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 goto out_ok; 3356 case VGET_HOLDCNT: 3357 vdrop(vp); 3358 goto out_ok; 3359 case VGET_NONE: 3360 break; 3361 } 3362 3363 __assert_unreachable(); 3364 3365 /* 3366 * This is a goto label should the cases above have more in common than 3367 * just the 'return' statement. 3368 */ 3369 out_ok: 3370 return; 3371 } 3372 3373 int 3374 vget(struct vnode *vp, int flags) 3375 { 3376 enum vgetstate vs; 3377 3378 vs = vget_prep(vp); 3379 return (vget_finish(vp, flags, vs)); 3380 } 3381 3382 int 3383 vget_finish(struct vnode *vp, int flags, enum vgetstate vs) 3384 { 3385 int error; 3386 3387 if ((flags & LK_INTERLOCK) != 0) 3388 ASSERT_VI_LOCKED(vp, __func__); 3389 else 3390 ASSERT_VI_UNLOCKED(vp, __func__); 3391 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); 3392 VNPASS(vp->v_holdcnt > 0, vp); 3393 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); 3394 3395 error = vn_lock(vp, flags); 3396 if (__predict_false(error != 0)) { 3397 vget_abort(vp, vs); 3398 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, 3399 vp); 3400 return (error); 3401 } 3402 3403 vget_finish_ref(vp, vs); 3404 return (0); 3405 } 3406 3407 void 3408 vget_finish_ref(struct vnode *vp, enum vgetstate vs) 3409 { 3410 int old; 3411 3412 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); 3413 VNPASS(vp->v_holdcnt > 0, vp); 3414 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); 3415 3416 if (vs == VGET_USECOUNT) 3417 return; 3418 3419 /* 3420 * We hold the vnode. If the usecount is 0 it will be utilized to keep 3421 * the vnode around. Otherwise someone else lended their hold count and 3422 * we have to drop ours. 3423 */ 3424 old = atomic_fetchadd_int(&vp->v_usecount, 1); 3425 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old)); 3426 if (old != 0) { 3427 #ifdef INVARIANTS 3428 old = atomic_fetchadd_int(&vp->v_holdcnt, -1); 3429 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); 3430 #else 3431 refcount_release(&vp->v_holdcnt); 3432 #endif 3433 } 3434 } 3435 3436 void 3437 vref(struct vnode *vp) 3438 { 3439 enum vgetstate vs; 3440 3441 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3442 vs = vget_prep(vp); 3443 vget_finish_ref(vp, vs); 3444 } 3445 3446 void 3447 vrefact(struct vnode *vp) 3448 { 3449 int old __diagused; 3450 3451 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3452 old = refcount_acquire(&vp->v_usecount); 3453 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); 3454 } 3455 3456 void 3457 vlazy(struct vnode *vp) 3458 { 3459 struct mount *mp; 3460 3461 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__)); 3462 3463 if ((vp->v_mflag & VMP_LAZYLIST) != 0) 3464 return; 3465 /* 3466 * We may get here for inactive routines after the vnode got doomed. 3467 */ 3468 if (VN_IS_DOOMED(vp)) 3469 return; 3470 mp = vp->v_mount; 3471 mtx_lock(&mp->mnt_listmtx); 3472 if ((vp->v_mflag & VMP_LAZYLIST) == 0) { 3473 vp->v_mflag |= VMP_LAZYLIST; 3474 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3475 mp->mnt_lazyvnodelistsize++; 3476 } 3477 mtx_unlock(&mp->mnt_listmtx); 3478 } 3479 3480 static void 3481 vunlazy(struct vnode *vp) 3482 { 3483 struct mount *mp; 3484 3485 ASSERT_VI_LOCKED(vp, __func__); 3486 VNPASS(!VN_IS_DOOMED(vp), vp); 3487 3488 mp = vp->v_mount; 3489 mtx_lock(&mp->mnt_listmtx); 3490 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 3491 /* 3492 * Don't remove the vnode from the lazy list if another thread 3493 * has increased the hold count. It may have re-enqueued the 3494 * vnode to the lazy list and is now responsible for its 3495 * removal. 3496 */ 3497 if (vp->v_holdcnt == 0) { 3498 vp->v_mflag &= ~VMP_LAZYLIST; 3499 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3500 mp->mnt_lazyvnodelistsize--; 3501 } 3502 mtx_unlock(&mp->mnt_listmtx); 3503 } 3504 3505 /* 3506 * This routine is only meant to be called from vgonel prior to dooming 3507 * the vnode. 3508 */ 3509 static void 3510 vunlazy_gone(struct vnode *vp) 3511 { 3512 struct mount *mp; 3513 3514 ASSERT_VOP_ELOCKED(vp, __func__); 3515 ASSERT_VI_LOCKED(vp, __func__); 3516 VNPASS(!VN_IS_DOOMED(vp), vp); 3517 3518 if (vp->v_mflag & VMP_LAZYLIST) { 3519 mp = vp->v_mount; 3520 mtx_lock(&mp->mnt_listmtx); 3521 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 3522 vp->v_mflag &= ~VMP_LAZYLIST; 3523 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3524 mp->mnt_lazyvnodelistsize--; 3525 mtx_unlock(&mp->mnt_listmtx); 3526 } 3527 } 3528 3529 static void 3530 vdefer_inactive(struct vnode *vp) 3531 { 3532 3533 ASSERT_VI_LOCKED(vp, __func__); 3534 VNPASS(vp->v_holdcnt > 0, vp); 3535 if (VN_IS_DOOMED(vp)) { 3536 vdropl(vp); 3537 return; 3538 } 3539 if (vp->v_iflag & VI_DEFINACT) { 3540 VNPASS(vp->v_holdcnt > 1, vp); 3541 vdropl(vp); 3542 return; 3543 } 3544 if (vp->v_usecount > 0) { 3545 vp->v_iflag &= ~VI_OWEINACT; 3546 vdropl(vp); 3547 return; 3548 } 3549 vlazy(vp); 3550 vp->v_iflag |= VI_DEFINACT; 3551 VI_UNLOCK(vp); 3552 atomic_add_long(&deferred_inact, 1); 3553 } 3554 3555 static void 3556 vdefer_inactive_unlocked(struct vnode *vp) 3557 { 3558 3559 VI_LOCK(vp); 3560 if ((vp->v_iflag & VI_OWEINACT) == 0) { 3561 vdropl(vp); 3562 return; 3563 } 3564 vdefer_inactive(vp); 3565 } 3566 3567 enum vput_op { VRELE, VPUT, VUNREF }; 3568 3569 /* 3570 * Handle ->v_usecount transitioning to 0. 3571 * 3572 * By releasing the last usecount we take ownership of the hold count which 3573 * provides liveness of the vnode, meaning we have to vdrop. 3574 * 3575 * For all vnodes we may need to perform inactive processing. It requires an 3576 * exclusive lock on the vnode, while it is legal to call here with only a 3577 * shared lock (or no locks). If locking the vnode in an expected manner fails, 3578 * inactive processing gets deferred to the syncer. 3579 */ 3580 static void 3581 vput_final(struct vnode *vp, enum vput_op func) 3582 { 3583 int error; 3584 bool want_unlock; 3585 3586 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3587 VNPASS(vp->v_holdcnt > 0, vp); 3588 3589 VI_LOCK(vp); 3590 3591 /* 3592 * By the time we got here someone else might have transitioned 3593 * the count back to > 0. 3594 */ 3595 if (vp->v_usecount > 0) 3596 goto out; 3597 3598 /* 3599 * If the vnode is doomed vgone already performed inactive processing 3600 * (if needed). 3601 */ 3602 if (VN_IS_DOOMED(vp)) 3603 goto out; 3604 3605 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0)) 3606 goto out; 3607 3608 if (vp->v_iflag & VI_DOINGINACT) 3609 goto out; 3610 3611 /* 3612 * Locking operations here will drop the interlock and possibly the 3613 * vnode lock, opening a window where the vnode can get doomed all the 3614 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to 3615 * perform inactive. 3616 */ 3617 vp->v_iflag |= VI_OWEINACT; 3618 want_unlock = false; 3619 error = 0; 3620 switch (func) { 3621 case VRELE: 3622 switch (VOP_ISLOCKED(vp)) { 3623 case LK_EXCLUSIVE: 3624 break; 3625 case LK_EXCLOTHER: 3626 case 0: 3627 want_unlock = true; 3628 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 3629 VI_LOCK(vp); 3630 break; 3631 default: 3632 /* 3633 * The lock has at least one sharer, but we have no way 3634 * to conclude whether this is us. Play it safe and 3635 * defer processing. 3636 */ 3637 error = EAGAIN; 3638 break; 3639 } 3640 break; 3641 case VPUT: 3642 want_unlock = true; 3643 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3644 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | 3645 LK_NOWAIT); 3646 VI_LOCK(vp); 3647 } 3648 break; 3649 case VUNREF: 3650 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3651 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); 3652 VI_LOCK(vp); 3653 } 3654 break; 3655 } 3656 if (error != 0) { 3657 vdefer_inactive(vp); 3658 return; 3659 } 3660 if (func == VUNREF) { 3661 VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp, 3662 ("recursive vunref")); 3663 vp->v_vflag |= VV_UNREF; 3664 } 3665 for (;;) { 3666 error = vinactive(vp); 3667 if (want_unlock) 3668 VOP_UNLOCK(vp); 3669 if (error != ERELOOKUP || !want_unlock) 3670 break; 3671 VOP_LOCK(vp, LK_EXCLUSIVE); 3672 } 3673 if (func == VUNREF) 3674 vp->v_vflag &= ~VV_UNREF; 3675 vdropl(vp); 3676 return; 3677 out: 3678 if (func == VPUT) 3679 VOP_UNLOCK(vp); 3680 vdropl(vp); 3681 } 3682 3683 /* 3684 * Decrement ->v_usecount for a vnode. 3685 * 3686 * Releasing the last use count requires additional processing, see vput_final 3687 * above for details. 3688 * 3689 * Comment above each variant denotes lock state on entry and exit. 3690 */ 3691 3692 /* 3693 * in: any 3694 * out: same as passed in 3695 */ 3696 void 3697 vrele(struct vnode *vp) 3698 { 3699 3700 ASSERT_VI_UNLOCKED(vp, __func__); 3701 if (!refcount_release(&vp->v_usecount)) 3702 return; 3703 vput_final(vp, VRELE); 3704 } 3705 3706 /* 3707 * in: locked 3708 * out: unlocked 3709 */ 3710 void 3711 vput(struct vnode *vp) 3712 { 3713 3714 ASSERT_VOP_LOCKED(vp, __func__); 3715 ASSERT_VI_UNLOCKED(vp, __func__); 3716 if (refcount_release_if_last(&vp->v_usecount)) { 3717 vput_final(vp, VPUT); 3718 return; 3719 } 3720 VOP_UNLOCK(vp); 3721 vrele(vp); 3722 } 3723 3724 /* 3725 * in: locked 3726 * out: locked 3727 */ 3728 void 3729 vunref(struct vnode *vp) 3730 { 3731 3732 ASSERT_VOP_LOCKED(vp, __func__); 3733 ASSERT_VI_UNLOCKED(vp, __func__); 3734 if (!refcount_release(&vp->v_usecount)) 3735 return; 3736 vput_final(vp, VUNREF); 3737 } 3738 3739 void 3740 vhold(struct vnode *vp) 3741 { 3742 int old; 3743 3744 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3745 old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3746 VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3747 ("%s: wrong hold count %d", __func__, old)); 3748 if (old == 0) 3749 vfs_freevnodes_dec(); 3750 } 3751 3752 void 3753 vholdnz(struct vnode *vp) 3754 { 3755 3756 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3757 #ifdef INVARIANTS 3758 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3759 VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3760 ("%s: wrong hold count %d", __func__, old)); 3761 #else 3762 atomic_add_int(&vp->v_holdcnt, 1); 3763 #endif 3764 } 3765 3766 /* 3767 * Grab a hold count unless the vnode is freed. 3768 * 3769 * Only use this routine if vfs smr is the only protection you have against 3770 * freeing the vnode. 3771 * 3772 * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag 3773 * is not set. After the flag is set the vnode becomes immutable to anyone but 3774 * the thread which managed to set the flag. 3775 * 3776 * It may be tempting to replace the loop with: 3777 * count = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3778 * if (count & VHOLD_NO_SMR) { 3779 * backpedal and error out; 3780 * } 3781 * 3782 * However, while this is more performant, it hinders debugging by eliminating 3783 * the previously mentioned invariant. 3784 */ 3785 bool 3786 vhold_smr(struct vnode *vp) 3787 { 3788 int count; 3789 3790 VFS_SMR_ASSERT_ENTERED(); 3791 3792 count = atomic_load_int(&vp->v_holdcnt); 3793 for (;;) { 3794 if (count & VHOLD_NO_SMR) { 3795 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, 3796 ("non-zero hold count with flags %d\n", count)); 3797 return (false); 3798 } 3799 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); 3800 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) { 3801 if (count == 0) 3802 vfs_freevnodes_dec(); 3803 return (true); 3804 } 3805 } 3806 } 3807 3808 /* 3809 * Hold a free vnode for recycling. 3810 * 3811 * Note: vnode_init references this comment. 3812 * 3813 * Attempts to recycle only need the global vnode list lock and have no use for 3814 * SMR. 3815 * 3816 * However, vnodes get inserted into the global list before they get fully 3817 * initialized and stay there until UMA decides to free the memory. This in 3818 * particular means the target can be found before it becomes usable and after 3819 * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to 3820 * VHOLD_NO_SMR. 3821 * 3822 * Note: the vnode may gain more references after we transition the count 0->1. 3823 */ 3824 static bool 3825 vhold_recycle_free(struct vnode *vp) 3826 { 3827 int count; 3828 3829 mtx_assert(&vnode_list_mtx, MA_OWNED); 3830 3831 count = atomic_load_int(&vp->v_holdcnt); 3832 for (;;) { 3833 if (count & VHOLD_NO_SMR) { 3834 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, 3835 ("non-zero hold count with flags %d\n", count)); 3836 return (false); 3837 } 3838 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); 3839 if (count > 0) { 3840 return (false); 3841 } 3842 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) { 3843 vfs_freevnodes_dec(); 3844 return (true); 3845 } 3846 } 3847 } 3848 3849 static void __noinline 3850 vdbatch_process(struct vdbatch *vd) 3851 { 3852 struct vnode *vp; 3853 int i; 3854 3855 mtx_assert(&vd->lock, MA_OWNED); 3856 MPASS(curthread->td_pinned > 0); 3857 MPASS(vd->index == VDBATCH_SIZE); 3858 3859 /* 3860 * Attempt to requeue the passed batch, but give up easily. 3861 * 3862 * Despite batching the mechanism is prone to transient *significant* 3863 * lock contention, where vnode_list_mtx becomes the primary bottleneck 3864 * if multiple CPUs get here (one real-world example is highly parallel 3865 * do-nothing make , which will stat *tons* of vnodes). Since it is 3866 * quasi-LRU (read: not that great even if fully honoured) provide an 3867 * option to just dodge the problem. Parties which don't like it are 3868 * welcome to implement something better. 3869 */ 3870 if (vnode_can_skip_requeue) { 3871 if (!mtx_trylock(&vnode_list_mtx)) { 3872 counter_u64_add(vnode_skipped_requeues, 1); 3873 critical_enter(); 3874 for (i = 0; i < VDBATCH_SIZE; i++) { 3875 vp = vd->tab[i]; 3876 vd->tab[i] = NULL; 3877 MPASS(vp->v_dbatchcpu != NOCPU); 3878 vp->v_dbatchcpu = NOCPU; 3879 } 3880 vd->index = 0; 3881 critical_exit(); 3882 return; 3883 3884 } 3885 /* fallthrough to locked processing */ 3886 } else { 3887 mtx_lock(&vnode_list_mtx); 3888 } 3889 3890 mtx_assert(&vnode_list_mtx, MA_OWNED); 3891 critical_enter(); 3892 for (i = 0; i < VDBATCH_SIZE; i++) { 3893 vp = vd->tab[i]; 3894 vd->tab[i] = NULL; 3895 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 3896 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist); 3897 MPASS(vp->v_dbatchcpu != NOCPU); 3898 vp->v_dbatchcpu = NOCPU; 3899 } 3900 mtx_unlock(&vnode_list_mtx); 3901 vd->index = 0; 3902 critical_exit(); 3903 } 3904 3905 static void 3906 vdbatch_enqueue(struct vnode *vp) 3907 { 3908 struct vdbatch *vd; 3909 3910 ASSERT_VI_LOCKED(vp, __func__); 3911 VNPASS(!VN_IS_DOOMED(vp), vp); 3912 3913 if (vp->v_dbatchcpu != NOCPU) { 3914 VI_UNLOCK(vp); 3915 return; 3916 } 3917 3918 sched_pin(); 3919 vd = DPCPU_PTR(vd); 3920 mtx_lock(&vd->lock); 3921 MPASS(vd->index < VDBATCH_SIZE); 3922 MPASS(vd->tab[vd->index] == NULL); 3923 /* 3924 * A hack: we depend on being pinned so that we know what to put in 3925 * ->v_dbatchcpu. 3926 */ 3927 vp->v_dbatchcpu = curcpu; 3928 vd->tab[vd->index] = vp; 3929 vd->index++; 3930 VI_UNLOCK(vp); 3931 if (vd->index == VDBATCH_SIZE) 3932 vdbatch_process(vd); 3933 mtx_unlock(&vd->lock); 3934 sched_unpin(); 3935 } 3936 3937 /* 3938 * This routine must only be called for vnodes which are about to be 3939 * deallocated. Supporting dequeue for arbitrary vndoes would require 3940 * validating that the locked batch matches. 3941 */ 3942 static void 3943 vdbatch_dequeue(struct vnode *vp) 3944 { 3945 struct vdbatch *vd; 3946 int i; 3947 short cpu; 3948 3949 VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp); 3950 3951 cpu = vp->v_dbatchcpu; 3952 if (cpu == NOCPU) 3953 return; 3954 3955 vd = DPCPU_ID_PTR(cpu, vd); 3956 mtx_lock(&vd->lock); 3957 for (i = 0; i < vd->index; i++) { 3958 if (vd->tab[i] != vp) 3959 continue; 3960 vp->v_dbatchcpu = NOCPU; 3961 vd->index--; 3962 vd->tab[i] = vd->tab[vd->index]; 3963 vd->tab[vd->index] = NULL; 3964 break; 3965 } 3966 mtx_unlock(&vd->lock); 3967 /* 3968 * Either we dequeued the vnode above or the target CPU beat us to it. 3969 */ 3970 MPASS(vp->v_dbatchcpu == NOCPU); 3971 } 3972 3973 /* 3974 * Drop the hold count of the vnode. 3975 * 3976 * It will only get freed if this is the last hold *and* it has been vgone'd. 3977 * 3978 * Because the vnode vm object keeps a hold reference on the vnode if 3979 * there is at least one resident non-cached page, the vnode cannot 3980 * leave the active list without the page cleanup done. 3981 */ 3982 static void __noinline 3983 vdropl_final(struct vnode *vp) 3984 { 3985 3986 ASSERT_VI_LOCKED(vp, __func__); 3987 VNPASS(VN_IS_DOOMED(vp), vp); 3988 /* 3989 * Set the VHOLD_NO_SMR flag. 3990 * 3991 * We may be racing against vhold_smr. If they win we can just pretend 3992 * we never got this far, they will vdrop later. 3993 */ 3994 if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) { 3995 vfs_freevnodes_inc(); 3996 VI_UNLOCK(vp); 3997 /* 3998 * We lost the aforementioned race. Any subsequent access is 3999 * invalid as they might have managed to vdropl on their own. 4000 */ 4001 return; 4002 } 4003 /* 4004 * Don't bump freevnodes as this one is going away. 4005 */ 4006 freevnode(vp); 4007 } 4008 4009 void 4010 vdrop(struct vnode *vp) 4011 { 4012 4013 ASSERT_VI_UNLOCKED(vp, __func__); 4014 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4015 if (refcount_release_if_not_last(&vp->v_holdcnt)) 4016 return; 4017 VI_LOCK(vp); 4018 vdropl(vp); 4019 } 4020 4021 static __always_inline void 4022 vdropl_impl(struct vnode *vp, bool enqueue) 4023 { 4024 4025 ASSERT_VI_LOCKED(vp, __func__); 4026 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4027 if (!refcount_release(&vp->v_holdcnt)) { 4028 VI_UNLOCK(vp); 4029 return; 4030 } 4031 VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp); 4032 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp); 4033 if (VN_IS_DOOMED(vp)) { 4034 vdropl_final(vp); 4035 return; 4036 } 4037 4038 vfs_freevnodes_inc(); 4039 if (vp->v_mflag & VMP_LAZYLIST) { 4040 vunlazy(vp); 4041 } 4042 4043 if (!enqueue) { 4044 VI_UNLOCK(vp); 4045 return; 4046 } 4047 4048 /* 4049 * Also unlocks the interlock. We can't assert on it as we 4050 * released our hold and by now the vnode might have been 4051 * freed. 4052 */ 4053 vdbatch_enqueue(vp); 4054 } 4055 4056 void 4057 vdropl(struct vnode *vp) 4058 { 4059 4060 vdropl_impl(vp, true); 4061 } 4062 4063 /* 4064 * vdrop a vnode when recycling 4065 * 4066 * This is a special case routine only to be used when recycling, differs from 4067 * regular vdrop by not requeieing the vnode on LRU. 4068 * 4069 * Consider a case where vtryrecycle continuously fails with all vnodes (due to 4070 * e.g., frozen writes on the filesystem), filling the batch and causing it to 4071 * be requeued. Then vnlru will end up revisiting the same vnodes. This is a 4072 * loop which can last for as long as writes are frozen. 4073 */ 4074 static void 4075 vdropl_recycle(struct vnode *vp) 4076 { 4077 4078 vdropl_impl(vp, false); 4079 } 4080 4081 static void 4082 vdrop_recycle(struct vnode *vp) 4083 { 4084 4085 VI_LOCK(vp); 4086 vdropl_recycle(vp); 4087 } 4088 4089 /* 4090 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 4091 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 4092 */ 4093 static int 4094 vinactivef(struct vnode *vp) 4095 { 4096 int error; 4097 4098 ASSERT_VOP_ELOCKED(vp, "vinactive"); 4099 ASSERT_VI_LOCKED(vp, "vinactive"); 4100 VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp); 4101 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4102 vp->v_iflag |= VI_DOINGINACT; 4103 vp->v_iflag &= ~VI_OWEINACT; 4104 VI_UNLOCK(vp); 4105 4106 /* 4107 * Before moving off the active list, we must be sure that any 4108 * modified pages are converted into the vnode's dirty 4109 * buffers, since these will no longer be checked once the 4110 * vnode is on the inactive list. 4111 * 4112 * The write-out of the dirty pages is asynchronous. At the 4113 * point that VOP_INACTIVE() is called, there could still be 4114 * pending I/O and dirty pages in the object. 4115 */ 4116 if ((vp->v_vflag & VV_NOSYNC) == 0) 4117 vnode_pager_clean_async(vp); 4118 4119 error = VOP_INACTIVE(vp); 4120 VI_LOCK(vp); 4121 VNPASS(vp->v_iflag & VI_DOINGINACT, vp); 4122 vp->v_iflag &= ~VI_DOINGINACT; 4123 return (error); 4124 } 4125 4126 int 4127 vinactive(struct vnode *vp) 4128 { 4129 4130 ASSERT_VOP_ELOCKED(vp, "vinactive"); 4131 ASSERT_VI_LOCKED(vp, "vinactive"); 4132 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4133 4134 if ((vp->v_iflag & VI_OWEINACT) == 0) 4135 return (0); 4136 if (vp->v_iflag & VI_DOINGINACT) 4137 return (0); 4138 if (vp->v_usecount > 0) { 4139 vp->v_iflag &= ~VI_OWEINACT; 4140 return (0); 4141 } 4142 return (vinactivef(vp)); 4143 } 4144 4145 /* 4146 * Remove any vnodes in the vnode table belonging to mount point mp. 4147 * 4148 * If FORCECLOSE is not specified, there should not be any active ones, 4149 * return error if any are found (nb: this is a user error, not a 4150 * system error). If FORCECLOSE is specified, detach any active vnodes 4151 * that are found. 4152 * 4153 * If WRITECLOSE is set, only flush out regular file vnodes open for 4154 * writing. 4155 * 4156 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 4157 * 4158 * `rootrefs' specifies the base reference count for the root vnode 4159 * of this filesystem. The root vnode is considered busy if its 4160 * v_usecount exceeds this value. On a successful return, vflush(, td) 4161 * will call vrele() on the root vnode exactly rootrefs times. 4162 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 4163 * be zero. 4164 */ 4165 #ifdef DIAGNOSTIC 4166 static int busyprt = 0; /* print out busy vnodes */ 4167 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 4168 #endif 4169 4170 int 4171 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 4172 { 4173 struct vnode *vp, *mvp, *rootvp = NULL; 4174 struct vattr vattr; 4175 int busy = 0, error; 4176 4177 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 4178 rootrefs, flags); 4179 if (rootrefs > 0) { 4180 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 4181 ("vflush: bad args")); 4182 /* 4183 * Get the filesystem root vnode. We can vput() it 4184 * immediately, since with rootrefs > 0, it won't go away. 4185 */ 4186 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 4187 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 4188 __func__, error); 4189 return (error); 4190 } 4191 vput(rootvp); 4192 } 4193 loop: 4194 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 4195 vholdl(vp); 4196 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 4197 if (error) { 4198 vdrop(vp); 4199 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 4200 goto loop; 4201 } 4202 /* 4203 * Skip over a vnodes marked VV_SYSTEM. 4204 */ 4205 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 4206 VOP_UNLOCK(vp); 4207 vdrop(vp); 4208 continue; 4209 } 4210 /* 4211 * If WRITECLOSE is set, flush out unlinked but still open 4212 * files (even if open only for reading) and regular file 4213 * vnodes open for writing. 4214 */ 4215 if (flags & WRITECLOSE) { 4216 vnode_pager_clean_async(vp); 4217 do { 4218 error = VOP_FSYNC(vp, MNT_WAIT, td); 4219 } while (error == ERELOOKUP); 4220 if (error != 0) { 4221 VOP_UNLOCK(vp); 4222 vdrop(vp); 4223 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 4224 return (error); 4225 } 4226 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 4227 VI_LOCK(vp); 4228 4229 if ((vp->v_type == VNON || 4230 (error == 0 && vattr.va_nlink > 0)) && 4231 (vp->v_writecount <= 0 || vp->v_type != VREG)) { 4232 VOP_UNLOCK(vp); 4233 vdropl(vp); 4234 continue; 4235 } 4236 } else 4237 VI_LOCK(vp); 4238 /* 4239 * With v_usecount == 0, all we need to do is clear out the 4240 * vnode data structures and we are done. 4241 * 4242 * If FORCECLOSE is set, forcibly close the vnode. 4243 */ 4244 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 4245 vgonel(vp); 4246 } else { 4247 busy++; 4248 #ifdef DIAGNOSTIC 4249 if (busyprt) 4250 vn_printf(vp, "vflush: busy vnode "); 4251 #endif 4252 } 4253 VOP_UNLOCK(vp); 4254 vdropl(vp); 4255 } 4256 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 4257 /* 4258 * If just the root vnode is busy, and if its refcount 4259 * is equal to `rootrefs', then go ahead and kill it. 4260 */ 4261 VI_LOCK(rootvp); 4262 KASSERT(busy > 0, ("vflush: not busy")); 4263 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 4264 ("vflush: usecount %d < rootrefs %d", 4265 rootvp->v_usecount, rootrefs)); 4266 if (busy == 1 && rootvp->v_usecount == rootrefs) { 4267 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 4268 vgone(rootvp); 4269 VOP_UNLOCK(rootvp); 4270 busy = 0; 4271 } else 4272 VI_UNLOCK(rootvp); 4273 } 4274 if (busy) { 4275 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 4276 busy); 4277 return (EBUSY); 4278 } 4279 for (; rootrefs > 0; rootrefs--) 4280 vrele(rootvp); 4281 return (0); 4282 } 4283 4284 /* 4285 * Recycle an unused vnode. 4286 */ 4287 int 4288 vrecycle(struct vnode *vp) 4289 { 4290 int recycled; 4291 4292 VI_LOCK(vp); 4293 recycled = vrecyclel(vp); 4294 VI_UNLOCK(vp); 4295 return (recycled); 4296 } 4297 4298 /* 4299 * vrecycle, with the vp interlock held. 4300 */ 4301 int 4302 vrecyclel(struct vnode *vp) 4303 { 4304 int recycled; 4305 4306 ASSERT_VOP_ELOCKED(vp, __func__); 4307 ASSERT_VI_LOCKED(vp, __func__); 4308 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4309 recycled = 0; 4310 if (vp->v_usecount == 0) { 4311 recycled = 1; 4312 vgonel(vp); 4313 } 4314 return (recycled); 4315 } 4316 4317 /* 4318 * Eliminate all activity associated with a vnode 4319 * in preparation for reuse. 4320 */ 4321 void 4322 vgone(struct vnode *vp) 4323 { 4324 VI_LOCK(vp); 4325 vgonel(vp); 4326 VI_UNLOCK(vp); 4327 } 4328 4329 /* 4330 * Notify upper mounts about reclaimed or unlinked vnode. 4331 */ 4332 void 4333 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event) 4334 { 4335 struct mount *mp; 4336 struct mount_upper_node *ump; 4337 4338 mp = atomic_load_ptr(&vp->v_mount); 4339 if (mp == NULL) 4340 return; 4341 if (TAILQ_EMPTY(&mp->mnt_notify)) 4342 return; 4343 4344 MNT_ILOCK(mp); 4345 mp->mnt_upper_pending++; 4346 KASSERT(mp->mnt_upper_pending > 0, 4347 ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending)); 4348 TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) { 4349 MNT_IUNLOCK(mp); 4350 switch (event) { 4351 case VFS_NOTIFY_UPPER_RECLAIM: 4352 VFS_RECLAIM_LOWERVP(ump->mp, vp); 4353 break; 4354 case VFS_NOTIFY_UPPER_UNLINK: 4355 VFS_UNLINK_LOWERVP(ump->mp, vp); 4356 break; 4357 } 4358 MNT_ILOCK(mp); 4359 } 4360 mp->mnt_upper_pending--; 4361 if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 && 4362 mp->mnt_upper_pending == 0) { 4363 mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER; 4364 wakeup(&mp->mnt_uppers); 4365 } 4366 MNT_IUNLOCK(mp); 4367 } 4368 4369 /* 4370 * vgone, with the vp interlock held. 4371 */ 4372 static void 4373 vgonel(struct vnode *vp) 4374 { 4375 struct thread *td; 4376 struct mount *mp; 4377 vm_object_t object; 4378 bool active, doinginact, oweinact; 4379 4380 ASSERT_VOP_ELOCKED(vp, "vgonel"); 4381 ASSERT_VI_LOCKED(vp, "vgonel"); 4382 VNASSERT(vp->v_holdcnt, vp, 4383 ("vgonel: vp %p has no reference.", vp)); 4384 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4385 td = curthread; 4386 4387 /* 4388 * Don't vgonel if we're already doomed. 4389 */ 4390 if (VN_IS_DOOMED(vp)) { 4391 VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \ 4392 vn_get_state(vp) == VSTATE_DEAD, vp); 4393 return; 4394 } 4395 /* 4396 * Paired with freevnode. 4397 */ 4398 vn_seqc_write_begin_locked(vp); 4399 vunlazy_gone(vp); 4400 vn_irflag_set_locked(vp, VIRF_DOOMED); 4401 vn_set_state(vp, VSTATE_DESTROYING); 4402 4403 /* 4404 * Check to see if the vnode is in use. If so, we have to 4405 * call VOP_CLOSE() and VOP_INACTIVE(). 4406 * 4407 * It could be that VOP_INACTIVE() requested reclamation, in 4408 * which case we should avoid recursion, so check 4409 * VI_DOINGINACT. This is not precise but good enough. 4410 */ 4411 active = vp->v_usecount > 0; 4412 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 4413 doinginact = (vp->v_iflag & VI_DOINGINACT) != 0; 4414 4415 /* 4416 * If we need to do inactive VI_OWEINACT will be set. 4417 */ 4418 if (vp->v_iflag & VI_DEFINACT) { 4419 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 4420 vp->v_iflag &= ~VI_DEFINACT; 4421 vdropl(vp); 4422 } else { 4423 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); 4424 VI_UNLOCK(vp); 4425 } 4426 cache_purge_vgone(vp); 4427 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); 4428 4429 /* 4430 * If purging an active vnode, it must be closed and 4431 * deactivated before being reclaimed. 4432 */ 4433 if (active) 4434 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 4435 if (!doinginact) { 4436 do { 4437 if (oweinact || active) { 4438 VI_LOCK(vp); 4439 vinactivef(vp); 4440 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 4441 VI_UNLOCK(vp); 4442 } 4443 } while (oweinact); 4444 } 4445 if (vp->v_type == VSOCK) 4446 vfs_unp_reclaim(vp); 4447 4448 /* 4449 * Clean out any buffers associated with the vnode. 4450 * If the flush fails, just toss the buffers. 4451 */ 4452 mp = NULL; 4453 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 4454 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 4455 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { 4456 while (vinvalbuf(vp, 0, 0, 0) != 0) 4457 ; 4458 } 4459 4460 BO_LOCK(&vp->v_bufobj); 4461 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && 4462 vp->v_bufobj.bo_dirty.bv_cnt == 0 && 4463 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && 4464 vp->v_bufobj.bo_clean.bv_cnt == 0, 4465 ("vp %p bufobj not invalidated", vp)); 4466 4467 /* 4468 * For VMIO bufobj, BO_DEAD is set later, or in 4469 * vm_object_terminate() after the object's page queue is 4470 * flushed. 4471 */ 4472 object = vp->v_bufobj.bo_object; 4473 if (object == NULL) 4474 vp->v_bufobj.bo_flag |= BO_DEAD; 4475 BO_UNLOCK(&vp->v_bufobj); 4476 4477 /* 4478 * Handle the VM part. Tmpfs handles v_object on its own (the 4479 * OBJT_VNODE check). Nullfs or other bypassing filesystems 4480 * should not touch the object borrowed from the lower vnode 4481 * (the handle check). 4482 */ 4483 if (object != NULL && object->type == OBJT_VNODE && 4484 object->handle == vp) 4485 vnode_destroy_vobject(vp); 4486 4487 /* 4488 * Reclaim the vnode. 4489 */ 4490 if (VOP_RECLAIM(vp)) 4491 panic("vgone: cannot reclaim"); 4492 if (mp != NULL) 4493 vn_finished_secondary_write(mp); 4494 VNASSERT(vp->v_object == NULL, vp, 4495 ("vop_reclaim left v_object vp=%p", vp)); 4496 /* 4497 * Clear the advisory locks and wake up waiting threads. 4498 */ 4499 if (vp->v_lockf != NULL) { 4500 (void)VOP_ADVLOCKPURGE(vp); 4501 vp->v_lockf = NULL; 4502 } 4503 /* 4504 * Delete from old mount point vnode list. 4505 */ 4506 if (vp->v_mount == NULL) { 4507 VI_LOCK(vp); 4508 } else { 4509 delmntque(vp); 4510 ASSERT_VI_LOCKED(vp, "vgonel 2"); 4511 } 4512 /* 4513 * Done with purge, reset to the standard lock and invalidate 4514 * the vnode. 4515 * 4516 * FIXME: this is buggy for vnode ops with custom locking primitives. 4517 * 4518 * vget used to be gated with a special flag serializing it against vgone, 4519 * which got lost in the process of SMP-ifying the VFS layer. 4520 * 4521 * Suppose a custom locking routine references ->v_data. 4522 * 4523 * Since now it is possible to start executing it as vgone is 4524 * progressing, this very well may crash as ->v_data gets invalidated 4525 * and memory used to back it is freed. 4526 */ 4527 vp->v_vnlock = &vp->v_lock; 4528 vp->v_op = &dead_vnodeops; 4529 vp->v_type = VBAD; 4530 vn_set_state(vp, VSTATE_DEAD); 4531 } 4532 4533 /* 4534 * Print out a description of a vnode. 4535 */ 4536 static const char *const vtypename[] = { 4537 [VNON] = "VNON", 4538 [VREG] = "VREG", 4539 [VDIR] = "VDIR", 4540 [VBLK] = "VBLK", 4541 [VCHR] = "VCHR", 4542 [VLNK] = "VLNK", 4543 [VSOCK] = "VSOCK", 4544 [VFIFO] = "VFIFO", 4545 [VBAD] = "VBAD", 4546 [VMARKER] = "VMARKER", 4547 }; 4548 _Static_assert(nitems(vtypename) == VLASTTYPE + 1, 4549 "vnode type name not added to vtypename"); 4550 4551 static const char *const vstatename[] = { 4552 [VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED", 4553 [VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED", 4554 [VSTATE_DESTROYING] = "VSTATE_DESTROYING", 4555 [VSTATE_DEAD] = "VSTATE_DEAD", 4556 }; 4557 _Static_assert(nitems(vstatename) == VLASTSTATE + 1, 4558 "vnode state name not added to vstatename"); 4559 4560 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0, 4561 "new hold count flag not added to vn_printf"); 4562 4563 void 4564 vn_printf(struct vnode *vp, const char *fmt, ...) 4565 { 4566 va_list ap; 4567 char buf[256], buf2[16]; 4568 u_long flags; 4569 u_int holdcnt; 4570 short irflag; 4571 4572 va_start(ap, fmt); 4573 vprintf(fmt, ap); 4574 va_end(ap); 4575 printf("%p: ", (void *)vp); 4576 printf("type %s state %s op %p\n", vtypename[vp->v_type], 4577 vstatename[vp->v_state], vp->v_op); 4578 holdcnt = atomic_load_int(&vp->v_holdcnt); 4579 printf(" usecount %d, writecount %d, refcount %d seqc users %d", 4580 vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS, 4581 vp->v_seqc_users); 4582 switch (vp->v_type) { 4583 case VDIR: 4584 printf(" mountedhere %p\n", vp->v_mountedhere); 4585 break; 4586 case VCHR: 4587 printf(" rdev %p\n", vp->v_rdev); 4588 break; 4589 case VSOCK: 4590 printf(" socket %p\n", vp->v_unpcb); 4591 break; 4592 case VFIFO: 4593 printf(" fifoinfo %p\n", vp->v_fifoinfo); 4594 break; 4595 default: 4596 printf("\n"); 4597 break; 4598 } 4599 buf[0] = '\0'; 4600 buf[1] = '\0'; 4601 if (holdcnt & VHOLD_NO_SMR) 4602 strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf)); 4603 printf(" hold count flags (%s)\n", buf + 1); 4604 4605 buf[0] = '\0'; 4606 buf[1] = '\0'; 4607 irflag = vn_irflag_read(vp); 4608 if (irflag & VIRF_DOOMED) 4609 strlcat(buf, "|VIRF_DOOMED", sizeof(buf)); 4610 if (irflag & VIRF_PGREAD) 4611 strlcat(buf, "|VIRF_PGREAD", sizeof(buf)); 4612 if (irflag & VIRF_MOUNTPOINT) 4613 strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf)); 4614 if (irflag & VIRF_TEXT_REF) 4615 strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf)); 4616 flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF); 4617 if (flags != 0) { 4618 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags); 4619 strlcat(buf, buf2, sizeof(buf)); 4620 } 4621 if (vp->v_vflag & VV_ROOT) 4622 strlcat(buf, "|VV_ROOT", sizeof(buf)); 4623 if (vp->v_vflag & VV_ISTTY) 4624 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 4625 if (vp->v_vflag & VV_NOSYNC) 4626 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 4627 if (vp->v_vflag & VV_ETERNALDEV) 4628 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 4629 if (vp->v_vflag & VV_CACHEDLABEL) 4630 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 4631 if (vp->v_vflag & VV_VMSIZEVNLOCK) 4632 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf)); 4633 if (vp->v_vflag & VV_COPYONWRITE) 4634 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 4635 if (vp->v_vflag & VV_SYSTEM) 4636 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 4637 if (vp->v_vflag & VV_PROCDEP) 4638 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 4639 if (vp->v_vflag & VV_DELETED) 4640 strlcat(buf, "|VV_DELETED", sizeof(buf)); 4641 if (vp->v_vflag & VV_MD) 4642 strlcat(buf, "|VV_MD", sizeof(buf)); 4643 if (vp->v_vflag & VV_FORCEINSMQ) 4644 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 4645 if (vp->v_vflag & VV_READLINK) 4646 strlcat(buf, "|VV_READLINK", sizeof(buf)); 4647 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 4648 VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM | 4649 VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK); 4650 if (flags != 0) { 4651 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 4652 strlcat(buf, buf2, sizeof(buf)); 4653 } 4654 if (vp->v_iflag & VI_MOUNT) 4655 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 4656 if (vp->v_iflag & VI_DOINGINACT) 4657 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 4658 if (vp->v_iflag & VI_OWEINACT) 4659 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 4660 if (vp->v_iflag & VI_DEFINACT) 4661 strlcat(buf, "|VI_DEFINACT", sizeof(buf)); 4662 if (vp->v_iflag & VI_FOPENING) 4663 strlcat(buf, "|VI_FOPENING", sizeof(buf)); 4664 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT | 4665 VI_OWEINACT | VI_DEFINACT | VI_FOPENING); 4666 if (flags != 0) { 4667 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 4668 strlcat(buf, buf2, sizeof(buf)); 4669 } 4670 if (vp->v_mflag & VMP_LAZYLIST) 4671 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf)); 4672 flags = vp->v_mflag & ~(VMP_LAZYLIST); 4673 if (flags != 0) { 4674 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags); 4675 strlcat(buf, buf2, sizeof(buf)); 4676 } 4677 printf(" flags (%s)", buf + 1); 4678 if (mtx_owned(VI_MTX(vp))) 4679 printf(" VI_LOCKed"); 4680 printf("\n"); 4681 if (vp->v_object != NULL) 4682 printf(" v_object %p ref %d pages %d " 4683 "cleanbuf %d dirtybuf %d\n", 4684 vp->v_object, vp->v_object->ref_count, 4685 vp->v_object->resident_page_count, 4686 vp->v_bufobj.bo_clean.bv_cnt, 4687 vp->v_bufobj.bo_dirty.bv_cnt); 4688 printf(" "); 4689 lockmgr_printinfo(vp->v_vnlock); 4690 if (vp->v_data != NULL) 4691 VOP_PRINT(vp); 4692 } 4693 4694 #ifdef DDB 4695 /* 4696 * List all of the locked vnodes in the system. 4697 * Called when debugging the kernel. 4698 */ 4699 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE) 4700 { 4701 struct mount *mp; 4702 struct vnode *vp; 4703 4704 /* 4705 * Note: because this is DDB, we can't obey the locking semantics 4706 * for these structures, which means we could catch an inconsistent 4707 * state and dereference a nasty pointer. Not much to be done 4708 * about that. 4709 */ 4710 db_printf("Locked vnodes\n"); 4711 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4712 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4713 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) 4714 vn_printf(vp, "vnode "); 4715 } 4716 } 4717 } 4718 4719 /* 4720 * Show details about the given vnode. 4721 */ 4722 DB_SHOW_COMMAND(vnode, db_show_vnode) 4723 { 4724 struct vnode *vp; 4725 4726 if (!have_addr) 4727 return; 4728 vp = (struct vnode *)addr; 4729 vn_printf(vp, "vnode "); 4730 } 4731 4732 /* 4733 * Show details about the given mount point. 4734 */ 4735 DB_SHOW_COMMAND(mount, db_show_mount) 4736 { 4737 struct mount *mp; 4738 struct vfsopt *opt; 4739 struct statfs *sp; 4740 struct vnode *vp; 4741 char buf[512]; 4742 uint64_t mflags; 4743 u_int flags; 4744 4745 if (!have_addr) { 4746 /* No address given, print short info about all mount points. */ 4747 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4748 db_printf("%p %s on %s (%s)\n", mp, 4749 mp->mnt_stat.f_mntfromname, 4750 mp->mnt_stat.f_mntonname, 4751 mp->mnt_stat.f_fstypename); 4752 if (db_pager_quit) 4753 break; 4754 } 4755 db_printf("\nMore info: show mount <addr>\n"); 4756 return; 4757 } 4758 4759 mp = (struct mount *)addr; 4760 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 4761 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 4762 4763 buf[0] = '\0'; 4764 mflags = mp->mnt_flag; 4765 #define MNT_FLAG(flag) do { \ 4766 if (mflags & (flag)) { \ 4767 if (buf[0] != '\0') \ 4768 strlcat(buf, ", ", sizeof(buf)); \ 4769 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 4770 mflags &= ~(flag); \ 4771 } \ 4772 } while (0) 4773 MNT_FLAG(MNT_RDONLY); 4774 MNT_FLAG(MNT_SYNCHRONOUS); 4775 MNT_FLAG(MNT_NOEXEC); 4776 MNT_FLAG(MNT_NOSUID); 4777 MNT_FLAG(MNT_NFS4ACLS); 4778 MNT_FLAG(MNT_UNION); 4779 MNT_FLAG(MNT_ASYNC); 4780 MNT_FLAG(MNT_SUIDDIR); 4781 MNT_FLAG(MNT_SOFTDEP); 4782 MNT_FLAG(MNT_NOSYMFOLLOW); 4783 MNT_FLAG(MNT_GJOURNAL); 4784 MNT_FLAG(MNT_MULTILABEL); 4785 MNT_FLAG(MNT_ACLS); 4786 MNT_FLAG(MNT_NOATIME); 4787 MNT_FLAG(MNT_NOCLUSTERR); 4788 MNT_FLAG(MNT_NOCLUSTERW); 4789 MNT_FLAG(MNT_SUJ); 4790 MNT_FLAG(MNT_EXRDONLY); 4791 MNT_FLAG(MNT_EXPORTED); 4792 MNT_FLAG(MNT_DEFEXPORTED); 4793 MNT_FLAG(MNT_EXPORTANON); 4794 MNT_FLAG(MNT_EXKERB); 4795 MNT_FLAG(MNT_EXPUBLIC); 4796 MNT_FLAG(MNT_LOCAL); 4797 MNT_FLAG(MNT_QUOTA); 4798 MNT_FLAG(MNT_ROOTFS); 4799 MNT_FLAG(MNT_USER); 4800 MNT_FLAG(MNT_IGNORE); 4801 MNT_FLAG(MNT_UPDATE); 4802 MNT_FLAG(MNT_DELEXPORT); 4803 MNT_FLAG(MNT_RELOAD); 4804 MNT_FLAG(MNT_FORCE); 4805 MNT_FLAG(MNT_SNAPSHOT); 4806 MNT_FLAG(MNT_BYFSID); 4807 MNT_FLAG(MNT_NAMEDATTR); 4808 #undef MNT_FLAG 4809 if (mflags != 0) { 4810 if (buf[0] != '\0') 4811 strlcat(buf, ", ", sizeof(buf)); 4812 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4813 "0x%016jx", mflags); 4814 } 4815 db_printf(" mnt_flag = %s\n", buf); 4816 4817 buf[0] = '\0'; 4818 flags = mp->mnt_kern_flag; 4819 #define MNT_KERN_FLAG(flag) do { \ 4820 if (flags & (flag)) { \ 4821 if (buf[0] != '\0') \ 4822 strlcat(buf, ", ", sizeof(buf)); \ 4823 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 4824 flags &= ~(flag); \ 4825 } \ 4826 } while (0) 4827 MNT_KERN_FLAG(MNTK_UNMOUNTF); 4828 MNT_KERN_FLAG(MNTK_ASYNC); 4829 MNT_KERN_FLAG(MNTK_SOFTDEP); 4830 MNT_KERN_FLAG(MNTK_NOMSYNC); 4831 MNT_KERN_FLAG(MNTK_DRAINING); 4832 MNT_KERN_FLAG(MNTK_REFEXPIRE); 4833 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 4834 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 4835 MNT_KERN_FLAG(MNTK_NO_IOPF); 4836 MNT_KERN_FLAG(MNTK_RECURSE); 4837 MNT_KERN_FLAG(MNTK_UPPER_WAITER); 4838 MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE); 4839 MNT_KERN_FLAG(MNTK_USES_BCACHE); 4840 MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG); 4841 MNT_KERN_FLAG(MNTK_FPLOOKUP); 4842 MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER); 4843 MNT_KERN_FLAG(MNTK_NOASYNC); 4844 MNT_KERN_FLAG(MNTK_UNMOUNT); 4845 MNT_KERN_FLAG(MNTK_MWAIT); 4846 MNT_KERN_FLAG(MNTK_SUSPEND); 4847 MNT_KERN_FLAG(MNTK_SUSPEND2); 4848 MNT_KERN_FLAG(MNTK_SUSPENDED); 4849 MNT_KERN_FLAG(MNTK_NULL_NOCACHE); 4850 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 4851 #undef MNT_KERN_FLAG 4852 if (flags != 0) { 4853 if (buf[0] != '\0') 4854 strlcat(buf, ", ", sizeof(buf)); 4855 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4856 "0x%08x", flags); 4857 } 4858 db_printf(" mnt_kern_flag = %s\n", buf); 4859 4860 db_printf(" mnt_opt = "); 4861 opt = TAILQ_FIRST(mp->mnt_opt); 4862 if (opt != NULL) { 4863 db_printf("%s", opt->name); 4864 opt = TAILQ_NEXT(opt, link); 4865 while (opt != NULL) { 4866 db_printf(", %s", opt->name); 4867 opt = TAILQ_NEXT(opt, link); 4868 } 4869 } 4870 db_printf("\n"); 4871 4872 sp = &mp->mnt_stat; 4873 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 4874 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 4875 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 4876 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 4877 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 4878 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 4879 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 4880 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 4881 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 4882 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 4883 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 4884 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 4885 4886 db_printf(" mnt_cred = { uid=%u ruid=%u", 4887 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 4888 if (jailed(mp->mnt_cred)) 4889 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 4890 db_printf(" }\n"); 4891 db_printf(" mnt_ref = %d (with %d in the struct)\n", 4892 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref); 4893 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 4894 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 4895 db_printf(" mnt_lazyvnodelistsize = %d\n", 4896 mp->mnt_lazyvnodelistsize); 4897 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n", 4898 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount); 4899 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 4900 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 4901 db_printf(" mnt_lockref = %d (with %d in the struct)\n", 4902 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref); 4903 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 4904 db_printf(" mnt_secondary_accwrites = %d\n", 4905 mp->mnt_secondary_accwrites); 4906 db_printf(" mnt_gjprovider = %s\n", 4907 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 4908 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops); 4909 4910 db_printf("\n\nList of active vnodes\n"); 4911 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4912 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) { 4913 vn_printf(vp, "vnode "); 4914 if (db_pager_quit) 4915 break; 4916 } 4917 } 4918 db_printf("\n\nList of inactive vnodes\n"); 4919 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4920 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) { 4921 vn_printf(vp, "vnode "); 4922 if (db_pager_quit) 4923 break; 4924 } 4925 } 4926 } 4927 #endif /* DDB */ 4928 4929 /* 4930 * Fill in a struct xvfsconf based on a struct vfsconf. 4931 */ 4932 static int 4933 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 4934 { 4935 struct xvfsconf xvfsp; 4936 4937 bzero(&xvfsp, sizeof(xvfsp)); 4938 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4939 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4940 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4941 xvfsp.vfc_flags = vfsp->vfc_flags; 4942 /* 4943 * These are unused in userland, we keep them 4944 * to not break binary compatibility. 4945 */ 4946 xvfsp.vfc_vfsops = NULL; 4947 xvfsp.vfc_next = NULL; 4948 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4949 } 4950 4951 #ifdef COMPAT_FREEBSD32 4952 struct xvfsconf32 { 4953 uint32_t vfc_vfsops; 4954 char vfc_name[MFSNAMELEN]; 4955 int32_t vfc_typenum; 4956 int32_t vfc_refcount; 4957 int32_t vfc_flags; 4958 uint32_t vfc_next; 4959 }; 4960 4961 static int 4962 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 4963 { 4964 struct xvfsconf32 xvfsp; 4965 4966 bzero(&xvfsp, sizeof(xvfsp)); 4967 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4968 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4969 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4970 xvfsp.vfc_flags = vfsp->vfc_flags; 4971 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4972 } 4973 #endif 4974 4975 /* 4976 * Top level filesystem related information gathering. 4977 */ 4978 static int 4979 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 4980 { 4981 struct vfsconf *vfsp; 4982 int error; 4983 4984 error = 0; 4985 vfsconf_slock(); 4986 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4987 #ifdef COMPAT_FREEBSD32 4988 if (req->flags & SCTL_MASK32) 4989 error = vfsconf2x32(req, vfsp); 4990 else 4991 #endif 4992 error = vfsconf2x(req, vfsp); 4993 if (error) 4994 break; 4995 } 4996 vfsconf_sunlock(); 4997 return (error); 4998 } 4999 5000 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | 5001 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, 5002 "S,xvfsconf", "List of all configured filesystems"); 5003 5004 #ifndef BURN_BRIDGES 5005 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 5006 5007 static int 5008 vfs_sysctl(SYSCTL_HANDLER_ARGS) 5009 { 5010 int *name = (int *)arg1 - 1; /* XXX */ 5011 u_int namelen = arg2 + 1; /* XXX */ 5012 struct vfsconf *vfsp; 5013 5014 log(LOG_WARNING, "userland calling deprecated sysctl, " 5015 "please rebuild world\n"); 5016 5017 #if 1 || defined(COMPAT_PRELITE2) 5018 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 5019 if (namelen == 1) 5020 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 5021 #endif 5022 5023 switch (name[1]) { 5024 case VFS_MAXTYPENUM: 5025 if (namelen != 2) 5026 return (ENOTDIR); 5027 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 5028 case VFS_CONF: 5029 if (namelen != 3) 5030 return (ENOTDIR); /* overloaded */ 5031 vfsconf_slock(); 5032 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 5033 if (vfsp->vfc_typenum == name[2]) 5034 break; 5035 } 5036 vfsconf_sunlock(); 5037 if (vfsp == NULL) 5038 return (EOPNOTSUPP); 5039 #ifdef COMPAT_FREEBSD32 5040 if (req->flags & SCTL_MASK32) 5041 return (vfsconf2x32(req, vfsp)); 5042 else 5043 #endif 5044 return (vfsconf2x(req, vfsp)); 5045 } 5046 return (EOPNOTSUPP); 5047 } 5048 5049 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | 5050 CTLFLAG_MPSAFE, vfs_sysctl, 5051 "Generic filesystem"); 5052 5053 #if 1 || defined(COMPAT_PRELITE2) 5054 5055 static int 5056 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 5057 { 5058 int error; 5059 struct vfsconf *vfsp; 5060 struct ovfsconf ovfs; 5061 5062 vfsconf_slock(); 5063 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 5064 bzero(&ovfs, sizeof(ovfs)); 5065 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 5066 strcpy(ovfs.vfc_name, vfsp->vfc_name); 5067 ovfs.vfc_index = vfsp->vfc_typenum; 5068 ovfs.vfc_refcount = vfsp->vfc_refcount; 5069 ovfs.vfc_flags = vfsp->vfc_flags; 5070 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 5071 if (error != 0) { 5072 vfsconf_sunlock(); 5073 return (error); 5074 } 5075 } 5076 vfsconf_sunlock(); 5077 return (0); 5078 } 5079 5080 #endif /* 1 || COMPAT_PRELITE2 */ 5081 #endif /* !BURN_BRIDGES */ 5082 5083 static void 5084 unmount_or_warn(struct mount *mp) 5085 { 5086 int error; 5087 5088 error = dounmount(mp, MNT_FORCE, curthread); 5089 if (error != 0) { 5090 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); 5091 if (error == EBUSY) 5092 printf("BUSY)\n"); 5093 else 5094 printf("%d)\n", error); 5095 } 5096 } 5097 5098 /* 5099 * Unmount all filesystems. The list is traversed in reverse order 5100 * of mounting to avoid dependencies. 5101 */ 5102 void 5103 vfs_unmountall(void) 5104 { 5105 struct mount *mp, *tmp; 5106 5107 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 5108 5109 /* 5110 * Since this only runs when rebooting, it is not interlocked. 5111 */ 5112 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { 5113 vfs_ref(mp); 5114 5115 /* 5116 * Forcibly unmounting "/dev" before "/" would prevent clean 5117 * unmount of the latter. 5118 */ 5119 if (mp == rootdevmp) 5120 continue; 5121 5122 unmount_or_warn(mp); 5123 } 5124 5125 if (rootdevmp != NULL) 5126 unmount_or_warn(rootdevmp); 5127 } 5128 5129 static void 5130 vfs_deferred_inactive(struct vnode *vp, int lkflags) 5131 { 5132 5133 ASSERT_VI_LOCKED(vp, __func__); 5134 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp); 5135 if ((vp->v_iflag & VI_OWEINACT) == 0) { 5136 vdropl(vp); 5137 return; 5138 } 5139 if (vn_lock(vp, lkflags) == 0) { 5140 VI_LOCK(vp); 5141 vinactive(vp); 5142 VOP_UNLOCK(vp); 5143 vdropl(vp); 5144 return; 5145 } 5146 vdefer_inactive_unlocked(vp); 5147 } 5148 5149 static int 5150 vfs_periodic_inactive_filter(struct vnode *vp, void *arg) 5151 { 5152 5153 return (vp->v_iflag & VI_DEFINACT); 5154 } 5155 5156 static void __noinline 5157 vfs_periodic_inactive(struct mount *mp, int flags) 5158 { 5159 struct vnode *vp, *mvp; 5160 int lkflags; 5161 5162 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 5163 if (flags != MNT_WAIT) 5164 lkflags |= LK_NOWAIT; 5165 5166 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) { 5167 if ((vp->v_iflag & VI_DEFINACT) == 0) { 5168 VI_UNLOCK(vp); 5169 continue; 5170 } 5171 vp->v_iflag &= ~VI_DEFINACT; 5172 vfs_deferred_inactive(vp, lkflags); 5173 } 5174 } 5175 5176 static inline bool 5177 vfs_want_msync(struct vnode *vp) 5178 { 5179 struct vm_object *obj; 5180 5181 /* 5182 * This test may be performed without any locks held. 5183 * We rely on vm_object's type stability. 5184 */ 5185 if (vp->v_vflag & VV_NOSYNC) 5186 return (false); 5187 obj = vp->v_object; 5188 return (obj != NULL && vm_object_mightbedirty(obj)); 5189 } 5190 5191 static int 5192 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused) 5193 { 5194 5195 if (vp->v_vflag & VV_NOSYNC) 5196 return (false); 5197 if (vp->v_iflag & VI_DEFINACT) 5198 return (true); 5199 return (vfs_want_msync(vp)); 5200 } 5201 5202 static void __noinline 5203 vfs_periodic_msync_inactive(struct mount *mp, int flags) 5204 { 5205 struct vnode *vp, *mvp; 5206 int lkflags; 5207 bool seen_defer; 5208 5209 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 5210 if (flags != MNT_WAIT) 5211 lkflags |= LK_NOWAIT; 5212 5213 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) { 5214 seen_defer = false; 5215 if (vp->v_iflag & VI_DEFINACT) { 5216 vp->v_iflag &= ~VI_DEFINACT; 5217 seen_defer = true; 5218 } 5219 if (!vfs_want_msync(vp)) { 5220 if (seen_defer) 5221 vfs_deferred_inactive(vp, lkflags); 5222 else 5223 VI_UNLOCK(vp); 5224 continue; 5225 } 5226 if (vget(vp, lkflags) == 0) { 5227 if ((vp->v_vflag & VV_NOSYNC) == 0) { 5228 if (flags == MNT_WAIT) 5229 vnode_pager_clean_sync(vp); 5230 else 5231 vnode_pager_clean_async(vp); 5232 } 5233 vput(vp); 5234 if (seen_defer) 5235 vdrop(vp); 5236 } else { 5237 if (seen_defer) 5238 vdefer_inactive_unlocked(vp); 5239 } 5240 } 5241 } 5242 5243 void 5244 vfs_periodic(struct mount *mp, int flags) 5245 { 5246 5247 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 5248 5249 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0) 5250 vfs_periodic_inactive(mp, flags); 5251 else 5252 vfs_periodic_msync_inactive(mp, flags); 5253 } 5254 5255 static void 5256 destroy_vpollinfo_free(struct vpollinfo *vi) 5257 { 5258 5259 knlist_destroy(&vi->vpi_selinfo.si_note); 5260 mtx_destroy(&vi->vpi_lock); 5261 free(vi, M_VNODEPOLL); 5262 } 5263 5264 static void 5265 destroy_vpollinfo(struct vpollinfo *vi) 5266 { 5267 KASSERT(TAILQ_EMPTY(&vi->vpi_inotify), 5268 ("%s: pollinfo %p has lingering watches", __func__, vi)); 5269 knlist_clear(&vi->vpi_selinfo.si_note, 1); 5270 seldrain(&vi->vpi_selinfo); 5271 destroy_vpollinfo_free(vi); 5272 } 5273 5274 /* 5275 * Initialize per-vnode helper structure to hold poll-related state. 5276 */ 5277 void 5278 v_addpollinfo(struct vnode *vp) 5279 { 5280 struct vpollinfo *vi; 5281 5282 if (atomic_load_ptr(&vp->v_pollinfo) != NULL) 5283 return; 5284 vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO); 5285 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 5286 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 5287 vfs_knlunlock, vfs_knl_assert_lock); 5288 TAILQ_INIT(&vi->vpi_inotify); 5289 VI_LOCK(vp); 5290 if (vp->v_pollinfo != NULL) { 5291 VI_UNLOCK(vp); 5292 destroy_vpollinfo_free(vi); 5293 return; 5294 } 5295 vp->v_pollinfo = vi; 5296 VI_UNLOCK(vp); 5297 } 5298 5299 /* 5300 * Record a process's interest in events which might happen to 5301 * a vnode. Because poll uses the historic select-style interface 5302 * internally, this routine serves as both the ``check for any 5303 * pending events'' and the ``record my interest in future events'' 5304 * functions. (These are done together, while the lock is held, 5305 * to avoid race conditions.) 5306 */ 5307 int 5308 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 5309 { 5310 5311 v_addpollinfo(vp); 5312 mtx_lock(&vp->v_pollinfo->vpi_lock); 5313 if (vp->v_pollinfo->vpi_revents & events) { 5314 /* 5315 * This leaves events we are not interested 5316 * in available for the other process which 5317 * which presumably had requested them 5318 * (otherwise they would never have been 5319 * recorded). 5320 */ 5321 events &= vp->v_pollinfo->vpi_revents; 5322 vp->v_pollinfo->vpi_revents &= ~events; 5323 5324 mtx_unlock(&vp->v_pollinfo->vpi_lock); 5325 return (events); 5326 } 5327 vp->v_pollinfo->vpi_events |= events; 5328 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 5329 mtx_unlock(&vp->v_pollinfo->vpi_lock); 5330 return (0); 5331 } 5332 5333 /* 5334 * Routine to create and manage a filesystem syncer vnode. 5335 */ 5336 #define sync_close ((int (*)(struct vop_close_args *))nullop) 5337 static int sync_fsync(struct vop_fsync_args *); 5338 static int sync_inactive(struct vop_inactive_args *); 5339 static int sync_reclaim(struct vop_reclaim_args *); 5340 5341 static struct vop_vector sync_vnodeops = { 5342 .vop_bypass = VOP_EOPNOTSUPP, 5343 .vop_close = sync_close, 5344 .vop_fsync = sync_fsync, 5345 .vop_getwritemount = vop_stdgetwritemount, 5346 .vop_inactive = sync_inactive, 5347 .vop_need_inactive = vop_stdneed_inactive, 5348 .vop_reclaim = sync_reclaim, 5349 .vop_lock1 = vop_stdlock, 5350 .vop_unlock = vop_stdunlock, 5351 .vop_islocked = vop_stdislocked, 5352 .vop_fplookup_vexec = VOP_EAGAIN, 5353 .vop_fplookup_symlink = VOP_EAGAIN, 5354 }; 5355 VFS_VOP_VECTOR_REGISTER(sync_vnodeops); 5356 5357 /* 5358 * Create a new filesystem syncer vnode for the specified mount point. 5359 */ 5360 void 5361 vfs_allocate_syncvnode(struct mount *mp) 5362 { 5363 struct vnode *vp; 5364 struct bufobj *bo; 5365 static long start, incr, next; 5366 int error; 5367 5368 /* Allocate a new vnode */ 5369 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 5370 if (error != 0) 5371 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 5372 vp->v_type = VNON; 5373 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5374 vp->v_vflag |= VV_FORCEINSMQ; 5375 error = insmntque1(vp, mp); 5376 if (error != 0) 5377 panic("vfs_allocate_syncvnode: insmntque() failed"); 5378 vp->v_vflag &= ~VV_FORCEINSMQ; 5379 vn_set_state(vp, VSTATE_CONSTRUCTED); 5380 VOP_UNLOCK(vp); 5381 /* 5382 * Place the vnode onto the syncer worklist. We attempt to 5383 * scatter them about on the list so that they will go off 5384 * at evenly distributed times even if all the filesystems 5385 * are mounted at once. 5386 */ 5387 next += incr; 5388 if (next == 0 || next > syncer_maxdelay) { 5389 start /= 2; 5390 incr /= 2; 5391 if (start == 0) { 5392 start = syncer_maxdelay / 2; 5393 incr = syncer_maxdelay; 5394 } 5395 next = start; 5396 } 5397 bo = &vp->v_bufobj; 5398 BO_LOCK(bo); 5399 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 5400 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 5401 mtx_lock(&sync_mtx); 5402 sync_vnode_count++; 5403 if (mp->mnt_syncer == NULL) { 5404 mp->mnt_syncer = vp; 5405 vp = NULL; 5406 } 5407 mtx_unlock(&sync_mtx); 5408 BO_UNLOCK(bo); 5409 if (vp != NULL) { 5410 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5411 vgone(vp); 5412 vput(vp); 5413 } 5414 } 5415 5416 void 5417 vfs_deallocate_syncvnode(struct mount *mp) 5418 { 5419 struct vnode *vp; 5420 5421 mtx_lock(&sync_mtx); 5422 vp = mp->mnt_syncer; 5423 if (vp != NULL) 5424 mp->mnt_syncer = NULL; 5425 mtx_unlock(&sync_mtx); 5426 if (vp != NULL) 5427 vrele(vp); 5428 } 5429 5430 /* 5431 * Do a lazy sync of the filesystem. 5432 */ 5433 static int 5434 sync_fsync(struct vop_fsync_args *ap) 5435 { 5436 struct vnode *syncvp = ap->a_vp; 5437 struct mount *mp = syncvp->v_mount; 5438 int error, save; 5439 struct bufobj *bo; 5440 5441 /* 5442 * We only need to do something if this is a lazy evaluation. 5443 */ 5444 if (ap->a_waitfor != MNT_LAZY) 5445 return (0); 5446 5447 /* 5448 * Move ourselves to the back of the sync list. 5449 */ 5450 bo = &syncvp->v_bufobj; 5451 BO_LOCK(bo); 5452 vn_syncer_add_to_worklist(bo, syncdelay); 5453 BO_UNLOCK(bo); 5454 5455 /* 5456 * Walk the list of vnodes pushing all that are dirty and 5457 * not already on the sync list. 5458 */ 5459 if (vfs_busy(mp, MBF_NOWAIT) != 0) 5460 return (0); 5461 VOP_UNLOCK(syncvp); 5462 save = curthread_pflags_set(TDP_SYNCIO); 5463 /* 5464 * The filesystem at hand may be idle with free vnodes stored in the 5465 * batch. Return them instead of letting them stay there indefinitely. 5466 */ 5467 vfs_periodic(mp, MNT_NOWAIT); 5468 error = VFS_SYNC(mp, MNT_LAZY); 5469 curthread_pflags_restore(save); 5470 vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY); 5471 vfs_unbusy(mp); 5472 return (error); 5473 } 5474 5475 /* 5476 * The syncer vnode is no referenced. 5477 */ 5478 static int 5479 sync_inactive(struct vop_inactive_args *ap) 5480 { 5481 5482 vgone(ap->a_vp); 5483 return (0); 5484 } 5485 5486 /* 5487 * The syncer vnode is no longer needed and is being decommissioned. 5488 * 5489 * Modifications to the worklist must be protected by sync_mtx. 5490 */ 5491 static int 5492 sync_reclaim(struct vop_reclaim_args *ap) 5493 { 5494 struct vnode *vp = ap->a_vp; 5495 struct bufobj *bo; 5496 5497 bo = &vp->v_bufobj; 5498 BO_LOCK(bo); 5499 mtx_lock(&sync_mtx); 5500 if (vp->v_mount->mnt_syncer == vp) 5501 vp->v_mount->mnt_syncer = NULL; 5502 if (bo->bo_flag & BO_ONWORKLST) { 5503 LIST_REMOVE(bo, bo_synclist); 5504 syncer_worklist_len--; 5505 sync_vnode_count--; 5506 bo->bo_flag &= ~BO_ONWORKLST; 5507 } 5508 mtx_unlock(&sync_mtx); 5509 BO_UNLOCK(bo); 5510 5511 return (0); 5512 } 5513 5514 int 5515 vn_need_pageq_flush(struct vnode *vp) 5516 { 5517 struct vm_object *obj; 5518 5519 obj = vp->v_object; 5520 return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 5521 vm_object_mightbedirty(obj)); 5522 } 5523 5524 /* 5525 * Check if vnode represents a disk device 5526 */ 5527 bool 5528 vn_isdisk_error(struct vnode *vp, int *errp) 5529 { 5530 int error; 5531 5532 if (vp->v_type != VCHR) { 5533 error = ENOTBLK; 5534 goto out; 5535 } 5536 error = 0; 5537 dev_lock(); 5538 if (vp->v_rdev == NULL) 5539 error = ENXIO; 5540 else if (vp->v_rdev->si_devsw == NULL) 5541 error = ENXIO; 5542 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 5543 error = ENOTBLK; 5544 dev_unlock(); 5545 out: 5546 *errp = error; 5547 return (error == 0); 5548 } 5549 5550 bool 5551 vn_isdisk(struct vnode *vp) 5552 { 5553 int error; 5554 5555 return (vn_isdisk_error(vp, &error)); 5556 } 5557 5558 /* 5559 * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see 5560 * the comment above cache_fplookup for details. 5561 */ 5562 int 5563 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred) 5564 { 5565 int error; 5566 5567 VFS_SMR_ASSERT_ENTERED(); 5568 5569 /* Check the owner. */ 5570 if (cred->cr_uid == file_uid) { 5571 if (file_mode & S_IXUSR) 5572 return (0); 5573 goto out_error; 5574 } 5575 5576 /* Otherwise, check the groups (first match) */ 5577 if (groupmember(file_gid, cred)) { 5578 if (file_mode & S_IXGRP) 5579 return (0); 5580 goto out_error; 5581 } 5582 5583 /* Otherwise, check everyone else. */ 5584 if (file_mode & S_IXOTH) 5585 return (0); 5586 out_error: 5587 /* 5588 * Permission check failed, but it is possible denial will get overwritten 5589 * (e.g., when root is traversing through a 700 directory owned by someone 5590 * else). 5591 * 5592 * vaccess() calls priv_check_cred which in turn can descent into MAC 5593 * modules overriding this result. It's quite unclear what semantics 5594 * are allowed for them to operate, thus for safety we don't call them 5595 * from within the SMR section. This also means if any such modules 5596 * are present, we have to let the regular lookup decide. 5597 */ 5598 error = priv_check_cred_vfs_lookup_nomac(cred); 5599 switch (error) { 5600 case 0: 5601 return (0); 5602 case EAGAIN: 5603 /* 5604 * MAC modules present. 5605 */ 5606 return (EAGAIN); 5607 case EPERM: 5608 return (EACCES); 5609 default: 5610 return (error); 5611 } 5612 } 5613 5614 /* 5615 * Common filesystem object access control check routine. Accepts a 5616 * vnode's type, "mode", uid and gid, requested access mode, and credentials. 5617 * Returns 0 on success, or an errno on failure. 5618 */ 5619 int 5620 vaccess(__enum_uint8(vtype) type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 5621 accmode_t accmode, struct ucred *cred) 5622 { 5623 accmode_t dac_granted; 5624 accmode_t priv_granted; 5625 5626 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 5627 ("invalid bit in accmode")); 5628 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 5629 ("VAPPEND without VWRITE")); 5630 5631 /* 5632 * Look for a normal, non-privileged way to access the file/directory 5633 * as requested. If it exists, go with that. 5634 */ 5635 5636 dac_granted = 0; 5637 5638 /* Check the owner. */ 5639 if (cred->cr_uid == file_uid) { 5640 dac_granted |= VADMIN; 5641 if (file_mode & S_IXUSR) 5642 dac_granted |= VEXEC; 5643 if (file_mode & S_IRUSR) 5644 dac_granted |= VREAD; 5645 if (file_mode & S_IWUSR) 5646 dac_granted |= (VWRITE | VAPPEND); 5647 5648 if ((accmode & dac_granted) == accmode) 5649 return (0); 5650 5651 goto privcheck; 5652 } 5653 5654 /* Otherwise, check the groups (first match) */ 5655 if (groupmember(file_gid, cred)) { 5656 if (file_mode & S_IXGRP) 5657 dac_granted |= VEXEC; 5658 if (file_mode & S_IRGRP) 5659 dac_granted |= VREAD; 5660 if (file_mode & S_IWGRP) 5661 dac_granted |= (VWRITE | VAPPEND); 5662 5663 if ((accmode & dac_granted) == accmode) 5664 return (0); 5665 5666 goto privcheck; 5667 } 5668 5669 /* Otherwise, check everyone else. */ 5670 if (file_mode & S_IXOTH) 5671 dac_granted |= VEXEC; 5672 if (file_mode & S_IROTH) 5673 dac_granted |= VREAD; 5674 if (file_mode & S_IWOTH) 5675 dac_granted |= (VWRITE | VAPPEND); 5676 if ((accmode & dac_granted) == accmode) 5677 return (0); 5678 5679 privcheck: 5680 /* 5681 * Build a privilege mask to determine if the set of privileges 5682 * satisfies the requirements when combined with the granted mask 5683 * from above. For each privilege, if the privilege is required, 5684 * bitwise or the request type onto the priv_granted mask. 5685 */ 5686 priv_granted = 0; 5687 5688 if (type == VDIR) { 5689 /* 5690 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 5691 * requests, instead of PRIV_VFS_EXEC. 5692 */ 5693 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5694 !priv_check_cred(cred, PRIV_VFS_LOOKUP)) 5695 priv_granted |= VEXEC; 5696 } else { 5697 /* 5698 * Ensure that at least one execute bit is on. Otherwise, 5699 * a privileged user will always succeed, and we don't want 5700 * this to happen unless the file really is executable. 5701 */ 5702 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5703 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 5704 !priv_check_cred(cred, PRIV_VFS_EXEC)) 5705 priv_granted |= VEXEC; 5706 } 5707 5708 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 5709 !priv_check_cred(cred, PRIV_VFS_READ)) 5710 priv_granted |= VREAD; 5711 5712 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 5713 !priv_check_cred(cred, PRIV_VFS_WRITE)) 5714 priv_granted |= (VWRITE | VAPPEND); 5715 5716 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 5717 !priv_check_cred(cred, PRIV_VFS_ADMIN)) 5718 priv_granted |= VADMIN; 5719 5720 if ((accmode & (priv_granted | dac_granted)) == accmode) { 5721 return (0); 5722 } 5723 5724 return ((accmode & VADMIN) ? EPERM : EACCES); 5725 } 5726 5727 /* 5728 * Credential check based on process requesting service, and per-attribute 5729 * permissions. 5730 */ 5731 int 5732 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 5733 struct thread *td, accmode_t accmode) 5734 { 5735 5736 /* 5737 * Kernel-invoked always succeeds. 5738 */ 5739 if (cred == NOCRED) 5740 return (0); 5741 5742 /* 5743 * Do not allow privileged processes in jail to directly manipulate 5744 * system attributes. 5745 */ 5746 switch (attrnamespace) { 5747 case EXTATTR_NAMESPACE_SYSTEM: 5748 /* Potentially should be: return (EPERM); */ 5749 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM)); 5750 case EXTATTR_NAMESPACE_USER: 5751 return (VOP_ACCESS(vp, accmode, cred, td)); 5752 default: 5753 return (EPERM); 5754 } 5755 } 5756 5757 #ifdef INVARIANTS 5758 void 5759 assert_vi_locked(struct vnode *vp, const char *str) 5760 { 5761 VNASSERT(mtx_owned(VI_MTX(vp)), vp, 5762 ("%s: vnode interlock is not locked but should be", str)); 5763 } 5764 5765 void 5766 assert_vi_unlocked(struct vnode *vp, const char *str) 5767 { 5768 VNASSERT(!mtx_owned(VI_MTX(vp)), vp, 5769 ("%s: vnode interlock is locked but should not be", str)); 5770 } 5771 5772 void 5773 assert_vop_locked(struct vnode *vp, const char *str) 5774 { 5775 bool locked; 5776 5777 if (KERNEL_PANICKED() || vp == NULL) 5778 return; 5779 5780 #ifdef WITNESS 5781 locked = !((vp->v_irflag & VIRF_CROSSMP) == 0 && 5782 witness_is_owned(&vp->v_vnlock->lock_object) == -1); 5783 #else 5784 int state = VOP_ISLOCKED(vp); 5785 locked = state != 0 && state != LK_EXCLOTHER; 5786 #endif 5787 VNASSERT(locked, vp, ("%s: vnode is not locked but should be", str)); 5788 } 5789 5790 void 5791 assert_vop_unlocked(struct vnode *vp, const char *str) 5792 { 5793 bool locked; 5794 5795 if (KERNEL_PANICKED() || vp == NULL) 5796 return; 5797 5798 #ifdef WITNESS 5799 locked = (vp->v_irflag & VIRF_CROSSMP) == 0 && 5800 witness_is_owned(&vp->v_vnlock->lock_object) == 1; 5801 #else 5802 locked = VOP_ISLOCKED(vp) == LK_EXCLUSIVE; 5803 #endif 5804 VNASSERT(!locked, vp, ("%s: vnode is locked but should not be", str)); 5805 } 5806 5807 void 5808 assert_vop_elocked(struct vnode *vp, const char *str) 5809 { 5810 bool locked; 5811 5812 if (KERNEL_PANICKED() || vp == NULL) 5813 return; 5814 5815 locked = VOP_ISLOCKED(vp) == LK_EXCLUSIVE; 5816 VNASSERT(locked, vp, 5817 ("%s: vnode is not exclusive locked but should be", str)); 5818 } 5819 #endif /* INVARIANTS */ 5820 5821 void 5822 vop_rename_fail(struct vop_rename_args *ap) 5823 { 5824 5825 if (ap->a_tvp != NULL) 5826 vput(ap->a_tvp); 5827 if (ap->a_tdvp == ap->a_tvp) 5828 vrele(ap->a_tdvp); 5829 else 5830 vput(ap->a_tdvp); 5831 vrele(ap->a_fdvp); 5832 vrele(ap->a_fvp); 5833 } 5834 5835 void 5836 vop_rename_pre(void *ap) 5837 { 5838 struct vop_rename_args *a = ap; 5839 5840 #ifdef INVARIANTS 5841 struct mount *tmp; 5842 5843 if (a->a_tvp) 5844 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 5845 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 5846 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 5847 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 5848 5849 /* Check the source (from). */ 5850 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 5851 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 5852 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 5853 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 5854 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 5855 5856 /* Check the target. */ 5857 if (a->a_tvp) 5858 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 5859 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 5860 5861 tmp = NULL; 5862 VOP_GETWRITEMOUNT(a->a_tdvp, &tmp); 5863 lockmgr_assert(&tmp->mnt_renamelock, KA_XLOCKED); 5864 vfs_rel(tmp); 5865 #endif 5866 /* 5867 * It may be tempting to add vn_seqc_write_begin/end calls here and 5868 * in vop_rename_post but that's not going to work out since some 5869 * filesystems relookup vnodes mid-rename. This is probably a bug. 5870 * 5871 * For now filesystems are expected to do the relevant calls after they 5872 * decide what vnodes to operate on. 5873 */ 5874 if (a->a_tdvp != a->a_fdvp) 5875 vhold(a->a_fdvp); 5876 if (a->a_tvp != a->a_fvp) 5877 vhold(a->a_fvp); 5878 vhold(a->a_tdvp); 5879 if (a->a_tvp) 5880 vhold(a->a_tvp); 5881 } 5882 5883 #ifdef INVARIANTS 5884 void 5885 vop_fplookup_vexec_debugpre(void *ap __unused) 5886 { 5887 5888 VFS_SMR_ASSERT_ENTERED(); 5889 } 5890 5891 void 5892 vop_fplookup_vexec_debugpost(void *ap, int rc) 5893 { 5894 struct vop_fplookup_vexec_args *a; 5895 struct vnode *vp; 5896 5897 a = ap; 5898 vp = a->a_vp; 5899 5900 VFS_SMR_ASSERT_ENTERED(); 5901 if (rc == EOPNOTSUPP) 5902 VNPASS(VN_IS_DOOMED(vp), vp); 5903 } 5904 5905 void 5906 vop_fplookup_symlink_debugpre(void *ap __unused) 5907 { 5908 5909 VFS_SMR_ASSERT_ENTERED(); 5910 } 5911 5912 void 5913 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused) 5914 { 5915 5916 VFS_SMR_ASSERT_ENTERED(); 5917 } 5918 5919 static void 5920 vop_fsync_debugprepost(struct vnode *vp, const char *name) 5921 { 5922 struct mount *mp; 5923 5924 if (vp->v_type == VCHR) 5925 ; 5926 /* 5927 * The shared vs. exclusive locking policy for fsync() 5928 * is actually determined by vp's write mount as indicated 5929 * by VOP_GETWRITEMOUNT(), which for stacked filesystems 5930 * may not be the same as vp->v_mount. However, if the 5931 * underlying filesystem which really handles the fsync() 5932 * supports shared locking, the stacked filesystem must also 5933 * be prepared for its VOP_FSYNC() operation to be called 5934 * with only a shared lock. On the other hand, if the 5935 * stacked filesystem claims support for shared write 5936 * locking but the underlying filesystem does not, and the 5937 * caller incorrectly uses a shared lock, this condition 5938 * should still be caught when the stacked filesystem 5939 * invokes VOP_FSYNC() on the underlying filesystem. 5940 */ 5941 else { 5942 mp = NULL; 5943 VOP_GETWRITEMOUNT(vp, &mp); 5944 if (vn_lktype_write(mp, vp) == LK_SHARED) 5945 ASSERT_VOP_LOCKED(vp, name); 5946 else 5947 ASSERT_VOP_ELOCKED(vp, name); 5948 if (mp != NULL) 5949 vfs_rel(mp); 5950 } 5951 } 5952 5953 void 5954 vop_fsync_debugpre(void *a) 5955 { 5956 struct vop_fsync_args *ap; 5957 5958 ap = a; 5959 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5960 } 5961 5962 void 5963 vop_fsync_debugpost(void *a, int rc __unused) 5964 { 5965 struct vop_fsync_args *ap; 5966 5967 ap = a; 5968 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5969 } 5970 5971 void 5972 vop_fdatasync_debugpre(void *a) 5973 { 5974 struct vop_fdatasync_args *ap; 5975 5976 ap = a; 5977 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5978 } 5979 5980 void 5981 vop_fdatasync_debugpost(void *a, int rc __unused) 5982 { 5983 struct vop_fdatasync_args *ap; 5984 5985 ap = a; 5986 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5987 } 5988 5989 void 5990 vop_strategy_debugpre(void *ap) 5991 { 5992 struct vop_strategy_args *a; 5993 struct buf *bp; 5994 5995 a = ap; 5996 bp = a->a_bp; 5997 5998 /* 5999 * Cluster ops lock their component buffers but not the IO container. 6000 */ 6001 if ((bp->b_flags & B_CLUSTER) != 0) 6002 return; 6003 6004 BUF_ASSERT_LOCKED(bp); 6005 } 6006 6007 void 6008 vop_lock_debugpre(void *ap) 6009 { 6010 struct vop_lock1_args *a = ap; 6011 6012 if ((a->a_flags & LK_INTERLOCK) == 0) 6013 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 6014 else 6015 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 6016 } 6017 6018 void 6019 vop_lock_debugpost(void *ap, int rc) 6020 { 6021 struct vop_lock1_args *a = ap; 6022 6023 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 6024 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) 6025 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 6026 } 6027 6028 void 6029 vop_unlock_debugpre(void *ap) 6030 { 6031 struct vop_unlock_args *a = ap; 6032 struct vnode *vp = a->a_vp; 6033 6034 VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp); 6035 ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK"); 6036 } 6037 6038 void 6039 vop_need_inactive_debugpre(void *ap) 6040 { 6041 struct vop_need_inactive_args *a = ap; 6042 6043 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 6044 } 6045 6046 void 6047 vop_need_inactive_debugpost(void *ap, int rc) 6048 { 6049 struct vop_need_inactive_args *a = ap; 6050 6051 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 6052 } 6053 #endif /* INVARIANTS */ 6054 6055 void 6056 vop_allocate_post(void *ap, int rc) 6057 { 6058 struct vop_allocate_args *a; 6059 6060 a = ap; 6061 if (rc == 0) 6062 INOTIFY(a->a_vp, IN_MODIFY); 6063 } 6064 6065 void 6066 vop_copy_file_range_post(void *ap, int rc) 6067 { 6068 struct vop_copy_file_range_args *a; 6069 6070 a = ap; 6071 if (rc == 0) { 6072 INOTIFY(a->a_invp, IN_ACCESS); 6073 INOTIFY(a->a_outvp, IN_MODIFY); 6074 } 6075 } 6076 6077 void 6078 vop_create_pre(void *ap) 6079 { 6080 struct vop_create_args *a; 6081 struct vnode *dvp; 6082 6083 a = ap; 6084 dvp = a->a_dvp; 6085 vn_seqc_write_begin(dvp); 6086 } 6087 6088 void 6089 vop_create_post(void *ap, int rc) 6090 { 6091 struct vop_create_args *a; 6092 struct vnode *dvp; 6093 6094 a = ap; 6095 dvp = a->a_dvp; 6096 vn_seqc_write_end(dvp); 6097 if (!rc) { 6098 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6099 INOTIFY_NAME(*a->a_vpp, dvp, a->a_cnp, IN_CREATE); 6100 } 6101 } 6102 6103 void 6104 vop_deallocate_post(void *ap, int rc) 6105 { 6106 struct vop_deallocate_args *a; 6107 6108 a = ap; 6109 if (rc == 0) 6110 INOTIFY(a->a_vp, IN_MODIFY); 6111 } 6112 6113 void 6114 vop_whiteout_pre(void *ap) 6115 { 6116 struct vop_whiteout_args *a; 6117 struct vnode *dvp; 6118 6119 a = ap; 6120 dvp = a->a_dvp; 6121 vn_seqc_write_begin(dvp); 6122 } 6123 6124 void 6125 vop_whiteout_post(void *ap, int rc) 6126 { 6127 struct vop_whiteout_args *a; 6128 struct vnode *dvp; 6129 6130 a = ap; 6131 dvp = a->a_dvp; 6132 vn_seqc_write_end(dvp); 6133 } 6134 6135 void 6136 vop_deleteextattr_pre(void *ap) 6137 { 6138 struct vop_deleteextattr_args *a; 6139 struct vnode *vp; 6140 6141 a = ap; 6142 vp = a->a_vp; 6143 vn_seqc_write_begin(vp); 6144 } 6145 6146 void 6147 vop_deleteextattr_post(void *ap, int rc) 6148 { 6149 struct vop_deleteextattr_args *a; 6150 struct vnode *vp; 6151 6152 a = ap; 6153 vp = a->a_vp; 6154 vn_seqc_write_end(vp); 6155 if (!rc) { 6156 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 6157 INOTIFY(vp, IN_ATTRIB); 6158 } 6159 } 6160 6161 void 6162 vop_link_pre(void *ap) 6163 { 6164 struct vop_link_args *a; 6165 struct vnode *vp, *tdvp; 6166 6167 a = ap; 6168 vp = a->a_vp; 6169 tdvp = a->a_tdvp; 6170 vn_seqc_write_begin(vp); 6171 vn_seqc_write_begin(tdvp); 6172 } 6173 6174 void 6175 vop_link_post(void *ap, int rc) 6176 { 6177 struct vop_link_args *a; 6178 struct vnode *vp, *tdvp; 6179 6180 a = ap; 6181 vp = a->a_vp; 6182 tdvp = a->a_tdvp; 6183 vn_seqc_write_end(vp); 6184 vn_seqc_write_end(tdvp); 6185 if (!rc) { 6186 VFS_KNOTE_LOCKED(vp, NOTE_LINK); 6187 VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE); 6188 INOTIFY_NAME(vp, tdvp, a->a_cnp, _IN_ATTRIB_LINKCOUNT); 6189 INOTIFY_NAME(vp, tdvp, a->a_cnp, IN_CREATE); 6190 } 6191 } 6192 6193 void 6194 vop_mkdir_pre(void *ap) 6195 { 6196 struct vop_mkdir_args *a; 6197 struct vnode *dvp; 6198 6199 a = ap; 6200 dvp = a->a_dvp; 6201 vn_seqc_write_begin(dvp); 6202 } 6203 6204 void 6205 vop_mkdir_post(void *ap, int rc) 6206 { 6207 struct vop_mkdir_args *a; 6208 struct vnode *dvp; 6209 6210 a = ap; 6211 dvp = a->a_dvp; 6212 vn_seqc_write_end(dvp); 6213 if (!rc) { 6214 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 6215 INOTIFY_NAME(*a->a_vpp, dvp, a->a_cnp, IN_CREATE); 6216 } 6217 } 6218 6219 #ifdef INVARIANTS 6220 void 6221 vop_mkdir_debugpost(void *ap, int rc) 6222 { 6223 struct vop_mkdir_args *a; 6224 6225 a = ap; 6226 if (!rc) 6227 cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp); 6228 } 6229 #endif 6230 6231 void 6232 vop_mknod_pre(void *ap) 6233 { 6234 struct vop_mknod_args *a; 6235 struct vnode *dvp; 6236 6237 a = ap; 6238 dvp = a->a_dvp; 6239 vn_seqc_write_begin(dvp); 6240 } 6241 6242 void 6243 vop_mknod_post(void *ap, int rc) 6244 { 6245 struct vop_mknod_args *a; 6246 struct vnode *dvp; 6247 6248 a = ap; 6249 dvp = a->a_dvp; 6250 vn_seqc_write_end(dvp); 6251 if (!rc) { 6252 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6253 INOTIFY_NAME(*a->a_vpp, dvp, a->a_cnp, IN_CREATE); 6254 } 6255 } 6256 6257 void 6258 vop_reclaim_post(void *ap, int rc) 6259 { 6260 struct vop_reclaim_args *a; 6261 struct vnode *vp; 6262 6263 a = ap; 6264 vp = a->a_vp; 6265 ASSERT_VOP_IN_SEQC(vp); 6266 if (!rc) { 6267 VFS_KNOTE_LOCKED(vp, NOTE_REVOKE); 6268 INOTIFY_REVOKE(vp); 6269 } 6270 } 6271 6272 void 6273 vop_remove_pre(void *ap) 6274 { 6275 struct vop_remove_args *a; 6276 struct vnode *dvp, *vp; 6277 6278 a = ap; 6279 dvp = a->a_dvp; 6280 vp = a->a_vp; 6281 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_UNLINK); 6282 vn_seqc_write_begin(dvp); 6283 vn_seqc_write_begin(vp); 6284 } 6285 6286 void 6287 vop_remove_post(void *ap, int rc) 6288 { 6289 struct vop_remove_args *a; 6290 struct vnode *dvp, *vp; 6291 6292 a = ap; 6293 dvp = a->a_dvp; 6294 vp = a->a_vp; 6295 vn_seqc_write_end(dvp); 6296 vn_seqc_write_end(vp); 6297 if (!rc) { 6298 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6299 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 6300 INOTIFY_NAME(vp, dvp, a->a_cnp, _IN_ATTRIB_LINKCOUNT); 6301 INOTIFY_NAME(vp, dvp, a->a_cnp, IN_DELETE); 6302 } 6303 } 6304 6305 void 6306 vop_rename_post(void *ap, int rc) 6307 { 6308 struct vop_rename_args *a = ap; 6309 long hint; 6310 6311 if (!rc) { 6312 hint = NOTE_WRITE; 6313 if (a->a_fdvp == a->a_tdvp) { 6314 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) 6315 hint |= NOTE_LINK; 6316 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 6317 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 6318 } else { 6319 hint |= NOTE_EXTEND; 6320 if (a->a_fvp->v_type == VDIR) 6321 hint |= NOTE_LINK; 6322 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 6323 6324 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && 6325 a->a_tvp->v_type == VDIR) 6326 hint &= ~NOTE_LINK; 6327 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 6328 } 6329 6330 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 6331 if (a->a_tvp) 6332 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 6333 INOTIFY_MOVE(a->a_fvp, a->a_fdvp, a->a_fcnp, a->a_tvp, 6334 a->a_tdvp, a->a_tcnp); 6335 } 6336 if (a->a_tdvp != a->a_fdvp) 6337 vdrop(a->a_fdvp); 6338 if (a->a_tvp != a->a_fvp) 6339 vdrop(a->a_fvp); 6340 vdrop(a->a_tdvp); 6341 if (a->a_tvp) 6342 vdrop(a->a_tvp); 6343 } 6344 6345 void 6346 vop_rmdir_pre(void *ap) 6347 { 6348 struct vop_rmdir_args *a; 6349 struct vnode *dvp, *vp; 6350 6351 a = ap; 6352 dvp = a->a_dvp; 6353 vp = a->a_vp; 6354 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_UNLINK); 6355 vn_seqc_write_begin(dvp); 6356 vn_seqc_write_begin(vp); 6357 } 6358 6359 void 6360 vop_rmdir_post(void *ap, int rc) 6361 { 6362 struct vop_rmdir_args *a; 6363 struct vnode *dvp, *vp; 6364 6365 a = ap; 6366 dvp = a->a_dvp; 6367 vp = a->a_vp; 6368 vn_seqc_write_end(dvp); 6369 vn_seqc_write_end(vp); 6370 if (!rc) { 6371 vp->v_vflag |= VV_UNLINKED; 6372 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 6373 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 6374 INOTIFY_NAME(vp, dvp, a->a_cnp, IN_DELETE); 6375 } 6376 } 6377 6378 void 6379 vop_setattr_pre(void *ap) 6380 { 6381 struct vop_setattr_args *a; 6382 struct vnode *vp; 6383 6384 a = ap; 6385 vp = a->a_vp; 6386 vn_seqc_write_begin(vp); 6387 } 6388 6389 void 6390 vop_setattr_post(void *ap, int rc) 6391 { 6392 struct vop_setattr_args *a; 6393 struct vnode *vp; 6394 6395 a = ap; 6396 vp = a->a_vp; 6397 vn_seqc_write_end(vp); 6398 if (!rc) { 6399 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 6400 INOTIFY(vp, IN_ATTRIB); 6401 } 6402 } 6403 6404 void 6405 vop_setacl_pre(void *ap) 6406 { 6407 struct vop_setacl_args *a; 6408 struct vnode *vp; 6409 6410 a = ap; 6411 vp = a->a_vp; 6412 vn_seqc_write_begin(vp); 6413 } 6414 6415 void 6416 vop_setacl_post(void *ap, int rc __unused) 6417 { 6418 struct vop_setacl_args *a; 6419 struct vnode *vp; 6420 6421 a = ap; 6422 vp = a->a_vp; 6423 vn_seqc_write_end(vp); 6424 } 6425 6426 void 6427 vop_setextattr_pre(void *ap) 6428 { 6429 struct vop_setextattr_args *a; 6430 struct vnode *vp; 6431 6432 a = ap; 6433 vp = a->a_vp; 6434 vn_seqc_write_begin(vp); 6435 } 6436 6437 void 6438 vop_setextattr_post(void *ap, int rc) 6439 { 6440 struct vop_setextattr_args *a; 6441 struct vnode *vp; 6442 6443 a = ap; 6444 vp = a->a_vp; 6445 vn_seqc_write_end(vp); 6446 if (!rc) { 6447 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 6448 INOTIFY(vp, IN_ATTRIB); 6449 } 6450 } 6451 6452 void 6453 vop_symlink_pre(void *ap) 6454 { 6455 struct vop_symlink_args *a; 6456 struct vnode *dvp; 6457 6458 a = ap; 6459 dvp = a->a_dvp; 6460 vn_seqc_write_begin(dvp); 6461 } 6462 6463 void 6464 vop_symlink_post(void *ap, int rc) 6465 { 6466 struct vop_symlink_args *a; 6467 struct vnode *dvp; 6468 6469 a = ap; 6470 dvp = a->a_dvp; 6471 vn_seqc_write_end(dvp); 6472 if (!rc) { 6473 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6474 INOTIFY_NAME(*a->a_vpp, dvp, a->a_cnp, IN_CREATE); 6475 } 6476 } 6477 6478 void 6479 vop_open_post(void *ap, int rc) 6480 { 6481 struct vop_open_args *a = ap; 6482 6483 if (!rc) { 6484 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); 6485 INOTIFY(a->a_vp, IN_OPEN); 6486 } 6487 } 6488 6489 void 6490 vop_close_post(void *ap, int rc) 6491 { 6492 struct vop_close_args *a = ap; 6493 6494 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ 6495 !VN_IS_DOOMED(a->a_vp))) { 6496 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 6497 NOTE_CLOSE_WRITE : NOTE_CLOSE); 6498 INOTIFY(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 6499 IN_CLOSE_WRITE : IN_CLOSE_NOWRITE); 6500 } 6501 } 6502 6503 void 6504 vop_read_post(void *ap, int rc) 6505 { 6506 struct vop_read_args *a = ap; 6507 6508 if (!rc) { 6509 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 6510 INOTIFY(a->a_vp, IN_ACCESS); 6511 } 6512 } 6513 6514 void 6515 vop_read_pgcache_post(void *ap, int rc) 6516 { 6517 struct vop_read_pgcache_args *a = ap; 6518 6519 if (!rc) 6520 VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ); 6521 } 6522 6523 static struct knlist fs_knlist; 6524 6525 static void 6526 vfs_event_init(void *arg) 6527 { 6528 knlist_init_mtx(&fs_knlist, NULL); 6529 } 6530 /* XXX - correct order? */ 6531 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 6532 6533 void 6534 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 6535 { 6536 6537 KNOTE_UNLOCKED(&fs_knlist, event); 6538 } 6539 6540 static int filt_fsattach(struct knote *kn); 6541 static void filt_fsdetach(struct knote *kn); 6542 static int filt_fsevent(struct knote *kn, long hint); 6543 6544 const struct filterops fs_filtops = { 6545 .f_isfd = 0, 6546 .f_attach = filt_fsattach, 6547 .f_detach = filt_fsdetach, 6548 .f_event = filt_fsevent, 6549 .f_copy = knote_triv_copy, 6550 }; 6551 6552 static int 6553 filt_fsattach(struct knote *kn) 6554 { 6555 6556 kn->kn_flags |= EV_CLEAR; 6557 knlist_add(&fs_knlist, kn, 0); 6558 return (0); 6559 } 6560 6561 static void 6562 filt_fsdetach(struct knote *kn) 6563 { 6564 6565 knlist_remove(&fs_knlist, kn, 0); 6566 } 6567 6568 static int 6569 filt_fsevent(struct knote *kn, long hint) 6570 { 6571 6572 kn->kn_fflags |= kn->kn_sfflags & hint; 6573 6574 return (kn->kn_fflags != 0); 6575 } 6576 6577 static int 6578 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 6579 { 6580 struct vfsidctl vc; 6581 int error; 6582 struct mount *mp; 6583 6584 if (req->newptr == NULL) 6585 return (EINVAL); 6586 error = SYSCTL_IN(req, &vc, sizeof(vc)); 6587 if (error) 6588 return (error); 6589 if (vc.vc_vers != VFS_CTL_VERS1) 6590 return (EINVAL); 6591 mp = vfs_getvfs(&vc.vc_fsid); 6592 if (mp == NULL) 6593 return (ENOENT); 6594 /* ensure that a specific sysctl goes to the right filesystem. */ 6595 if (strcmp(vc.vc_fstypename, "*") != 0 && 6596 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 6597 vfs_rel(mp); 6598 return (EINVAL); 6599 } 6600 VCTLTOREQ(&vc, req); 6601 error = VFS_SYSCTL(mp, vc.vc_op, req); 6602 vfs_rel(mp); 6603 return (error); 6604 } 6605 6606 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR, 6607 NULL, 0, sysctl_vfs_ctl, "", 6608 "Sysctl by fsid"); 6609 6610 /* 6611 * Function to initialize a va_filerev field sensibly. 6612 * XXX: Wouldn't a random number make a lot more sense ?? 6613 */ 6614 u_quad_t 6615 init_va_filerev(void) 6616 { 6617 struct bintime bt; 6618 6619 getbinuptime(&bt); 6620 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 6621 } 6622 6623 static int filt_vfsread(struct knote *kn, long hint); 6624 static int filt_vfswrite(struct knote *kn, long hint); 6625 static int filt_vfsvnode(struct knote *kn, long hint); 6626 static void filt_vfsdetach(struct knote *kn); 6627 static int filt_vfsdump(struct proc *p, struct knote *kn, 6628 struct kinfo_knote *kin); 6629 static int filt_vfscopy(struct knote *kn, struct proc *p1); 6630 6631 static const struct filterops vfsread_filtops = { 6632 .f_isfd = 1, 6633 .f_detach = filt_vfsdetach, 6634 .f_event = filt_vfsread, 6635 .f_userdump = filt_vfsdump, 6636 .f_copy = filt_vfscopy, 6637 }; 6638 static const struct filterops vfswrite_filtops = { 6639 .f_isfd = 1, 6640 .f_detach = filt_vfsdetach, 6641 .f_event = filt_vfswrite, 6642 .f_userdump = filt_vfsdump, 6643 .f_copy = filt_vfscopy, 6644 }; 6645 static const struct filterops vfsvnode_filtops = { 6646 .f_isfd = 1, 6647 .f_detach = filt_vfsdetach, 6648 .f_event = filt_vfsvnode, 6649 .f_userdump = filt_vfsdump, 6650 .f_copy = filt_vfscopy, 6651 }; 6652 6653 static void 6654 vfs_knllock(void *arg) 6655 { 6656 struct vnode *vp = arg; 6657 6658 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 6659 } 6660 6661 static void 6662 vfs_knlunlock(void *arg) 6663 { 6664 struct vnode *vp = arg; 6665 6666 VOP_UNLOCK(vp); 6667 } 6668 6669 static void 6670 vfs_knl_assert_lock(void *arg, int what) 6671 { 6672 #ifdef INVARIANTS 6673 struct vnode *vp = arg; 6674 6675 if (what == LA_LOCKED) 6676 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 6677 else 6678 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 6679 #endif 6680 } 6681 6682 int 6683 vfs_kqfilter(struct vop_kqfilter_args *ap) 6684 { 6685 struct vnode *vp = ap->a_vp; 6686 struct knote *kn = ap->a_kn; 6687 struct knlist *knl; 6688 6689 KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ && 6690 kn->kn_filter != EVFILT_WRITE), 6691 ("READ/WRITE filter on a FIFO leaked through")); 6692 switch (kn->kn_filter) { 6693 case EVFILT_READ: 6694 kn->kn_fop = &vfsread_filtops; 6695 break; 6696 case EVFILT_WRITE: 6697 kn->kn_fop = &vfswrite_filtops; 6698 break; 6699 case EVFILT_VNODE: 6700 kn->kn_fop = &vfsvnode_filtops; 6701 break; 6702 default: 6703 return (EINVAL); 6704 } 6705 6706 kn->kn_hook = (caddr_t)vp; 6707 6708 v_addpollinfo(vp); 6709 if (vp->v_pollinfo == NULL) 6710 return (ENOMEM); 6711 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 6712 vhold(vp); 6713 knlist_add(knl, kn, 0); 6714 6715 return (0); 6716 } 6717 6718 /* 6719 * Detach knote from vnode 6720 */ 6721 static void 6722 filt_vfsdetach(struct knote *kn) 6723 { 6724 struct vnode *vp = (struct vnode *)kn->kn_hook; 6725 6726 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 6727 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 6728 vdrop(vp); 6729 } 6730 6731 /*ARGSUSED*/ 6732 static int 6733 filt_vfsread(struct knote *kn, long hint) 6734 { 6735 struct vnode *vp = (struct vnode *)kn->kn_hook; 6736 off_t size; 6737 int res; 6738 6739 /* 6740 * filesystem is gone, so set the EOF flag and schedule 6741 * the knote for deletion. 6742 */ 6743 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6744 VI_LOCK(vp); 6745 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6746 VI_UNLOCK(vp); 6747 return (1); 6748 } 6749 6750 if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0) 6751 return (0); 6752 6753 VI_LOCK(vp); 6754 kn->kn_data = size - kn->kn_fp->f_offset; 6755 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; 6756 VI_UNLOCK(vp); 6757 return (res); 6758 } 6759 6760 /*ARGSUSED*/ 6761 static int 6762 filt_vfswrite(struct knote *kn, long hint) 6763 { 6764 struct vnode *vp = (struct vnode *)kn->kn_hook; 6765 6766 VI_LOCK(vp); 6767 6768 /* 6769 * filesystem is gone, so set the EOF flag and schedule 6770 * the knote for deletion. 6771 */ 6772 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) 6773 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6774 6775 kn->kn_data = 0; 6776 VI_UNLOCK(vp); 6777 return (1); 6778 } 6779 6780 static int 6781 filt_vfsvnode(struct knote *kn, long hint) 6782 { 6783 struct vnode *vp = (struct vnode *)kn->kn_hook; 6784 int res; 6785 6786 VI_LOCK(vp); 6787 if (kn->kn_sfflags & hint) 6788 kn->kn_fflags |= hint; 6789 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6790 kn->kn_flags |= EV_EOF; 6791 VI_UNLOCK(vp); 6792 return (1); 6793 } 6794 res = (kn->kn_fflags != 0); 6795 VI_UNLOCK(vp); 6796 return (res); 6797 } 6798 6799 static int 6800 filt_vfsdump(struct proc *p, struct knote *kn, struct kinfo_knote *kin) 6801 { 6802 struct vattr va; 6803 struct vnode *vp; 6804 char *fullpath, *freepath; 6805 int error; 6806 6807 kin->knt_extdata = KNOTE_EXTDATA_VNODE; 6808 6809 vp = kn->kn_fp->f_vnode; 6810 kin->knt_vnode.knt_vnode_type = vntype_to_kinfo(vp->v_type); 6811 6812 va.va_fsid = VNOVAL; 6813 vn_lock(vp, LK_SHARED | LK_RETRY); 6814 error = VOP_GETATTR(vp, &va, curthread->td_ucred); 6815 VOP_UNLOCK(vp); 6816 if (error != 0) 6817 return (error); 6818 kin->knt_vnode.knt_vnode_fsid = va.va_fsid; 6819 kin->knt_vnode.knt_vnode_fileid = va.va_fileid; 6820 6821 freepath = NULL; 6822 fullpath = "-"; 6823 error = vn_fullpath(vp, &fullpath, &freepath); 6824 if (error == 0) { 6825 strlcpy(kin->knt_vnode.knt_vnode_fullpath, fullpath, 6826 sizeof(kin->knt_vnode.knt_vnode_fullpath)); 6827 } 6828 if (freepath != NULL) 6829 free(freepath, M_TEMP); 6830 6831 return (0); 6832 } 6833 6834 static int 6835 filt_vfscopy(struct knote *kn, struct proc *p1) 6836 { 6837 struct vnode *vp; 6838 6839 vp = (struct vnode *)kn->kn_hook; 6840 vhold(vp); 6841 return (0); 6842 } 6843 6844 int 6845 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 6846 { 6847 int error; 6848 6849 if (dp->d_reclen > ap->a_uio->uio_resid) 6850 return (ENAMETOOLONG); 6851 error = uiomove(dp, dp->d_reclen, ap->a_uio); 6852 if (error) { 6853 if (ap->a_ncookies != NULL) { 6854 if (ap->a_cookies != NULL) 6855 free(ap->a_cookies, M_TEMP); 6856 ap->a_cookies = NULL; 6857 *ap->a_ncookies = 0; 6858 } 6859 return (error); 6860 } 6861 if (ap->a_ncookies == NULL) 6862 return (0); 6863 6864 KASSERT(ap->a_cookies, 6865 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 6866 6867 *ap->a_cookies = realloc(*ap->a_cookies, 6868 (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO); 6869 (*ap->a_cookies)[*ap->a_ncookies] = off; 6870 *ap->a_ncookies += 1; 6871 return (0); 6872 } 6873 6874 /* 6875 * The purpose of this routine is to remove granularity from accmode_t, 6876 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 6877 * VADMIN and VAPPEND. 6878 * 6879 * If it returns 0, the caller is supposed to continue with the usual 6880 * access checks using 'accmode' as modified by this routine. If it 6881 * returns nonzero value, the caller is supposed to return that value 6882 * as errno. 6883 * 6884 * Note that after this routine runs, accmode may be zero. 6885 */ 6886 int 6887 vfs_unixify_accmode(accmode_t *accmode) 6888 { 6889 /* 6890 * There is no way to specify explicit "deny" rule using 6891 * file mode or POSIX.1e ACLs. 6892 */ 6893 if (*accmode & VEXPLICIT_DENY) { 6894 *accmode = 0; 6895 return (0); 6896 } 6897 6898 /* 6899 * None of these can be translated into usual access bits. 6900 * Also, the common case for NFSv4 ACLs is to not contain 6901 * either of these bits. Caller should check for VWRITE 6902 * on the containing directory instead. 6903 */ 6904 if (*accmode & (VDELETE_CHILD | VDELETE)) 6905 return (EPERM); 6906 6907 if (*accmode & VADMIN_PERMS) { 6908 *accmode &= ~VADMIN_PERMS; 6909 *accmode |= VADMIN; 6910 } 6911 6912 /* 6913 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 6914 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 6915 */ 6916 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 6917 6918 return (0); 6919 } 6920 6921 /* 6922 * Clear out a doomed vnode (if any) and replace it with a new one as long 6923 * as the fs is not being unmounted. Return the root vnode to the caller. 6924 */ 6925 static int __noinline 6926 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp) 6927 { 6928 struct vnode *vp; 6929 int error; 6930 6931 restart: 6932 if (mp->mnt_rootvnode != NULL) { 6933 MNT_ILOCK(mp); 6934 vp = mp->mnt_rootvnode; 6935 if (vp != NULL) { 6936 if (!VN_IS_DOOMED(vp)) { 6937 vrefact(vp); 6938 MNT_IUNLOCK(mp); 6939 error = vn_lock(vp, flags); 6940 if (error == 0) { 6941 *vpp = vp; 6942 return (0); 6943 } 6944 vrele(vp); 6945 goto restart; 6946 } 6947 /* 6948 * Clear the old one. 6949 */ 6950 mp->mnt_rootvnode = NULL; 6951 } 6952 MNT_IUNLOCK(mp); 6953 if (vp != NULL) { 6954 vfs_op_barrier_wait(mp); 6955 vrele(vp); 6956 } 6957 } 6958 error = VFS_CACHEDROOT(mp, flags, vpp); 6959 if (error != 0) 6960 return (error); 6961 if (mp->mnt_vfs_ops == 0) { 6962 MNT_ILOCK(mp); 6963 if (mp->mnt_vfs_ops != 0) { 6964 MNT_IUNLOCK(mp); 6965 return (0); 6966 } 6967 if (mp->mnt_rootvnode == NULL) { 6968 vrefact(*vpp); 6969 mp->mnt_rootvnode = *vpp; 6970 } else { 6971 if (mp->mnt_rootvnode != *vpp) { 6972 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) { 6973 panic("%s: mismatch between vnode returned " 6974 " by VFS_CACHEDROOT and the one cached " 6975 " (%p != %p)", 6976 __func__, *vpp, mp->mnt_rootvnode); 6977 } 6978 } 6979 } 6980 MNT_IUNLOCK(mp); 6981 } 6982 return (0); 6983 } 6984 6985 int 6986 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp) 6987 { 6988 struct mount_pcpu *mpcpu; 6989 struct vnode *vp; 6990 int error; 6991 6992 if (!vfs_op_thread_enter(mp, mpcpu)) 6993 return (vfs_cache_root_fallback(mp, flags, vpp)); 6994 vp = atomic_load_ptr(&mp->mnt_rootvnode); 6995 if (vp == NULL || VN_IS_DOOMED(vp)) { 6996 vfs_op_thread_exit(mp, mpcpu); 6997 return (vfs_cache_root_fallback(mp, flags, vpp)); 6998 } 6999 vrefact(vp); 7000 vfs_op_thread_exit(mp, mpcpu); 7001 error = vn_lock(vp, flags); 7002 if (error != 0) { 7003 vrele(vp); 7004 return (vfs_cache_root_fallback(mp, flags, vpp)); 7005 } 7006 *vpp = vp; 7007 return (0); 7008 } 7009 7010 struct vnode * 7011 vfs_cache_root_clear(struct mount *mp) 7012 { 7013 struct vnode *vp; 7014 7015 /* 7016 * ops > 0 guarantees there is nobody who can see this vnode 7017 */ 7018 MPASS(mp->mnt_vfs_ops > 0); 7019 vp = mp->mnt_rootvnode; 7020 if (vp != NULL) 7021 vn_seqc_write_begin(vp); 7022 mp->mnt_rootvnode = NULL; 7023 return (vp); 7024 } 7025 7026 void 7027 vfs_cache_root_set(struct mount *mp, struct vnode *vp) 7028 { 7029 7030 MPASS(mp->mnt_vfs_ops > 0); 7031 vrefact(vp); 7032 mp->mnt_rootvnode = vp; 7033 } 7034 7035 /* 7036 * These are helper functions for filesystems to traverse all 7037 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 7038 * 7039 * This interface replaces MNT_VNODE_FOREACH. 7040 */ 7041 7042 struct vnode * 7043 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 7044 { 7045 struct vnode *vp; 7046 7047 maybe_yield(); 7048 MNT_ILOCK(mp); 7049 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7050 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL; 7051 vp = TAILQ_NEXT(vp, v_nmntvnodes)) { 7052 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 7053 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 7054 continue; 7055 VI_LOCK(vp); 7056 if (VN_IS_DOOMED(vp)) { 7057 VI_UNLOCK(vp); 7058 continue; 7059 } 7060 break; 7061 } 7062 if (vp == NULL) { 7063 __mnt_vnode_markerfree_all(mvp, mp); 7064 /* MNT_IUNLOCK(mp); -- done in above function */ 7065 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 7066 return (NULL); 7067 } 7068 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 7069 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 7070 MNT_IUNLOCK(mp); 7071 return (vp); 7072 } 7073 7074 struct vnode * 7075 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 7076 { 7077 struct vnode *vp; 7078 7079 *mvp = vn_alloc_marker(mp); 7080 MNT_ILOCK(mp); 7081 MNT_REF(mp); 7082 7083 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 7084 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 7085 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 7086 continue; 7087 VI_LOCK(vp); 7088 if (VN_IS_DOOMED(vp)) { 7089 VI_UNLOCK(vp); 7090 continue; 7091 } 7092 break; 7093 } 7094 if (vp == NULL) { 7095 MNT_REL(mp); 7096 MNT_IUNLOCK(mp); 7097 vn_free_marker(*mvp); 7098 *mvp = NULL; 7099 return (NULL); 7100 } 7101 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 7102 MNT_IUNLOCK(mp); 7103 return (vp); 7104 } 7105 7106 void 7107 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 7108 { 7109 7110 if (*mvp == NULL) { 7111 MNT_IUNLOCK(mp); 7112 return; 7113 } 7114 7115 mtx_assert(MNT_MTX(mp), MA_OWNED); 7116 7117 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7118 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 7119 MNT_REL(mp); 7120 MNT_IUNLOCK(mp); 7121 vn_free_marker(*mvp); 7122 *mvp = NULL; 7123 } 7124 7125 /* 7126 * These are helper functions for filesystems to traverse their 7127 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h 7128 */ 7129 static void 7130 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 7131 { 7132 7133 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7134 7135 MNT_ILOCK(mp); 7136 MNT_REL(mp); 7137 MNT_IUNLOCK(mp); 7138 vn_free_marker(*mvp); 7139 *mvp = NULL; 7140 } 7141 7142 /* 7143 * Relock the mp mount vnode list lock with the vp vnode interlock in the 7144 * conventional lock order during mnt_vnode_next_lazy iteration. 7145 * 7146 * On entry, the mount vnode list lock is held and the vnode interlock is not. 7147 * The list lock is dropped and reacquired. On success, both locks are held. 7148 * On failure, the mount vnode list lock is held but the vnode interlock is 7149 * not, and the procedure may have yielded. 7150 */ 7151 static bool 7152 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp, 7153 struct vnode *vp) 7154 { 7155 7156 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER && 7157 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp, 7158 ("%s: bad marker", __func__)); 7159 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp, 7160 ("%s: inappropriate vnode", __func__)); 7161 ASSERT_VI_UNLOCKED(vp, __func__); 7162 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 7163 7164 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist); 7165 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist); 7166 7167 /* 7168 * Note we may be racing against vdrop which transitioned the hold 7169 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine, 7170 * if we are the only user after we get the interlock we will just 7171 * vdrop. 7172 */ 7173 vhold(vp); 7174 mtx_unlock(&mp->mnt_listmtx); 7175 VI_LOCK(vp); 7176 if (VN_IS_DOOMED(vp)) { 7177 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); 7178 goto out_lost; 7179 } 7180 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 7181 /* 7182 * There is nothing to do if we are the last user. 7183 */ 7184 if (!refcount_release_if_not_last(&vp->v_holdcnt)) 7185 goto out_lost; 7186 mtx_lock(&mp->mnt_listmtx); 7187 return (true); 7188 out_lost: 7189 vdropl(vp); 7190 maybe_yield(); 7191 mtx_lock(&mp->mnt_listmtx); 7192 return (false); 7193 } 7194 7195 static struct vnode * 7196 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 7197 void *cbarg) 7198 { 7199 struct vnode *vp; 7200 7201 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 7202 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7203 restart: 7204 vp = TAILQ_NEXT(*mvp, v_lazylist); 7205 while (vp != NULL) { 7206 if (vp->v_type == VMARKER) { 7207 vp = TAILQ_NEXT(vp, v_lazylist); 7208 continue; 7209 } 7210 /* 7211 * See if we want to process the vnode. Note we may encounter a 7212 * long string of vnodes we don't care about and hog the list 7213 * as a result. Check for it and requeue the marker. 7214 */ 7215 VNPASS(!VN_IS_DOOMED(vp), vp); 7216 if (!cb(vp, cbarg)) { 7217 if (!should_yield()) { 7218 vp = TAILQ_NEXT(vp, v_lazylist); 7219 continue; 7220 } 7221 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, 7222 v_lazylist); 7223 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, 7224 v_lazylist); 7225 mtx_unlock(&mp->mnt_listmtx); 7226 kern_yield(PRI_USER); 7227 mtx_lock(&mp->mnt_listmtx); 7228 goto restart; 7229 } 7230 /* 7231 * Try-lock because this is the wrong lock order. 7232 */ 7233 if (!VI_TRYLOCK(vp) && 7234 !mnt_vnode_next_lazy_relock(*mvp, mp, vp)) 7235 goto restart; 7236 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 7237 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 7238 ("alien vnode on the lazy list %p %p", vp, mp)); 7239 VNPASS(vp->v_mount == mp, vp); 7240 VNPASS(!VN_IS_DOOMED(vp), vp); 7241 break; 7242 } 7243 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 7244 7245 /* Check if we are done */ 7246 if (vp == NULL) { 7247 mtx_unlock(&mp->mnt_listmtx); 7248 mnt_vnode_markerfree_lazy(mvp, mp); 7249 return (NULL); 7250 } 7251 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist); 7252 mtx_unlock(&mp->mnt_listmtx); 7253 ASSERT_VI_LOCKED(vp, "lazy iter"); 7254 return (vp); 7255 } 7256 7257 struct vnode * 7258 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 7259 void *cbarg) 7260 { 7261 7262 maybe_yield(); 7263 mtx_lock(&mp->mnt_listmtx); 7264 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 7265 } 7266 7267 struct vnode * 7268 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 7269 void *cbarg) 7270 { 7271 struct vnode *vp; 7272 7273 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist)) 7274 return (NULL); 7275 7276 *mvp = vn_alloc_marker(mp); 7277 MNT_ILOCK(mp); 7278 MNT_REF(mp); 7279 MNT_IUNLOCK(mp); 7280 7281 mtx_lock(&mp->mnt_listmtx); 7282 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist); 7283 if (vp == NULL) { 7284 mtx_unlock(&mp->mnt_listmtx); 7285 mnt_vnode_markerfree_lazy(mvp, mp); 7286 return (NULL); 7287 } 7288 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist); 7289 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 7290 } 7291 7292 void 7293 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 7294 { 7295 7296 if (*mvp == NULL) 7297 return; 7298 7299 mtx_lock(&mp->mnt_listmtx); 7300 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 7301 mtx_unlock(&mp->mnt_listmtx); 7302 mnt_vnode_markerfree_lazy(mvp, mp); 7303 } 7304 7305 int 7306 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp) 7307 { 7308 7309 if ((cnp->cn_flags & NOEXECCHECK) != 0) { 7310 cnp->cn_flags &= ~NOEXECCHECK; 7311 return (0); 7312 } 7313 7314 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread)); 7315 } 7316 7317 /* 7318 * Do not use this variant unless you have means other than the hold count 7319 * to prevent the vnode from getting freed. 7320 */ 7321 void 7322 vn_seqc_write_begin_locked(struct vnode *vp) 7323 { 7324 7325 ASSERT_VI_LOCKED(vp, __func__); 7326 VNPASS(vp->v_holdcnt > 0, vp); 7327 VNPASS(vp->v_seqc_users >= 0, vp); 7328 vp->v_seqc_users++; 7329 if (vp->v_seqc_users == 1) 7330 seqc_sleepable_write_begin(&vp->v_seqc); 7331 } 7332 7333 void 7334 vn_seqc_write_begin(struct vnode *vp) 7335 { 7336 7337 VI_LOCK(vp); 7338 vn_seqc_write_begin_locked(vp); 7339 VI_UNLOCK(vp); 7340 } 7341 7342 void 7343 vn_seqc_write_end_locked(struct vnode *vp) 7344 { 7345 7346 ASSERT_VI_LOCKED(vp, __func__); 7347 VNPASS(vp->v_seqc_users > 0, vp); 7348 vp->v_seqc_users--; 7349 if (vp->v_seqc_users == 0) 7350 seqc_sleepable_write_end(&vp->v_seqc); 7351 } 7352 7353 void 7354 vn_seqc_write_end(struct vnode *vp) 7355 { 7356 7357 VI_LOCK(vp); 7358 vn_seqc_write_end_locked(vp); 7359 VI_UNLOCK(vp); 7360 } 7361 7362 /* 7363 * Special case handling for allocating and freeing vnodes. 7364 * 7365 * The counter remains unchanged on free so that a doomed vnode will 7366 * keep testing as in modify as long as it is accessible with SMR. 7367 */ 7368 static void 7369 vn_seqc_init(struct vnode *vp) 7370 { 7371 7372 vp->v_seqc = 0; 7373 vp->v_seqc_users = 0; 7374 } 7375 7376 static void 7377 vn_seqc_write_end_free(struct vnode *vp) 7378 { 7379 7380 VNPASS(seqc_in_modify(vp->v_seqc), vp); 7381 VNPASS(vp->v_seqc_users == 1, vp); 7382 } 7383 7384 void 7385 vn_irflag_set_locked(struct vnode *vp, short toset) 7386 { 7387 short flags; 7388 7389 ASSERT_VI_LOCKED(vp, __func__); 7390 flags = vn_irflag_read(vp); 7391 VNASSERT((flags & toset) == 0, vp, 7392 ("%s: some of the passed flags already set (have %d, passed %d)\n", 7393 __func__, flags, toset)); 7394 atomic_store_short(&vp->v_irflag, flags | toset); 7395 } 7396 7397 void 7398 vn_irflag_set(struct vnode *vp, short toset) 7399 { 7400 7401 VI_LOCK(vp); 7402 vn_irflag_set_locked(vp, toset); 7403 VI_UNLOCK(vp); 7404 } 7405 7406 void 7407 vn_irflag_set_cond_locked(struct vnode *vp, short toset) 7408 { 7409 short flags; 7410 7411 ASSERT_VI_LOCKED(vp, __func__); 7412 flags = vn_irflag_read(vp); 7413 atomic_store_short(&vp->v_irflag, flags | toset); 7414 } 7415 7416 void 7417 vn_irflag_set_cond(struct vnode *vp, short toset) 7418 { 7419 7420 VI_LOCK(vp); 7421 vn_irflag_set_cond_locked(vp, toset); 7422 VI_UNLOCK(vp); 7423 } 7424 7425 void 7426 vn_irflag_unset_locked(struct vnode *vp, short tounset) 7427 { 7428 short flags; 7429 7430 ASSERT_VI_LOCKED(vp, __func__); 7431 flags = vn_irflag_read(vp); 7432 VNASSERT((flags & tounset) == tounset, vp, 7433 ("%s: some of the passed flags not set (have %d, passed %d)\n", 7434 __func__, flags, tounset)); 7435 atomic_store_short(&vp->v_irflag, flags & ~tounset); 7436 } 7437 7438 void 7439 vn_irflag_unset(struct vnode *vp, short tounset) 7440 { 7441 7442 VI_LOCK(vp); 7443 vn_irflag_unset_locked(vp, tounset); 7444 VI_UNLOCK(vp); 7445 } 7446 7447 int 7448 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred) 7449 { 7450 struct vattr vattr; 7451 int error; 7452 7453 ASSERT_VOP_LOCKED(vp, __func__); 7454 error = VOP_GETATTR(vp, &vattr, cred); 7455 if (__predict_true(error == 0)) { 7456 if (vattr.va_size <= OFF_MAX) 7457 *size = vattr.va_size; 7458 else 7459 error = EFBIG; 7460 } 7461 return (error); 7462 } 7463 7464 int 7465 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred) 7466 { 7467 int error; 7468 7469 VOP_LOCK(vp, LK_SHARED); 7470 error = vn_getsize_locked(vp, size, cred); 7471 VOP_UNLOCK(vp); 7472 return (error); 7473 } 7474 7475 #ifdef INVARIANTS 7476 void 7477 vn_set_state_validate(struct vnode *vp, __enum_uint8(vstate) state) 7478 { 7479 7480 switch (vp->v_state) { 7481 case VSTATE_UNINITIALIZED: 7482 switch (state) { 7483 case VSTATE_CONSTRUCTED: 7484 case VSTATE_DESTROYING: 7485 return; 7486 default: 7487 break; 7488 } 7489 break; 7490 case VSTATE_CONSTRUCTED: 7491 ASSERT_VOP_ELOCKED(vp, __func__); 7492 switch (state) { 7493 case VSTATE_DESTROYING: 7494 return; 7495 default: 7496 break; 7497 } 7498 break; 7499 case VSTATE_DESTROYING: 7500 ASSERT_VOP_ELOCKED(vp, __func__); 7501 switch (state) { 7502 case VSTATE_DEAD: 7503 return; 7504 default: 7505 break; 7506 } 7507 break; 7508 case VSTATE_DEAD: 7509 switch (state) { 7510 case VSTATE_UNINITIALIZED: 7511 return; 7512 default: 7513 break; 7514 } 7515 break; 7516 } 7517 7518 vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state); 7519 panic("invalid state transition %d -> %d\n", vp->v_state, state); 7520 } 7521 #endif 7522