can_skip_requeue) { 3861 if (!mtx_trylock(&vnode_list_mtx)) { 3862 counter_u64_add(vnode_skipped_requeues, 1); 3863 critical_enter(); 3864 for (i = 0; i < VDBATCH_SIZE; i++) { 3865 vp = vd->tab[i]; 3866 vd->tab[i] = NULL; 3867 MPASS(vp->v_dbatchcpu != NOCPU); 3868 vp->v_dbatchcpu = NOCPU; 3869 } 3870 vd->index = 0; 3871 critical_exit(); 3872 return; 3873 3874 } 3875 /* fallthrough to locked processing */ 3876 } else { 3877 mtx_lock(&vnode_list_mtx); 3878 } 3879 3880 mtx_assert(&vnode_list_mtx, MA_OWNED); 3881 critical_enter(); 3882 for (i = 0; i < VDBATCH_SIZE; i++) { 3883 vp = vd->tab[i]; 3884 vd->tab[i] = NULL; 3885 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 3886 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist); 3887 MPASS(vp->v_dbatchcpu != NOCPU); 3888 vp->v_dbatchcpu = NOCPU; 3889 } 3890 mtx_unlock(&vnode_list_mtx); 3891 vd->index = 0; 3892 critical_exit(); 3893 } 3894 3895 static void 3896 vdbatch_enqueue(struct vnode *vp) 3897 { 3898 struct vdbatch *vd; 3899 3900 ASSERT_VI_LOCKED(vp, __func__); 3901 VNPASS(!VN_IS_DOOMED(vp), vp); 3902 3903 if (vp->v_dbatchcpu != NOCPU) { 3904 VI_UNLOCK(vp); 3905 return; 3906 } 3907 3908 sched_pin(); 3909 vd = DPCPU_PTR(vd); 3910 mtx_lock(&vd->lock); 3911 MPASS(vd->index < VDBATCH_SIZE); 3912 MPASS(vd->tab[vd->index] == NULL); 3913 /* 3914 * A hack: we depend on being pinned so that we know what to put in 3915 * ->v_dbatchcpu. 3916 */ 3917 vp->v_dbatchcpu = curcpu; 3918 vd->tab[vd->index] = vp; 3919 vd->index++; 3920 VI_UNLOCK(vp); 3921 if (vd->index == VDBATCH_SIZE) 3922 vdbatch_process(vd); 3923 mtx_unlock(&vd->lock); 3924 sched_unpin(); 3925 } 3926 3927 /* 3928 * This routine must only be called for vnodes which are about to be 3929 * deallocated. Supporting dequeue for arbitrary vndoes would require 3930 * validating that the locked batch matches. 3931 */ 3932 static void 3933 vdbatch_dequeue(struct vnode *vp) 3934 { 3935 struct vdbatch *vd; 3936 int i; 3937 short cpu; 3938 3939 VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp); 3940 3941 cpu = vp->v_dbatchcpu; 3942 if (cpu == NOCPU) 3943 return; 3944 3945 vd = DPCPU_ID_PTR(cpu, vd); 3946 mtx_lock(&vd->lock); 3947 for (i = 0; i < vd->index; i++) { 3948 if (vd->tab[i] != vp) 3949 continue; 3950 vp->v_dbatchcpu = NOCPU; 3951 vd->index--; 3952 vd->tab[i] = vd->tab[vd->index]; 3953 vd->tab[vd->index] = NULL; 3954 break; 3955 } 3956 mtx_unlock(&vd->lock); 3957 /* 3958 * Either we dequeued the vnode above or the target CPU beat us to it. 3959 */ 3960 MPASS(vp->v_dbatchcpu == NOCPU); 3961 } 3962 3963 /* 3964 * Drop the hold count of the vnode. 3965 * 3966 * It will only get freed if this is the last hold *and* it has been vgone'd. 3967 * 3968 * Because the vnode vm object keeps a hold reference on the vnode if 3969 * there is at least one resident non-cached page, the vnode cannot 3970 * leave the active list without the page cleanup done. 3971 */ 3972 static void __noinline 3973 vdropl_final(struct vnode *vp) 3974 { 3975 3976 ASSERT_VI_LOCKED(vp, __func__); 3977 VNPASS(VN_IS_DOOMED(vp), vp); 3978 /* 3979 * Set the VHOLD_NO_SMR flag. 3980 * 3981 * We may be racing against vhold_smr. If they win we can just pretend 3982 * we never got this far, they will vdrop later. 3983 */ 3984 if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) { 3985 vfs_freevnodes_inc(); 3986 VI_UNLOCK(vp); 3987 /* 3988 * We lost the aforementioned race. Any subsequent access is 3989 * invalid as they might have managed to vdropl on their own. 3990 */ 3991 return; 3992 } 3993 /* 3994 * Don't bump freevnodes as this one is going away. 3995 */ 3996 freevnode(vp); 3997 } 3998 3999 void 4000 vdrop(struct vnode *vp) 4001 { 4002 4003 ASSERT_VI_UNLOCKED(vp, __func__); 4004 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4005 if (refcount_release_if_not_last(&vp->v_holdcnt)) 4006 return; 4007 VI_LOCK(vp); 4008 vdropl(vp); 4009 } 4010 4011 static __always_inline void 4012 vdropl_impl(struct vnode *vp, bool enqueue) 4013 { 4014 4015 ASSERT_VI_LOCKED(vp, __func__); 4016 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4017 if (!refcount_release(&vp->v_holdcnt)) { 4018 VI_UNLOCK(vp); 4019 return; 4020 } 4021 VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp); 4022 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp); 4023 if (VN_IS_DOOMED(vp)) { 4024 vdropl_final(vp); 4025 return; 4026 } 4027 4028 vfs_freevnodes_inc(); 4029 if (vp->v_mflag & VMP_LAZYLIST) { 4030 vunlazy(vp); 4031 } 4032 4033 if (!enqueue) { 4034 VI_UNLOCK(vp); 4035 return; 4036 } 4037 4038 /* 4039 * Also unlocks the interlock. We can't assert on it as we 4040 * released our hold and by now the vnode might have been 4041 * freed. 4042 */ 4043 vdbatch_enqueue(vp); 4044 } 4045 4046 void 4047 vdropl(struct vnode *vp) 4048 { 4049 4050 vdropl_impl(vp, true); 4051 } 4052 4053 /* 4054 * vdrop a vnode when recycling 4055 * 4056 * This is a special case routine only to be used when recycling, differs from 4057 * regular vdrop by not requeieing the vnode on LRU. 4058 * 4059 * Consider a case where vtryrecycle continuously fails with all vnodes (due to 4060 * e.g., frozen writes on the filesystem), filling the batch and causing it to 4061 * be requeued. Then vnlru will end up revisiting the same vnodes. This is a 4062 * loop which can last for as long as writes are frozen. 4063 */ 4064 static void 4065 vdropl_recycle(struct vnode *vp) 4066 { 4067 4068 vdropl_impl(vp, false); 4069 } 4070 4071 static void 4072 vdrop_recycle(struct vnode *vp) 4073 { 4074 4075 VI_LOCK(vp); 4076 vdropl_recycle(vp); 4077 } 4078 4079 /* 4080 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 4081 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 4082 */ 4083 static int 4084 vinactivef(struct vnode *vp) 4085 { 4086 int error; 4087 4088 ASSERT_VOP_ELOCKED(vp, "vinactive"); 4089 ASSERT_VI_LOCKED(vp, "vinactive"); 4090 VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp); 4091 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4092 vp->v_iflag |= VI_DOINGINACT; 4093 vp->v_iflag &= ~VI_OWEINACT; 4094 VI_UNLOCK(vp); 4095 4096 /* 4097 * Before moving off the active list, we must be sure that any 4098 * modified pages are converted into the vnode's dirty 4099 * buffers, since these will no longer be checked once the 4100 * vnode is on the inactive list. 4101 * 4102 * The write-out of the dirty pages is asynchronous. At the 4103 * point that VOP_INACTIVE() is called, there could still be 4104 * pending I/O and dirty pages in the object. 4105 */ 4106 if ((vp->v_vflag & VV_NOSYNC) == 0) 4107 vnode_pager_clean_async(vp); 4108 4109 error = VOP_INACTIVE(vp); 4110 VI_LOCK(vp); 4111 VNPASS(vp->v_iflag & VI_DOINGINACT, vp); 4112 vp->v_iflag &= ~VI_DOINGINACT; 4113 return (error); 4114 } 4115 4116 int 4117 vinactive(struct vnode *vp) 4118 { 4119 4120 ASSERT_VOP_ELOCKED(vp, "vinactive"); 4121 ASSERT_VI_LOCKED(vp, "vinactive"); 4122 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4123 4124 if ((vp->v_iflag & VI_OWEINACT) == 0) 4125 return (0); 4126 if (vp->v_iflag & VI_DOINGINACT) 4127 return (0); 4128 if (vp->v_usecount > 0) { 4129 vp->v_iflag &= ~VI_OWEINACT; 4130 return (0); 4131 } 4132 return (vinactivef(vp)); 4133 } 4134 4135 /* 4136 * Remove any vnodes in the vnode table belonging to mount point mp. 4137 * 4138 * If FORCECLOSE is not specified, there should not be any active ones, 4139 * return error if any are found (nb: this is a user error, not a 4140 * system error). If FORCECLOSE is specified, detach any active vnodes 4141 * that are found. 4142 * 4143 * If WRITECLOSE is set, only flush out regular file vnodes open for 4144 * writing. 4145 * 4146 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 4147 * 4148 * `rootrefs' specifies the base reference count for the root vnode 4149 * of this filesystem. The root vnode is considered busy if its 4150 * v_usecount exceeds this value. On a successful return, vflush(, td) 4151 * will call vrele() on the root vnode exactly rootrefs times. 4152 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 4153 * be zero. 4154 */ 4155 #ifdef DIAGNOSTIC 4156 static int busyprt = 0; /* print out busy vnodes */ 4157 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 4158 #endif 4159 4160 int 4161 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 4162 { 4163 struct vnode *vp, *mvp, *rootvp = NULL; 4164 struct vattr vattr; 4165 int busy = 0, error; 4166 4167 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 4168 rootrefs, flags); 4169 if (rootrefs > 0) { 4170 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 4171 ("vflush: bad args")); 4172 /* 4173 * Get the filesystem root vnode. We can vput() it 4174 * immediately, since with rootrefs > 0, it won't go away. 4175 */ 4176 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 4177 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 4178 __func__, error); 4179 return (error); 4180 } 4181 vput(rootvp); 4182 } 4183 loop: 4184 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 4185 vholdl(vp); 4186 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 4187 if (error) { 4188 vdrop(vp); 4189 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 4190 goto loop; 4191 } 4192 /* 4193 * Skip over a vnodes marked VV_SYSTEM. 4194 */ 4195 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 4196 VOP_UNLOCK(vp); 4197 vdrop(vp); 4198 continue; 4199 } 4200 /* 4201 * If WRITECLOSE is set, flush out unlinked but still open 4202 * files (even if open only for reading) and regular file 4203 * vnodes open for writing. 4204 */ 4205 if (flags & WRITECLOSE) { 4206 vnode_pager_clean_async(vp); 4207 do { 4208 error = VOP_FSYNC(vp, MNT_WAIT, td); 4209 } while (error == ERELOOKUP); 4210 if (error != 0) { 4211 VOP_UNLOCK(vp); 4212 vdrop(vp); 4213 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 4214 return (error); 4215 } 4216 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 4217 VI_LOCK(vp); 4218 4219 if ((vp->v_type == VNON || 4220 (error == 0 && vattr.va_nlink > 0)) && 4221 (vp->v_writecount <= 0 || vp->v_type != VREG)) { 4222 VOP_UNLOCK(vp); 4223 vdropl(vp); 4224 continue; 4225 } 4226 } else 4227 VI_LOCK(vp); 4228 /* 4229 * With v_usecount == 0, all we need to do is clear out the 4230 * vnode data structures and we are done. 4231 * 4232 * If FORCECLOSE is set, forcibly close the vnode. 4233 */ 4234 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 4235 vgonel(vp); 4236 } else { 4237 busy++; 4238 #ifdef DIAGNOSTIC 4239 if (busyprt) 4240 vn_printf(vp, "vflush: busy vnode "); 4241 #endif 4242 } 4243 VOP_UNLOCK(vp); 4244 vdropl(vp); 4245 } 4246 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 4247 /* 4248 * If just the root vnode is busy, and if its refcount 4249 * is equal to `rootrefs', then go ahead and kill it. 4250 */ 4251 VI_LOCK(rootvp); 4252 KASSERT(busy > 0, ("vflush: not busy")); 4253 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 4254 ("vflush: usecount %d < rootrefs %d", 4255 rootvp->v_usecount, rootrefs)); 4256 if (busy == 1 && rootvp->v_usecount == rootrefs) { 4257 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 4258 vgone(rootvp); 4259 VOP_UNLOCK(rootvp); 4260 busy = 0; 4261 } else 4262 VI_UNLOCK(rootvp); 4263 } 4264 if (busy) { 4265 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 4266 busy); 4267 return (EBUSY); 4268 } 4269 for (; rootrefs > 0; rootrefs--) 4270 vrele(rootvp); 4271 return (0); 4272 } 4273 4274 /* 4275 * Recycle an unused vnode. 4276 */ 4277 int 4278 vrecycle(struct vnode *vp) 4279 { 4280 int recycled; 4281 4282 VI_LOCK(vp); 4283 recycled = vrecyclel(vp); 4284 VI_UNLOCK(vp); 4285 return (recycled); 4286 } 4287 4288 /* 4289 * vrecycle, with the vp interlock held. 4290 */ 4291 int 4292 vrecyclel(struct vnode *vp) 4293 { 4294 int recycled; 4295 4296 ASSERT_VOP_ELOCKED(vp, __func__); 4297 ASSERT_VI_LOCKED(vp, __func__); 4298 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4299 recycled = 0; 4300 if (vp->v_usecount == 0) { 4301 recycled = 1; 4302 vgonel(vp); 4303 } 4304 return (recycled); 4305 } 4306 4307 /* 4308 * Eliminate all activity associated with a vnode 4309 * in preparation for reuse. 4310 */ 4311 void 4312 vgone(struct vnode *vp) 4313 { 4314 VI_LOCK(vp); 4315 vgonel(vp); 4316 VI_UNLOCK(vp); 4317 } 4318 4319 /* 4320 * Notify upper mounts about reclaimed or unlinked vnode. 4321 */ 4322 void 4323 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event) 4324 { 4325 struct mount *mp; 4326 struct mount_upper_node *ump; 4327 4328 mp = atomic_load_ptr(&vp->v_mount); 4329 if (mp == NULL) 4330 return; 4331 if (TAILQ_EMPTY(&mp->mnt_notify)) 4332 return; 4333 4334 MNT_ILOCK(mp); 4335 mp->mnt_upper_pending++; 4336 KASSERT(mp->mnt_upper_pending > 0, 4337 ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending)); 4338 TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) { 4339 MNT_IUNLOCK(mp); 4340 switch (event) { 4341 case VFS_NOTIFY_UPPER_RECLAIM: 4342 VFS_RECLAIM_LOWERVP(ump->mp, vp); 4343 break; 4344 case VFS_NOTIFY_UPPER_UNLINK: 4345 VFS_UNLINK_LOWERVP(ump->mp, vp); 4346 break; 4347 } 4348 MNT_ILOCK(mp); 4349 } 4350 mp->mnt_upper_pending--; 4351 if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 && 4352 mp->mnt_upper_pending == 0) { 4353 mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER; 4354 wakeup(&mp->mnt_uppers); 4355 } 4356 MNT_IUNLOCK(mp); 4357 } 4358 4359 /* 4360 * vgone, with the vp interlock held. 4361 */ 4362 static void 4363 vgonel(struct vnode *vp) 4364 { 4365 struct thread *td; 4366 struct mount *mp; 4367 vm_object_t object; 4368 bool active, doinginact, oweinact; 4369 4370 ASSERT_VOP_ELOCKED(vp, "vgonel"); 4371 ASSERT_VI_LOCKED(vp, "vgonel"); 4372 VNASSERT(vp->v_holdcnt, vp, 4373 ("vgonel: vp %p has no reference.", vp)); 4374 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4375 td = curthread; 4376 4377 /* 4378 * Don't vgonel if we're already doomed. 4379 */ 4380 if (VN_IS_DOOMED(vp)) { 4381 VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \ 4382 vn_get_state(vp) == VSTATE_DEAD, vp); 4383 return; 4384 } 4385 /* 4386 * Paired with freevnode. 4387 */ 4388 vn_seqc_write_begin_locked(vp); 4389 vunlazy_gone(vp); 4390 vn_irflag_set_locked(vp, VIRF_DOOMED); 4391 vn_set_state(vp, VSTATE_DESTROYING); 4392 4393 /* 4394 * Check to see if the vnode is in use. If so, we have to 4395 * call VOP_CLOSE() and VOP_INACTIVE(). 4396 * 4397 * It could be that VOP_INACTIVE() requested reclamation, in 4398 * which case we should avoid recursion, so check 4399 * VI_DOINGINACT. This is not precise but good enough. 4400 */ 4401 active = vp->v_usecount > 0; 4402 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 4403 doinginact = (vp->v_iflag & VI_DOINGINACT) != 0; 4404 4405 /* 4406 * If we need to do inactive VI_OWEINACT will be set. 4407 */ 4408 if (vp->v_iflag & VI_DEFINACT) { 4409 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 4410 vp->v_iflag &= ~VI_DEFINACT; 4411 vdropl(vp); 4412 } else { 4413 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); 4414 VI_UNLOCK(vp); 4415 } 4416 cache_purge_vgone(vp); 4417 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); 4418 4419 /* 4420 * If purging an active vnode, it must be closed and 4421 * deactivated before being reclaimed. 4422 */ 4423 if (active) 4424 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 4425 if (!doinginact) { 4426 do { 4427 if (oweinact || active) { 4428 VI_LOCK(vp); 4429 vinactivef(vp); 4430 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 4431 VI_UNLOCK(vp); 4432 } 4433 } while (oweinact); 4434 } 4435 if (vp->v_type == VSOCK) 4436 vfs_unp_reclaim(vp); 4437 4438 /* 4439 * Clean out any buffers associated with the vnode. 4440 * If the flush fails, just toss the buffers. 4441 */ 4442 mp = NULL; 4443 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 4444 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 4445 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { 4446 while (vinvalbuf(vp, 0, 0, 0) != 0) 4447 ; 4448 } 4449 4450 BO_LOCK(&vp->v_bufobj); 4451 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && 4452 vp->v_bufobj.bo_dirty.bv_cnt == 0 && 4453 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && 4454 vp->v_bufobj.bo_clean.bv_cnt == 0, 4455 ("vp %p bufobj not invalidated", vp)); 4456 4457 /* 4458 * For VMIO bufobj, BO_DEAD is set later, or in 4459 * vm_object_terminate() after the object's page queue is 4460 * flushed. 4461 */ 4462 object = vp->v_bufobj.bo_object; 4463 if (object == NULL) 4464 vp->v_bufobj.bo_flag |= BO_DEAD; 4465 BO_UNLOCK(&vp->v_bufobj); 4466 4467 /* 4468 * Handle the VM part. Tmpfs handles v_object on its own (the 4469 * OBJT_VNODE check). Nullfs or other bypassing filesystems 4470 * should not touch the object borrowed from the lower vnode 4471 * (the handle check). 4472 */ 4473 if (object != NULL && object->type == OBJT_VNODE && 4474 object->handle == vp) 4475 vnode_destroy_vobject(vp); 4476 4477 /* 4478 * Reclaim the vnode. 4479 */ 4480 if (VOP_RECLAIM(vp)) 4481 panic("vgone: cannot reclaim"); 4482 if (mp != NULL) 4483 vn_finished_secondary_write(mp); 4484 VNASSERT(vp->v_object == NULL, vp, 4485 ("vop_reclaim left v_object vp=%p", vp)); 4486 /* 4487 * Clear the advisory locks and wake up waiting threads. 4488 */ 4489 if (vp->v_lockf != NULL) { 4490 (void)VOP_ADVLOCKPURGE(vp); 4491 vp->v_lockf = NULL; 4492 } 4493 /* 4494 * Delete from old mount point vnode list. 4495 */ 4496 if (vp->v_mount == NULL) { 4497 VI_LOCK(vp); 4498 } else { 4499 delmntque(vp); 4500 ASSERT_VI_LOCKED(vp, "vgonel 2"); 4501 } 4502 /* 4503 * Done with purge, reset to the standard lock and invalidate 4504 * the vnode. 4505 * 4506 * FIXME: this is buggy for vnode ops with custom locking primitives. 4507 * 4508 * vget used to be gated with a special flag serializing it against vgone, 4509 * which got lost in the process of SMP-ifying the VFS layer. 4510 * 4511 * Suppose a custom locking routine references ->v_data. 4512 * 4513 * Since now it is possible to start executing it as vgone is 4514 * progressing, this very well may crash as ->v_data gets invalidated 4515 * and memory used to back it is freed. 4516 */ 4517 vp->v_vnlock = &vp->v_lock; 4518 vp->v_op = &dead_vnodeops; 4519 vp->v_type = VBAD; 4520 vn_set_state(vp, VSTATE_DEAD); 4521 } 4522 4523 /* 4524 * Print out a description of a vnode. 4525 */ 4526 static const char *const vtypename[] = { 4527 [VNON] = "VNON", 4528 [VREG] = "VREG", 4529 [VDIR] = "VDIR", 4530 [VBLK] = "VBLK", 4531 [VCHR] = "VCHR", 4532 [VLNK] = "VLNK", 4533 [VSOCK] = "VSOCK", 4534 [VFIFO] = "VFIFO", 4535 [VBAD] = "VBAD", 4536 [VMARKER] = "VMARKER", 4537 }; 4538 _Static_assert(nitems(vtypename) == VLASTTYPE + 1, 4539 "vnode type name not added to vtypename"); 4540 4541 static const char *const vstatename[] = { 4542 [VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED", 4543 [VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED", 4544 [VSTATE_DESTROYING] = "VSTATE_DESTROYING", 4545 [VSTATE_DEAD] = "VSTATE_DEAD", 4546 }; 4547 _Static_assert(nitems(vstatename) == VLASTSTATE + 1, 4548 "vnode state name not added to vstatename"); 4549 4550 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0, 4551 "new hold count flag not added to vn_printf"); 4552 4553 void 4554 vn_printf(struct vnode *vp, const char *fmt, ...) 4555 { 4556 va_list ap; 4557 char buf[256], buf2[16]; 4558 u_long flags; 4559 u_int holdcnt; 4560 short irflag; 4561 4562 va_start(ap, fmt); 4563 vprintf(fmt, ap); 4564 va_end(ap); 4565 printf("%p: ", (void *)vp); 4566 printf("type %s state %s op %p\n", vtypename[vp->v_type], 4567 vstatename[vp->v_state], vp->v_op); 4568 holdcnt = atomic_load_int(&vp->v_holdcnt); 4569 printf(" usecount %d, writecount %d, refcount %d seqc users %d", 4570 vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS, 4571 vp->v_seqc_users); 4572 switch (vp->v_type) { 4573 case VDIR: 4574 printf(" mountedhere %p\n", vp->v_mountedhere); 4575 break; 4576 case VCHR: 4577 printf(" rdev %p\n", vp->v_rdev); 4578 break; 4579 case VSOCK: 4580 printf(" socket %p\n", vp->v_unpcb); 4581 break; 4582 case VFIFO: 4583 printf(" fifoinfo %p\n", vp->v_fifoinfo); 4584 break; 4585 default: 4586 printf("\n"); 4587 break; 4588 } 4589 buf[0] = '\0'; 4590 buf[1] = '\0'; 4591 if (holdcnt & VHOLD_NO_SMR) 4592 strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf)); 4593 printf(" hold count flags (%s)\n", buf + 1); 4594 4595 buf[0] = '\0'; 4596 buf[1] = '\0'; 4597 irflag = vn_irflag_read(vp); 4598 if (irflag & VIRF_DOOMED) 4599 strlcat(buf, "|VIRF_DOOMED", sizeof(buf)); 4600 if (irflag & VIRF_PGREAD) 4601 strlcat(buf, "|VIRF_PGREAD", sizeof(buf)); 4602 if (irflag & VIRF_MOUNTPOINT) 4603 strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf)); 4604 if (irflag & VIRF_TEXT_REF) 4605 strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf)); 4606 flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF); 4607 if (flags != 0) { 4608 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags); 4609 strlcat(buf, buf2, sizeof(buf)); 4610 } 4611 if (vp->v_vflag & VV_ROOT) 4612 strlcat(buf, "|VV_ROOT", sizeof(buf)); 4613 if (vp->v_vflag & VV_ISTTY) 4614 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 4615 if (vp->v_vflag & VV_NOSYNC) 4616 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 4617 if (vp->v_vflag & VV_ETERNALDEV) 4618 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 4619 if (vp->v_vflag & VV_CACHEDLABEL) 4620 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 4621 if (vp->v_vflag & VV_VMSIZEVNLOCK) 4622 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf)); 4623 if (vp->v_vflag & VV_COPYONWRITE) 4624 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 4625 if (vp->v_vflag & VV_SYSTEM) 4626 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 4627 if (vp->v_vflag & VV_PROCDEP) 4628 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 4629 if (vp->v_vflag & VV_DELETED) 4630 strlcat(buf, "|VV_DELETED", sizeof(buf)); 4631 if (vp->v_vflag & VV_MD) 4632 strlcat(buf, "|VV_MD", sizeof(buf)); 4633 if (vp->v_vflag & VV_FORCEINSMQ) 4634 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 4635 if (vp->v_vflag & VV_READLINK) 4636 strlcat(buf, "|VV_READLINK", sizeof(buf)); 4637 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 4638 VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM | 4639 VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK); 4640 if (flags != 0) { 4641 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 4642 strlcat(buf, buf2, sizeof(buf)); 4643 } 4644 if (vp->v_iflag & VI_MOUNT) 4645 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 4646 if (vp->v_iflag & VI_DOINGINACT) 4647 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 4648 if (vp->v_iflag & VI_OWEINACT) 4649 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 4650 if (vp->v_iflag & VI_DEFINACT) 4651 strlcat(buf, "|VI_DEFINACT", sizeof(buf)); 4652 if (vp->v_iflag & VI_FOPENING) 4653 strlcat(buf, "|VI_FOPENING", sizeof(buf)); 4654 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT | 4655 VI_OWEINACT | VI_DEFINACT | VI_FOPENING); 4656 if (flags != 0) { 4657 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 4658 strlcat(buf, buf2, sizeof(buf)); 4659 } 4660 if (vp->v_mflag & VMP_LAZYLIST) 4661 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf)); 4662 flags = vp->v_mflag & ~(VMP_LAZYLIST); 4663 if (flags != 0) { 4664 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags); 4665 strlcat(buf, buf2, sizeof(buf)); 4666 } 4667 printf(" flags (%s)", buf + 1); 4668 if (mtx_owned(VI_MTX(vp))) 4669 printf(" VI_LOCKed"); 4670 printf("\n"); 4671 if (vp->v_object != NULL) 4672 printf(" v_object %p ref %d pages %d " 4673 "cleanbuf %d dirtybuf %d\n", 4674 vp->v_object, vp->v_object->ref_count, 4675 vp->v_object->resident_page_count, 4676 vp->v_bufobj.bo_clean.bv_cnt, 4677 vp->v_bufobj.bo_dirty.bv_cnt); 4678 printf(" "); 4679 lockmgr_printinfo(vp->v_vnlock); 4680 if (vp->v_data != NULL) 4681 VOP_PRINT(vp); 4682 } 4683 4684 #ifdef DDB 4685 /* 4686 * List all of the locked vnodes in the system. 4687 * Called when debugging the kernel. 4688 */ 4689 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE) 4690 { 4691 struct mount *mp; 4692 struct vnode *vp; 4693 4694 /* 4695 * Note: because this is DDB, we can't obey the locking semantics 4696 * for these structures, which means we could catch an inconsistent 4697 * state and dereference a nasty pointer. Not much to be done 4698 * about that. 4699 */ 4700 db_printf("Locked vnodes\n"); 4701 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4702 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4703 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) 4704 vn_printf(vp, "vnode "); 4705 } 4706 } 4707 } 4708 4709 /* 4710 * Show details about the given vnode. 4711 */ 4712 DB_SHOW_COMMAND(vnode, db_show_vnode) 4713 { 4714 struct vnode *vp; 4715 4716 if (!have_addr) 4717 return; 4718 vp = (struct vnode *)addr; 4719 vn_printf(vp, "vnode "); 4720 } 4721 4722 /* 4723 * Show details about the given mount point. 4724 */ 4725 DB_SHOW_COMMAND(mount, db_show_mount) 4726 { 4727 struct mount *mp; 4728 struct vfsopt *opt; 4729 struct statfs *sp; 4730 struct vnode *vp; 4731 char buf[512]; 4732 uint64_t mflags; 4733 u_int flags; 4734 4735 if (!have_addr) { 4736 /* No address given, print short info about all mount points. */ 4737 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4738 db_printf("%p %s on %s (%s)\n", mp, 4739 mp->mnt_stat.f_mntfromname, 4740 mp->mnt_stat.f_mntonname, 4741 mp->mnt_stat.f_fstypename); 4742 if (db_pager_quit) 4743 break; 4744 } 4745 db_printf("\nMore info: show mount <addr>\n"); 4746 return; 4747 } 4748 4749 mp = (struct mount *)addr; 4750 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 4751 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 4752 4753 buf[0] = '\0'; 4754 mflags = mp->mnt_flag; 4755 #define MNT_FLAG(flag) do { \ 4756 if (mflags & (flag)) { \ 4757 if (buf[0] != '\0') \ 4758 strlcat(buf, ", ", sizeof(buf)); \ 4759 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 4760 mflags &= ~(flag); \ 4761 } \ 4762 } while (0) 4763 MNT_FLAG(MNT_RDONLY); 4764 MNT_FLAG(MNT_SYNCHRONOUS); 4765 MNT_FLAG(MNT_NOEXEC); 4766 MNT_FLAG(MNT_NOSUID); 4767 MNT_FLAG(MNT_NFS4ACLS); 4768 MNT_FLAG(MNT_UNION); 4769 MNT_FLAG(MNT_ASYNC); 4770 MNT_FLAG(MNT_SUIDDIR); 4771 MNT_FLAG(MNT_SOFTDEP); 4772 MNT_FLAG(MNT_NOSYMFOLLOW); 4773 MNT_FLAG(MNT_GJOURNAL); 4774 MNT_FLAG(MNT_MULTILABEL); 4775 MNT_FLAG(MNT_ACLS); 4776 MNT_FLAG(MNT_NOATIME); 4777 MNT_FLAG(MNT_NOCLUSTERR); 4778 MNT_FLAG(MNT_NOCLUSTERW); 4779 MNT_FLAG(MNT_SUJ); 4780 MNT_FLAG(MNT_EXRDONLY); 4781 MNT_FLAG(MNT_EXPORTED); 4782 MNT_FLAG(MNT_DEFEXPORTED); 4783 MNT_FLAG(MNT_EXPORTANON); 4784 MNT_FLAG(MNT_EXKERB); 4785 MNT_FLAG(MNT_EXPUBLIC); 4786 MNT_FLAG(MNT_LOCAL); 4787 MNT_FLAG(MNT_QUOTA); 4788 MNT_FLAG(MNT_ROOTFS); 4789 MNT_FLAG(MNT_USER); 4790 MNT_FLAG(MNT_IGNORE); 4791 MNT_FLAG(MNT_UPDATE); 4792 MNT_FLAG(MNT_DELEXPORT); 4793 MNT_FLAG(MNT_RELOAD); 4794 MNT_FLAG(MNT_FORCE); 4795 MNT_FLAG(MNT_SNAPSHOT); 4796 MNT_FLAG(MNT_BYFSID); 4797 MNT_FLAG(MNT_NAMEDATTR); 4798 #undef MNT_FLAG 4799 if (mflags != 0) { 4800 if (buf[0] != '\0') 4801 strlcat(buf, ", ", sizeof(buf)); 4802 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4803 "0x%016jx", mflags); 4804 } 4805 db_printf(" mnt_flag = %s\n", buf); 4806 4807 buf[0] = '\0'; 4808 flags = mp->mnt_kern_flag; 4809 #define MNT_KERN_FLAG(flag) do { \ 4810 if (flags & (flag)) { \ 4811 if (buf[0] != '\0') \ 4812 strlcat(buf, ", ", sizeof(buf)); \ 4813 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 4814 flags &= ~(flag); \ 4815 } \ 4816 } while (0) 4817 MNT_KERN_FLAG(MNTK_UNMOUNTF); 4818 MNT_KERN_FLAG(MNTK_ASYNC); 4819 MNT_KERN_FLAG(MNTK_SOFTDEP); 4820 MNT_KERN_FLAG(MNTK_NOMSYNC); 4821 MNT_KERN_FLAG(MNTK_DRAINING); 4822 MNT_KERN_FLAG(MNTK_REFEXPIRE); 4823 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 4824 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 4825 MNT_KERN_FLAG(MNTK_NO_IOPF); 4826 MNT_KERN_FLAG(MNTK_RECURSE); 4827 MNT_KERN_FLAG(MNTK_UPPER_WAITER); 4828 MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE); 4829 MNT_KERN_FLAG(MNTK_USES_BCACHE); 4830 MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG); 4831 MNT_KERN_FLAG(MNTK_FPLOOKUP); 4832 MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER); 4833 MNT_KERN_FLAG(MNTK_NOASYNC); 4834 MNT_KERN_FLAG(MNTK_UNMOUNT); 4835 MNT_KERN_FLAG(MNTK_MWAIT); 4836 MNT_KERN_FLAG(MNTK_SUSPEND); 4837 MNT_KERN_FLAG(MNTK_SUSPEND2); 4838 MNT_KERN_FLAG(MNTK_SUSPENDED); 4839 MNT_KERN_FLAG(MNTK_NULL_NOCACHE); 4840 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 4841 #undef MNT_KERN_FLAG 4842 if (flags != 0) { 4843 if (buf[0] != '\0') 4844 strlcat(buf, ", ", sizeof(buf)); 4845 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4846 "0x%08x", flags); 4847 } 4848 db_printf(" mnt_kern_flag = %s\n", buf); 4849 4850 db_printf(" mnt_opt = "); 4851 opt = TAILQ_FIRST(mp->mnt_opt); 4852 if (opt != NULL) { 4853 db_printf("%s", opt->name); 4854 opt = TAILQ_NEXT(opt, link); 4855 while (opt != NULL) { 4856 db_printf(", %s", opt->name); 4857 opt = TAILQ_NEXT(opt, link); 4858 } 4859 } 4860 db_printf("\n"); 4861 4862 sp = &mp->mnt_stat; 4863 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 4864 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 4865 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 4866 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 4867 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 4868 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 4869 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 4870 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 4871 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 4872 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 4873 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 4874 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 4875 4876 db_printf(" mnt_cred = { uid=%u ruid=%u", 4877 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 4878 if (jailed(mp->mnt_cred)) 4879 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 4880 db_printf(" }\n"); 4881 db_printf(" mnt_ref = %d (with %d in the struct)\n", 4882 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref); 4883 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 4884 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 4885 db_printf(" mnt_lazyvnodelistsize = %d\n", 4886 mp->mnt_lazyvnodelistsize); 4887 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n", 4888 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount); 4889 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 4890 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 4891 db_printf(" mnt_lockref = %d (with %d in the struct)\n", 4892 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref); 4893 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 4894 db_printf(" mnt_secondary_accwrites = %d\n", 4895 mp->mnt_secondary_accwrites); 4896 db_printf(" mnt_gjprovider = %s\n", 4897 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 4898 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops); 4899 4900 db_printf("\n\nList of active vnodes\n"); 4901 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4902 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) { 4903 vn_printf(vp, "vnode "); 4904 if (db_pager_quit) 4905 break; 4906 } 4907 } 4908 db_printf("\n\nList of inactive vnodes\n"); 4909 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4910 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) { 4911 vn_printf(vp, "vnode "); 4912 if (db_pager_quit) 4913 break; 4914 } 4915 } 4916 } 4917 #endif /* DDB */ 4918 4919 /* 4920 * Fill in a struct xvfsconf based on a struct vfsconf. 4921 */ 4922 static int 4923 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 4924 { 4925 struct xvfsconf xvfsp; 4926 4927 bzero(&xvfsp, sizeof(xvfsp)); 4928 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4929 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4930 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4931 xvfsp.vfc_flags = vfsp->vfc_flags; 4932 /* 4933 * These are unused in userland, we keep them 4934 * to not break binary compatibility. 4935 */ 4936 xvfsp.vfc_vfsops = NULL; 4937 xvfsp.vfc_next = NULL; 4938 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4939 } 4940 4941 #ifdef COMPAT_FREEBSD32 4942 struct xvfsconf32 { 4943 uint32_t vfc_vfsops; 4944 char vfc_name[MFSNAMELEN]; 4945 int32_t vfc_typenum; 4946 int32_t vfc_refcount; 4947 int32_t vfc_flags; 4948 uint32_t vfc_next; 4949 }; 4950 4951 static int 4952 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 4953 { 4954 struct xvfsconf32 xvfsp; 4955 4956 bzero(&xvfsp, sizeof(xvfsp)); 4957 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4958 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4959 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4960 xvfsp.vfc_flags = vfsp->vfc_flags; 4961 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4962 } 4963 #endif 4964 4965 /* 4966 * Top level filesystem related information gathering. 4967 */ 4968 static int 4969 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 4970 { 4971 struct vfsconf *vfsp; 4972 int error; 4973 4974 error = 0; 4975 vfsconf_slock(); 4976 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4977 #ifdef COMPAT_FREEBSD32 4978 if (req->flags & SCTL_MASK32) 4979 error = vfsconf2x32(req, vfsp); 4980 else 4981 #endif 4982 error = vfsconf2x(req, vfsp); 4983 if (error) 4984 break; 4985 } 4986 vfsconf_sunlock(); 4987 return (error); 4988 } 4989 4990 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | 4991 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, 4992 "S,xvfsconf", "List of all configured filesystems"); 4993 4994 #ifndef BURN_BRIDGES 4995 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 4996 4997 static int 4998 vfs_sysctl(SYSCTL_HANDLER_ARGS) 4999 { 5000 int *name = (int *)arg1 - 1; /* XXX */ 5001 u_int namelen = arg2 + 1; /* XXX */ 5002 struct vfsconf *vfsp; 5003 5004 log(LOG_WARNING, "userland calling deprecated sysctl, " 5005 "please rebuild world\n"); 5006 5007 #if 1 || defined(COMPAT_PRELITE2) 5008 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 5009 if (namelen == 1) 5010 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 5011 #endif 5012 5013 switch (name[1]) { 5014 case VFS_MAXTYPENUM: 5015 if (namelen != 2) 5016 return (ENOTDIR); 5017 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 5018 case VFS_CONF: 5019 if (namelen != 3) 5020 return (ENOTDIR); /* overloaded */ 5021 vfsconf_slock(); 5022 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 5023 if (vfsp->vfc_typenum == name[2]) 5024 break; 5025 } 5026 vfsconf_sunlock(); 5027 if (vfsp == NULL) 5028 return (EOPNOTSUPP); 5029 #ifdef COMPAT_FREEBSD32 5030 if (req->flags & SCTL_MASK32) 5031 return (vfsconf2x32(req, vfsp)); 5032 else 5033 #endif 5034 return (vfsconf2x(req, vfsp)); 5035 } 5036 return (EOPNOTSUPP); 5037 } 5038 5039 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | 5040 CTLFLAG_MPSAFE, vfs_sysctl, 5041 "Generic filesystem"); 5042 5043 #if 1 || defined(COMPAT_PRELITE2) 5044 5045 static int 5046 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 5047 { 5048 int error; 5049 struct vfsconf *vfsp; 5050 struct ovfsconf ovfs; 5051 5052 vfsconf_slock(); 5053 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 5054 bzero(&ovfs, sizeof(ovfs)); 5055 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 5056 strcpy(ovfs.vfc_name, vfsp->vfc_name); 5057 ovfs.vfc_index = vfsp->vfc_typenum; 5058 ovfs.vfc_refcount = vfsp->vfc_refcount; 5059 ovfs.vfc_flags = vfsp->vfc_flags; 5060 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 5061 if (error != 0) { 5062 vfsconf_sunlock(); 5063 return (error); 5064 } 5065 } 5066 vfsconf_sunlock(); 5067 return (0); 5068 } 5069 5070 #endif /* 1 || COMPAT_PRELITE2 */ 5071 #endif /* !BURN_BRIDGES */ 5072 5073 static void 5074 unmount_or_warn(struct mount *mp) 5075 { 5076 int error; 5077 5078 error = dounmount(mp, MNT_FORCE, curthread); 5079 if (error != 0) { 5080 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); 5081 if (error == EBUSY) 5082 printf("BUSY)\n"); 5083 else 5084 printf("%d)\n", error); 5085 } 5086 } 5087 5088 /* 5089 * Unmount all filesystems. The list is traversed in reverse order 5090 * of mounting to avoid dependencies. 5091 */ 5092 void 5093 vfs_unmountall(void) 5094 { 5095 struct mount *mp, *tmp; 5096 5097 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 5098 5099 /* 5100 * Since this only runs when rebooting, it is not interlocked. 5101 */ 5102 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { 5103 vfs_ref(mp); 5104 5105 /* 5106 * Forcibly unmounting "/dev" before "/" would prevent clean 5107 * unmount of the latter. 5108 */ 5109 if (mp == rootdevmp) 5110 continue; 5111 5112 unmount_or_warn(mp); 5113 } 5114 5115 if (rootdevmp != NULL) 5116 unmount_or_warn(rootdevmp); 5117 } 5118 5119 static void 5120 vfs_deferred_inactive(struct vnode *vp, int lkflags) 5121 { 5122 5123 ASSERT_VI_LOCKED(vp, __func__); 5124 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp); 5125 if ((vp->v_iflag & VI_OWEINACT) == 0) { 5126 vdropl(vp); 5127 return; 5128 } 5129 if (vn_lock(vp, lkflags) == 0) { 5130 VI_LOCK(vp); 5131 vinactive(vp); 5132 VOP_UNLOCK(vp); 5133 vdropl(vp); 5134 return; 5135 } 5136 vdefer_inactive_unlocked(vp); 5137 } 5138 5139 static int 5140 vfs_periodic_inactive_filter(struct vnode *vp, void *arg) 5141 { 5142 5143 return (vp->v_iflag & VI_DEFINACT); 5144 } 5145 5146 static void __noinline 5147 vfs_periodic_inactive(struct mount *mp, int flags) 5148 { 5149 struct vnode *vp, *mvp; 5150 int lkflags; 5151 5152 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 5153 if (flags != MNT_WAIT) 5154 lkflags |= LK_NOWAIT; 5155 5156 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) { 5157 if ((vp->v_iflag & VI_DEFINACT) == 0) { 5158 VI_UNLOCK(vp); 5159 continue; 5160 } 5161 vp->v_iflag &= ~VI_DEFINACT; 5162 vfs_deferred_inactive(vp, lkflags); 5163 } 5164 } 5165 5166 static inline bool 5167 vfs_want_msync(struct vnode *vp) 5168 { 5169 struct vm_object *obj; 5170 5171 /* 5172 * This test may be performed without any locks held. 5173 * We rely on vm_object's type stability. 5174 */ 5175 if (vp->v_vflag & VV_NOSYNC) 5176 return (false); 5177 obj = vp->v_object; 5178 return (obj != NULL && vm_object_mightbedirty(obj)); 5179 } 5180 5181 static int 5182 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused) 5183 { 5184 5185 if (vp->v_vflag & VV_NOSYNC) 5186 return (false); 5187 if (vp->v_iflag & VI_DEFINACT) 5188 return (true); 5189 return (vfs_want_msync(vp)); 5190 } 5191 5192 static void __noinline 5193 vfs_periodic_msync_inactive(struct mount *mp, int flags) 5194 { 5195 struct vnode *vp, *mvp; 5196 int lkflags; 5197 bool seen_defer; 5198 5199 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 5200 if (flags != MNT_WAIT) 5201 lkflags |= LK_NOWAIT; 5202 5203 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) { 5204 seen_defer = false; 5205 if (vp->v_iflag & VI_DEFINACT) { 5206 vp->v_iflag &= ~VI_DEFINACT; 5207 seen_defer = true; 5208 } 5209 if (!vfs_want_msync(vp)) { 5210 if (seen_defer) 5211 vfs_deferred_inactive(vp, lkflags); 5212 else 5213 VI_UNLOCK(vp); 5214 continue; 5215 } 5216 if (vget(vp, lkflags) == 0) { 5217 if ((vp->v_vflag & VV_NOSYNC) == 0) { 5218 if (flags == MNT_WAIT) 5219 vnode_pager_clean_sync(vp); 5220 else 5221 vnode_pager_clean_async(vp); 5222 } 5223 vput(vp); 5224 if (seen_defer) 5225 vdrop(vp); 5226 } else { 5227 if (seen_defer) 5228 vdefer_inactive_unlocked(vp); 5229 } 5230 } 5231 } 5232 5233 void 5234 vfs_periodic(struct mount *mp, int flags) 5235 { 5236 5237 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 5238 5239 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0) 5240 vfs_periodic_inactive(mp, flags); 5241 else 5242 vfs_periodic_msync_inactive(mp, flags); 5243 } 5244 5245 static void 5246 destroy_vpollinfo_free(struct vpollinfo *vi) 5247 { 5248 5249 knlist_destroy(&vi->vpi_selinfo.si_note); 5250 mtx_destroy(&vi->vpi_lock); 5251 free(vi, M_VNODEPOLL); 5252 } 5253 5254 static void 5255 destroy_vpollinfo(struct vpollinfo *vi) 5256 { 5257 KASSERT(TAILQ_EMPTY(&vi->vpi_inotify), 5258 ("%s: pollinfo %p has lingering watches", __func__, vi)); 5259 knlist_clear(&vi->vpi_selinfo.si_note, 1); 5260 seldrain(&vi->vpi_selinfo); 5261 destroy_vpollinfo_free(vi); 5262 } 5263 5264 /* 5265 * Initialize per-vnode helper structure to hold poll-related state. 5266 */ 5267 void 5268 v_addpollinfo(struct vnode *vp) 5269 { 5270 struct vpollinfo *vi; 5271 5272 if (atomic_load_ptr(&vp->v_pollinfo) != NULL) 5273 return; 5274 vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO); 5275 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 5276 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 5277 vfs_knlunlock, vfs_knl_assert_lock); 5278 TAILQ_INIT(&vi->vpi_inotify); 5279 VI_LOCK(vp); 5280 if (vp->v_pollinfo != NULL) { 5281 VI_UNLOCK(vp); 5282 destroy_vpollinfo_free(vi); 5283 return; 5284 } 5285 vp->v_pollinfo = vi; 5286 VI_UNLOCK(vp); 5287 } 5288 5289 /* 5290 * Record a process's interest in events which might happen to 5291 * a vnode. Because poll uses the historic select-style interface 5292 * internally, this routine serves as both the ``check for any 5293 * pending events'' and the ``record my interest in future events'' 5294 * functions. (These are done together, while the lock is held, 5295 * to avoid race conditions.) 5296 */ 5297 int 5298 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 5299 { 5300 5301 v_addpollinfo(vp); 5302 mtx_lock(&vp->v_pollinfo->vpi_lock); 5303 if (vp->v_pollinfo->vpi_revents & events) { 5304 /* 5305 * This leaves events we are not interested 5306 * in available for the other process which 5307 * which presumably had requested them 5308 * (otherwise they would never have been 5309 * recorded). 5310 */ 5311 events &= vp->v_pollinfo->vpi_revents; 5312 vp->v_pollinfo->vpi_revents &= ~events; 5313 5314 mtx_unlock(&vp->v_pollinfo->vpi_lock); 5315 return (events); 5316 } 5317 vp->v_pollinfo->vpi_events |= events; 5318 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 5319 mtx_unlock(&vp->v_pollinfo->vpi_lock); 5320 return (0); 5321 } 5322 5323 /* 5324 * Routine to create and manage a filesystem syncer vnode. 5325 */ 5326 #define sync_close ((int (*)(struct vop_close_args *))nullop) 5327 static int sync_fsync(struct vop_fsync_args *); 5328 static int sync_inactive(struct vop_inactive_args *); 5329 static int sync_reclaim(struct vop_reclaim_args *); 5330 5331 static struct vop_vector sync_vnodeops = { 5332 .vop_bypass = VOP_EOPNOTSUPP, 5333 .vop_close = sync_close, 5334 .vop_fsync = sync_fsync, 5335 .vop_getwritemount = vop_stdgetwritemount, 5336 .vop_inactive = sync_inactive, 5337 .vop_need_inactive = vop_stdneed_inactive, 5338 .vop_reclaim = sync_reclaim, 5339 .vop_lock1 = vop_stdlock, 5340 .vop_unlock = vop_stdunlock, 5341 .vop_islocked = vop_stdislocked, 5342 .vop_fplookup_vexec = VOP_EAGAIN, 5343 .vop_fplookup_symlink = VOP_EAGAIN, 5344 }; 5345 VFS_VOP_VECTOR_REGISTER(sync_vnodeops); 5346 5347 /* 5348 * Create a new filesystem syncer vnode for the specified mount point. 5349 */ 5350 void 5351 vfs_allocate_syncvnode(struct mount *mp) 5352 { 5353 struct vnode *vp; 5354 struct bufobj *bo; 5355 static long start, incr, next; 5356 int error; 5357 5358 /* Allocate a new vnode */ 5359 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 5360 if (error != 0) 5361 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 5362 vp->v_type = VNON; 5363 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5364 vp->v_vflag |= VV_FORCEINSMQ; 5365 error = insmntque1(vp, mp); 5366 if (error != 0) 5367 panic("vfs_allocate_syncvnode: insmntque() failed"); 5368 vp->v_vflag &= ~VV_FORCEINSMQ; 5369 vn_set_state(vp, VSTATE_CONSTRUCTED); 5370 VOP_UNLOCK(vp); 5371 /* 5372 * Place the vnode onto the syncer worklist. We attempt to 5373 * scatter them about on the list so that they will go off 5374 * at evenly distributed times even if all the filesystems 5375 * are mounted at once. 5376 */ 5377 next += incr; 5378 if (next == 0 || next > syncer_maxdelay) { 5379 start /= 2; 5380 incr /= 2; 5381 if (start == 0) { 5382 start = syncer_maxdelay / 2; 5383 incr = syncer_maxdelay; 5384 } 5385 next = start; 5386 } 5387 bo = &vp->v_bufobj; 5388 BO_LOCK(bo); 5389 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 5390 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 5391 mtx_lock(&sync_mtx); 5392 sync_vnode_count++; 5393 if (mp->mnt_syncer == NULL) { 5394 mp->mnt_syncer = vp; 5395 vp = NULL; 5396 } 5397 mtx_unlock(&sync_mtx); 5398 BO_UNLOCK(bo); 5399 if (vp != NULL) { 5400 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5401 vgone(vp); 5402 vput(vp); 5403 } 5404 } 5405 5406 void 5407 vfs_deallocate_syncvnode(struct mount *mp) 5408 { 5409 struct vnode *vp; 5410 5411 mtx_lock(&sync_mtx); 5412 vp = mp->mnt_syncer; 5413 if (vp != NULL) 5414 mp->mnt_syncer = NULL; 5415 mtx_unlock(&sync_mtx); 5416 if (vp != NULL) 5417 vrele(vp); 5418 } 5419 5420 /* 5421 * Do a lazy sync of the filesystem. 5422 */ 5423 static int 5424 sync_fsync(struct vop_fsync_args *ap) 5425 { 5426 struct vnode *syncvp = ap->a_vp; 5427 struct mount *mp = syncvp->v_mount; 5428 int error, save; 5429 struct bufobj *bo; 5430 5431 /* 5432 * We only need to do something if this is a lazy evaluation. 5433 */ 5434 if (ap->a_waitfor != MNT_LAZY) 5435 return (0); 5436 5437 /* 5438 * Move ourselves to the back of the sync list. 5439 */ 5440 bo = &syncvp->v_bufobj; 5441 BO_LOCK(bo); 5442 vn_syncer_add_to_worklist(bo, syncdelay); 5443 BO_UNLOCK(bo); 5444 5445 /* 5446 * Walk the list of vnodes pushing all that are dirty and 5447 * not already on the sync list. 5448 */ 5449 if (vfs_busy(mp, MBF_NOWAIT) != 0) 5450 return (0); 5451 VOP_UNLOCK(syncvp); 5452 save = curthread_pflags_set(TDP_SYNCIO); 5453 /* 5454 * The filesystem at hand may be idle with free vnodes stored in the 5455 * batch. Return them instead of letting them stay there indefinitely. 5456 */ 5457 vfs_periodic(mp, MNT_NOWAIT); 5458 error = VFS_SYNC(mp, MNT_LAZY); 5459 curthread_pflags_restore(save); 5460 vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY); 5461 vfs_unbusy(mp); 5462 return (error); 5463 } 5464 5465 /* 5466 * The syncer vnode is no referenced. 5467 */ 5468 static int 5469 sync_inactive(struct vop_inactive_args *ap) 5470 { 5471 5472 vgone(ap->a_vp); 5473 return (0); 5474 } 5475 5476 /* 5477 * The syncer vnode is no longer needed and is being decommissioned. 5478 * 5479 * Modifications to the worklist must be protected by sync_mtx. 5480 */ 5481 static int 5482 sync_reclaim(struct vop_reclaim_args *ap) 5483 { 5484 struct vnode *vp = ap->a_vp; 5485 struct bufobj *bo; 5486 5487 bo = &vp->v_bufobj; 5488 BO_LOCK(bo); 5489 mtx_lock(&sync_mtx); 5490 if (vp->v_mount->mnt_syncer == vp) 5491 vp->v_mount->mnt_syncer = NULL; 5492 if (bo->bo_flag & BO_ONWORKLST) { 5493 LIST_REMOVE(bo, bo_synclist); 5494 syncer_worklist_len--; 5495 sync_vnode_count--; 5496 bo->bo_flag &= ~BO_ONWORKLST; 5497 } 5498 mtx_unlock(&sync_mtx); 5499 BO_UNLOCK(bo); 5500 5501 return (0); 5502 } 5503 5504 int 5505 vn_need_pageq_flush(struct vnode *vp) 5506 { 5507 struct vm_object *obj; 5508 5509 obj = vp->v_object; 5510 return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 5511 vm_object_mightbedirty(obj)); 5512 } 5513 5514 /* 5515 * Check if vnode represents a disk device 5516 */ 5517 bool 5518 vn_isdisk_error(struct vnode *vp, int *errp) 5519 { 5520 int error; 5521 5522 if (vp->v_type != VCHR) { 5523 error = ENOTBLK; 5524 goto out; 5525 } 5526 error = 0; 5527 dev_lock(); 5528 if (vp->v_rdev == NULL) 5529 error = ENXIO; 5530 else if (vp->v_rdev->si_devsw == NULL) 5531 error = ENXIO; 5532 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 5533 error = ENOTBLK; 5534 dev_unlock(); 5535 out: 5536 *errp = error; 5537 return (error == 0); 5538 } 5539 5540 bool 5541 vn_isdisk(struct vnode *vp) 5542 { 5543 int error; 5544 5545 return (vn_isdisk_error(vp, &error)); 5546 } 5547 5548 /* 5549 * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see 5550 * the comment above cache_fplookup for details. 5551 */ 5552 int 5553 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred) 5554 { 5555 int error; 5556 5557 VFS_SMR_ASSERT_ENTERED(); 5558 5559 /* Check the owner. */ 5560 if (cred->cr_uid == file_uid) { 5561 if (file_mode & S_IXUSR) 5562 return (0); 5563 goto out_error; 5564 } 5565 5566 /* Otherwise, check the groups (first match) */ 5567 if (groupmember(file_gid, cred)) { 5568 if (file_mode & S_IXGRP) 5569 return (0); 5570 goto out_error; 5571 } 5572 5573 /* Otherwise, check everyone else. */ 5574 if (file_mode & S_IXOTH) 5575 return (0); 5576 out_error: 5577 /* 5578 * Permission check failed, but it is possible denial will get overwritten 5579 * (e.g., when root is traversing through a 700 directory owned by someone 5580 * else). 5581 * 5582 * vaccess() calls priv_check_cred which in turn can descent into MAC 5583 * modules overriding this result. It's quite unclear what semantics 5584 * are allowed for them to operate, thus for safety we don't call them 5585 * from within the SMR section. This also means if any such modules 5586 * are present, we have to let the regular lookup decide. 5587 */ 5588 error = priv_check_cred_vfs_lookup_nomac(cred); 5589 switch (error) { 5590 case 0: 5591 return (0); 5592 case EAGAIN: 5593 /* 5594 * MAC modules present. 5595 */ 5596 return (EAGAIN); 5597 case EPERM: 5598 return (EACCES); 5599 default: 5600 return (error); 5601 } 5602 } 5603 5604 /* 5605 * Common filesystem object access control check routine. Accepts a 5606 * vnode's type, "mode", uid and gid, requested access mode, and credentials. 5607 * Returns 0 on success, or an errno on failure. 5608 */ 5609 int 5610 vaccess(__enum_uint8(vtype) type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 5611 accmode_t accmode, struct ucred *cred) 5612 { 5613 accmode_t dac_granted; 5614 accmode_t priv_granted; 5615 5616 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 5617 ("invalid bit in accmode")); 5618 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 5619 ("VAPPEND without VWRITE")); 5620 5621 /* 5622 * Look for a normal, non-privileged way to access the file/directory 5623 * as requested. If it exists, go with that. 5624 */ 5625 5626 dac_granted = 0; 5627 5628 /* Check the owner. */ 5629 if (cred->cr_uid == file_uid) { 5630 dac_granted |= VADMIN; 5631 if (file_mode & S_IXUSR) 5632 dac_granted |= VEXEC; 5633 if (file_mode & S_IRUSR) 5634 dac_granted |= VREAD; 5635 if (file_mode & S_IWUSR) 5636 dac_granted |= (VWRITE | VAPPEND); 5637 5638 if ((accmode & dac_granted) == accmode) 5639 return (0); 5640 5641 goto privcheck; 5642 } 5643 5644 /* Otherwise, check the groups (first match) */ 5645 if (groupmember(file_gid, cred)) { 5646 if (file_mode & S_IXGRP) 5647 dac_granted |= VEXEC; 5648 if (file_mode & S_IRGRP) 5649 dac_granted |= VREAD; 5650 if (file_mode & S_IWGRP) 5651 dac_granted |= (VWRITE | VAPPEND); 5652 5653 if ((accmode & dac_granted) == accmode) 5654 return (0); 5655 5656 goto privcheck; 5657 } 5658 5659 /* Otherwise, check everyone else. */ 5660 if (file_mode & S_IXOTH) 5661 dac_granted |= VEXEC; 5662 if (file_mode & S_IROTH) 5663 dac_granted |= VREAD; 5664 if (file_mode & S_IWOTH) 5665 dac_granted |= (VWRITE | VAPPEND); 5666 if ((accmode & dac_granted) == accmode) 5667 return (0); 5668 5669 privcheck: 5670 /* 5671 * Build a privilege mask to determine if the set of privileges 5672 * satisfies the requirements when combined with the granted mask 5673 * from above. For each privilege, if the privilege is required, 5674 * bitwise or the request type onto the priv_granted mask. 5675 */ 5676 priv_granted = 0; 5677 5678 if (type == VDIR) { 5679 /* 5680 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 5681 * requests, instead of PRIV_VFS_EXEC. 5682 */ 5683 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5684 !priv_check_cred(cred, PRIV_VFS_LOOKUP)) 5685 priv_granted |= VEXEC; 5686 } else { 5687 /* 5688 * Ensure that at least one execute bit is on. Otherwise, 5689 * a privileged user will always succeed, and we don't want 5690 * this to happen unless the file really is executable. 5691 */ 5692 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5693 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 5694 !priv_check_cred(cred, PRIV_VFS_EXEC)) 5695 priv_granted |= VEXEC; 5696 } 5697 5698 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 5699 !priv_check_cred(cred, PRIV_VFS_READ)) 5700 priv_granted |= VREAD; 5701 5702 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 5703 !priv_check_cred(cred, PRIV_VFS_WRITE)) 5704 priv_granted |= (VWRITE | VAPPEND); 5705 5706 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 5707 !priv_check_cred(cred, PRIV_VFS_ADMIN)) 5708 priv_granted |= VADMIN; 5709 5710 if ((accmode & (priv_granted | dac_granted)) == accmode) { 5711 return (0); 5712 } 5713 5714 return ((accmode & VADMIN) ? EPERM : EACCES); 5715 } 5716 5717 /* 5718 * Credential check based on process requesting service, and per-attribute 5719 * permissions. 5720 */ 5721 int 5722 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 5723 struct thread *td, accmode_t accmode) 5724 { 5725 5726 /* 5727 * Kernel-invoked always succeeds. 5728 */ 5729 if (cred == NOCRED) 5730 return (0); 5731 5732 /* 5733 * Do not allow privileged processes in jail to directly manipulate 5734 * system attributes. 5735 */ 5736 switch (attrnamespace) { 5737 case EXTATTR_NAMESPACE_SYSTEM: 5738 /* Potentially should be: return (EPERM); */ 5739 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM)); 5740 case EXTATTR_NAMESPACE_USER: 5741 return (VOP_ACCESS(vp, accmode, cred, td)); 5742 default: 5743 return (EPERM); 5744 } 5745 } 5746 5747 #ifdef INVARIANTS 5748 void 5749 assert_vi_locked(struct vnode *vp, const char *str) 5750 { 5751 VNASSERT(mtx_owned(VI_MTX(vp)), vp, 5752 ("%s: vnode interlock is not locked but should be", str)); 5753 } 5754 5755 void 5756 assert_vi_unlocked(struct vnode *vp, const char *str) 5757 { 5758 VNASSERT(!mtx_owned(VI_MTX(vp)), vp, 5759 ("%s: vnode interlock is locked but should not be", str)); 5760 } 5761 5762 void 5763 assert_vop_locked(struct vnode *vp, const char *str) 5764 { 5765 bool locked; 5766 5767 if (KERNEL_PANICKED() || vp == NULL) 5768 return; 5769 5770 #ifdef WITNESS 5771 locked = !((vp->v_irflag & VIRF_CROSSMP) == 0 && 5772 witness_is_owned(&vp->v_vnlock->lock_object) == -1); 5773 #else 5774 int state = VOP_ISLOCKED(vp); 5775 locked = state != 0 && state != LK_EXCLOTHER; 5776 #endif 5777 VNASSERT(locked, vp, ("%s: vnode is not locked but should be", str)); 5778 } 5779 5780 void 5781 assert_vop_unlocked(struct vnode *vp, const char *str) 5782 { 5783 bool locked; 5784 5785 if (KERNEL_PANICKED() || vp == NULL) 5786 return; 5787 5788 #ifdef WITNESS 5789 locked = (vp->v_irflag & VIRF_CROSSMP) == 0 && 5790 witness_is_owned(&vp->v_vnlock->lock_object) == 1; 5791 #else 5792 locked = VOP_ISLOCKED(vp) == LK_EXCLUSIVE; 5793 #endif 5794 VNASSERT(!locked, vp, ("%s: vnode is locked but should not be", str)); 5795 } 5796 5797 void 5798 assert_vop_elocked(struct vnode *vp, const char *str) 5799 { 5800 bool locked; 5801 5802 if (KERNEL_PANICKED() || vp == NULL) 5803 return; 5804 5805 locked = VOP_ISLOCKED(vp) == LK_EXCLUSIVE; 5806 VNASSERT(locked, vp, 5807 ("%s: vnode is not exclusive locked but should be", str)); 5808 } 5809 #endif /* INVARIANTS */ 5810 5811 void 5812 vop_rename_fail(struct vop_rename_args *ap) 5813 { 5814 5815 if (ap->a_tvp != NULL) 5816 vput(ap->a_tvp); 5817 if (ap->a_tdvp == ap->a_tvp) 5818 vrele(ap->a_tdvp); 5819 else 5820 vput(ap->a_tdvp); 5821 vrele(ap->a_fdvp); 5822 vrele(ap->a_fvp); 5823 } 5824 5825 void 5826 vop_rename_pre(void *ap) 5827 { 5828 struct vop_rename_args *a = ap; 5829 5830 #ifdef INVARIANTS 5831 struct mount *tmp; 5832 5833 if (a->a_tvp) 5834 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 5835 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 5836 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 5837 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 5838 5839 /* Check the source (from). */ 5840 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 5841 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 5842 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 5843 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 5844 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 5845 5846 /* Check the target. */ 5847 if (a->a_tvp) 5848 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 5849 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 5850 5851 tmp = NULL; 5852 VOP_GETWRITEMOUNT(a->a_tdvp, &tmp); 5853 lockmgr_assert(&tmp->mnt_renamelock, KA_XLOCKED); 5854 vfs_rel(tmp); 5855 #endif 5856 /* 5857 * It may be tempting to add vn_seqc_write_begin/end calls here and 5858 * in vop_rename_post but that's not going to work out since some 5859 * filesystems relookup vnodes mid-rename. This is probably a bug. 5860 * 5861 * For now filesystems are expected to do the relevant calls after they 5862 * decide what vnodes to operate on. 5863 */ 5864 if (a->a_tdvp != a->a_fdvp) 5865 vhold(a->a_fdvp); 5866 if (a->a_tvp != a->a_fvp) 5867 vhold(a->a_fvp); 5868 vhold(a->a_tdvp); 5869 if (a->a_tvp) 5870 vhold(a->a_tvp); 5871 } 5872 5873 #ifdef INVARIANTS 5874 void 5875 vop_fplookup_vexec_debugpre(void *ap __unused) 5876 { 5877 5878 VFS_SMR_ASSERT_ENTERED(); 5879 } 5880 5881 void 5882 vop_fplookup_vexec_debugpost(void *ap, int rc) 5883 { 5884 struct vop_fplookup_vexec_args *a; 5885 struct vnode *vp; 5886 5887 a = ap; 5888 vp = a->a_vp; 5889 5890 VFS_SMR_ASSERT_ENTERED(); 5891 if (rc == EOPNOTSUPP) 5892 VNPASS(VN_IS_DOOMED(vp), vp); 5893 } 5894 5895 void 5896 vop_fplookup_symlink_debugpre(void *ap __unused) 5897 { 5898 5899 VFS_SMR_ASSERT_ENTERED(); 5900 } 5901 5902 void 5903 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused) 5904 { 5905 5906 VFS_SMR_ASSERT_ENTERED(); 5907 } 5908 5909 static void 5910 vop_fsync_debugprepost(struct vnode *vp, const char *name) 5911 { 5912 struct mount *mp; 5913 5914 if (vp->v_type == VCHR) 5915 ; 5916 /* 5917 * The shared vs. exclusive locking policy for fsync() 5918 * is actually determined by vp's write mount as indicated 5919 * by VOP_GETWRITEMOUNT(), which for stacked filesystems 5920 * may not be the same as vp->v_mount. However, if the 5921 * underlying filesystem which really handles the fsync() 5922 * supports shared locking, the stacked filesystem must also 5923 * be prepared for its VOP_FSYNC() operation to be called 5924 * with only a shared lock. On the other hand, if the 5925 * stacked filesystem claims support for shared write 5926 * locking but the underlying filesystem does not, and the 5927 * caller incorrectly uses a shared lock, this condition 5928 * should still be caught when the stacked filesystem 5929 * invokes VOP_FSYNC() on the underlying filesystem. 5930 */ 5931 else { 5932 mp = NULL; 5933 VOP_GETWRITEMOUNT(vp, &mp); 5934 if (vn_lktype_write(mp, vp) == LK_SHARED) 5935 ASSERT_VOP_LOCKED(vp, name); 5936 else 5937 ASSERT_VOP_ELOCKED(vp, name); 5938 if (mp != NULL) 5939 vfs_rel(mp); 5940 } 5941 } 5942 5943 void 5944 vop_fsync_debugpre(void *a) 5945 { 5946 struct vop_fsync_args *ap; 5947 5948 ap = a; 5949 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5950 } 5951 5952 void 5953 vop_fsync_debugpost(void *a, int rc __unused) 5954 { 5955 struct vop_fsync_args *ap; 5956 5957 ap = a; 5958 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5959 } 5960 5961 void 5962 vop_fdatasync_debugpre(void *a) 5963 { 5964 struct vop_fdatasync_args *ap; 5965 5966 ap = a; 5967 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5968 } 5969 5970 void 5971 vop_fdatasync_debugpost(void *a, int rc __unused) 5972 { 5973 struct vop_fdatasync_args *ap; 5974 5975 ap = a; 5976 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5977 } 5978 5979 void 5980 vop_strategy_debugpre(void *ap) 5981 { 5982 struct vop_strategy_args *a; 5983 struct buf *bp; 5984 5985 a = ap; 5986 bp = a->a_bp; 5987 5988 /* 5989 * Cluster ops lock their component buffers but not the IO container. 5990 */ 5991 if ((bp->b_flags & B_CLUSTER) != 0) 5992 return; 5993 5994 BUF_ASSERT_LOCKED(bp); 5995 } 5996 5997 void 5998 vop_lock_debugpre(void *ap) 5999 { 6000 struct vop_lock1_args *a = ap; 6001 6002 if ((a->a_flags & LK_INTERLOCK) == 0) 6003 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 6004 else 6005 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 6006 } 6007 6008 void 6009 vop_lock_debugpost(void *ap, int rc) 6010 { 6011 struct vop_lock1_args *a = ap; 6012 6013 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 6014 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) 6015 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 6016 } 6017 6018 void 6019 vop_unlock_debugpre(void *ap) 6020 { 6021 struct vop_unlock_args *a = ap; 6022 struct vnode *vp = a->a_vp; 6023 6024 VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp); 6025 ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK"); 6026 } 6027 6028 void 6029 vop_need_inactive_debugpre(void *ap) 6030 { 6031 struct vop_need_inactive_args *a = ap; 6032 6033 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 6034 } 6035 6036 void 6037 vop_need_inactive_debugpost(void *ap, int rc) 6038 { 6039 struct vop_need_inactive_args *a = ap; 6040 6041 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 6042 } 6043 #endif /* INVARIANTS */ 6044 6045 void 6046 vop_allocate_post(void *ap, int rc) 6047 { 6048 struct vop_allocate_args *a; 6049 6050 a = ap; 6051 if (rc == 0) 6052 INOTIFY(a->a_vp, IN_MODIFY); 6053 } 6054 6055 void 6056 vop_copy_file_range_post(void *ap, int rc) 6057 { 6058 struct vop_copy_file_range_args *a; 6059 6060 a = ap; 6061 if (rc == 0) { 6062 INOTIFY(a->a_invp, IN_ACCESS); 6063 INOTIFY(a->a_outvp, IN_MODIFY); 6064 } 6065 } 6066 6067 void 6068 vop_create_pre(void *ap) 6069 { 6070 struct vop_create_args *a; 6071 struct vnode *dvp; 6072 6073 a = ap; 6074 dvp = a->a_dvp; 6075 vn_seqc_write_begin(dvp); 6076 } 6077 6078 void 6079 vop_create_post(void *ap, int rc) 6080 { 6081 struct vop_create_args *a; 6082 struct vnode *dvp; 6083 6084 a = ap; 6085 dvp = a->a_dvp; 6086 vn_seqc_write_end(dvp); 6087 if (!rc) { 6088 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6089 INOTIFY_NAME(*a->a_vpp, dvp, a->a_cnp, IN_CREATE); 6090 } 6091 } 6092 6093 void 6094 vop_deallocate_post(void *ap, int rc) 6095 { 6096 struct vop_deallocate_args *a; 6097 6098 a = ap; 6099 if (rc == 0) 6100 INOTIFY(a->a_vp, IN_MODIFY); 6101 } 6102 6103 void 6104 vop_whiteout_pre(void *ap) 6105 { 6106 struct vop_whiteout_args *a; 6107 struct vnode *dvp; 6108 6109 a = ap; 6110 dvp = a->a_dvp; 6111 vn_seqc_write_begin(dvp); 6112 } 6113 6114 void 6115 vop_whiteout_post(void *ap, int rc) 6116 { 6117 struct vop_whiteout_args *a; 6118 struct vnode *dvp; 6119 6120 a = ap; 6121 dvp = a->a_dvp; 6122 vn_seqc_write_end(dvp); 6123 } 6124 6125 void 6126 vop_deleteextattr_pre(void *ap) 6127 { 6128 struct vop_deleteextattr_args *a; 6129 struct vnode *vp; 6130 6131 a = ap; 6132 vp = a->a_vp; 6133 vn_seqc_write_begin(vp); 6134 } 6135 6136 void 6137 vop_deleteextattr_post(void *ap, int rc) 6138 { 6139 struct vop_deleteextattr_args *a; 6140 struct vnode *vp; 6141 6142 a = ap; 6143 vp = a->a_vp; 6144 vn_seqc_write_end(vp); 6145 if (!rc) { 6146 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 6147 INOTIFY(vp, IN_ATTRIB); 6148 } 6149 } 6150 6151 void 6152 vop_link_pre(void *ap) 6153 { 6154 struct vop_link_args *a; 6155 struct vnode *vp, *tdvp; 6156 6157 a = ap; 6158 vp = a->a_vp; 6159 tdvp = a->a_tdvp; 6160 vn_seqc_write_begin(vp); 6161 vn_seqc_write_begin(tdvp); 6162 } 6163 6164 void 6165 vop_link_post(void *ap, int rc) 6166 { 6167 struct vop_link_args *a; 6168 struct vnode *vp, *tdvp; 6169 6170 a = ap; 6171 vp = a->a_vp; 6172 tdvp = a->a_tdvp; 6173 vn_seqc_write_end(vp); 6174 vn_seqc_write_end(tdvp); 6175 if (!rc) { 6176 VFS_KNOTE_LOCKED(vp, NOTE_LINK); 6177 VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE); 6178 INOTIFY_NAME(vp, tdvp, a->a_cnp, _IN_ATTRIB_LINKCOUNT); 6179 INOTIFY_NAME(vp, tdvp, a->a_cnp, IN_CREATE); 6180 } 6181 } 6182 6183 void 6184 vop_mkdir_pre(void *ap) 6185 { 6186 struct vop_mkdir_args *a; 6187 struct vnode *dvp; 6188 6189 a = ap; 6190 dvp = a->a_dvp; 6191 vn_seqc_write_begin(dvp); 6192 } 6193 6194 void 6195 vop_mkdir_post(void *ap, int rc) 6196 { 6197 struct vop_mkdir_args *a; 6198 struct vnode *dvp; 6199 6200 a = ap; 6201 dvp = a->a_dvp; 6202 vn_seqc_write_end(dvp); 6203 if (!rc) { 6204 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 6205 INOTIFY_NAME(*a->a_vpp, dvp, a->a_cnp, IN_CREATE); 6206 } 6207 } 6208 6209 #ifdef INVARIANTS 6210 void 6211 vop_mkdir_debugpost(void *ap, int rc) 6212 { 6213 struct vop_mkdir_args *a; 6214 6215 a = ap; 6216 if (!rc) 6217 cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp); 6218 } 6219 #endif 6220 6221 void 6222 vop_mknod_pre(void *ap) 6223 { 6224 struct vop_mknod_args *a; 6225 struct vnode *dvp; 6226 6227 a = ap; 6228 dvp = a->a_dvp; 6229 vn_seqc_write_begin(dvp); 6230 } 6231 6232 void 6233 vop_mknod_post(void *ap, int rc) 6234 { 6235 struct vop_mknod_args *a; 6236 struct vnode *dvp; 6237 6238 a = ap; 6239 dvp = a->a_dvp; 6240 vn_seqc_write_end(dvp); 6241 if (!rc) { 6242 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6243 INOTIFY_NAME(*a->a_vpp, dvp, a->a_cnp, IN_CREATE); 6244 } 6245 } 6246 6247 void 6248 vop_reclaim_post(void *ap, int rc) 6249 { 6250 struct vop_reclaim_args *a; 6251 struct vnode *vp; 6252 6253 a = ap; 6254 vp = a->a_vp; 6255 ASSERT_VOP_IN_SEQC(vp); 6256 if (!rc) { 6257 VFS_KNOTE_LOCKED(vp, NOTE_REVOKE); 6258 INOTIFY_REVOKE(vp); 6259 } 6260 } 6261 6262 void 6263 vop_remove_pre(void *ap) 6264 { 6265 struct vop_remove_args *a; 6266 struct vnode *dvp, *vp; 6267 6268 a = ap; 6269 dvp = a->a_dvp; 6270 vp = a->a_vp; 6271 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_UNLINK); 6272 vn_seqc_write_begin(dvp); 6273 vn_seqc_write_begin(vp); 6274 } 6275 6276 void 6277 vop_remove_post(void *ap, int rc) 6278 { 6279 struct vop_remove_args *a; 6280 struct vnode *dvp, *vp; 6281 6282 a = ap; 6283 dvp = a->a_dvp; 6284 vp = a->a_vp; 6285 vn_seqc_write_end(dvp); 6286 vn_seqc_write_end(vp); 6287 if (!rc) { 6288 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6289 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 6290 INOTIFY_NAME(vp, dvp, a->a_cnp, _IN_ATTRIB_LINKCOUNT); 6291 INOTIFY_NAME(vp, dvp, a->a_cnp, IN_DELETE); 6292 } 6293 } 6294 6295 void 6296 vop_rename_post(void *ap, int rc) 6297 { 6298 struct vop_rename_args *a = ap; 6299 long hint; 6300 6301 if (!rc) { 6302 hint = NOTE_WRITE; 6303 if (a->a_fdvp == a->a_tdvp) { 6304 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) 6305 hint |= NOTE_LINK; 6306 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 6307 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 6308 } else { 6309 hint |= NOTE_EXTEND; 6310 if (a->a_fvp->v_type == VDIR) 6311 hint |= NOTE_LINK; 6312 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 6313 6314 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && 6315 a->a_tvp->v_type == VDIR) 6316 hint &= ~NOTE_LINK; 6317 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 6318 } 6319 6320 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 6321 if (a->a_tvp) 6322 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 6323 INOTIFY_MOVE(a->a_fvp, a->a_fdvp, a->a_fcnp, a->a_tvp, 6324 a->a_tdvp, a->a_tcnp); 6325 } 6326 if (a->a_tdvp != a->a_fdvp) 6327 vdrop(a->a_fdvp); 6328 if (a->a_tvp != a->a_fvp) 6329 vdrop(a->a_fvp); 6330 vdrop(a->a_tdvp); 6331 if (a->a_tvp) 6332 vdrop(a->a_tvp); 6333 } 6334 6335 void 6336 vop_rmdir_pre(void *ap) 6337 { 6338 struct vop_rmdir_args *a; 6339 struct vnode *dvp, *vp; 6340 6341 a = ap; 6342 dvp = a->a_dvp; 6343 vp = a->a_vp; 6344 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_UNLINK); 6345 vn_seqc_write_begin(dvp); 6346 vn_seqc_write_begin(vp); 6347 } 6348 6349 void 6350 vop_rmdir_post(void *ap, int rc) 6351 { 6352 struct vop_rmdir_args *a; 6353 struct vnode *dvp, *vp; 6354 6355 a = ap; 6356 dvp = a->a_dvp; 6357 vp = a->a_vp; 6358 vn_seqc_write_end(dvp); 6359 vn_seqc_write_end(vp); 6360 if (!rc) { 6361 vp->v_vflag |= VV_UNLINKED; 6362 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 6363 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 6364 INOTIFY_NAME(vp, dvp, a->a_cnp, IN_DELETE); 6365 } 6366 } 6367 6368 void 6369 vop_setattr_pre(void *ap) 6370 { 6371 struct vop_setattr_args *a; 6372 struct vnode *vp; 6373 6374 a = ap; 6375 vp = a->a_vp; 6376 vn_seqc_write_begin(vp); 6377 } 6378 6379 void 6380 vop_setattr_post(void *ap, int rc) 6381 { 6382 struct vop_setattr_args *a; 6383 struct vnode *vp; 6384 6385 a = ap; 6386 vp = a->a_vp; 6387 vn_seqc_write_end(vp); 6388 if (!rc) { 6389 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 6390 INOTIFY(vp, IN_ATTRIB); 6391 } 6392 } 6393 6394 void 6395 vop_setacl_pre(void *ap) 6396 { 6397 struct vop_setacl_args *a; 6398 struct vnode *vp; 6399 6400 a = ap; 6401 vp = a->a_vp; 6402 vn_seqc_write_begin(vp); 6403 } 6404 6405 void 6406 vop_setacl_post(void *ap, int rc __unused) 6407 { 6408 struct vop_setacl_args *a; 6409 struct vnode *vp; 6410 6411 a = ap; 6412 vp = a->a_vp; 6413 vn_seqc_write_end(vp); 6414 } 6415 6416 void 6417 vop_setextattr_pre(void *ap) 6418 { 6419 struct vop_setextattr_args *a; 6420 struct vnode *vp; 6421 6422 a = ap; 6423 vp = a->a_vp; 6424 vn_seqc_write_begin(vp); 6425 } 6426 6427 void 6428 vop_setextattr_post(void *ap, int rc) 6429 { 6430 struct vop_setextattr_args *a; 6431 struct vnode *vp; 6432 6433 a = ap; 6434 vp = a->a_vp; 6435 vn_seqc_write_end(vp); 6436 if (!rc) { 6437 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 6438 INOTIFY(vp, IN_ATTRIB); 6439 } 6440 } 6441 6442 void 6443 vop_symlink_pre(void *ap) 6444 { 6445 struct vop_symlink_args *a; 6446 struct vnode *dvp; 6447 6448 a = ap; 6449 dvp = a->a_dvp; 6450 vn_seqc_write_begin(dvp); 6451 } 6452 6453 void 6454 vop_symlink_post(void *ap, int rc) 6455 { 6456 struct vop_symlink_args *a; 6457 struct vnode *dvp; 6458 6459 a = ap; 6460 dvp = a->a_dvp; 6461 vn_seqc_write_end(dvp); 6462 if (!rc) { 6463 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6464 INOTIFY_NAME(*a->a_vpp, dvp, a->a_cnp, IN_CREATE); 6465 } 6466 } 6467 6468 void 6469 vop_open_post(void *ap, int rc) 6470 { 6471 struct vop_open_args *a = ap; 6472 6473 if (!rc) { 6474 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); 6475 INOTIFY(a->a_vp, IN_OPEN); 6476 } 6477 } 6478 6479 void 6480 vop_close_post(void *ap, int rc) 6481 { 6482 struct vop_close_args *a = ap; 6483 6484 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ 6485 !VN_IS_DOOMED(a->a_vp))) { 6486 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 6487 NOTE_CLOSE_WRITE : NOTE_CLOSE); 6488 INOTIFY(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 6489 IN_CLOSE_WRITE : IN_CLOSE_NOWRITE); 6490 } 6491 } 6492 6493 void 6494 vop_read_post(void *ap, int rc) 6495 { 6496 struct vop_read_args *a = ap; 6497 6498 if (!rc) { 6499 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 6500 INOTIFY(a->a_vp, IN_ACCESS); 6501 } 6502 } 6503 6504 void 6505 vop_read_pgcache_post(void *ap, int rc) 6506 { 6507 struct vop_read_pgcache_args *a = ap; 6508 6509 if (!rc) 6510 VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ); 6511 } 6512 6513 static struct knlist fs_knlist; 6514 6515 static void 6516 vfs_event_init(void *arg) 6517 { 6518 knlist_init_mtx(&fs_knlist, NULL); 6519 } 6520 /* XXX - correct order? */ 6521 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 6522 6523 void 6524 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 6525 { 6526 6527 KNOTE_UNLOCKED(&fs_knlist, event); 6528 } 6529 6530 static int filt_fsattach(struct knote *kn); 6531 static void filt_fsdetach(struct knote *kn); 6532 static int filt_fsevent(struct knote *kn, long hint); 6533 6534 const struct filterops fs_filtops = { 6535 .f_isfd = 0, 6536 .f_attach = filt_fsattach, 6537 .f_detach = filt_fsdetach, 6538 .f_event = filt_fsevent, 6539 }; 6540 6541 static int 6542 filt_fsattach(struct knote *kn) 6543 { 6544 6545 kn->kn_flags |= EV_CLEAR; 6546 knlist_add(&fs_knlist, kn, 0); 6547 return (0); 6548 } 6549 6550 static void 6551 filt_fsdetach(struct knote *kn) 6552 { 6553 6554 knlist_remove(&fs_knlist, kn, 0); 6555 } 6556 6557 static int 6558 filt_fsevent(struct knote *kn, long hint) 6559 { 6560 6561 kn->kn_fflags |= kn->kn_sfflags & hint; 6562 6563 return (kn->kn_fflags != 0); 6564 } 6565 6566 static int 6567 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 6568 { 6569 struct vfsidctl vc; 6570 int error; 6571 struct mount *mp; 6572 6573 if (req->newptr == NULL) 6574 return (EINVAL); 6575 error = SYSCTL_IN(req, &vc, sizeof(vc)); 6576 if (error) 6577 return (error); 6578 if (vc.vc_vers != VFS_CTL_VERS1) 6579 return (EINVAL); 6580 mp = vfs_getvfs(&vc.vc_fsid); 6581 if (mp == NULL) 6582 return (ENOENT); 6583 /* ensure that a specific sysctl goes to the right filesystem. */ 6584 if (strcmp(vc.vc_fstypename, "*") != 0 && 6585 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 6586 vfs_rel(mp); 6587 return (EINVAL); 6588 } 6589 VCTLTOREQ(&vc, req); 6590 error = VFS_SYSCTL(mp, vc.vc_op, req); 6591 vfs_rel(mp); 6592 return (error); 6593 } 6594 6595 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR, 6596 NULL, 0, sysctl_vfs_ctl, "", 6597 "Sysctl by fsid"); 6598 6599 /* 6600 * Function to initialize a va_filerev field sensibly. 6601 * XXX: Wouldn't a random number make a lot more sense ?? 6602 */ 6603 u_quad_t 6604 init_va_filerev(void) 6605 { 6606 struct bintime bt; 6607 6608 getbinuptime(&bt); 6609 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 6610 } 6611 6612 static int filt_vfsread(struct knote *kn, long hint); 6613 static int filt_vfswrite(struct knote *kn, long hint); 6614 static int filt_vfsvnode(struct knote *kn, long hint); 6615 static void filt_vfsdetach(struct knote *kn); 6616 static int filt_vfsdump(struct proc *p, struct knote *kn, 6617 struct kinfo_knote *kin); 6618 6619 static const struct filterops vfsread_filtops = { 6620 .f_isfd = 1, 6621 .f_detach = filt_vfsdetach, 6622 .f_event = filt_vfsread, 6623 .f_userdump = filt_vfsdump, 6624 }; 6625 static const struct filterops vfswrite_filtops = { 6626 .f_isfd = 1, 6627 .f_detach = filt_vfsdetach, 6628 .f_event = filt_vfswrite, 6629 .f_userdump = filt_vfsdump, 6630 }; 6631 static const struct filterops vfsvnode_filtops = { 6632 .f_isfd = 1, 6633 .f_detach = filt_vfsdetach, 6634 .f_event = filt_vfsvnode, 6635 .f_userdump = filt_vfsdump, 6636 }; 6637 6638 static void 6639 vfs_knllock(void *arg) 6640 { 6641 struct vnode *vp = arg; 6642 6643 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 6644 } 6645 6646 static void 6647 vfs_knlunlock(void *arg) 6648 { 6649 struct vnode *vp = arg; 6650 6651 VOP_UNLOCK(vp); 6652 } 6653 6654 static void 6655 vfs_knl_assert_lock(void *arg, int what) 6656 { 6657 #ifdef INVARIANTS 6658 struct vnode *vp = arg; 6659 6660 if (what == LA_LOCKED) 6661 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 6662 else 6663 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 6664 #endif 6665 } 6666 6667 int 6668 vfs_kqfilter(struct vop_kqfilter_args *ap) 6669 { 6670 struct vnode *vp = ap->a_vp; 6671 struct knote *kn = ap->a_kn; 6672 struct knlist *knl; 6673 6674 KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ && 6675 kn->kn_filter != EVFILT_WRITE), 6676 ("READ/WRITE filter on a FIFO leaked through")); 6677 switch (kn->kn_filter) { 6678 case EVFILT_READ: 6679 kn->kn_fop = &vfsread_filtops; 6680 break; 6681 case EVFILT_WRITE: 6682 kn->kn_fop = &vfswrite_filtops; 6683 break; 6684 case EVFILT_VNODE: 6685 kn->kn_fop = &vfsvnode_filtops; 6686 break; 6687 default: 6688 return (EINVAL); 6689 } 6690 6691 kn->kn_hook = (caddr_t)vp; 6692 6693 v_addpollinfo(vp); 6694 if (vp->v_pollinfo == NULL) 6695 return (ENOMEM); 6696 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 6697 vhold(vp); 6698 knlist_add(knl, kn, 0); 6699 6700 return (0); 6701 } 6702 6703 /* 6704 * Detach knote from vnode 6705 */ 6706 static void 6707 filt_vfsdetach(struct knote *kn) 6708 { 6709 struct vnode *vp = (struct vnode *)kn->kn_hook; 6710 6711 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 6712 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 6713 vdrop(vp); 6714 } 6715 6716 /*ARGSUSED*/ 6717 static int 6718 filt_vfsread(struct knote *kn, long hint) 6719 { 6720 struct vnode *vp = (struct vnode *)kn->kn_hook; 6721 off_t size; 6722 int res; 6723 6724 /* 6725 * filesystem is gone, so set the EOF flag and schedule 6726 * the knote for deletion. 6727 */ 6728 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6729 VI_LOCK(vp); 6730 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6731 VI_UNLOCK(vp); 6732 return (1); 6733 } 6734 6735 if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0) 6736 return (0); 6737 6738 VI_LOCK(vp); 6739 kn->kn_data = size - kn->kn_fp->f_offset; 6740 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; 6741 VI_UNLOCK(vp); 6742 return (res); 6743 } 6744 6745 /*ARGSUSED*/ 6746 static int 6747 filt_vfswrite(struct knote *kn, long hint) 6748 { 6749 struct vnode *vp = (struct vnode *)kn->kn_hook; 6750 6751 VI_LOCK(vp); 6752 6753 /* 6754 * filesystem is gone, so set the EOF flag and schedule 6755 * the knote for deletion. 6756 */ 6757 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) 6758 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6759 6760 kn->kn_data = 0; 6761 VI_UNLOCK(vp); 6762 return (1); 6763 } 6764 6765 static int 6766 filt_vfsvnode(struct knote *kn, long hint) 6767 { 6768 struct vnode *vp = (struct vnode *)kn->kn_hook; 6769 int res; 6770 6771 VI_LOCK(vp); 6772 if (kn->kn_sfflags & hint) 6773 kn->kn_fflags |= hint; 6774 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6775 kn->kn_flags |= EV_EOF; 6776 VI_UNLOCK(vp); 6777 return (1); 6778 } 6779 res = (kn->kn_fflags != 0); 6780 VI_UNLOCK(vp); 6781 return (res); 6782 } 6783 6784 static int 6785 filt_vfsdump(struct proc *p, struct knote *kn, struct kinfo_knote *kin) 6786 { 6787 struct vattr va; 6788 struct vnode *vp; 6789 char *fullpath, *freepath; 6790 int error; 6791 6792 kin->knt_extdata = KNOTE_EXTDATA_VNODE; 6793 6794 vp = kn->kn_fp->f_vnode; 6795 kin->knt_vnode.knt_vnode_type = vntype_to_kinfo(vp->v_type); 6796 6797 va.va_fsid = VNOVAL; 6798 vn_lock(vp, LK_SHARED | LK_RETRY); 6799 error = VOP_GETATTR(vp, &va, curthread->td_ucred); 6800 VOP_UNLOCK(vp); 6801 if (error != 0) 6802 return (error); 6803 kin->knt_vnode.knt_vnode_fsid = va.va_fsid; 6804 kin->knt_vnode.knt_vnode_fileid = va.va_fileid; 6805 6806 freepath = NULL; 6807 fullpath = "-"; 6808 error = vn_fullpath(vp, &fullpath, &freepath); 6809 if (error == 0) { 6810 strlcpy(kin->knt_vnode.knt_vnode_fullpath, fullpath, 6811 sizeof(kin->knt_vnode.knt_vnode_fullpath)); 6812 } 6813 if (freepath != NULL) 6814 free(freepath, M_TEMP); 6815 6816 return (0); 6817 } 6818 6819 int 6820 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 6821 { 6822 int error; 6823 6824 if (dp->d_reclen > ap->a_uio->uio_resid) 6825 return (ENAMETOOLONG); 6826 error = uiomove(dp, dp->d_reclen, ap->a_uio); 6827 if (error) { 6828 if (ap->a_ncookies != NULL) { 6829 if (ap->a_cookies != NULL) 6830 free(ap->a_cookies, M_TEMP); 6831 ap->a_cookies = NULL; 6832 *ap->a_ncookies = 0; 6833 } 6834 return (error); 6835 } 6836 if (ap->a_ncookies == NULL) 6837 return (0); 6838 6839 KASSERT(ap->a_cookies, 6840 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 6841 6842 *ap->a_cookies = realloc(*ap->a_cookies, 6843 (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO); 6844 (*ap->a_cookies)[*ap->a_ncookies] = off; 6845 *ap->a_ncookies += 1; 6846 return (0); 6847 } 6848 6849 /* 6850 * The purpose of this routine is to remove granularity from accmode_t, 6851 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 6852 * VADMIN and VAPPEND. 6853 * 6854 * If it returns 0, the caller is supposed to continue with the usual 6855 * access checks using 'accmode' as modified by this routine. If it 6856 * returns nonzero value, the caller is supposed to return that value 6857 * as errno. 6858 * 6859 * Note that after this routine runs, accmode may be zero. 6860 */ 6861 int 6862 vfs_unixify_accmode(accmode_t *accmode) 6863 { 6864 /* 6865 * There is no way to specify explicit "deny" rule using 6866 * file mode or POSIX.1e ACLs. 6867 */ 6868 if (*accmode & VEXPLICIT_DENY) { 6869 *accmode = 0; 6870 return (0); 6871 } 6872 6873 /* 6874 * None of these can be translated into usual access bits. 6875 * Also, the common case for NFSv4 ACLs is to not contain 6876 * either of these bits. Caller should check for VWRITE 6877 * on the containing directory instead. 6878 */ 6879 if (*accmode & (VDELETE_CHILD | VDELETE)) 6880 return (EPERM); 6881 6882 if (*accmode & VADMIN_PERMS) { 6883 *accmode &= ~VADMIN_PERMS; 6884 *accmode |= VADMIN; 6885 } 6886 6887 /* 6888 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 6889 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 6890 */ 6891 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 6892 6893 return (0); 6894 } 6895 6896 /* 6897 * Clear out a doomed vnode (if any) and replace it with a new one as long 6898 * as the fs is not being unmounted. Return the root vnode to the caller. 6899 */ 6900 static int __noinline 6901 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp) 6902 { 6903 struct vnode *vp; 6904 int error; 6905 6906 restart: 6907 if (mp->mnt_rootvnode != NULL) { 6908 MNT_ILOCK(mp); 6909 vp = mp->mnt_rootvnode; 6910 if (vp != NULL) { 6911 if (!VN_IS_DOOMED(vp)) { 6912 vrefact(vp); 6913 MNT_IUNLOCK(mp); 6914 error = vn_lock(vp, flags); 6915 if (error == 0) { 6916 *vpp = vp; 6917 return (0); 6918 } 6919 vrele(vp); 6920 goto restart; 6921 } 6922 /* 6923 * Clear the old one. 6924 */ 6925 mp->mnt_rootvnode = NULL; 6926 } 6927 MNT_IUNLOCK(mp); 6928 if (vp != NULL) { 6929 vfs_op_barrier_wait(mp); 6930 vrele(vp); 6931 } 6932 } 6933 error = VFS_CACHEDROOT(mp, flags, vpp); 6934 if (error != 0) 6935 return (error); 6936 if (mp->mnt_vfs_ops == 0) { 6937 MNT_ILOCK(mp); 6938 if (mp->mnt_vfs_ops != 0) { 6939 MNT_IUNLOCK(mp); 6940 return (0); 6941 } 6942 if (mp->mnt_rootvnode == NULL) { 6943 vrefact(*vpp); 6944 mp->mnt_rootvnode = *vpp; 6945 } else { 6946 if (mp->mnt_rootvnode != *vpp) { 6947 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) { 6948 panic("%s: mismatch between vnode returned " 6949 " by VFS_CACHEDROOT and the one cached " 6950 " (%p != %p)", 6951 __func__, *vpp, mp->mnt_rootvnode); 6952 } 6953 } 6954 } 6955 MNT_IUNLOCK(mp); 6956 } 6957 return (0); 6958 } 6959 6960 int 6961 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp) 6962 { 6963 struct mount_pcpu *mpcpu; 6964 struct vnode *vp; 6965 int error; 6966 6967 if (!vfs_op_thread_enter(mp, mpcpu)) 6968 return (vfs_cache_root_fallback(mp, flags, vpp)); 6969 vp = atomic_load_ptr(&mp->mnt_rootvnode); 6970 if (vp == NULL || VN_IS_DOOMED(vp)) { 6971 vfs_op_thread_exit(mp, mpcpu); 6972 return (vfs_cache_root_fallback(mp, flags, vpp)); 6973 } 6974 vrefact(vp); 6975 vfs_op_thread_exit(mp, mpcpu); 6976 error = vn_lock(vp, flags); 6977 if (error != 0) { 6978 vrele(vp); 6979 return (vfs_cache_root_fallback(mp, flags, vpp)); 6980 } 6981 *vpp = vp; 6982 return (0); 6983 } 6984 6985 struct vnode * 6986 vfs_cache_root_clear(struct mount *mp) 6987 { 6988 struct vnode *vp; 6989 6990 /* 6991 * ops > 0 guarantees there is nobody who can see this vnode 6992 */ 6993 MPASS(mp->mnt_vfs_ops > 0); 6994 vp = mp->mnt_rootvnode; 6995 if (vp != NULL) 6996 vn_seqc_write_begin(vp); 6997 mp->mnt_rootvnode = NULL; 6998 return (vp); 6999 } 7000 7001 void 7002 vfs_cache_root_set(struct mount *mp, struct vnode *vp) 7003 { 7004 7005 MPASS(mp->mnt_vfs_ops > 0); 7006 vrefact(vp); 7007 mp->mnt_rootvnode = vp; 7008 } 7009 7010 /* 7011 * These are helper functions for filesystems to traverse all 7012 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 7013 * 7014 * This interface replaces MNT_VNODE_FOREACH. 7015 */ 7016 7017 struct vnode * 7018 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 7019 { 7020 struct vnode *vp; 7021 7022 maybe_yield(); 7023 MNT_ILOCK(mp); 7024 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7025 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL; 7026 vp = TAILQ_NEXT(vp, v_nmntvnodes)) { 7027 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 7028 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 7029 continue; 7030 VI_LOCK(vp); 7031 if (VN_IS_DOOMED(vp)) { 7032 VI_UNLOCK(vp); 7033 continue; 7034 } 7035 break; 7036 } 7037 if (vp == NULL) { 7038 __mnt_vnode_markerfree_all(mvp, mp); 7039 /* MNT_IUNLOCK(mp); -- done in above function */ 7040 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 7041 return (NULL); 7042 } 7043 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 7044 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 7045 MNT_IUNLOCK(mp); 7046 return (vp); 7047 } 7048 7049 struct vnode * 7050 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 7051 { 7052 struct vnode *vp; 7053 7054 *mvp = vn_alloc_marker(mp); 7055 MNT_ILOCK(mp); 7056 MNT_REF(mp); 7057 7058 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 7059 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 7060 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 7061 continue; 7062 VI_LOCK(vp); 7063 if (VN_IS_DOOMED(vp)) { 7064 VI_UNLOCK(vp); 7065 continue; 7066 } 7067 break; 7068 } 7069 if (vp == NULL) { 7070 MNT_REL(mp); 7071 MNT_IUNLOCK(mp); 7072 vn_free_marker(*mvp); 7073 *mvp = NULL; 7074 return (NULL); 7075 } 7076 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 7077 MNT_IUNLOCK(mp); 7078 return (vp); 7079 } 7080 7081 void 7082 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 7083 { 7084 7085 if (*mvp == NULL) { 7086 MNT_IUNLOCK(mp); 7087 return; 7088 } 7089 7090 mtx_assert(MNT_MTX(mp), MA_OWNED); 7091 7092 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7093 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 7094 MNT_REL(mp); 7095 MNT_IUNLOCK(mp); 7096 vn_free_marker(*mvp); 7097 *mvp = NULL; 7098 } 7099 7100 /* 7101 * These are helper functions for filesystems to traverse their 7102 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h 7103 */ 7104 static void 7105 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 7106 { 7107 7108 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7109 7110 MNT_ILOCK(mp); 7111 MNT_REL(mp); 7112 MNT_IUNLOCK(mp); 7113 vn_free_marker(*mvp); 7114 *mvp = NULL; 7115 } 7116 7117 /* 7118 * Relock the mp mount vnode list lock with the vp vnode interlock in the 7119 * conventional lock order during mnt_vnode_next_lazy iteration. 7120 * 7121 * On entry, the mount vnode list lock is held and the vnode interlock is not. 7122 * The list lock is dropped and reacquired. On success, both locks are held. 7123 * On failure, the mount vnode list lock is held but the vnode interlock is 7124 * not, and the procedure may have yielded. 7125 */ 7126 static bool 7127 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp, 7128 struct vnode *vp) 7129 { 7130 7131 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER && 7132 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp, 7133 ("%s: bad marker", __func__)); 7134 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp, 7135 ("%s: inappropriate vnode", __func__)); 7136 ASSERT_VI_UNLOCKED(vp, __func__); 7137 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 7138 7139 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist); 7140 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist); 7141 7142 /* 7143 * Note we may be racing against vdrop which transitioned the hold 7144 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine, 7145 * if we are the only user after we get the interlock we will just 7146 * vdrop. 7147 */ 7148 vhold(vp); 7149 mtx_unlock(&mp->mnt_listmtx); 7150 VI_LOCK(vp); 7151 if (VN_IS_DOOMED(vp)) { 7152 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); 7153 goto out_lost; 7154 } 7155 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 7156 /* 7157 * There is nothing to do if we are the last user. 7158 */ 7159 if (!refcount_release_if_not_last(&vp->v_holdcnt)) 7160 goto out_lost; 7161 mtx_lock(&mp->mnt_listmtx); 7162 return (true); 7163 out_lost: 7164 vdropl(vp); 7165 maybe_yield(); 7166 mtx_lock(&mp->mnt_listmtx); 7167 return (false); 7168 } 7169 7170 static struct vnode * 7171 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 7172 void *cbarg) 7173 { 7174 struct vnode *vp; 7175 7176 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 7177 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7178 restart: 7179 vp = TAILQ_NEXT(*mvp, v_lazylist); 7180 while (vp != NULL) { 7181 if (vp->v_type == VMARKER) { 7182 vp = TAILQ_NEXT(vp, v_lazylist); 7183 continue; 7184 } 7185 /* 7186 * See if we want to process the vnode. Note we may encounter a 7187 * long string of vnodes we don't care about and hog the list 7188 * as a result. Check for it and requeue the marker. 7189 */ 7190 VNPASS(!VN_IS_DOOMED(vp), vp); 7191 if (!cb(vp, cbarg)) { 7192 if (!should_yield()) { 7193 vp = TAILQ_NEXT(vp, v_lazylist); 7194 continue; 7195 } 7196 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, 7197 v_lazylist); 7198 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, 7199 v_lazylist); 7200 mtx_unlock(&mp->mnt_listmtx); 7201 kern_yield(PRI_USER); 7202 mtx_lock(&mp->mnt_listmtx); 7203 goto restart; 7204 } 7205 /* 7206 * Try-lock because this is the wrong lock order. 7207 */ 7208 if (!VI_TRYLOCK(vp) && 7209 !mnt_vnode_next_lazy_relock(*mvp, mp, vp)) 7210 goto restart; 7211 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 7212 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 7213 ("alien vnode on the lazy list %p %p", vp, mp)); 7214 VNPASS(vp->v_mount == mp, vp); 7215 VNPASS(!VN_IS_DOOMED(vp), vp); 7216 break; 7217 } 7218 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 7219 7220 /* Check if we are done */ 7221 if (vp == NULL) { 7222 mtx_unlock(&mp->mnt_listmtx); 7223 mnt_vnode_markerfree_lazy(mvp, mp); 7224 return (NULL); 7225 } 7226 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist); 7227 mtx_unlock(&mp->mnt_listmtx); 7228 ASSERT_VI_LOCKED(vp, "lazy iter"); 7229 return (vp); 7230 } 7231 7232 struct vnode * 7233 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 7234 void *cbarg) 7235 { 7236 7237 maybe_yield(); 7238 mtx_lock(&mp->mnt_listmtx); 7239 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 7240 } 7241 7242 struct vnode * 7243 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 7244 void *cbarg) 7245 { 7246 struct vnode *vp; 7247 7248 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist)) 7249 return (NULL); 7250 7251 *mvp = vn_alloc_marker(mp); 7252 MNT_ILOCK(mp); 7253 MNT_REF(mp); 7254 MNT_IUNLOCK(mp); 7255 7256 mtx_lock(&mp->mnt_listmtx); 7257 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist); 7258 if (vp == NULL) { 7259 mtx_unlock(&mp->mnt_listmtx); 7260 mnt_vnode_markerfree_lazy(mvp, mp); 7261 return (NULL); 7262 } 7263 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist); 7264 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 7265 } 7266 7267 void 7268 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 7269 { 7270 7271 if (*mvp == NULL) 7272 return; 7273 7274 mtx_lock(&mp->mnt_listmtx); 7275 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 7276 mtx_unlock(&mp->mnt_listmtx); 7277 mnt_vnode_markerfree_lazy(mvp, mp); 7278 } 7279 7280 int 7281 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp) 7282 { 7283 7284 if ((cnp->cn_flags & NOEXECCHECK) != 0) { 7285 cnp->cn_flags &= ~NOEXECCHECK; 7286 return (0); 7287 } 7288 7289 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread)); 7290 } 7291 7292 /* 7293 * Do not use this variant unless you have means other than the hold count 7294 * to prevent the vnode from getting freed. 7295 */ 7296 void 7297 vn_seqc_write_begin_locked(struct vnode *vp) 7298 { 7299 7300 ASSERT_VI_LOCKED(vp, __func__); 7301 VNPASS(vp->v_holdcnt > 0, vp); 7302 VNPASS(vp->v_seqc_users >= 0, vp); 7303 vp->v_seqc_users++; 7304 if (vp->v_seqc_users == 1) 7305 seqc_sleepable_write_begin(&vp->v_seqc); 7306 } 7307 7308 void 7309 vn_seqc_write_begin(struct vnode *vp) 7310 { 7311 7312 VI_LOCK(vp); 7313 vn_seqc_write_begin_locked(vp); 7314 VI_UNLOCK(vp); 7315 } 7316 7317 void 7318 vn_seqc_write_end_locked(struct vnode *vp) 7319 { 7320 7321 ASSERT_VI_LOCKED(vp, __func__); 7322 VNPASS(vp->v_seqc_users > 0, vp); 7323 vp->v_seqc_users--; 7324 if (vp->v_seqc_users == 0) 7325 seqc_sleepable_write_end(&vp->v_seqc); 7326 } 7327 7328 void 7329 vn_seqc_write_end(struct vnode *vp) 7330 { 7331 7332 VI_LOCK(vp); 7333 vn_seqc_write_end_locked(vp); 7334 VI_UNLOCK(vp); 7335 } 7336 7337 /* 7338 * Special case handling for allocating and freeing vnodes. 7339 * 7340 * The counter remains unchanged on free so that a doomed vnode will 7341 * keep testing as in modify as long as it is accessible with SMR. 7342 */ 7343 static void 7344 vn_seqc_init(struct vnode *vp) 7345 { 7346 7347 vp->v_seqc = 0; 7348 vp->v_seqc_users = 0; 7349 } 7350 7351 static void 7352 vn_seqc_write_end_free(struct vnode *vp) 7353 { 7354 7355 VNPASS(seqc_in_modify(vp->v_seqc), vp); 7356 VNPASS(vp->v_seqc_users == 1, vp); 7357 } 7358 7359 void 7360 vn_irflag_set_locked(struct vnode *vp, short toset) 7361 { 7362 short flags; 7363 7364 ASSERT_VI_LOCKED(vp, __func__); 7365 flags = vn_irflag_read(vp); 7366 VNASSERT((flags & toset) == 0, vp, 7367 ("%s: some of the passed flags already set (have %d, passed %d)\n", 7368 __func__, flags, toset)); 7369 atomic_store_short(&vp->v_irflag, flags | toset); 7370 } 7371 7372 void 7373 vn_irflag_set(struct vnode *vp, short toset) 7374 { 7375 7376 VI_LOCK(vp); 7377 vn_irflag_set_locked(vp, toset); 7378 VI_UNLOCK(vp); 7379 } 7380 7381 void 7382 vn_irflag_set_cond_locked(struct vnode *vp, short toset) 7383 { 7384 short flags; 7385 7386 ASSERT_VI_LOCKED(vp, __func__); 7387 flags = vn_irflag_read(vp); 7388 atomic_store_short(&vp->v_irflag, flags | toset); 7389 } 7390 7391 void 7392 vn_irflag_set_cond(struct vnode *vp, short toset) 7393 { 7394 7395 VI_LOCK(vp); 7396 vn_irflag_set_cond_locked(vp, toset); 7397 VI_UNLOCK(vp); 7398 } 7399 7400 void 7401 vn_irflag_unset_locked(struct vnode *vp, short tounset) 7402 { 7403 short flags; 7404 7405 ASSERT_VI_LOCKED(vp, __func__); 7406 flags = vn_irflag_read(vp); 7407 VNASSERT((flags & tounset) == tounset, vp, 7408 ("%s: some of the passed flags not set (have %d, passed %d)\n", 7409 __func__, flags, tounset)); 7410 atomic_store_short(&vp->v_irflag, flags & ~tounset); 7411 } 7412 7413 void 7414 vn_irflag_unset(struct vnode *vp, short tounset) 7415 { 7416 7417 VI_LOCK(vp); 7418 vn_irflag_unset_locked(vp, tounset); 7419 VI_UNLOCK(vp); 7420 } 7421 7422 int 7423 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred) 7424 { 7425 struct vattr vattr; 7426 int error; 7427 7428 ASSERT_VOP_LOCKED(vp, __func__); 7429 error = VOP_GETATTR(vp, &vattr, cred); 7430 if (__predict_true(error == 0)) { 7431 if (vattr.va_size <= OFF_MAX) 7432 *size = vattr.va_size; 7433 else 7434 error = EFBIG; 7435 } 7436 return (error); 7437 } 7438 7439 int 7440 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred) 7441 { 7442 int error; 7443 7444 VOP_LOCK(vp, LK_SHARED); 7445 error = vn_getsize_locked(vp, size, cred); 7446 VOP_UNLOCK(vp); 7447 return (error); 7448 } 7449 7450 #ifdef INVARIANTS 7451 void 7452 vn_set_state_validate(struct vnode *vp, __enum_uint8(vstate) state) 7453 { 7454 7455 switch (vp->v_state) { 7456 case VSTATE_UNINITIALIZED: 7457 switch (state) { 7458 case VSTATE_CONSTRUCTED: 7459 case VSTATE_DESTROYING: 7460 return; 7461 default: 7462 break; 7463 } 7464 break; 7465 case VSTATE_CONSTRUCTED: 7466 ASSERT_VOP_ELOCKED(vp, __func__); 7467 switch (state) { 7468 case VSTATE_DESTROYING: 7469 return; 7470 default: 7471 break; 7472 } 7473 break; 7474 case VSTATE_DESTROYING: 7475 ASSERT_VOP_ELOCKED(vp, __func__); 7476 switch (state) { 7477 case VSTATE_DEAD: 7478 return; 7479 default: 7480 break; 7481 } 7482 break; 7483 case VSTATE_DEAD: 7484 switch (state) { 7485 case VSTATE_UNINITIALIZED: 7486 return; 7487 default: 7488 break; 7489 } 7490 break; 7491 } 7492 7493 vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state); 7494 panic("invalid state transition %d -> %d\n", vp->v_state, state); 7495 } 7496 #endif 7497