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 <sys/cdefs.h> 42 #include "opt_ddb.h" 43 #include "opt_watchdog.h" 44 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/asan.h> 48 #include <sys/bio.h> 49 #include <sys/buf.h> 50 #include <sys/capsicum.h> 51 #include <sys/condvar.h> 52 #include <sys/conf.h> 53 #include <sys/counter.h> 54 #include <sys/dirent.h> 55 #include <sys/event.h> 56 #include <sys/eventhandler.h> 57 #include <sys/extattr.h> 58 #include <sys/file.h> 59 #include <sys/fcntl.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/sysctl.h> 81 #include <sys/syslog.h> 82 #include <sys/vmmeter.h> 83 #include <sys/vnode.h> 84 #include <sys/watchdog.h> 85 86 #include <machine/stdarg.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 #if defined(DEBUG_VFS_LOCKS) && (!defined(INVARIANTS) || !defined(WITNESS)) 101 #error DEBUG_VFS_LOCKS requires INVARIANTS and WITNESS 102 #endif 103 104 #ifdef DDB 105 #include <ddb/ddb.h> 106 #endif 107 108 static void delmntque(struct vnode *vp); 109 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, 110 int slpflag, int slptimeo); 111 static void syncer_shutdown(void *arg, int howto); 112 static int vtryrecycle(struct vnode *vp, bool isvnlru); 113 static void v_init_counters(struct vnode *); 114 static void vn_seqc_init(struct vnode *); 115 static void vn_seqc_write_end_free(struct vnode *vp); 116 static void vgonel(struct vnode *); 117 static bool vhold_recycle_free(struct vnode *); 118 static void vdropl_recycle(struct vnode *vp); 119 static void vdrop_recycle(struct vnode *vp); 120 static void vfs_knllock(void *arg); 121 static void vfs_knlunlock(void *arg); 122 static void vfs_knl_assert_lock(void *arg, int what); 123 static void destroy_vpollinfo(struct vpollinfo *vi); 124 static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, 125 daddr_t startlbn, daddr_t endlbn); 126 static void vnlru_recalc(void); 127 128 static SYSCTL_NODE(_vfs, OID_AUTO, vnode, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 129 "vnode configuration and statistics"); 130 static SYSCTL_NODE(_vfs_vnode, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 131 "vnode configuration"); 132 static SYSCTL_NODE(_vfs_vnode, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 133 "vnode statistics"); 134 static SYSCTL_NODE(_vfs_vnode, OID_AUTO, vnlru, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 135 "vnode recycling"); 136 137 /* 138 * Number of vnodes in existence. Increased whenever getnewvnode() 139 * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode. 140 */ 141 static u_long __exclusive_cache_line numvnodes; 142 143 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, 144 "Number of vnodes in existence (legacy)"); 145 SYSCTL_ULONG(_vfs_vnode_stats, OID_AUTO, count, CTLFLAG_RD, &numvnodes, 0, 146 "Number of vnodes in existence"); 147 148 static counter_u64_t vnodes_created; 149 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created, 150 "Number of vnodes created by getnewvnode (legacy)"); 151 SYSCTL_COUNTER_U64(_vfs_vnode_stats, OID_AUTO, created, CTLFLAG_RD, &vnodes_created, 152 "Number of vnodes created by getnewvnode"); 153 154 /* 155 * Conversion tables for conversion from vnode types to inode formats 156 * and back. 157 */ 158 __enum_uint8(vtype) iftovt_tab[16] = { 159 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 160 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON 161 }; 162 int vttoif_tab[10] = { 163 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 164 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT 165 }; 166 167 /* 168 * List of allocates vnodes in the system. 169 */ 170 static TAILQ_HEAD(freelst, vnode) vnode_list; 171 static struct vnode *vnode_list_free_marker; 172 static struct vnode *vnode_list_reclaim_marker; 173 174 /* 175 * "Free" vnode target. Free vnodes are rarely completely free, but are 176 * just ones that are cheap to recycle. Usually they are for files which 177 * have been stat'd but not read; these usually have inode and namecache 178 * data attached to them. This target is the preferred minimum size of a 179 * sub-cache consisting mostly of such files. The system balances the size 180 * of this sub-cache with its complement to try to prevent either from 181 * thrashing while the other is relatively inactive. The targets express 182 * a preference for the best balance. 183 * 184 * "Above" this target there are 2 further targets (watermarks) related 185 * to recyling of free vnodes. In the best-operating case, the cache is 186 * exactly full, the free list has size between vlowat and vhiwat above the 187 * free target, and recycling from it and normal use maintains this state. 188 * Sometimes the free list is below vlowat or even empty, but this state 189 * is even better for immediate use provided the cache is not full. 190 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free 191 * ones) to reach one of these states. The watermarks are currently hard- 192 * coded as 4% and 9% of the available space higher. These and the default 193 * of 25% for wantfreevnodes are too large if the memory size is large. 194 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim 195 * whenever vnlru_proc() becomes active. 196 */ 197 static long wantfreevnodes; 198 static long __exclusive_cache_line freevnodes; 199 static long freevnodes_old; 200 201 static u_long recycles_count; 202 SYSCTL_ULONG(_vfs, OID_AUTO, recycles, CTLFLAG_RD | CTLFLAG_STATS, &recycles_count, 0, 203 "Number of vnodes recycled to meet vnode cache targets (legacy)"); 204 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, recycles, CTLFLAG_RD | CTLFLAG_STATS, 205 &recycles_count, 0, 206 "Number of vnodes recycled to meet vnode cache targets"); 207 208 static u_long recycles_free_count; 209 SYSCTL_ULONG(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD | CTLFLAG_STATS, 210 &recycles_free_count, 0, 211 "Number of free vnodes recycled to meet vnode cache targets (legacy)"); 212 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, recycles_free, CTLFLAG_RD | CTLFLAG_STATS, 213 &recycles_free_count, 0, 214 "Number of free vnodes recycled to meet vnode cache targets"); 215 216 static counter_u64_t direct_recycles_free_count; 217 SYSCTL_COUNTER_U64(_vfs_vnode_vnlru, OID_AUTO, direct_recycles_free, CTLFLAG_RD, 218 &direct_recycles_free_count, 219 "Number of free vnodes recycled by vn_alloc callers to meet vnode cache targets"); 220 221 static counter_u64_t vnode_skipped_requeues; 222 SYSCTL_COUNTER_U64(_vfs_vnode_stats, OID_AUTO, skipped_requeues, CTLFLAG_RD, &vnode_skipped_requeues, 223 "Number of times LRU requeue was skipped due to lock contention"); 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 * When shutting down the syncer, run it at four times normal speed. 323 */ 324 #define SYNCER_SHUTDOWN_SPEEDUP 4 325 static int sync_vnode_count; 326 static int syncer_worklist_len; 327 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 328 syncer_state; 329 330 /* Target for maximum number of vnodes. */ 331 u_long desiredvnodes; 332 static u_long gapvnodes; /* gap between wanted and desired */ 333 static u_long vhiwat; /* enough extras after expansion */ 334 static u_long vlowat; /* minimal extras before expansion */ 335 static bool vstir; /* nonzero to stir non-free vnodes */ 336 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */ 337 338 static u_long vnlru_read_freevnodes(void); 339 340 /* 341 * Note that no attempt is made to sanitize these parameters. 342 */ 343 static int 344 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS) 345 { 346 u_long val; 347 int error; 348 349 val = desiredvnodes; 350 error = sysctl_handle_long(oidp, &val, 0, req); 351 if (error != 0 || req->newptr == NULL) 352 return (error); 353 354 if (val == desiredvnodes) 355 return (0); 356 mtx_lock(&vnode_list_mtx); 357 desiredvnodes = val; 358 wantfreevnodes = desiredvnodes / 4; 359 vnlru_recalc(); 360 mtx_unlock(&vnode_list_mtx); 361 /* 362 * XXX There is no protection against multiple threads changing 363 * desiredvnodes at the same time. Locking above only helps vnlru and 364 * getnewvnode. 365 */ 366 vfs_hash_changesize(desiredvnodes); 367 cache_changesize(desiredvnodes); 368 return (0); 369 } 370 371 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes, 372 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes, 373 "LU", "Target for maximum number of vnodes (legacy)"); 374 SYSCTL_PROC(_vfs_vnode_param, OID_AUTO, limit, 375 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes, 376 "LU", "Target for maximum number of vnodes"); 377 378 static int 379 sysctl_freevnodes(SYSCTL_HANDLER_ARGS) 380 { 381 u_long rfreevnodes; 382 383 rfreevnodes = vnlru_read_freevnodes(); 384 return (sysctl_handle_long(oidp, &rfreevnodes, 0, req)); 385 } 386 387 SYSCTL_PROC(_vfs, OID_AUTO, freevnodes, 388 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_freevnodes, 389 "LU", "Number of \"free\" vnodes (legacy)"); 390 SYSCTL_PROC(_vfs_vnode_stats, OID_AUTO, free, 391 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_freevnodes, 392 "LU", "Number of \"free\" vnodes"); 393 394 static int 395 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS) 396 { 397 u_long val; 398 int error; 399 400 val = wantfreevnodes; 401 error = sysctl_handle_long(oidp, &val, 0, req); 402 if (error != 0 || req->newptr == NULL) 403 return (error); 404 405 if (val == wantfreevnodes) 406 return (0); 407 mtx_lock(&vnode_list_mtx); 408 wantfreevnodes = val; 409 vnlru_recalc(); 410 mtx_unlock(&vnode_list_mtx); 411 return (0); 412 } 413 414 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes, 415 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes, 416 "LU", "Target for minimum number of \"free\" vnodes (legacy)"); 417 SYSCTL_PROC(_vfs_vnode_param, OID_AUTO, wantfree, 418 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes, 419 "LU", "Target for minimum number of \"free\" vnodes"); 420 421 static int vnlru_nowhere; 422 SYSCTL_INT(_vfs_vnode_vnlru, OID_AUTO, failed_runs, CTLFLAG_RD | CTLFLAG_STATS, 423 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 424 425 static int 426 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS) 427 { 428 struct vnode *vp; 429 struct nameidata nd; 430 char *buf; 431 unsigned long ndflags; 432 int error; 433 434 if (req->newptr == NULL) 435 return (EINVAL); 436 if (req->newlen >= PATH_MAX) 437 return (E2BIG); 438 439 buf = malloc(PATH_MAX, M_TEMP, M_WAITOK); 440 error = SYSCTL_IN(req, buf, req->newlen); 441 if (error != 0) 442 goto out; 443 444 buf[req->newlen] = '\0'; 445 446 ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1; 447 NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf); 448 if ((error = namei(&nd)) != 0) 449 goto out; 450 vp = nd.ni_vp; 451 452 if (VN_IS_DOOMED(vp)) { 453 /* 454 * This vnode is being recycled. Return != 0 to let the caller 455 * know that the sysctl had no effect. Return EAGAIN because a 456 * subsequent call will likely succeed (since namei will create 457 * a new vnode if necessary) 458 */ 459 error = EAGAIN; 460 goto putvnode; 461 } 462 463 vgone(vp); 464 putvnode: 465 vput(vp); 466 NDFREE_PNBUF(&nd); 467 out: 468 free(buf, M_TEMP); 469 return (error); 470 } 471 472 static int 473 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS) 474 { 475 struct thread *td = curthread; 476 struct vnode *vp; 477 struct file *fp; 478 int error; 479 int fd; 480 481 if (req->newptr == NULL) 482 return (EBADF); 483 484 error = sysctl_handle_int(oidp, &fd, 0, req); 485 if (error != 0) 486 return (error); 487 error = getvnode(curthread, fd, &cap_fcntl_rights, &fp); 488 if (error != 0) 489 return (error); 490 vp = fp->f_vnode; 491 492 error = vn_lock(vp, LK_EXCLUSIVE); 493 if (error != 0) 494 goto drop; 495 496 vgone(vp); 497 VOP_UNLOCK(vp); 498 drop: 499 fdrop(fp, td); 500 return (error); 501 } 502 503 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode, 504 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, 505 sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname"); 506 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode, 507 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, 508 sysctl_ftry_reclaim_vnode, "I", 509 "Try to reclaim a vnode by its file descriptor"); 510 511 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */ 512 #define vnsz2log 8 513 #ifndef DEBUG_LOCKS 514 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log && 515 sizeof(struct vnode) < 1UL << (vnsz2log + 1), 516 "vnsz2log needs to be updated"); 517 #endif 518 519 /* 520 * Support for the bufobj clean & dirty pctrie. 521 */ 522 static void * 523 buf_trie_alloc(struct pctrie *ptree) 524 { 525 return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT)); 526 } 527 528 static void 529 buf_trie_free(struct pctrie *ptree, void *node) 530 { 531 uma_zfree_smr(buf_trie_zone, node); 532 } 533 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free, 534 buf_trie_smr); 535 536 /* 537 * Initialize the vnode management data structures. 538 * 539 * Reevaluate the following cap on the number of vnodes after the physical 540 * memory size exceeds 512GB. In the limit, as the physical memory size 541 * grows, the ratio of the memory size in KB to vnodes approaches 64:1. 542 */ 543 #ifndef MAXVNODES_MAX 544 #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */ 545 #endif 546 547 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker"); 548 549 static struct vnode * 550 vn_alloc_marker(struct mount *mp) 551 { 552 struct vnode *vp; 553 554 vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 555 vp->v_type = VMARKER; 556 vp->v_mount = mp; 557 558 return (vp); 559 } 560 561 static void 562 vn_free_marker(struct vnode *vp) 563 { 564 565 MPASS(vp->v_type == VMARKER); 566 free(vp, M_VNODE_MARKER); 567 } 568 569 #ifdef KASAN 570 static int 571 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused) 572 { 573 kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0); 574 return (0); 575 } 576 577 static void 578 vnode_dtor(void *mem, int size, void *arg __unused) 579 { 580 size_t end1, end2, off1, off2; 581 582 _Static_assert(offsetof(struct vnode, v_vnodelist) < 583 offsetof(struct vnode, v_dbatchcpu), 584 "KASAN marks require updating"); 585 586 off1 = offsetof(struct vnode, v_vnodelist); 587 off2 = offsetof(struct vnode, v_dbatchcpu); 588 end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist); 589 end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu); 590 591 /* 592 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even 593 * after the vnode has been freed. Try to get some KASAN coverage by 594 * marking everything except those two fields as invalid. Because 595 * KASAN's tracking is not byte-granular, any preceding fields sharing 596 * the same 8-byte aligned word must also be marked valid. 597 */ 598 599 /* Handle the area from the start until v_vnodelist... */ 600 off1 = rounddown2(off1, KASAN_SHADOW_SCALE); 601 kasan_mark(mem, off1, off1, KASAN_UMA_FREED); 602 603 /* ... then the area between v_vnodelist and v_dbatchcpu ... */ 604 off1 = roundup2(end1, KASAN_SHADOW_SCALE); 605 off2 = rounddown2(off2, KASAN_SHADOW_SCALE); 606 if (off2 > off1) 607 kasan_mark((void *)((char *)mem + off1), off2 - off1, 608 off2 - off1, KASAN_UMA_FREED); 609 610 /* ... and finally the area from v_dbatchcpu to the end. */ 611 off2 = roundup2(end2, KASAN_SHADOW_SCALE); 612 kasan_mark((void *)((char *)mem + off2), size - off2, size - off2, 613 KASAN_UMA_FREED); 614 } 615 #endif /* KASAN */ 616 617 /* 618 * Initialize a vnode as it first enters the zone. 619 */ 620 static int 621 vnode_init(void *mem, int size, int flags) 622 { 623 struct vnode *vp; 624 625 vp = mem; 626 bzero(vp, size); 627 /* 628 * Setup locks. 629 */ 630 vp->v_vnlock = &vp->v_lock; 631 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 632 /* 633 * By default, don't allow shared locks unless filesystems opt-in. 634 */ 635 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT, 636 LK_NOSHARE | LK_IS_VNODE); 637 /* 638 * Initialize bufobj. 639 */ 640 bufobj_init(&vp->v_bufobj, vp); 641 /* 642 * Initialize namecache. 643 */ 644 cache_vnode_init(vp); 645 /* 646 * Initialize rangelocks. 647 */ 648 rangelock_init(&vp->v_rl); 649 650 vp->v_dbatchcpu = NOCPU; 651 652 vp->v_state = VSTATE_DEAD; 653 654 /* 655 * Check vhold_recycle_free for an explanation. 656 */ 657 vp->v_holdcnt = VHOLD_NO_SMR; 658 vp->v_type = VNON; 659 mtx_lock(&vnode_list_mtx); 660 TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist); 661 mtx_unlock(&vnode_list_mtx); 662 return (0); 663 } 664 665 /* 666 * Free a vnode when it is cleared from the zone. 667 */ 668 static void 669 vnode_fini(void *mem, int size) 670 { 671 struct vnode *vp; 672 struct bufobj *bo; 673 674 vp = mem; 675 vdbatch_dequeue(vp); 676 mtx_lock(&vnode_list_mtx); 677 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 678 mtx_unlock(&vnode_list_mtx); 679 rangelock_destroy(&vp->v_rl); 680 lockdestroy(vp->v_vnlock); 681 mtx_destroy(&vp->v_interlock); 682 bo = &vp->v_bufobj; 683 rw_destroy(BO_LOCKPTR(bo)); 684 685 kasan_mark(mem, size, size, 0); 686 } 687 688 /* 689 * Provide the size of NFS nclnode and NFS fh for calculation of the 690 * vnode memory consumption. The size is specified directly to 691 * eliminate dependency on NFS-private header. 692 * 693 * Other filesystems may use bigger or smaller (like UFS and ZFS) 694 * private inode data, but the NFS-based estimation is ample enough. 695 * Still, we care about differences in the size between 64- and 32-bit 696 * platforms. 697 * 698 * Namecache structure size is heuristically 699 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1. 700 */ 701 #ifdef _LP64 702 #define NFS_NCLNODE_SZ (528 + 64) 703 #define NC_SZ 148 704 #else 705 #define NFS_NCLNODE_SZ (360 + 32) 706 #define NC_SZ 92 707 #endif 708 709 static void 710 vntblinit(void *dummy __unused) 711 { 712 struct vdbatch *vd; 713 uma_ctor ctor; 714 uma_dtor dtor; 715 int cpu, physvnodes, virtvnodes; 716 717 /* 718 * Desiredvnodes is a function of the physical memory size and the 719 * kernel's heap size. Generally speaking, it scales with the 720 * physical memory size. The ratio of desiredvnodes to the physical 721 * memory size is 1:16 until desiredvnodes exceeds 98,304. 722 * Thereafter, the 723 * marginal ratio of desiredvnodes to the physical memory size is 724 * 1:64. However, desiredvnodes is limited by the kernel's heap 725 * size. The memory required by desiredvnodes vnodes and vm objects 726 * must not exceed 1/10th of the kernel's heap size. 727 */ 728 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 + 729 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64; 730 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) + 731 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ)); 732 desiredvnodes = min(physvnodes, virtvnodes); 733 if (desiredvnodes > MAXVNODES_MAX) { 734 if (bootverbose) 735 printf("Reducing kern.maxvnodes %lu -> %lu\n", 736 desiredvnodes, MAXVNODES_MAX); 737 desiredvnodes = MAXVNODES_MAX; 738 } 739 wantfreevnodes = desiredvnodes / 4; 740 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 741 TAILQ_INIT(&vnode_list); 742 mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF); 743 /* 744 * The lock is taken to appease WITNESS. 745 */ 746 mtx_lock(&vnode_list_mtx); 747 vnlru_recalc(); 748 mtx_unlock(&vnode_list_mtx); 749 vnode_list_free_marker = vn_alloc_marker(NULL); 750 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist); 751 vnode_list_reclaim_marker = vn_alloc_marker(NULL); 752 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist); 753 754 #ifdef KASAN 755 ctor = vnode_ctor; 756 dtor = vnode_dtor; 757 #else 758 ctor = NULL; 759 dtor = NULL; 760 #endif 761 vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor, 762 vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN); 763 uma_zone_set_smr(vnode_zone, vfs_smr); 764 765 /* 766 * Preallocate enough nodes to support one-per buf so that 767 * we can not fail an insert. reassignbuf() callers can not 768 * tolerate the insertion failure. 769 */ 770 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(), 771 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, 772 UMA_ZONE_NOFREE | UMA_ZONE_SMR); 773 buf_trie_smr = uma_zone_get_smr(buf_trie_zone); 774 uma_prealloc(buf_trie_zone, nbuf); 775 776 vnodes_created = counter_u64_alloc(M_WAITOK); 777 direct_recycles_free_count = counter_u64_alloc(M_WAITOK); 778 vnode_skipped_requeues = counter_u64_alloc(M_WAITOK); 779 780 /* 781 * Initialize the filesystem syncer. 782 */ 783 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 784 &syncer_mask); 785 syncer_maxdelay = syncer_mask + 1; 786 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 787 cv_init(&sync_wakeup, "syncer"); 788 789 CPU_FOREACH(cpu) { 790 vd = DPCPU_ID_PTR((cpu), vd); 791 bzero(vd, sizeof(*vd)); 792 mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF); 793 } 794 } 795 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL); 796 797 /* 798 * Mark a mount point as busy. Used to synchronize access and to delay 799 * unmounting. Eventually, mountlist_mtx is not released on failure. 800 * 801 * vfs_busy() is a custom lock, it can block the caller. 802 * vfs_busy() only sleeps if the unmount is active on the mount point. 803 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any 804 * vnode belonging to mp. 805 * 806 * Lookup uses vfs_busy() to traverse mount points. 807 * root fs var fs 808 * / vnode lock A / vnode lock (/var) D 809 * /var vnode lock B /log vnode lock(/var/log) E 810 * vfs_busy lock C vfs_busy lock F 811 * 812 * Within each file system, the lock order is C->A->B and F->D->E. 813 * 814 * When traversing across mounts, the system follows that lock order: 815 * 816 * C->A->B 817 * | 818 * +->F->D->E 819 * 820 * The lookup() process for namei("/var") illustrates the process: 821 * 1. VOP_LOOKUP() obtains B while A is held 822 * 2. vfs_busy() obtains a shared lock on F while A and B are held 823 * 3. vput() releases lock on B 824 * 4. vput() releases lock on A 825 * 5. VFS_ROOT() obtains lock on D while shared lock on F is held 826 * 6. vfs_unbusy() releases shared lock on F 827 * 7. vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A. 828 * Attempt to lock A (instead of vp_crossmp) while D is held would 829 * violate the global order, causing deadlocks. 830 * 831 * dounmount() locks B while F is drained. Note that for stacked 832 * filesystems, D and B in the example above may be the same lock, 833 * which introdues potential lock order reversal deadlock between 834 * dounmount() and step 5 above. These filesystems may avoid the LOR 835 * by setting VV_CROSSLOCK on the covered vnode so that lock B will 836 * remain held until after step 5. 837 */ 838 int 839 vfs_busy(struct mount *mp, int flags) 840 { 841 struct mount_pcpu *mpcpu; 842 843 MPASS((flags & ~MBF_MASK) == 0); 844 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags); 845 846 if (vfs_op_thread_enter(mp, mpcpu)) { 847 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 848 MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0); 849 MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0); 850 vfs_mp_count_add_pcpu(mpcpu, ref, 1); 851 vfs_mp_count_add_pcpu(mpcpu, lockref, 1); 852 vfs_op_thread_exit(mp, mpcpu); 853 if (flags & MBF_MNTLSTLOCK) 854 mtx_unlock(&mountlist_mtx); 855 return (0); 856 } 857 858 MNT_ILOCK(mp); 859 vfs_assert_mount_counters(mp); 860 MNT_REF(mp); 861 /* 862 * If mount point is currently being unmounted, sleep until the 863 * mount point fate is decided. If thread doing the unmounting fails, 864 * it will clear MNTK_UNMOUNT flag before waking us up, indicating 865 * that this mount point has survived the unmount attempt and vfs_busy 866 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE 867 * flag in addition to MNTK_UNMOUNT, indicating that mount point is 868 * about to be really destroyed. vfs_busy needs to release its 869 * reference on the mount point in this case and return with ENOENT, 870 * telling the caller the mount it tried to busy is no longer valid. 871 */ 872 while (mp->mnt_kern_flag & MNTK_UNMOUNT) { 873 KASSERT(TAILQ_EMPTY(&mp->mnt_uppers), 874 ("%s: non-empty upper mount list with pending unmount", 875 __func__)); 876 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) { 877 MNT_REL(mp); 878 MNT_IUNLOCK(mp); 879 CTR1(KTR_VFS, "%s: failed busying before sleeping", 880 __func__); 881 return (ENOENT); 882 } 883 if (flags & MBF_MNTLSTLOCK) 884 mtx_unlock(&mountlist_mtx); 885 mp->mnt_kern_flag |= MNTK_MWAIT; 886 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0); 887 if (flags & MBF_MNTLSTLOCK) 888 mtx_lock(&mountlist_mtx); 889 MNT_ILOCK(mp); 890 } 891 if (flags & MBF_MNTLSTLOCK) 892 mtx_unlock(&mountlist_mtx); 893 mp->mnt_lockref++; 894 MNT_IUNLOCK(mp); 895 return (0); 896 } 897 898 /* 899 * Free a busy filesystem. 900 */ 901 void 902 vfs_unbusy(struct mount *mp) 903 { 904 struct mount_pcpu *mpcpu; 905 int c; 906 907 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 908 909 if (vfs_op_thread_enter(mp, mpcpu)) { 910 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 911 vfs_mp_count_sub_pcpu(mpcpu, lockref, 1); 912 vfs_mp_count_sub_pcpu(mpcpu, ref, 1); 913 vfs_op_thread_exit(mp, mpcpu); 914 return; 915 } 916 917 MNT_ILOCK(mp); 918 vfs_assert_mount_counters(mp); 919 MNT_REL(mp); 920 c = --mp->mnt_lockref; 921 if (mp->mnt_vfs_ops == 0) { 922 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 923 MNT_IUNLOCK(mp); 924 return; 925 } 926 if (c < 0) 927 vfs_dump_mount_counters(mp); 928 if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) { 929 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT); 930 CTR1(KTR_VFS, "%s: waking up waiters", __func__); 931 mp->mnt_kern_flag &= ~MNTK_DRAINING; 932 wakeup(&mp->mnt_lockref); 933 } 934 MNT_IUNLOCK(mp); 935 } 936 937 /* 938 * Lookup a mount point by filesystem identifier. 939 */ 940 struct mount * 941 vfs_getvfs(fsid_t *fsid) 942 { 943 struct mount *mp; 944 945 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 946 mtx_lock(&mountlist_mtx); 947 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 948 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) { 949 vfs_ref(mp); 950 mtx_unlock(&mountlist_mtx); 951 return (mp); 952 } 953 } 954 mtx_unlock(&mountlist_mtx); 955 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 956 return ((struct mount *) 0); 957 } 958 959 /* 960 * Lookup a mount point by filesystem identifier, busying it before 961 * returning. 962 * 963 * To avoid congestion on mountlist_mtx, implement simple direct-mapped 964 * cache for popular filesystem identifiers. The cache is lockess, using 965 * the fact that struct mount's are never freed. In worst case we may 966 * get pointer to unmounted or even different filesystem, so we have to 967 * check what we got, and go slow way if so. 968 */ 969 struct mount * 970 vfs_busyfs(fsid_t *fsid) 971 { 972 #define FSID_CACHE_SIZE 256 973 typedef struct mount * volatile vmp_t; 974 static vmp_t cache[FSID_CACHE_SIZE]; 975 struct mount *mp; 976 int error; 977 uint32_t hash; 978 979 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 980 hash = fsid->val[0] ^ fsid->val[1]; 981 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1); 982 mp = cache[hash]; 983 if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0) 984 goto slow; 985 if (vfs_busy(mp, 0) != 0) { 986 cache[hash] = NULL; 987 goto slow; 988 } 989 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) 990 return (mp); 991 else 992 vfs_unbusy(mp); 993 994 slow: 995 mtx_lock(&mountlist_mtx); 996 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 997 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) { 998 error = vfs_busy(mp, MBF_MNTLSTLOCK); 999 if (error) { 1000 cache[hash] = NULL; 1001 mtx_unlock(&mountlist_mtx); 1002 return (NULL); 1003 } 1004 cache[hash] = mp; 1005 return (mp); 1006 } 1007 } 1008 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 1009 mtx_unlock(&mountlist_mtx); 1010 return ((struct mount *) 0); 1011 } 1012 1013 /* 1014 * Check if a user can access privileged mount options. 1015 */ 1016 int 1017 vfs_suser(struct mount *mp, struct thread *td) 1018 { 1019 int error; 1020 1021 if (jailed(td->td_ucred)) { 1022 /* 1023 * If the jail of the calling thread lacks permission for 1024 * this type of file system, deny immediately. 1025 */ 1026 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag)) 1027 return (EPERM); 1028 1029 /* 1030 * If the file system was mounted outside the jail of the 1031 * calling thread, deny immediately. 1032 */ 1033 if (prison_check(td->td_ucred, mp->mnt_cred) != 0) 1034 return (EPERM); 1035 } 1036 1037 /* 1038 * If file system supports delegated administration, we don't check 1039 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified 1040 * by the file system itself. 1041 * If this is not the user that did original mount, we check for 1042 * the PRIV_VFS_MOUNT_OWNER privilege. 1043 */ 1044 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) && 1045 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 1046 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) 1047 return (error); 1048 } 1049 return (0); 1050 } 1051 1052 /* 1053 * Get a new unique fsid. Try to make its val[0] unique, since this value 1054 * will be used to create fake device numbers for stat(). Also try (but 1055 * not so hard) make its val[0] unique mod 2^16, since some emulators only 1056 * support 16-bit device numbers. We end up with unique val[0]'s for the 1057 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 1058 * 1059 * Keep in mind that several mounts may be running in parallel. Starting 1060 * the search one past where the previous search terminated is both a 1061 * micro-optimization and a defense against returning the same fsid to 1062 * different mounts. 1063 */ 1064 void 1065 vfs_getnewfsid(struct mount *mp) 1066 { 1067 static uint16_t mntid_base; 1068 struct mount *nmp; 1069 fsid_t tfsid; 1070 int mtype; 1071 1072 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 1073 mtx_lock(&mntid_mtx); 1074 mtype = mp->mnt_vfc->vfc_typenum; 1075 tfsid.val[1] = mtype; 1076 mtype = (mtype & 0xFF) << 24; 1077 for (;;) { 1078 tfsid.val[0] = makedev(255, 1079 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 1080 mntid_base++; 1081 if ((nmp = vfs_getvfs(&tfsid)) == NULL) 1082 break; 1083 vfs_rel(nmp); 1084 } 1085 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 1086 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 1087 mtx_unlock(&mntid_mtx); 1088 } 1089 1090 /* 1091 * Knob to control the precision of file timestamps: 1092 * 1093 * 0 = seconds only; nanoseconds zeroed. 1094 * 1 = seconds and nanoseconds, accurate within 1/HZ. 1095 * 2 = seconds and nanoseconds, truncated to microseconds. 1096 * >=3 = seconds and nanoseconds, maximum precision. 1097 */ 1098 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 1099 1100 static int timestamp_precision = TSP_USEC; 1101 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 1102 ×tamp_precision, 0, "File timestamp precision (0: seconds, " 1103 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, " 1104 "3+: sec + ns (max. precision))"); 1105 1106 /* 1107 * Get a current timestamp. 1108 */ 1109 void 1110 vfs_timestamp(struct timespec *tsp) 1111 { 1112 struct timeval tv; 1113 1114 switch (timestamp_precision) { 1115 case TSP_SEC: 1116 tsp->tv_sec = time_second; 1117 tsp->tv_nsec = 0; 1118 break; 1119 case TSP_HZ: 1120 getnanotime(tsp); 1121 break; 1122 case TSP_USEC: 1123 microtime(&tv); 1124 TIMEVAL_TO_TIMESPEC(&tv, tsp); 1125 break; 1126 case TSP_NSEC: 1127 default: 1128 nanotime(tsp); 1129 break; 1130 } 1131 } 1132 1133 /* 1134 * Set vnode attributes to VNOVAL 1135 */ 1136 void 1137 vattr_null(struct vattr *vap) 1138 { 1139 1140 vap->va_type = VNON; 1141 vap->va_size = VNOVAL; 1142 vap->va_bytes = VNOVAL; 1143 vap->va_mode = VNOVAL; 1144 vap->va_nlink = VNOVAL; 1145 vap->va_uid = VNOVAL; 1146 vap->va_gid = VNOVAL; 1147 vap->va_fsid = VNOVAL; 1148 vap->va_fileid = VNOVAL; 1149 vap->va_blocksize = VNOVAL; 1150 vap->va_rdev = VNOVAL; 1151 vap->va_atime.tv_sec = VNOVAL; 1152 vap->va_atime.tv_nsec = VNOVAL; 1153 vap->va_mtime.tv_sec = VNOVAL; 1154 vap->va_mtime.tv_nsec = VNOVAL; 1155 vap->va_ctime.tv_sec = VNOVAL; 1156 vap->va_ctime.tv_nsec = VNOVAL; 1157 vap->va_birthtime.tv_sec = VNOVAL; 1158 vap->va_birthtime.tv_nsec = VNOVAL; 1159 vap->va_flags = VNOVAL; 1160 vap->va_gen = VNOVAL; 1161 vap->va_vaflags = 0; 1162 } 1163 1164 /* 1165 * Try to reduce the total number of vnodes. 1166 * 1167 * This routine (and its user) are buggy in at least the following ways: 1168 * - all parameters were picked years ago when RAM sizes were significantly 1169 * smaller 1170 * - it can pick vnodes based on pages used by the vm object, but filesystems 1171 * like ZFS don't use it making the pick broken 1172 * - since ZFS has its own aging policy it gets partially combated by this one 1173 * - a dedicated method should be provided for filesystems to let them decide 1174 * whether the vnode should be recycled 1175 * 1176 * This routine is called when we have too many vnodes. It attempts 1177 * to free <count> vnodes and will potentially free vnodes that still 1178 * have VM backing store (VM backing store is typically the cause 1179 * of a vnode blowout so we want to do this). Therefore, this operation 1180 * is not considered cheap. 1181 * 1182 * A number of conditions may prevent a vnode from being reclaimed. 1183 * the buffer cache may have references on the vnode, a directory 1184 * vnode may still have references due to the namei cache representing 1185 * underlying files, or the vnode may be in active use. It is not 1186 * desirable to reuse such vnodes. These conditions may cause the 1187 * number of vnodes to reach some minimum value regardless of what 1188 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 1189 * 1190 * @param reclaim_nc_src Only reclaim directories with outgoing namecache 1191 * entries if this argument is strue 1192 * @param trigger Only reclaim vnodes with fewer than this many resident 1193 * pages. 1194 * @param target How many vnodes to reclaim. 1195 * @return The number of vnodes that were reclaimed. 1196 */ 1197 static int 1198 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target) 1199 { 1200 struct vnode *vp, *mvp; 1201 struct mount *mp; 1202 struct vm_object *object; 1203 u_long done; 1204 bool retried; 1205 1206 mtx_assert(&vnode_list_mtx, MA_OWNED); 1207 1208 retried = false; 1209 done = 0; 1210 1211 mvp = vnode_list_reclaim_marker; 1212 restart: 1213 vp = mvp; 1214 while (done < target) { 1215 vp = TAILQ_NEXT(vp, v_vnodelist); 1216 if (__predict_false(vp == NULL)) 1217 break; 1218 1219 if (__predict_false(vp->v_type == VMARKER)) 1220 continue; 1221 1222 /* 1223 * If it's been deconstructed already, it's still 1224 * referenced, or it exceeds the trigger, skip it. 1225 * Also skip free vnodes. We are trying to make space 1226 * for more free vnodes, not reduce their count. 1227 */ 1228 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 || 1229 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src))) 1230 goto next_iter; 1231 1232 if (vp->v_type == VBAD || vp->v_type == VNON) 1233 goto next_iter; 1234 1235 object = atomic_load_ptr(&vp->v_object); 1236 if (object == NULL || object->resident_page_count > trigger) { 1237 goto next_iter; 1238 } 1239 1240 /* 1241 * Handle races against vnode allocation. Filesystems lock the 1242 * vnode some time after it gets returned from getnewvnode, 1243 * despite type and hold count being manipulated earlier. 1244 * Resorting to checking v_mount restores guarantees present 1245 * before the global list was reworked to contain all vnodes. 1246 */ 1247 if (!VI_TRYLOCK(vp)) 1248 goto next_iter; 1249 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) { 1250 VI_UNLOCK(vp); 1251 goto next_iter; 1252 } 1253 if (vp->v_mount == NULL) { 1254 VI_UNLOCK(vp); 1255 goto next_iter; 1256 } 1257 vholdl(vp); 1258 VI_UNLOCK(vp); 1259 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1260 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1261 mtx_unlock(&vnode_list_mtx); 1262 1263 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1264 vdrop_recycle(vp); 1265 goto next_iter_unlocked; 1266 } 1267 if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) { 1268 vdrop_recycle(vp); 1269 vn_finished_write(mp); 1270 goto next_iter_unlocked; 1271 } 1272 1273 VI_LOCK(vp); 1274 if (vp->v_usecount > 0 || 1275 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 1276 (vp->v_object != NULL && vp->v_object->handle == vp && 1277 vp->v_object->resident_page_count > trigger)) { 1278 VOP_UNLOCK(vp); 1279 vdropl_recycle(vp); 1280 vn_finished_write(mp); 1281 goto next_iter_unlocked; 1282 } 1283 recycles_count++; 1284 vgonel(vp); 1285 VOP_UNLOCK(vp); 1286 vdropl_recycle(vp); 1287 vn_finished_write(mp); 1288 done++; 1289 next_iter_unlocked: 1290 maybe_yield(); 1291 mtx_lock(&vnode_list_mtx); 1292 goto restart; 1293 next_iter: 1294 MPASS(vp->v_type != VMARKER); 1295 if (!should_yield()) 1296 continue; 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 kern_yield(PRI_USER); 1301 mtx_lock(&vnode_list_mtx); 1302 goto restart; 1303 } 1304 if (done == 0 && !retried) { 1305 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1306 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist); 1307 retried = true; 1308 goto restart; 1309 } 1310 return (done); 1311 } 1312 1313 static int max_free_per_call = 10000; 1314 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_free_per_call, 0, 1315 "limit on vnode free requests per call to the vnlru_free routine (legacy)"); 1316 SYSCTL_INT(_vfs_vnode_vnlru, OID_AUTO, max_free_per_call, CTLFLAG_RW, 1317 &max_free_per_call, 0, 1318 "limit on vnode free requests per call to the vnlru_free routine"); 1319 1320 /* 1321 * Attempt to recycle requested amount of free vnodes. 1322 */ 1323 static int 1324 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp, bool isvnlru) 1325 { 1326 struct vnode *vp; 1327 struct mount *mp; 1328 int ocount; 1329 bool retried; 1330 1331 mtx_assert(&vnode_list_mtx, MA_OWNED); 1332 if (count > max_free_per_call) 1333 count = max_free_per_call; 1334 if (count == 0) { 1335 mtx_unlock(&vnode_list_mtx); 1336 return (0); 1337 } 1338 ocount = count; 1339 retried = false; 1340 vp = mvp; 1341 for (;;) { 1342 vp = TAILQ_NEXT(vp, v_vnodelist); 1343 if (__predict_false(vp == NULL)) { 1344 /* 1345 * The free vnode marker can be past eligible vnodes: 1346 * 1. if vdbatch_process trylock failed 1347 * 2. if vtryrecycle failed 1348 * 1349 * If so, start the scan from scratch. 1350 */ 1351 if (!retried && vnlru_read_freevnodes() > 0) { 1352 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1353 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist); 1354 vp = mvp; 1355 retried = true; 1356 continue; 1357 } 1358 1359 /* 1360 * Give up 1361 */ 1362 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1363 TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist); 1364 mtx_unlock(&vnode_list_mtx); 1365 break; 1366 } 1367 if (__predict_false(vp->v_type == VMARKER)) 1368 continue; 1369 if (vp->v_holdcnt > 0) 1370 continue; 1371 /* 1372 * Don't recycle if our vnode is from different type 1373 * of mount point. Note that mp is type-safe, the 1374 * check does not reach unmapped address even if 1375 * vnode is reclaimed. 1376 */ 1377 if (mnt_op != NULL && (mp = vp->v_mount) != NULL && 1378 mp->mnt_op != mnt_op) { 1379 continue; 1380 } 1381 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) { 1382 continue; 1383 } 1384 if (!vhold_recycle_free(vp)) 1385 continue; 1386 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1387 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1388 mtx_unlock(&vnode_list_mtx); 1389 /* 1390 * FIXME: ignores the return value, meaning it may be nothing 1391 * got recycled but it claims otherwise to the caller. 1392 * 1393 * Originally the value started being ignored in 2005 with 1394 * 114a1006a8204aa156e1f9ad6476cdff89cada7f . 1395 * 1396 * Respecting the value can run into significant stalls if most 1397 * vnodes belong to one file system and it has writes 1398 * suspended. In presence of many threads and millions of 1399 * vnodes they keep contending on the vnode_list_mtx lock only 1400 * to find vnodes they can't recycle. 1401 * 1402 * The solution would be to pre-check if the vnode is likely to 1403 * be recycle-able, but it needs to happen with the 1404 * vnode_list_mtx lock held. This runs into a problem where 1405 * VOP_GETWRITEMOUNT (currently needed to find out about if 1406 * writes are frozen) can take locks which LOR against it. 1407 * 1408 * Check nullfs for one example (null_getwritemount). 1409 */ 1410 vtryrecycle(vp, isvnlru); 1411 count--; 1412 if (count == 0) { 1413 break; 1414 } 1415 mtx_lock(&vnode_list_mtx); 1416 vp = mvp; 1417 } 1418 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1419 return (ocount - count); 1420 } 1421 1422 /* 1423 * XXX: returns without vnode_list_mtx locked! 1424 */ 1425 static int 1426 vnlru_free_locked_direct(int count) 1427 { 1428 int ret; 1429 1430 mtx_assert(&vnode_list_mtx, MA_OWNED); 1431 ret = vnlru_free_impl(count, NULL, vnode_list_free_marker, false); 1432 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1433 return (ret); 1434 } 1435 1436 static int 1437 vnlru_free_locked_vnlru(int count) 1438 { 1439 int ret; 1440 1441 mtx_assert(&vnode_list_mtx, MA_OWNED); 1442 ret = vnlru_free_impl(count, NULL, vnode_list_free_marker, true); 1443 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1444 return (ret); 1445 } 1446 1447 static int 1448 vnlru_free_vnlru(int count) 1449 { 1450 1451 mtx_lock(&vnode_list_mtx); 1452 return (vnlru_free_locked_vnlru(count)); 1453 } 1454 1455 void 1456 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp) 1457 { 1458 1459 MPASS(mnt_op != NULL); 1460 MPASS(mvp != NULL); 1461 VNPASS(mvp->v_type == VMARKER, mvp); 1462 mtx_lock(&vnode_list_mtx); 1463 vnlru_free_impl(count, mnt_op, mvp, true); 1464 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1465 } 1466 1467 struct vnode * 1468 vnlru_alloc_marker(void) 1469 { 1470 struct vnode *mvp; 1471 1472 mvp = vn_alloc_marker(NULL); 1473 mtx_lock(&vnode_list_mtx); 1474 TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist); 1475 mtx_unlock(&vnode_list_mtx); 1476 return (mvp); 1477 } 1478 1479 void 1480 vnlru_free_marker(struct vnode *mvp) 1481 { 1482 mtx_lock(&vnode_list_mtx); 1483 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1484 mtx_unlock(&vnode_list_mtx); 1485 vn_free_marker(mvp); 1486 } 1487 1488 static void 1489 vnlru_recalc(void) 1490 { 1491 1492 mtx_assert(&vnode_list_mtx, MA_OWNED); 1493 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100); 1494 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */ 1495 vlowat = vhiwat / 2; 1496 } 1497 1498 /* 1499 * Attempt to recycle vnodes in a context that is always safe to block. 1500 * Calling vlrurecycle() from the bowels of filesystem code has some 1501 * interesting deadlock problems. 1502 */ 1503 static struct proc *vnlruproc; 1504 static int vnlruproc_sig; 1505 static u_long vnlruproc_kicks; 1506 1507 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, kicks, CTLFLAG_RD, &vnlruproc_kicks, 0, 1508 "Number of times vnlru awakened due to vnode shortage"); 1509 1510 #define VNLRU_COUNT_SLOP 100 1511 1512 /* 1513 * The main freevnodes counter is only updated when a counter local to CPU 1514 * diverges from 0 by more than VNLRU_FREEVNODES_SLOP. CPUs are conditionally 1515 * walked to compute a more accurate total. 1516 * 1517 * Note: the actual value at any given moment can still exceed slop, but it 1518 * should not be by significant margin in practice. 1519 */ 1520 #define VNLRU_FREEVNODES_SLOP 126 1521 1522 static void __noinline 1523 vfs_freevnodes_rollup(int8_t *lfreevnodes) 1524 { 1525 1526 atomic_add_long(&freevnodes, *lfreevnodes); 1527 *lfreevnodes = 0; 1528 critical_exit(); 1529 } 1530 1531 static __inline void 1532 vfs_freevnodes_inc(void) 1533 { 1534 int8_t *lfreevnodes; 1535 1536 critical_enter(); 1537 lfreevnodes = PCPU_PTR(vfs_freevnodes); 1538 (*lfreevnodes)++; 1539 if (__predict_false(*lfreevnodes == VNLRU_FREEVNODES_SLOP)) 1540 vfs_freevnodes_rollup(lfreevnodes); 1541 else 1542 critical_exit(); 1543 } 1544 1545 static __inline void 1546 vfs_freevnodes_dec(void) 1547 { 1548 int8_t *lfreevnodes; 1549 1550 critical_enter(); 1551 lfreevnodes = PCPU_PTR(vfs_freevnodes); 1552 (*lfreevnodes)--; 1553 if (__predict_false(*lfreevnodes == -VNLRU_FREEVNODES_SLOP)) 1554 vfs_freevnodes_rollup(lfreevnodes); 1555 else 1556 critical_exit(); 1557 } 1558 1559 static u_long 1560 vnlru_read_freevnodes(void) 1561 { 1562 long slop, rfreevnodes, rfreevnodes_old; 1563 int cpu; 1564 1565 rfreevnodes = atomic_load_long(&freevnodes); 1566 rfreevnodes_old = atomic_load_long(&freevnodes_old); 1567 1568 if (rfreevnodes > rfreevnodes_old) 1569 slop = rfreevnodes - rfreevnodes_old; 1570 else 1571 slop = rfreevnodes_old - rfreevnodes; 1572 if (slop < VNLRU_FREEVNODES_SLOP) 1573 return (rfreevnodes >= 0 ? rfreevnodes : 0); 1574 CPU_FOREACH(cpu) { 1575 rfreevnodes += cpuid_to_pcpu[cpu]->pc_vfs_freevnodes; 1576 } 1577 atomic_store_long(&freevnodes_old, rfreevnodes); 1578 return (freevnodes_old >= 0 ? freevnodes_old : 0); 1579 } 1580 1581 static bool 1582 vnlru_under(u_long rnumvnodes, u_long limit) 1583 { 1584 u_long rfreevnodes, space; 1585 1586 if (__predict_false(rnumvnodes > desiredvnodes)) 1587 return (true); 1588 1589 space = desiredvnodes - rnumvnodes; 1590 if (space < limit) { 1591 rfreevnodes = vnlru_read_freevnodes(); 1592 if (rfreevnodes > wantfreevnodes) 1593 space += rfreevnodes - wantfreevnodes; 1594 } 1595 return (space < limit); 1596 } 1597 1598 static void 1599 vnlru_kick_locked(void) 1600 { 1601 1602 mtx_assert(&vnode_list_mtx, MA_OWNED); 1603 if (vnlruproc_sig == 0) { 1604 vnlruproc_sig = 1; 1605 vnlruproc_kicks++; 1606 wakeup(vnlruproc); 1607 } 1608 } 1609 1610 static void 1611 vnlru_kick_cond(void) 1612 { 1613 1614 if (vnlru_read_freevnodes() > wantfreevnodes) 1615 return; 1616 1617 if (vnlruproc_sig) 1618 return; 1619 mtx_lock(&vnode_list_mtx); 1620 vnlru_kick_locked(); 1621 mtx_unlock(&vnode_list_mtx); 1622 } 1623 1624 static void 1625 vnlru_proc_sleep(void) 1626 { 1627 1628 if (vnlruproc_sig) { 1629 vnlruproc_sig = 0; 1630 wakeup(&vnlruproc_sig); 1631 } 1632 msleep(vnlruproc, &vnode_list_mtx, PVFS|PDROP, "vlruwt", hz); 1633 } 1634 1635 /* 1636 * A lighter version of the machinery below. 1637 * 1638 * Tries to reach goals only by recycling free vnodes and does not invoke 1639 * uma_reclaim(UMA_RECLAIM_DRAIN). 1640 * 1641 * This works around pathological behavior in vnlru in presence of tons of free 1642 * vnodes, but without having to rewrite the machinery at this time. Said 1643 * behavior boils down to continuously trying to reclaim all kinds of vnodes 1644 * (cycling through all levels of "force") when the count is transiently above 1645 * limit. This happens a lot when all vnodes are used up and vn_alloc 1646 * speculatively increments the counter. 1647 * 1648 * Sample testcase: vnode limit 8388608, 20 separate directory trees each with 1649 * 1 million files in total and 20 find(1) processes stating them in parallel 1650 * (one per each tree). 1651 * 1652 * On a kernel with only stock machinery this needs anywhere between 60 and 120 1653 * seconds to execute (time varies *wildly* between runs). With the workaround 1654 * it consistently stays around 20 seconds [it got further down with later 1655 * changes]. 1656 * 1657 * That is to say the entire thing needs a fundamental redesign (most notably 1658 * to accommodate faster recycling), the above only tries to get it ouf the way. 1659 * 1660 * Return values are: 1661 * -1 -- fallback to regular vnlru loop 1662 * 0 -- do nothing, go to sleep 1663 * >0 -- recycle this many vnodes 1664 */ 1665 static long 1666 vnlru_proc_light_pick(void) 1667 { 1668 u_long rnumvnodes, rfreevnodes; 1669 1670 if (vstir || vnlruproc_sig == 1) 1671 return (-1); 1672 1673 rnumvnodes = atomic_load_long(&numvnodes); 1674 rfreevnodes = vnlru_read_freevnodes(); 1675 1676 /* 1677 * vnode limit might have changed and now we may be at a significant 1678 * excess. Bail if we can't sort it out with free vnodes. 1679 * 1680 * Due to atomic updates the count can legitimately go above 1681 * the limit for a short period, don't bother doing anything in 1682 * that case. 1683 */ 1684 if (rnumvnodes > desiredvnodes + VNLRU_COUNT_SLOP + 10) { 1685 if (rnumvnodes - rfreevnodes >= desiredvnodes || 1686 rfreevnodes <= wantfreevnodes) { 1687 return (-1); 1688 } 1689 1690 return (rnumvnodes - desiredvnodes); 1691 } 1692 1693 /* 1694 * Don't try to reach wantfreevnodes target if there are too few vnodes 1695 * to begin with. 1696 */ 1697 if (rnumvnodes < wantfreevnodes) { 1698 return (0); 1699 } 1700 1701 if (rfreevnodes < wantfreevnodes) { 1702 return (-1); 1703 } 1704 1705 return (0); 1706 } 1707 1708 static bool 1709 vnlru_proc_light(void) 1710 { 1711 long freecount; 1712 1713 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1714 1715 freecount = vnlru_proc_light_pick(); 1716 if (freecount == -1) 1717 return (false); 1718 1719 if (freecount != 0) { 1720 vnlru_free_vnlru(freecount); 1721 } 1722 1723 mtx_lock(&vnode_list_mtx); 1724 vnlru_proc_sleep(); 1725 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1726 return (true); 1727 } 1728 1729 static u_long uma_reclaim_calls; 1730 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, uma_reclaim_calls, CTLFLAG_RD | CTLFLAG_STATS, 1731 &uma_reclaim_calls, 0, "Number of calls to uma_reclaim"); 1732 1733 static void 1734 vnlru_proc(void) 1735 { 1736 u_long rnumvnodes, rfreevnodes, target; 1737 unsigned long onumvnodes; 1738 int done, force, trigger, usevnodes; 1739 bool reclaim_nc_src, want_reread; 1740 1741 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc, 1742 SHUTDOWN_PRI_FIRST); 1743 1744 force = 0; 1745 want_reread = false; 1746 for (;;) { 1747 kproc_suspend_check(vnlruproc); 1748 1749 if (force == 0 && vnlru_proc_light()) 1750 continue; 1751 1752 mtx_lock(&vnode_list_mtx); 1753 rnumvnodes = atomic_load_long(&numvnodes); 1754 1755 if (want_reread) { 1756 force = vnlru_under(numvnodes, vhiwat) ? 1 : 0; 1757 want_reread = false; 1758 } 1759 1760 /* 1761 * If numvnodes is too large (due to desiredvnodes being 1762 * adjusted using its sysctl, or emergency growth), first 1763 * try to reduce it by discarding free vnodes. 1764 */ 1765 if (rnumvnodes > desiredvnodes + 10) { 1766 vnlru_free_locked_vnlru(rnumvnodes - desiredvnodes); 1767 mtx_lock(&vnode_list_mtx); 1768 rnumvnodes = atomic_load_long(&numvnodes); 1769 } 1770 /* 1771 * Sleep if the vnode cache is in a good state. This is 1772 * when it is not over-full and has space for about a 4% 1773 * or 9% expansion (by growing its size or inexcessively 1774 * reducing free vnode count). Otherwise, try to reclaim 1775 * space for a 10% expansion. 1776 */ 1777 if (vstir && force == 0) { 1778 force = 1; 1779 vstir = false; 1780 } 1781 if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) { 1782 vnlru_proc_sleep(); 1783 continue; 1784 } 1785 rfreevnodes = vnlru_read_freevnodes(); 1786 1787 onumvnodes = rnumvnodes; 1788 /* 1789 * Calculate parameters for recycling. These are the same 1790 * throughout the loop to give some semblance of fairness. 1791 * The trigger point is to avoid recycling vnodes with lots 1792 * of resident pages. We aren't trying to free memory; we 1793 * are trying to recycle or at least free vnodes. 1794 */ 1795 if (rnumvnodes <= desiredvnodes) 1796 usevnodes = rnumvnodes - rfreevnodes; 1797 else 1798 usevnodes = rnumvnodes; 1799 if (usevnodes <= 0) 1800 usevnodes = 1; 1801 /* 1802 * The trigger value is chosen to give a conservatively 1803 * large value to ensure that it alone doesn't prevent 1804 * making progress. The value can easily be so large that 1805 * it is effectively infinite in some congested and 1806 * misconfigured cases, and this is necessary. Normally 1807 * it is about 8 to 100 (pages), which is quite large. 1808 */ 1809 trigger = vm_cnt.v_page_count * 2 / usevnodes; 1810 if (force < 2) 1811 trigger = vsmalltrigger; 1812 reclaim_nc_src = force >= 3; 1813 target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1); 1814 target = target / 10 + 1; 1815 done = vlrureclaim(reclaim_nc_src, trigger, target); 1816 mtx_unlock(&vnode_list_mtx); 1817 /* 1818 * Total number of vnodes can transiently go slightly above the 1819 * limit (see vn_alloc_hard), no need to call uma_reclaim if 1820 * this happens. 1821 */ 1822 if (onumvnodes + VNLRU_COUNT_SLOP + 1000 > desiredvnodes && 1823 numvnodes <= desiredvnodes) { 1824 uma_reclaim_calls++; 1825 uma_reclaim(UMA_RECLAIM_DRAIN); 1826 } 1827 if (done == 0) { 1828 if (force == 0 || force == 1) { 1829 force = 2; 1830 continue; 1831 } 1832 if (force == 2) { 1833 force = 3; 1834 continue; 1835 } 1836 want_reread = true; 1837 force = 0; 1838 vnlru_nowhere++; 1839 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 1840 } else { 1841 want_reread = true; 1842 kern_yield(PRI_USER); 1843 } 1844 } 1845 } 1846 1847 static struct kproc_desc vnlru_kp = { 1848 "vnlru", 1849 vnlru_proc, 1850 &vnlruproc 1851 }; 1852 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, 1853 &vnlru_kp); 1854 1855 /* 1856 * Routines having to do with the management of the vnode table. 1857 */ 1858 1859 /* 1860 * Try to recycle a freed vnode. 1861 */ 1862 static int 1863 vtryrecycle(struct vnode *vp, bool isvnlru) 1864 { 1865 struct mount *vnmp; 1866 1867 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 1868 VNPASS(vp->v_holdcnt > 0, vp); 1869 /* 1870 * This vnode may found and locked via some other list, if so we 1871 * can't recycle it yet. 1872 */ 1873 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 1874 CTR2(KTR_VFS, 1875 "%s: impossible to recycle, vp %p lock is already held", 1876 __func__, vp); 1877 vdrop_recycle(vp); 1878 return (EWOULDBLOCK); 1879 } 1880 /* 1881 * Don't recycle if its filesystem is being suspended. 1882 */ 1883 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 1884 VOP_UNLOCK(vp); 1885 CTR2(KTR_VFS, 1886 "%s: impossible to recycle, cannot start the write for %p", 1887 __func__, vp); 1888 vdrop_recycle(vp); 1889 return (EBUSY); 1890 } 1891 /* 1892 * If we got this far, we need to acquire the interlock and see if 1893 * anyone picked up this vnode from another list. If not, we will 1894 * mark it with DOOMED via vgonel() so that anyone who does find it 1895 * will skip over it. 1896 */ 1897 VI_LOCK(vp); 1898 if (vp->v_usecount) { 1899 VOP_UNLOCK(vp); 1900 vdropl_recycle(vp); 1901 vn_finished_write(vnmp); 1902 CTR2(KTR_VFS, 1903 "%s: impossible to recycle, %p is already referenced", 1904 __func__, vp); 1905 return (EBUSY); 1906 } 1907 if (!VN_IS_DOOMED(vp)) { 1908 if (isvnlru) 1909 recycles_free_count++; 1910 else 1911 counter_u64_add(direct_recycles_free_count, 1); 1912 vgonel(vp); 1913 } 1914 VOP_UNLOCK(vp); 1915 vdropl_recycle(vp); 1916 vn_finished_write(vnmp); 1917 return (0); 1918 } 1919 1920 /* 1921 * Allocate a new vnode. 1922 * 1923 * The operation never returns an error. Returning an error was disabled 1924 * in r145385 (dated 2005) with the following comment: 1925 * 1926 * XXX Not all VFS_VGET/ffs_vget callers check returns. 1927 * 1928 * Given the age of this commit (almost 15 years at the time of writing this 1929 * comment) restoring the ability to fail requires a significant audit of 1930 * all codepaths. 1931 * 1932 * The routine can try to free a vnode or stall for up to 1 second waiting for 1933 * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation. 1934 */ 1935 static u_long vn_alloc_cyclecount; 1936 static u_long vn_alloc_sleeps; 1937 1938 SYSCTL_ULONG(_vfs_vnode_stats, OID_AUTO, alloc_sleeps, CTLFLAG_RD, &vn_alloc_sleeps, 0, 1939 "Number of times vnode allocation blocked waiting on vnlru"); 1940 1941 static struct vnode * __noinline 1942 vn_alloc_hard(struct mount *mp, u_long rnumvnodes, bool bumped) 1943 { 1944 u_long rfreevnodes; 1945 1946 if (bumped) { 1947 if (rnumvnodes > desiredvnodes + VNLRU_COUNT_SLOP) { 1948 atomic_subtract_long(&numvnodes, 1); 1949 bumped = false; 1950 } 1951 } 1952 1953 mtx_lock(&vnode_list_mtx); 1954 1955 if (vn_alloc_cyclecount != 0) { 1956 rnumvnodes = atomic_load_long(&numvnodes); 1957 if (rnumvnodes + 1 < desiredvnodes) { 1958 vn_alloc_cyclecount = 0; 1959 mtx_unlock(&vnode_list_mtx); 1960 goto alloc; 1961 } 1962 1963 rfreevnodes = vnlru_read_freevnodes(); 1964 if (rfreevnodes < wantfreevnodes) { 1965 if (vn_alloc_cyclecount++ >= rfreevnodes) { 1966 vn_alloc_cyclecount = 0; 1967 vstir = true; 1968 } 1969 } else { 1970 vn_alloc_cyclecount = 0; 1971 } 1972 } 1973 1974 /* 1975 * Grow the vnode cache if it will not be above its target max after 1976 * growing. Otherwise, if there is at least one free vnode, try to 1977 * reclaim 1 item from it before growing the cache (possibly above its 1978 * target max if the reclamation failed or is delayed). 1979 */ 1980 if (vnlru_free_locked_direct(1) > 0) 1981 goto alloc; 1982 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 1983 if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 1984 /* 1985 * Wait for space for a new vnode. 1986 */ 1987 if (bumped) { 1988 atomic_subtract_long(&numvnodes, 1); 1989 bumped = false; 1990 } 1991 mtx_lock(&vnode_list_mtx); 1992 vnlru_kick_locked(); 1993 vn_alloc_sleeps++; 1994 msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz); 1995 if (atomic_load_long(&numvnodes) + 1 > desiredvnodes && 1996 vnlru_read_freevnodes() > 1) 1997 vnlru_free_locked_direct(1); 1998 else 1999 mtx_unlock(&vnode_list_mtx); 2000 } 2001 alloc: 2002 mtx_assert(&vnode_list_mtx, MA_NOTOWNED); 2003 if (!bumped) 2004 atomic_add_long(&numvnodes, 1); 2005 vnlru_kick_cond(); 2006 return (uma_zalloc_smr(vnode_zone, M_WAITOK)); 2007 } 2008 2009 static struct vnode * 2010 vn_alloc(struct mount *mp) 2011 { 2012 u_long rnumvnodes; 2013 2014 if (__predict_false(vn_alloc_cyclecount != 0)) 2015 return (vn_alloc_hard(mp, 0, false)); 2016 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; 2017 if (__predict_false(vnlru_under(rnumvnodes, vlowat))) { 2018 return (vn_alloc_hard(mp, rnumvnodes, true)); 2019 } 2020 2021 return (uma_zalloc_smr(vnode_zone, M_WAITOK)); 2022 } 2023 2024 static void 2025 vn_free(struct vnode *vp) 2026 { 2027 2028 atomic_subtract_long(&numvnodes, 1); 2029 uma_zfree_smr(vnode_zone, vp); 2030 } 2031 2032 /* 2033 * Allocate a new vnode. 2034 */ 2035 int 2036 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, 2037 struct vnode **vpp) 2038 { 2039 struct vnode *vp; 2040 struct thread *td; 2041 struct lock_object *lo; 2042 2043 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag); 2044 2045 KASSERT(vops->registered, 2046 ("%s: not registered vector op %p\n", __func__, vops)); 2047 cache_validate_vop_vector(mp, vops); 2048 2049 td = curthread; 2050 if (td->td_vp_reserved != NULL) { 2051 vp = td->td_vp_reserved; 2052 td->td_vp_reserved = NULL; 2053 } else { 2054 vp = vn_alloc(mp); 2055 } 2056 counter_u64_add(vnodes_created, 1); 2057 2058 vn_set_state(vp, VSTATE_UNINITIALIZED); 2059 2060 /* 2061 * Locks are given the generic name "vnode" when created. 2062 * Follow the historic practice of using the filesystem 2063 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc. 2064 * 2065 * Locks live in a witness group keyed on their name. Thus, 2066 * when a lock is renamed, it must also move from the witness 2067 * group of its old name to the witness group of its new name. 2068 * 2069 * The change only needs to be made when the vnode moves 2070 * from one filesystem type to another. We ensure that each 2071 * filesystem use a single static name pointer for its tag so 2072 * that we can compare pointers rather than doing a strcmp(). 2073 */ 2074 lo = &vp->v_vnlock->lock_object; 2075 #ifdef WITNESS 2076 if (lo->lo_name != tag) { 2077 #endif 2078 lo->lo_name = tag; 2079 #ifdef WITNESS 2080 WITNESS_DESTROY(lo); 2081 WITNESS_INIT(lo, tag); 2082 } 2083 #endif 2084 /* 2085 * By default, don't allow shared locks unless filesystems opt-in. 2086 */ 2087 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE; 2088 /* 2089 * Finalize various vnode identity bits. 2090 */ 2091 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp)); 2092 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp)); 2093 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp)); 2094 vp->v_type = VNON; 2095 vp->v_op = vops; 2096 vp->v_irflag = 0; 2097 v_init_counters(vp); 2098 vn_seqc_init(vp); 2099 vp->v_bufobj.bo_ops = &buf_ops_bio; 2100 #ifdef DIAGNOSTIC 2101 if (mp == NULL && vops != &dead_vnodeops) 2102 printf("NULL mp in getnewvnode(9), tag %s\n", tag); 2103 #endif 2104 #ifdef MAC 2105 mac_vnode_init(vp); 2106 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 2107 mac_vnode_associate_singlelabel(mp, vp); 2108 #endif 2109 if (mp != NULL) { 2110 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize; 2111 } 2112 2113 /* 2114 * For the filesystems which do not use vfs_hash_insert(), 2115 * still initialize v_hash to have vfs_hash_index() useful. 2116 * E.g., nullfs uses vfs_hash_index() on the lower vnode for 2117 * its own hashing. 2118 */ 2119 vp->v_hash = (uintptr_t)vp >> vnsz2log; 2120 2121 *vpp = vp; 2122 return (0); 2123 } 2124 2125 void 2126 getnewvnode_reserve(void) 2127 { 2128 struct thread *td; 2129 2130 td = curthread; 2131 MPASS(td->td_vp_reserved == NULL); 2132 td->td_vp_reserved = vn_alloc(NULL); 2133 } 2134 2135 void 2136 getnewvnode_drop_reserve(void) 2137 { 2138 struct thread *td; 2139 2140 td = curthread; 2141 if (td->td_vp_reserved != NULL) { 2142 vn_free(td->td_vp_reserved); 2143 td->td_vp_reserved = NULL; 2144 } 2145 } 2146 2147 static void __noinline 2148 freevnode(struct vnode *vp) 2149 { 2150 struct bufobj *bo; 2151 2152 /* 2153 * The vnode has been marked for destruction, so free it. 2154 * 2155 * The vnode will be returned to the zone where it will 2156 * normally remain until it is needed for another vnode. We 2157 * need to cleanup (or verify that the cleanup has already 2158 * been done) any residual data left from its current use 2159 * so as not to contaminate the freshly allocated vnode. 2160 */ 2161 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp); 2162 /* 2163 * Paired with vgone. 2164 */ 2165 vn_seqc_write_end_free(vp); 2166 2167 bo = &vp->v_bufobj; 2168 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 2169 VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp); 2170 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 2171 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 2172 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 2173 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 2174 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp, 2175 ("clean blk trie not empty")); 2176 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 2177 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp, 2178 ("dirty blk trie not empty")); 2179 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp, 2180 ("Dangling rangelock waiters")); 2181 VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp, 2182 ("Leaked inactivation")); 2183 VI_UNLOCK(vp); 2184 cache_assert_no_entries(vp); 2185 2186 #ifdef MAC 2187 mac_vnode_destroy(vp); 2188 #endif 2189 if (vp->v_pollinfo != NULL) { 2190 /* 2191 * Use LK_NOWAIT to shut up witness about the lock. We may get 2192 * here while having another vnode locked when trying to 2193 * satisfy a lookup and needing to recycle. 2194 */ 2195 VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT); 2196 destroy_vpollinfo(vp->v_pollinfo); 2197 VOP_UNLOCK(vp); 2198 vp->v_pollinfo = NULL; 2199 } 2200 vp->v_mountedhere = NULL; 2201 vp->v_unpcb = NULL; 2202 vp->v_rdev = NULL; 2203 vp->v_fifoinfo = NULL; 2204 vp->v_iflag = 0; 2205 vp->v_vflag = 0; 2206 bo->bo_flag = 0; 2207 vn_free(vp); 2208 } 2209 2210 /* 2211 * Delete from old mount point vnode list, if on one. 2212 */ 2213 static void 2214 delmntque(struct vnode *vp) 2215 { 2216 struct mount *mp; 2217 2218 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); 2219 2220 mp = vp->v_mount; 2221 MNT_ILOCK(mp); 2222 VI_LOCK(vp); 2223 vp->v_mount = NULL; 2224 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 2225 ("bad mount point vnode list size")); 2226 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2227 mp->mnt_nvnodelistsize--; 2228 MNT_REL(mp); 2229 MNT_IUNLOCK(mp); 2230 /* 2231 * The caller expects the interlock to be still held. 2232 */ 2233 ASSERT_VI_LOCKED(vp, __func__); 2234 } 2235 2236 static int 2237 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr) 2238 { 2239 2240 KASSERT(vp->v_mount == NULL, 2241 ("insmntque: vnode already on per mount vnode list")); 2242 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 2243 if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) { 2244 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); 2245 } else { 2246 KASSERT(!dtr, 2247 ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup", 2248 __func__)); 2249 } 2250 2251 /* 2252 * We acquire the vnode interlock early to ensure that the 2253 * vnode cannot be recycled by another process releasing a 2254 * holdcnt on it before we get it on both the vnode list 2255 * and the active vnode list. The mount mutex protects only 2256 * manipulation of the vnode list and the vnode freelist 2257 * mutex protects only manipulation of the active vnode list. 2258 * Hence the need to hold the vnode interlock throughout. 2259 */ 2260 MNT_ILOCK(mp); 2261 VI_LOCK(vp); 2262 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 && 2263 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || 2264 mp->mnt_nvnodelistsize == 0)) && 2265 (vp->v_vflag & VV_FORCEINSMQ) == 0) { 2266 VI_UNLOCK(vp); 2267 MNT_IUNLOCK(mp); 2268 if (dtr) { 2269 vp->v_data = NULL; 2270 vp->v_op = &dead_vnodeops; 2271 vgone(vp); 2272 vput(vp); 2273 } 2274 return (EBUSY); 2275 } 2276 vp->v_mount = mp; 2277 MNT_REF(mp); 2278 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2279 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 2280 ("neg mount point vnode list size")); 2281 mp->mnt_nvnodelistsize++; 2282 VI_UNLOCK(vp); 2283 MNT_IUNLOCK(mp); 2284 return (0); 2285 } 2286 2287 /* 2288 * Insert into list of vnodes for the new mount point, if available. 2289 * insmntque() reclaims the vnode on insertion failure, insmntque1() 2290 * leaves handling of the vnode to the caller. 2291 */ 2292 int 2293 insmntque(struct vnode *vp, struct mount *mp) 2294 { 2295 return (insmntque1_int(vp, mp, true)); 2296 } 2297 2298 int 2299 insmntque1(struct vnode *vp, struct mount *mp) 2300 { 2301 return (insmntque1_int(vp, mp, false)); 2302 } 2303 2304 /* 2305 * Flush out and invalidate all buffers associated with a bufobj 2306 * Called with the underlying object locked. 2307 */ 2308 int 2309 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) 2310 { 2311 int error; 2312 2313 BO_LOCK(bo); 2314 if (flags & V_SAVE) { 2315 error = bufobj_wwait(bo, slpflag, slptimeo); 2316 if (error) { 2317 BO_UNLOCK(bo); 2318 return (error); 2319 } 2320 if (bo->bo_dirty.bv_cnt > 0) { 2321 BO_UNLOCK(bo); 2322 do { 2323 error = BO_SYNC(bo, MNT_WAIT); 2324 } while (error == ERELOOKUP); 2325 if (error != 0) 2326 return (error); 2327 BO_LOCK(bo); 2328 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) { 2329 BO_UNLOCK(bo); 2330 return (EBUSY); 2331 } 2332 } 2333 } 2334 /* 2335 * If you alter this loop please notice that interlock is dropped and 2336 * reacquired in flushbuflist. Special care is needed to ensure that 2337 * no race conditions occur from this. 2338 */ 2339 do { 2340 error = flushbuflist(&bo->bo_clean, 2341 flags, bo, slpflag, slptimeo); 2342 if (error == 0 && !(flags & V_CLEANONLY)) 2343 error = flushbuflist(&bo->bo_dirty, 2344 flags, bo, slpflag, slptimeo); 2345 if (error != 0 && error != EAGAIN) { 2346 BO_UNLOCK(bo); 2347 return (error); 2348 } 2349 } while (error != 0); 2350 2351 /* 2352 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 2353 * have write I/O in-progress but if there is a VM object then the 2354 * VM object can also have read-I/O in-progress. 2355 */ 2356 do { 2357 bufobj_wwait(bo, 0, 0); 2358 if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) { 2359 BO_UNLOCK(bo); 2360 vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx"); 2361 BO_LOCK(bo); 2362 } 2363 } while (bo->bo_numoutput > 0); 2364 BO_UNLOCK(bo); 2365 2366 /* 2367 * Destroy the copy in the VM cache, too. 2368 */ 2369 if (bo->bo_object != NULL && 2370 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) { 2371 VM_OBJECT_WLOCK(bo->bo_object); 2372 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? 2373 OBJPR_CLEANONLY : 0); 2374 VM_OBJECT_WUNLOCK(bo->bo_object); 2375 } 2376 2377 #ifdef INVARIANTS 2378 BO_LOCK(bo); 2379 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO | 2380 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 || 2381 bo->bo_clean.bv_cnt > 0)) 2382 panic("vinvalbuf: flush failed"); 2383 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 && 2384 bo->bo_dirty.bv_cnt > 0) 2385 panic("vinvalbuf: flush dirty failed"); 2386 BO_UNLOCK(bo); 2387 #endif 2388 return (0); 2389 } 2390 2391 /* 2392 * Flush out and invalidate all buffers associated with a vnode. 2393 * Called with the underlying object locked. 2394 */ 2395 int 2396 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 2397 { 2398 2399 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 2400 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 2401 if (vp->v_object != NULL && vp->v_object->handle != vp) 2402 return (0); 2403 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); 2404 } 2405 2406 /* 2407 * Flush out buffers on the specified list. 2408 * 2409 */ 2410 static int 2411 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 2412 int slptimeo) 2413 { 2414 struct buf *bp, *nbp; 2415 int retval, error; 2416 daddr_t lblkno; 2417 b_xflags_t xflags; 2418 2419 ASSERT_BO_WLOCKED(bo); 2420 2421 retval = 0; 2422 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 2423 /* 2424 * If we are flushing both V_NORMAL and V_ALT buffers then 2425 * do not skip any buffers. If we are flushing only V_NORMAL 2426 * buffers then skip buffers marked as BX_ALTDATA. If we are 2427 * flushing only V_ALT buffers then skip buffers not marked 2428 * as BX_ALTDATA. 2429 */ 2430 if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) && 2431 (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) || 2432 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) { 2433 continue; 2434 } 2435 if (nbp != NULL) { 2436 lblkno = nbp->b_lblkno; 2437 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN); 2438 } 2439 retval = EAGAIN; 2440 error = BUF_TIMELOCK(bp, 2441 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), 2442 "flushbuf", slpflag, slptimeo); 2443 if (error) { 2444 BO_LOCK(bo); 2445 return (error != ENOLCK ? error : EAGAIN); 2446 } 2447 KASSERT(bp->b_bufobj == bo, 2448 ("bp %p wrong b_bufobj %p should be %p", 2449 bp, bp->b_bufobj, bo)); 2450 /* 2451 * XXX Since there are no node locks for NFS, I 2452 * believe there is a slight chance that a delayed 2453 * write will occur while sleeping just above, so 2454 * check for it. 2455 */ 2456 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 2457 (flags & V_SAVE)) { 2458 bremfree(bp); 2459 bp->b_flags |= B_ASYNC; 2460 bwrite(bp); 2461 BO_LOCK(bo); 2462 return (EAGAIN); /* XXX: why not loop ? */ 2463 } 2464 bremfree(bp); 2465 bp->b_flags |= (B_INVAL | B_RELBUF); 2466 bp->b_flags &= ~B_ASYNC; 2467 brelse(bp); 2468 BO_LOCK(bo); 2469 if (nbp == NULL) 2470 break; 2471 nbp = gbincore(bo, lblkno); 2472 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2473 != xflags) 2474 break; /* nbp invalid */ 2475 } 2476 return (retval); 2477 } 2478 2479 int 2480 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn) 2481 { 2482 struct buf *bp; 2483 int error; 2484 daddr_t lblkno; 2485 2486 ASSERT_BO_LOCKED(bo); 2487 2488 for (lblkno = startn;;) { 2489 again: 2490 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno); 2491 if (bp == NULL || bp->b_lblkno >= endn || 2492 bp->b_lblkno < startn) 2493 break; 2494 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 2495 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0); 2496 if (error != 0) { 2497 BO_RLOCK(bo); 2498 if (error == ENOLCK) 2499 goto again; 2500 return (error); 2501 } 2502 KASSERT(bp->b_bufobj == bo, 2503 ("bp %p wrong b_bufobj %p should be %p", 2504 bp, bp->b_bufobj, bo)); 2505 lblkno = bp->b_lblkno + 1; 2506 if ((bp->b_flags & B_MANAGED) == 0) 2507 bremfree(bp); 2508 bp->b_flags |= B_RELBUF; 2509 /* 2510 * In the VMIO case, use the B_NOREUSE flag to hint that the 2511 * pages backing each buffer in the range are unlikely to be 2512 * reused. Dirty buffers will have the hint applied once 2513 * they've been written. 2514 */ 2515 if ((bp->b_flags & B_VMIO) != 0) 2516 bp->b_flags |= B_NOREUSE; 2517 brelse(bp); 2518 BO_RLOCK(bo); 2519 } 2520 return (0); 2521 } 2522 2523 /* 2524 * Truncate a file's buffer and pages to a specified length. This 2525 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 2526 * sync activity. 2527 */ 2528 int 2529 vtruncbuf(struct vnode *vp, off_t length, int blksize) 2530 { 2531 struct buf *bp, *nbp; 2532 struct bufobj *bo; 2533 daddr_t startlbn; 2534 2535 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__, 2536 vp, blksize, (uintmax_t)length); 2537 2538 /* 2539 * Round up to the *next* lbn. 2540 */ 2541 startlbn = howmany(length, blksize); 2542 2543 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 2544 2545 bo = &vp->v_bufobj; 2546 restart_unlocked: 2547 BO_LOCK(bo); 2548 2549 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN) 2550 ; 2551 2552 if (length > 0) { 2553 /* 2554 * Write out vnode metadata, e.g. indirect blocks. 2555 */ 2556 restartsync: 2557 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2558 if (bp->b_lblkno >= 0) 2559 continue; 2560 /* 2561 * Since we hold the vnode lock this should only 2562 * fail if we're racing with the buf daemon. 2563 */ 2564 if (BUF_LOCK(bp, 2565 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2566 BO_LOCKPTR(bo)) == ENOLCK) 2567 goto restart_unlocked; 2568 2569 VNASSERT((bp->b_flags & B_DELWRI), vp, 2570 ("buf(%p) on dirty queue without DELWRI", bp)); 2571 2572 bremfree(bp); 2573 bawrite(bp); 2574 BO_LOCK(bo); 2575 goto restartsync; 2576 } 2577 } 2578 2579 bufobj_wwait(bo, 0, 0); 2580 BO_UNLOCK(bo); 2581 vnode_pager_setsize(vp, length); 2582 2583 return (0); 2584 } 2585 2586 /* 2587 * Invalidate the cached pages of a file's buffer within the range of block 2588 * numbers [startlbn, endlbn). 2589 */ 2590 void 2591 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn, 2592 int blksize) 2593 { 2594 struct bufobj *bo; 2595 off_t start, end; 2596 2597 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range"); 2598 2599 start = blksize * startlbn; 2600 end = blksize * endlbn; 2601 2602 bo = &vp->v_bufobj; 2603 BO_LOCK(bo); 2604 MPASS(blksize == bo->bo_bsize); 2605 2606 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN) 2607 ; 2608 2609 BO_UNLOCK(bo); 2610 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1)); 2611 } 2612 2613 static int 2614 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, 2615 daddr_t startlbn, daddr_t endlbn) 2616 { 2617 struct buf *bp, *nbp; 2618 bool anyfreed; 2619 2620 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked"); 2621 ASSERT_BO_LOCKED(bo); 2622 2623 do { 2624 anyfreed = false; 2625 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 2626 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) 2627 continue; 2628 if (BUF_LOCK(bp, 2629 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2630 BO_LOCKPTR(bo)) == ENOLCK) { 2631 BO_LOCK(bo); 2632 return (EAGAIN); 2633 } 2634 2635 bremfree(bp); 2636 bp->b_flags |= B_INVAL | B_RELBUF; 2637 bp->b_flags &= ~B_ASYNC; 2638 brelse(bp); 2639 anyfreed = true; 2640 2641 BO_LOCK(bo); 2642 if (nbp != NULL && 2643 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 2644 nbp->b_vp != vp || 2645 (nbp->b_flags & B_DELWRI) != 0)) 2646 return (EAGAIN); 2647 } 2648 2649 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2650 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) 2651 continue; 2652 if (BUF_LOCK(bp, 2653 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2654 BO_LOCKPTR(bo)) == ENOLCK) { 2655 BO_LOCK(bo); 2656 return (EAGAIN); 2657 } 2658 bremfree(bp); 2659 bp->b_flags |= B_INVAL | B_RELBUF; 2660 bp->b_flags &= ~B_ASYNC; 2661 brelse(bp); 2662 anyfreed = true; 2663 2664 BO_LOCK(bo); 2665 if (nbp != NULL && 2666 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 2667 (nbp->b_vp != vp) || 2668 (nbp->b_flags & B_DELWRI) == 0)) 2669 return (EAGAIN); 2670 } 2671 } while (anyfreed); 2672 return (0); 2673 } 2674 2675 static void 2676 buf_vlist_remove(struct buf *bp) 2677 { 2678 struct bufv *bv; 2679 b_xflags_t flags; 2680 2681 flags = bp->b_xflags; 2682 2683 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2684 ASSERT_BO_WLOCKED(bp->b_bufobj); 2685 KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 && 2686 (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN), 2687 ("%s: buffer %p has invalid queue state", __func__, bp)); 2688 2689 if ((flags & BX_VNDIRTY) != 0) 2690 bv = &bp->b_bufobj->bo_dirty; 2691 else 2692 bv = &bp->b_bufobj->bo_clean; 2693 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno); 2694 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 2695 bv->bv_cnt--; 2696 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 2697 } 2698 2699 /* 2700 * Add the buffer to the sorted clean or dirty block list. Return zero on 2701 * success, EEXIST if a buffer with this identity already exists, or another 2702 * error on allocation failure. 2703 */ 2704 static inline int 2705 buf_vlist_find_or_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 2706 { 2707 struct bufv *bv; 2708 struct buf *n; 2709 int error; 2710 2711 ASSERT_BO_WLOCKED(bo); 2712 KASSERT((bo->bo_flag & BO_NOBUFS) == 0, 2713 ("buf_vlist_add: bo %p does not allow bufs", bo)); 2714 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0, 2715 ("dead bo %p", bo)); 2716 KASSERT((bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) == xflags, 2717 ("buf_vlist_add: b_xflags %#x not set on bp %p", xflags, bp)); 2718 2719 if (xflags & BX_VNDIRTY) 2720 bv = &bo->bo_dirty; 2721 else 2722 bv = &bo->bo_clean; 2723 2724 error = BUF_PCTRIE_INSERT_LOOKUP_LE(&bv->bv_root, bp, &n); 2725 if (n == NULL) { 2726 KASSERT(error != EEXIST, 2727 ("buf_vlist_add: EEXIST but no existing buf found: bp %p", 2728 bp)); 2729 } else { 2730 KASSERT((uint64_t)n->b_lblkno <= (uint64_t)bp->b_lblkno, 2731 ("buf_vlist_add: out of order insert/lookup: bp %p n %p", 2732 bp, n)); 2733 KASSERT((n->b_lblkno == bp->b_lblkno) == (error == EEXIST), 2734 ("buf_vlist_add: inconsistent result for existing buf: " 2735 "error %d bp %p n %p", error, bp, n)); 2736 } 2737 if (error != 0) 2738 return (error); 2739 2740 /* Keep the list ordered. */ 2741 if (n == NULL) { 2742 KASSERT(TAILQ_EMPTY(&bv->bv_hd) || 2743 (uint64_t)bp->b_lblkno < 2744 (uint64_t)TAILQ_FIRST(&bv->bv_hd)->b_lblkno, 2745 ("buf_vlist_add: queue order: " 2746 "%p should be before first %p", 2747 bp, TAILQ_FIRST(&bv->bv_hd))); 2748 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs); 2749 } else { 2750 KASSERT(TAILQ_NEXT(n, b_bobufs) == NULL || 2751 (uint64_t)bp->b_lblkno < 2752 (uint64_t)TAILQ_NEXT(n, b_bobufs)->b_lblkno, 2753 ("buf_vlist_add: queue order: " 2754 "%p should be before next %p", 2755 bp, TAILQ_NEXT(n, b_bobufs))); 2756 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs); 2757 } 2758 2759 bv->bv_cnt++; 2760 return (0); 2761 } 2762 2763 /* 2764 * Add the buffer to the sorted clean or dirty block list. 2765 * 2766 * NOTE: xflags is passed as a constant, optimizing this inline function! 2767 */ 2768 static void 2769 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 2770 { 2771 int error; 2772 2773 KASSERT((bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) == 0, 2774 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 2775 bp->b_xflags |= xflags; 2776 error = buf_vlist_find_or_add(bp, bo, xflags); 2777 if (error) 2778 panic("buf_vlist_add: error=%d", error); 2779 } 2780 2781 /* 2782 * Look up a buffer using the buffer tries. 2783 */ 2784 struct buf * 2785 gbincore(struct bufobj *bo, daddr_t lblkno) 2786 { 2787 struct buf *bp; 2788 2789 ASSERT_BO_LOCKED(bo); 2790 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno); 2791 if (bp != NULL) 2792 return (bp); 2793 return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno)); 2794 } 2795 2796 /* 2797 * Look up a buf using the buffer tries, without the bufobj lock. This relies 2798 * on SMR for safe lookup, and bufs being in a no-free zone to provide type 2799 * stability of the result. Like other lockless lookups, the found buf may 2800 * already be invalid by the time this function returns. 2801 */ 2802 struct buf * 2803 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno) 2804 { 2805 struct buf *bp; 2806 2807 ASSERT_BO_UNLOCKED(bo); 2808 bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno); 2809 if (bp != NULL) 2810 return (bp); 2811 return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno)); 2812 } 2813 2814 /* 2815 * Associate a buffer with a vnode. 2816 */ 2817 int 2818 bgetvp(struct vnode *vp, struct buf *bp) 2819 { 2820 struct bufobj *bo; 2821 int error; 2822 2823 bo = &vp->v_bufobj; 2824 ASSERT_BO_UNLOCKED(bo); 2825 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 2826 2827 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 2828 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 2829 ("bgetvp: bp already attached! %p", bp)); 2830 2831 /* 2832 * Add the buf to the vnode's clean list unless we lost a race and find 2833 * an existing buf in either dirty or clean. 2834 */ 2835 bp->b_vp = vp; 2836 bp->b_bufobj = bo; 2837 bp->b_xflags |= BX_VNCLEAN; 2838 error = EEXIST; 2839 BO_LOCK(bo); 2840 if (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, bp->b_lblkno) == NULL) 2841 error = buf_vlist_find_or_add(bp, bo, BX_VNCLEAN); 2842 BO_UNLOCK(bo); 2843 if (__predict_true(error == 0)) { 2844 vhold(vp); 2845 return (0); 2846 } 2847 if (error != EEXIST) 2848 panic("bgetvp: buf_vlist_add error: %d", error); 2849 bp->b_vp = NULL; 2850 bp->b_bufobj = NULL; 2851 bp->b_xflags &= ~BX_VNCLEAN; 2852 return (error); 2853 } 2854 2855 /* 2856 * Disassociate a buffer from a vnode. 2857 */ 2858 void 2859 brelvp(struct buf *bp) 2860 { 2861 struct bufobj *bo; 2862 struct vnode *vp; 2863 2864 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2865 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 2866 2867 /* 2868 * Delete from old vnode list, if on one. 2869 */ 2870 vp = bp->b_vp; /* XXX */ 2871 bo = bp->b_bufobj; 2872 BO_LOCK(bo); 2873 buf_vlist_remove(bp); 2874 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2875 bo->bo_flag &= ~BO_ONWORKLST; 2876 mtx_lock(&sync_mtx); 2877 LIST_REMOVE(bo, bo_synclist); 2878 syncer_worklist_len--; 2879 mtx_unlock(&sync_mtx); 2880 } 2881 bp->b_vp = NULL; 2882 bp->b_bufobj = NULL; 2883 BO_UNLOCK(bo); 2884 vdrop(vp); 2885 } 2886 2887 /* 2888 * Add an item to the syncer work queue. 2889 */ 2890 static void 2891 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 2892 { 2893 int slot; 2894 2895 ASSERT_BO_WLOCKED(bo); 2896 2897 mtx_lock(&sync_mtx); 2898 if (bo->bo_flag & BO_ONWORKLST) 2899 LIST_REMOVE(bo, bo_synclist); 2900 else { 2901 bo->bo_flag |= BO_ONWORKLST; 2902 syncer_worklist_len++; 2903 } 2904 2905 if (delay > syncer_maxdelay - 2) 2906 delay = syncer_maxdelay - 2; 2907 slot = (syncer_delayno + delay) & syncer_mask; 2908 2909 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 2910 mtx_unlock(&sync_mtx); 2911 } 2912 2913 static int 2914 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 2915 { 2916 int error, len; 2917 2918 mtx_lock(&sync_mtx); 2919 len = syncer_worklist_len - sync_vnode_count; 2920 mtx_unlock(&sync_mtx); 2921 error = SYSCTL_OUT(req, &len, sizeof(len)); 2922 return (error); 2923 } 2924 2925 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, 2926 CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0, 2927 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 2928 2929 static struct proc *updateproc; 2930 static void sched_sync(void); 2931 static struct kproc_desc up_kp = { 2932 "syncer", 2933 sched_sync, 2934 &updateproc 2935 }; 2936 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); 2937 2938 static int 2939 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) 2940 { 2941 struct vnode *vp; 2942 struct mount *mp; 2943 2944 *bo = LIST_FIRST(slp); 2945 if (*bo == NULL) 2946 return (0); 2947 vp = bo2vnode(*bo); 2948 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) 2949 return (1); 2950 /* 2951 * We use vhold in case the vnode does not 2952 * successfully sync. vhold prevents the vnode from 2953 * going away when we unlock the sync_mtx so that 2954 * we can acquire the vnode interlock. 2955 */ 2956 vholdl(vp); 2957 mtx_unlock(&sync_mtx); 2958 VI_UNLOCK(vp); 2959 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2960 vdrop(vp); 2961 mtx_lock(&sync_mtx); 2962 return (*bo == LIST_FIRST(slp)); 2963 } 2964 MPASSERT(mp == NULL || (curthread->td_pflags & TDP_IGNSUSP) != 0 || 2965 (mp->mnt_kern_flag & MNTK_SUSPENDED) == 0, mp, 2966 ("suspended mp syncing vp %p", vp)); 2967 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2968 (void) VOP_FSYNC(vp, MNT_LAZY, td); 2969 VOP_UNLOCK(vp); 2970 vn_finished_write(mp); 2971 BO_LOCK(*bo); 2972 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { 2973 /* 2974 * Put us back on the worklist. The worklist 2975 * routine will remove us from our current 2976 * position and then add us back in at a later 2977 * position. 2978 */ 2979 vn_syncer_add_to_worklist(*bo, syncdelay); 2980 } 2981 BO_UNLOCK(*bo); 2982 vdrop(vp); 2983 mtx_lock(&sync_mtx); 2984 return (0); 2985 } 2986 2987 static int first_printf = 1; 2988 2989 /* 2990 * System filesystem synchronizer daemon. 2991 */ 2992 static void 2993 sched_sync(void) 2994 { 2995 struct synclist *next, *slp; 2996 struct bufobj *bo; 2997 long starttime; 2998 struct thread *td = curthread; 2999 int last_work_seen; 3000 int net_worklist_len; 3001 int syncer_final_iter; 3002 int error; 3003 3004 last_work_seen = 0; 3005 syncer_final_iter = 0; 3006 syncer_state = SYNCER_RUNNING; 3007 starttime = time_uptime; 3008 td->td_pflags |= TDP_NORUNNINGBUF; 3009 3010 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 3011 SHUTDOWN_PRI_LAST); 3012 3013 mtx_lock(&sync_mtx); 3014 for (;;) { 3015 if (syncer_state == SYNCER_FINAL_DELAY && 3016 syncer_final_iter == 0) { 3017 mtx_unlock(&sync_mtx); 3018 kproc_suspend_check(td->td_proc); 3019 mtx_lock(&sync_mtx); 3020 } 3021 net_worklist_len = syncer_worklist_len - sync_vnode_count; 3022 if (syncer_state != SYNCER_RUNNING && 3023 starttime != time_uptime) { 3024 if (first_printf) { 3025 printf("\nSyncing disks, vnodes remaining... "); 3026 first_printf = 0; 3027 } 3028 printf("%d ", net_worklist_len); 3029 } 3030 starttime = time_uptime; 3031 3032 /* 3033 * Push files whose dirty time has expired. Be careful 3034 * of interrupt race on slp queue. 3035 * 3036 * Skip over empty worklist slots when shutting down. 3037 */ 3038 do { 3039 slp = &syncer_workitem_pending[syncer_delayno]; 3040 syncer_delayno += 1; 3041 if (syncer_delayno == syncer_maxdelay) 3042 syncer_delayno = 0; 3043 next = &syncer_workitem_pending[syncer_delayno]; 3044 /* 3045 * If the worklist has wrapped since the 3046 * it was emptied of all but syncer vnodes, 3047 * switch to the FINAL_DELAY state and run 3048 * for one more second. 3049 */ 3050 if (syncer_state == SYNCER_SHUTTING_DOWN && 3051 net_worklist_len == 0 && 3052 last_work_seen == syncer_delayno) { 3053 syncer_state = SYNCER_FINAL_DELAY; 3054 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 3055 } 3056 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 3057 syncer_worklist_len > 0); 3058 3059 /* 3060 * Keep track of the last time there was anything 3061 * on the worklist other than syncer vnodes. 3062 * Return to the SHUTTING_DOWN state if any 3063 * new work appears. 3064 */ 3065 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 3066 last_work_seen = syncer_delayno; 3067 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 3068 syncer_state = SYNCER_SHUTTING_DOWN; 3069 while (!LIST_EMPTY(slp)) { 3070 error = sync_vnode(slp, &bo, td); 3071 if (error == 1) { 3072 LIST_REMOVE(bo, bo_synclist); 3073 LIST_INSERT_HEAD(next, bo, bo_synclist); 3074 continue; 3075 } 3076 3077 if (first_printf == 0) { 3078 /* 3079 * Drop the sync mutex, because some watchdog 3080 * drivers need to sleep while patting 3081 */ 3082 mtx_unlock(&sync_mtx); 3083 wdog_kern_pat(WD_LASTVAL); 3084 mtx_lock(&sync_mtx); 3085 } 3086 } 3087 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 3088 syncer_final_iter--; 3089 /* 3090 * The variable rushjob allows the kernel to speed up the 3091 * processing of the filesystem syncer process. A rushjob 3092 * value of N tells the filesystem syncer to process the next 3093 * N seconds worth of work on its queue ASAP. Currently rushjob 3094 * is used by the soft update code to speed up the filesystem 3095 * syncer process when the incore state is getting so far 3096 * ahead of the disk that the kernel memory pool is being 3097 * threatened with exhaustion. 3098 */ 3099 if (rushjob > 0) { 3100 rushjob -= 1; 3101 continue; 3102 } 3103 /* 3104 * Just sleep for a short period of time between 3105 * iterations when shutting down to allow some I/O 3106 * to happen. 3107 * 3108 * If it has taken us less than a second to process the 3109 * current work, then wait. Otherwise start right over 3110 * again. We can still lose time if any single round 3111 * takes more than two seconds, but it does not really 3112 * matter as we are just trying to generally pace the 3113 * filesystem activity. 3114 */ 3115 if (syncer_state != SYNCER_RUNNING || 3116 time_uptime == starttime) { 3117 thread_lock(td); 3118 sched_prio(td, PPAUSE); 3119 thread_unlock(td); 3120 } 3121 if (syncer_state != SYNCER_RUNNING) 3122 cv_timedwait(&sync_wakeup, &sync_mtx, 3123 hz / SYNCER_SHUTDOWN_SPEEDUP); 3124 else if (time_uptime == starttime) 3125 cv_timedwait(&sync_wakeup, &sync_mtx, hz); 3126 } 3127 } 3128 3129 /* 3130 * Request the syncer daemon to speed up its work. 3131 * We never push it to speed up more than half of its 3132 * normal turn time, otherwise it could take over the cpu. 3133 */ 3134 int 3135 speedup_syncer(void) 3136 { 3137 int ret = 0; 3138 3139 mtx_lock(&sync_mtx); 3140 if (rushjob < syncdelay / 2) { 3141 rushjob += 1; 3142 stat_rush_requests += 1; 3143 ret = 1; 3144 } 3145 mtx_unlock(&sync_mtx); 3146 cv_broadcast(&sync_wakeup); 3147 return (ret); 3148 } 3149 3150 /* 3151 * Tell the syncer to speed up its work and run though its work 3152 * list several times, then tell it to shut down. 3153 */ 3154 static void 3155 syncer_shutdown(void *arg, int howto) 3156 { 3157 3158 if (howto & RB_NOSYNC) 3159 return; 3160 mtx_lock(&sync_mtx); 3161 syncer_state = SYNCER_SHUTTING_DOWN; 3162 rushjob = 0; 3163 mtx_unlock(&sync_mtx); 3164 cv_broadcast(&sync_wakeup); 3165 kproc_shutdown(arg, howto); 3166 } 3167 3168 void 3169 syncer_suspend(void) 3170 { 3171 3172 syncer_shutdown(updateproc, 0); 3173 } 3174 3175 void 3176 syncer_resume(void) 3177 { 3178 3179 mtx_lock(&sync_mtx); 3180 first_printf = 1; 3181 syncer_state = SYNCER_RUNNING; 3182 mtx_unlock(&sync_mtx); 3183 cv_broadcast(&sync_wakeup); 3184 kproc_resume(updateproc); 3185 } 3186 3187 /* 3188 * Move the buffer between the clean and dirty lists of its vnode. 3189 */ 3190 void 3191 reassignbuf(struct buf *bp) 3192 { 3193 struct vnode *vp; 3194 struct bufobj *bo; 3195 int delay; 3196 #ifdef INVARIANTS 3197 struct bufv *bv; 3198 #endif 3199 3200 vp = bp->b_vp; 3201 bo = bp->b_bufobj; 3202 3203 KASSERT((bp->b_flags & B_PAGING) == 0, 3204 ("%s: cannot reassign paging buffer %p", __func__, bp)); 3205 3206 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 3207 bp, bp->b_vp, bp->b_flags); 3208 3209 BO_LOCK(bo); 3210 buf_vlist_remove(bp); 3211 3212 /* 3213 * If dirty, put on list of dirty buffers; otherwise insert onto list 3214 * of clean buffers. 3215 */ 3216 if (bp->b_flags & B_DELWRI) { 3217 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 3218 switch (vp->v_type) { 3219 case VDIR: 3220 delay = dirdelay; 3221 break; 3222 case VCHR: 3223 delay = metadelay; 3224 break; 3225 default: 3226 delay = filedelay; 3227 } 3228 vn_syncer_add_to_worklist(bo, delay); 3229 } 3230 buf_vlist_add(bp, bo, BX_VNDIRTY); 3231 } else { 3232 buf_vlist_add(bp, bo, BX_VNCLEAN); 3233 3234 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 3235 mtx_lock(&sync_mtx); 3236 LIST_REMOVE(bo, bo_synclist); 3237 syncer_worklist_len--; 3238 mtx_unlock(&sync_mtx); 3239 bo->bo_flag &= ~BO_ONWORKLST; 3240 } 3241 } 3242 #ifdef INVARIANTS 3243 bv = &bo->bo_clean; 3244 bp = TAILQ_FIRST(&bv->bv_hd); 3245 KASSERT(bp == NULL || bp->b_bufobj == bo, 3246 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3247 bp = TAILQ_LAST(&bv->bv_hd, buflists); 3248 KASSERT(bp == NULL || bp->b_bufobj == bo, 3249 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3250 bv = &bo->bo_dirty; 3251 bp = TAILQ_FIRST(&bv->bv_hd); 3252 KASSERT(bp == NULL || bp->b_bufobj == bo, 3253 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3254 bp = TAILQ_LAST(&bv->bv_hd, buflists); 3255 KASSERT(bp == NULL || bp->b_bufobj == bo, 3256 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3257 #endif 3258 BO_UNLOCK(bo); 3259 } 3260 3261 static void 3262 v_init_counters(struct vnode *vp) 3263 { 3264 3265 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0, 3266 vp, ("%s called for an initialized vnode", __FUNCTION__)); 3267 ASSERT_VI_UNLOCKED(vp, __FUNCTION__); 3268 3269 refcount_init(&vp->v_holdcnt, 1); 3270 refcount_init(&vp->v_usecount, 1); 3271 } 3272 3273 /* 3274 * Get a usecount on a vnode. 3275 * 3276 * vget and vget_finish may fail to lock the vnode if they lose a race against 3277 * it being doomed. LK_RETRY can be passed in flags to lock it anyway. 3278 * 3279 * Consumers which don't guarantee liveness of the vnode can use SMR to 3280 * try to get a reference. Note this operation can fail since the vnode 3281 * may be awaiting getting freed by the time they get to it. 3282 */ 3283 enum vgetstate 3284 vget_prep_smr(struct vnode *vp) 3285 { 3286 enum vgetstate vs; 3287 3288 VFS_SMR_ASSERT_ENTERED(); 3289 3290 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 3291 vs = VGET_USECOUNT; 3292 } else { 3293 if (vhold_smr(vp)) 3294 vs = VGET_HOLDCNT; 3295 else 3296 vs = VGET_NONE; 3297 } 3298 return (vs); 3299 } 3300 3301 enum vgetstate 3302 vget_prep(struct vnode *vp) 3303 { 3304 enum vgetstate vs; 3305 3306 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 3307 vs = VGET_USECOUNT; 3308 } else { 3309 vhold(vp); 3310 vs = VGET_HOLDCNT; 3311 } 3312 return (vs); 3313 } 3314 3315 void 3316 vget_abort(struct vnode *vp, enum vgetstate vs) 3317 { 3318 3319 switch (vs) { 3320 case VGET_USECOUNT: 3321 vrele(vp); 3322 break; 3323 case VGET_HOLDCNT: 3324 vdrop(vp); 3325 break; 3326 default: 3327 __assert_unreachable(); 3328 } 3329 } 3330 3331 int 3332 vget(struct vnode *vp, int flags) 3333 { 3334 enum vgetstate vs; 3335 3336 vs = vget_prep(vp); 3337 return (vget_finish(vp, flags, vs)); 3338 } 3339 3340 int 3341 vget_finish(struct vnode *vp, int flags, enum vgetstate vs) 3342 { 3343 int error; 3344 3345 if ((flags & LK_INTERLOCK) != 0) 3346 ASSERT_VI_LOCKED(vp, __func__); 3347 else 3348 ASSERT_VI_UNLOCKED(vp, __func__); 3349 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); 3350 VNPASS(vp->v_holdcnt > 0, vp); 3351 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); 3352 3353 error = vn_lock(vp, flags); 3354 if (__predict_false(error != 0)) { 3355 vget_abort(vp, vs); 3356 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, 3357 vp); 3358 return (error); 3359 } 3360 3361 vget_finish_ref(vp, vs); 3362 return (0); 3363 } 3364 3365 void 3366 vget_finish_ref(struct vnode *vp, enum vgetstate vs) 3367 { 3368 int old; 3369 3370 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); 3371 VNPASS(vp->v_holdcnt > 0, vp); 3372 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); 3373 3374 if (vs == VGET_USECOUNT) 3375 return; 3376 3377 /* 3378 * We hold the vnode. If the usecount is 0 it will be utilized to keep 3379 * the vnode around. Otherwise someone else lended their hold count and 3380 * we have to drop ours. 3381 */ 3382 old = atomic_fetchadd_int(&vp->v_usecount, 1); 3383 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old)); 3384 if (old != 0) { 3385 #ifdef INVARIANTS 3386 old = atomic_fetchadd_int(&vp->v_holdcnt, -1); 3387 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); 3388 #else 3389 refcount_release(&vp->v_holdcnt); 3390 #endif 3391 } 3392 } 3393 3394 void 3395 vref(struct vnode *vp) 3396 { 3397 enum vgetstate vs; 3398 3399 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3400 vs = vget_prep(vp); 3401 vget_finish_ref(vp, vs); 3402 } 3403 3404 void 3405 vrefact(struct vnode *vp) 3406 { 3407 int old __diagused; 3408 3409 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3410 old = refcount_acquire(&vp->v_usecount); 3411 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); 3412 } 3413 3414 void 3415 vlazy(struct vnode *vp) 3416 { 3417 struct mount *mp; 3418 3419 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__)); 3420 3421 if ((vp->v_mflag & VMP_LAZYLIST) != 0) 3422 return; 3423 /* 3424 * We may get here for inactive routines after the vnode got doomed. 3425 */ 3426 if (VN_IS_DOOMED(vp)) 3427 return; 3428 mp = vp->v_mount; 3429 mtx_lock(&mp->mnt_listmtx); 3430 if ((vp->v_mflag & VMP_LAZYLIST) == 0) { 3431 vp->v_mflag |= VMP_LAZYLIST; 3432 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3433 mp->mnt_lazyvnodelistsize++; 3434 } 3435 mtx_unlock(&mp->mnt_listmtx); 3436 } 3437 3438 static void 3439 vunlazy(struct vnode *vp) 3440 { 3441 struct mount *mp; 3442 3443 ASSERT_VI_LOCKED(vp, __func__); 3444 VNPASS(!VN_IS_DOOMED(vp), vp); 3445 3446 mp = vp->v_mount; 3447 mtx_lock(&mp->mnt_listmtx); 3448 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 3449 /* 3450 * Don't remove the vnode from the lazy list if another thread 3451 * has increased the hold count. It may have re-enqueued the 3452 * vnode to the lazy list and is now responsible for its 3453 * removal. 3454 */ 3455 if (vp->v_holdcnt == 0) { 3456 vp->v_mflag &= ~VMP_LAZYLIST; 3457 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3458 mp->mnt_lazyvnodelistsize--; 3459 } 3460 mtx_unlock(&mp->mnt_listmtx); 3461 } 3462 3463 /* 3464 * This routine is only meant to be called from vgonel prior to dooming 3465 * the vnode. 3466 */ 3467 static void 3468 vunlazy_gone(struct vnode *vp) 3469 { 3470 struct mount *mp; 3471 3472 ASSERT_VOP_ELOCKED(vp, __func__); 3473 ASSERT_VI_LOCKED(vp, __func__); 3474 VNPASS(!VN_IS_DOOMED(vp), vp); 3475 3476 if (vp->v_mflag & VMP_LAZYLIST) { 3477 mp = vp->v_mount; 3478 mtx_lock(&mp->mnt_listmtx); 3479 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 3480 vp->v_mflag &= ~VMP_LAZYLIST; 3481 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3482 mp->mnt_lazyvnodelistsize--; 3483 mtx_unlock(&mp->mnt_listmtx); 3484 } 3485 } 3486 3487 static void 3488 vdefer_inactive(struct vnode *vp) 3489 { 3490 3491 ASSERT_VI_LOCKED(vp, __func__); 3492 VNPASS(vp->v_holdcnt > 0, vp); 3493 if (VN_IS_DOOMED(vp)) { 3494 vdropl(vp); 3495 return; 3496 } 3497 if (vp->v_iflag & VI_DEFINACT) { 3498 VNPASS(vp->v_holdcnt > 1, vp); 3499 vdropl(vp); 3500 return; 3501 } 3502 if (vp->v_usecount > 0) { 3503 vp->v_iflag &= ~VI_OWEINACT; 3504 vdropl(vp); 3505 return; 3506 } 3507 vlazy(vp); 3508 vp->v_iflag |= VI_DEFINACT; 3509 VI_UNLOCK(vp); 3510 atomic_add_long(&deferred_inact, 1); 3511 } 3512 3513 static void 3514 vdefer_inactive_unlocked(struct vnode *vp) 3515 { 3516 3517 VI_LOCK(vp); 3518 if ((vp->v_iflag & VI_OWEINACT) == 0) { 3519 vdropl(vp); 3520 return; 3521 } 3522 vdefer_inactive(vp); 3523 } 3524 3525 enum vput_op { VRELE, VPUT, VUNREF }; 3526 3527 /* 3528 * Handle ->v_usecount transitioning to 0. 3529 * 3530 * By releasing the last usecount we take ownership of the hold count which 3531 * provides liveness of the vnode, meaning we have to vdrop. 3532 * 3533 * For all vnodes we may need to perform inactive processing. It requires an 3534 * exclusive lock on the vnode, while it is legal to call here with only a 3535 * shared lock (or no locks). If locking the vnode in an expected manner fails, 3536 * inactive processing gets deferred to the syncer. 3537 * 3538 * XXX Some filesystems pass in an exclusively locked vnode and strongly depend 3539 * on the lock being held all the way until VOP_INACTIVE. This in particular 3540 * happens with UFS which adds half-constructed vnodes to the hash, where they 3541 * can be found by other code. 3542 */ 3543 static void 3544 vput_final(struct vnode *vp, enum vput_op func) 3545 { 3546 int error; 3547 bool want_unlock; 3548 3549 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3550 VNPASS(vp->v_holdcnt > 0, vp); 3551 3552 VI_LOCK(vp); 3553 3554 /* 3555 * By the time we got here someone else might have transitioned 3556 * the count back to > 0. 3557 */ 3558 if (vp->v_usecount > 0) 3559 goto out; 3560 3561 /* 3562 * If the vnode is doomed vgone already performed inactive processing 3563 * (if needed). 3564 */ 3565 if (VN_IS_DOOMED(vp)) 3566 goto out; 3567 3568 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0)) 3569 goto out; 3570 3571 if (vp->v_iflag & VI_DOINGINACT) 3572 goto out; 3573 3574 /* 3575 * Locking operations here will drop the interlock and possibly the 3576 * vnode lock, opening a window where the vnode can get doomed all the 3577 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to 3578 * perform inactive. 3579 */ 3580 vp->v_iflag |= VI_OWEINACT; 3581 want_unlock = false; 3582 error = 0; 3583 switch (func) { 3584 case VRELE: 3585 switch (VOP_ISLOCKED(vp)) { 3586 case LK_EXCLUSIVE: 3587 break; 3588 case LK_EXCLOTHER: 3589 case 0: 3590 want_unlock = true; 3591 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 3592 VI_LOCK(vp); 3593 break; 3594 default: 3595 /* 3596 * The lock has at least one sharer, but we have no way 3597 * to conclude whether this is us. Play it safe and 3598 * defer processing. 3599 */ 3600 error = EAGAIN; 3601 break; 3602 } 3603 break; 3604 case VPUT: 3605 want_unlock = true; 3606 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3607 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | 3608 LK_NOWAIT); 3609 VI_LOCK(vp); 3610 } 3611 break; 3612 case VUNREF: 3613 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3614 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); 3615 VI_LOCK(vp); 3616 } 3617 break; 3618 } 3619 if (error == 0) { 3620 if (func == VUNREF) { 3621 VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp, 3622 ("recursive vunref")); 3623 vp->v_vflag |= VV_UNREF; 3624 } 3625 for (;;) { 3626 error = vinactive(vp); 3627 if (want_unlock) 3628 VOP_UNLOCK(vp); 3629 if (error != ERELOOKUP || !want_unlock) 3630 break; 3631 VOP_LOCK(vp, LK_EXCLUSIVE); 3632 } 3633 if (func == VUNREF) 3634 vp->v_vflag &= ~VV_UNREF; 3635 vdropl(vp); 3636 } else { 3637 vdefer_inactive(vp); 3638 } 3639 return; 3640 out: 3641 if (func == VPUT) 3642 VOP_UNLOCK(vp); 3643 vdropl(vp); 3644 } 3645 3646 /* 3647 * Decrement ->v_usecount for a vnode. 3648 * 3649 * Releasing the last use count requires additional processing, see vput_final 3650 * above for details. 3651 * 3652 * Comment above each variant denotes lock state on entry and exit. 3653 */ 3654 3655 /* 3656 * in: any 3657 * out: same as passed in 3658 */ 3659 void 3660 vrele(struct vnode *vp) 3661 { 3662 3663 ASSERT_VI_UNLOCKED(vp, __func__); 3664 if (!refcount_release(&vp->v_usecount)) 3665 return; 3666 vput_final(vp, VRELE); 3667 } 3668 3669 /* 3670 * in: locked 3671 * out: unlocked 3672 */ 3673 void 3674 vput(struct vnode *vp) 3675 { 3676 3677 ASSERT_VOP_LOCKED(vp, __func__); 3678 ASSERT_VI_UNLOCKED(vp, __func__); 3679 if (!refcount_release(&vp->v_usecount)) { 3680 VOP_UNLOCK(vp); 3681 return; 3682 } 3683 vput_final(vp, VPUT); 3684 } 3685 3686 /* 3687 * in: locked 3688 * out: locked 3689 */ 3690 void 3691 vunref(struct vnode *vp) 3692 { 3693 3694 ASSERT_VOP_LOCKED(vp, __func__); 3695 ASSERT_VI_UNLOCKED(vp, __func__); 3696 if (!refcount_release(&vp->v_usecount)) 3697 return; 3698 vput_final(vp, VUNREF); 3699 } 3700 3701 void 3702 vhold(struct vnode *vp) 3703 { 3704 int old; 3705 3706 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3707 old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3708 VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3709 ("%s: wrong hold count %d", __func__, old)); 3710 if (old == 0) 3711 vfs_freevnodes_dec(); 3712 } 3713 3714 void 3715 vholdnz(struct vnode *vp) 3716 { 3717 3718 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3719 #ifdef INVARIANTS 3720 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3721 VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3722 ("%s: wrong hold count %d", __func__, old)); 3723 #else 3724 atomic_add_int(&vp->v_holdcnt, 1); 3725 #endif 3726 } 3727 3728 /* 3729 * Grab a hold count unless the vnode is freed. 3730 * 3731 * Only use this routine if vfs smr is the only protection you have against 3732 * freeing the vnode. 3733 * 3734 * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag 3735 * is not set. After the flag is set the vnode becomes immutable to anyone but 3736 * the thread which managed to set the flag. 3737 * 3738 * It may be tempting to replace the loop with: 3739 * count = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3740 * if (count & VHOLD_NO_SMR) { 3741 * backpedal and error out; 3742 * } 3743 * 3744 * However, while this is more performant, it hinders debugging by eliminating 3745 * the previously mentioned invariant. 3746 */ 3747 bool 3748 vhold_smr(struct vnode *vp) 3749 { 3750 int count; 3751 3752 VFS_SMR_ASSERT_ENTERED(); 3753 3754 count = atomic_load_int(&vp->v_holdcnt); 3755 for (;;) { 3756 if (count & VHOLD_NO_SMR) { 3757 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, 3758 ("non-zero hold count with flags %d\n", count)); 3759 return (false); 3760 } 3761 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); 3762 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) { 3763 if (count == 0) 3764 vfs_freevnodes_dec(); 3765 return (true); 3766 } 3767 } 3768 } 3769 3770 /* 3771 * Hold a free vnode for recycling. 3772 * 3773 * Note: vnode_init references this comment. 3774 * 3775 * Attempts to recycle only need the global vnode list lock and have no use for 3776 * SMR. 3777 * 3778 * However, vnodes get inserted into the global list before they get fully 3779 * initialized and stay there until UMA decides to free the memory. This in 3780 * particular means the target can be found before it becomes usable and after 3781 * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to 3782 * VHOLD_NO_SMR. 3783 * 3784 * Note: the vnode may gain more references after we transition the count 0->1. 3785 */ 3786 static bool 3787 vhold_recycle_free(struct vnode *vp) 3788 { 3789 int count; 3790 3791 mtx_assert(&vnode_list_mtx, MA_OWNED); 3792 3793 count = atomic_load_int(&vp->v_holdcnt); 3794 for (;;) { 3795 if (count & VHOLD_NO_SMR) { 3796 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, 3797 ("non-zero hold count with flags %d\n", count)); 3798 return (false); 3799 } 3800 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); 3801 if (count > 0) { 3802 return (false); 3803 } 3804 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) { 3805 vfs_freevnodes_dec(); 3806 return (true); 3807 } 3808 } 3809 } 3810 3811 static void __noinline 3812 vdbatch_process(struct vdbatch *vd) 3813 { 3814 struct vnode *vp; 3815 int i; 3816 3817 mtx_assert(&vd->lock, MA_OWNED); 3818 MPASS(curthread->td_pinned > 0); 3819 MPASS(vd->index == VDBATCH_SIZE); 3820 3821 /* 3822 * Attempt to requeue the passed batch, but give up easily. 3823 * 3824 * Despite batching the mechanism is prone to transient *significant* 3825 * lock contention, where vnode_list_mtx becomes the primary bottleneck 3826 * if multiple CPUs get here (one real-world example is highly parallel 3827 * do-nothing make , which will stat *tons* of vnodes). Since it is 3828 * quasi-LRU (read: not that great even if fully honoured) just dodge 3829 * the problem. Parties which don't like it are welcome to implement 3830 * something better. 3831 */ 3832 critical_enter(); 3833 if (mtx_trylock(&vnode_list_mtx)) { 3834 for (i = 0; i < VDBATCH_SIZE; i++) { 3835 vp = vd->tab[i]; 3836 vd->tab[i] = NULL; 3837 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 3838 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist); 3839 MPASS(vp->v_dbatchcpu != NOCPU); 3840 vp->v_dbatchcpu = NOCPU; 3841 } 3842 mtx_unlock(&vnode_list_mtx); 3843 } else { 3844 counter_u64_add(vnode_skipped_requeues, 1); 3845 3846 for (i = 0; i < VDBATCH_SIZE; i++) { 3847 vp = vd->tab[i]; 3848 vd->tab[i] = NULL; 3849 MPASS(vp->v_dbatchcpu != NOCPU); 3850 vp->v_dbatchcpu = NOCPU; 3851 } 3852 } 3853 vd->index = 0; 3854 critical_exit(); 3855 } 3856 3857 static void 3858 vdbatch_enqueue(struct vnode *vp) 3859 { 3860 struct vdbatch *vd; 3861 3862 ASSERT_VI_LOCKED(vp, __func__); 3863 VNPASS(!VN_IS_DOOMED(vp), vp); 3864 3865 if (vp->v_dbatchcpu != NOCPU) { 3866 VI_UNLOCK(vp); 3867 return; 3868 } 3869 3870 sched_pin(); 3871 vd = DPCPU_PTR(vd); 3872 mtx_lock(&vd->lock); 3873 MPASS(vd->index < VDBATCH_SIZE); 3874 MPASS(vd->tab[vd->index] == NULL); 3875 /* 3876 * A hack: we depend on being pinned so that we know what to put in 3877 * ->v_dbatchcpu. 3878 */ 3879 vp->v_dbatchcpu = curcpu; 3880 vd->tab[vd->index] = vp; 3881 vd->index++; 3882 VI_UNLOCK(vp); 3883 if (vd->index == VDBATCH_SIZE) 3884 vdbatch_process(vd); 3885 mtx_unlock(&vd->lock); 3886 sched_unpin(); 3887 } 3888 3889 /* 3890 * This routine must only be called for vnodes which are about to be 3891 * deallocated. Supporting dequeue for arbitrary vndoes would require 3892 * validating that the locked batch matches. 3893 */ 3894 static void 3895 vdbatch_dequeue(struct vnode *vp) 3896 { 3897 struct vdbatch *vd; 3898 int i; 3899 short cpu; 3900 3901 VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp); 3902 3903 cpu = vp->v_dbatchcpu; 3904 if (cpu == NOCPU) 3905 return; 3906 3907 vd = DPCPU_ID_PTR(cpu, vd); 3908 mtx_lock(&vd->lock); 3909 for (i = 0; i < vd->index; i++) { 3910 if (vd->tab[i] != vp) 3911 continue; 3912 vp->v_dbatchcpu = NOCPU; 3913 vd->index--; 3914 vd->tab[i] = vd->tab[vd->index]; 3915 vd->tab[vd->index] = NULL; 3916 break; 3917 } 3918 mtx_unlock(&vd->lock); 3919 /* 3920 * Either we dequeued the vnode above or the target CPU beat us to it. 3921 */ 3922 MPASS(vp->v_dbatchcpu == NOCPU); 3923 } 3924 3925 /* 3926 * Drop the hold count of the vnode. 3927 * 3928 * It will only get freed if this is the last hold *and* it has been vgone'd. 3929 * 3930 * Because the vnode vm object keeps a hold reference on the vnode if 3931 * there is at least one resident non-cached page, the vnode cannot 3932 * leave the active list without the page cleanup done. 3933 */ 3934 static void __noinline 3935 vdropl_final(struct vnode *vp) 3936 { 3937 3938 ASSERT_VI_LOCKED(vp, __func__); 3939 VNPASS(VN_IS_DOOMED(vp), vp); 3940 /* 3941 * Set the VHOLD_NO_SMR flag. 3942 * 3943 * We may be racing against vhold_smr. If they win we can just pretend 3944 * we never got this far, they will vdrop later. 3945 */ 3946 if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) { 3947 vfs_freevnodes_inc(); 3948 VI_UNLOCK(vp); 3949 /* 3950 * We lost the aforementioned race. Any subsequent access is 3951 * invalid as they might have managed to vdropl on their own. 3952 */ 3953 return; 3954 } 3955 /* 3956 * Don't bump freevnodes as this one is going away. 3957 */ 3958 freevnode(vp); 3959 } 3960 3961 void 3962 vdrop(struct vnode *vp) 3963 { 3964 3965 ASSERT_VI_UNLOCKED(vp, __func__); 3966 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3967 if (refcount_release_if_not_last(&vp->v_holdcnt)) 3968 return; 3969 VI_LOCK(vp); 3970 vdropl(vp); 3971 } 3972 3973 static void __always_inline 3974 vdropl_impl(struct vnode *vp, bool enqueue) 3975 { 3976 3977 ASSERT_VI_LOCKED(vp, __func__); 3978 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3979 if (!refcount_release(&vp->v_holdcnt)) { 3980 VI_UNLOCK(vp); 3981 return; 3982 } 3983 VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp); 3984 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp); 3985 if (VN_IS_DOOMED(vp)) { 3986 vdropl_final(vp); 3987 return; 3988 } 3989 3990 vfs_freevnodes_inc(); 3991 if (vp->v_mflag & VMP_LAZYLIST) { 3992 vunlazy(vp); 3993 } 3994 3995 if (!enqueue) { 3996 VI_UNLOCK(vp); 3997 return; 3998 } 3999 4000 /* 4001 * Also unlocks the interlock. We can't assert on it as we 4002 * released our hold and by now the vnode might have been 4003 * freed. 4004 */ 4005 vdbatch_enqueue(vp); 4006 } 4007 4008 void 4009 vdropl(struct vnode *vp) 4010 { 4011 4012 vdropl_impl(vp, true); 4013 } 4014 4015 /* 4016 * vdrop a vnode when recycling 4017 * 4018 * This is a special case routine only to be used when recycling, differs from 4019 * regular vdrop by not requeieing the vnode on LRU. 4020 * 4021 * Consider a case where vtryrecycle continuously fails with all vnodes (due to 4022 * e.g., frozen writes on the filesystem), filling the batch and causing it to 4023 * be requeued. Then vnlru will end up revisiting the same vnodes. This is a 4024 * loop which can last for as long as writes are frozen. 4025 */ 4026 static void 4027 vdropl_recycle(struct vnode *vp) 4028 { 4029 4030 vdropl_impl(vp, false); 4031 } 4032 4033 static void 4034 vdrop_recycle(struct vnode *vp) 4035 { 4036 4037 VI_LOCK(vp); 4038 vdropl_recycle(vp); 4039 } 4040 4041 /* 4042 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 4043 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 4044 */ 4045 static int 4046 vinactivef(struct vnode *vp) 4047 { 4048 int error; 4049 4050 ASSERT_VOP_ELOCKED(vp, "vinactive"); 4051 ASSERT_VI_LOCKED(vp, "vinactive"); 4052 VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp); 4053 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4054 vp->v_iflag |= VI_DOINGINACT; 4055 vp->v_iflag &= ~VI_OWEINACT; 4056 VI_UNLOCK(vp); 4057 4058 /* 4059 * Before moving off the active list, we must be sure that any 4060 * modified pages are converted into the vnode's dirty 4061 * buffers, since these will no longer be checked once the 4062 * vnode is on the inactive list. 4063 * 4064 * The write-out of the dirty pages is asynchronous. At the 4065 * point that VOP_INACTIVE() is called, there could still be 4066 * pending I/O and dirty pages in the object. 4067 */ 4068 if ((vp->v_vflag & VV_NOSYNC) == 0) 4069 vnode_pager_clean_async(vp); 4070 4071 error = VOP_INACTIVE(vp); 4072 VI_LOCK(vp); 4073 VNPASS(vp->v_iflag & VI_DOINGINACT, vp); 4074 vp->v_iflag &= ~VI_DOINGINACT; 4075 return (error); 4076 } 4077 4078 int 4079 vinactive(struct vnode *vp) 4080 { 4081 4082 ASSERT_VOP_ELOCKED(vp, "vinactive"); 4083 ASSERT_VI_LOCKED(vp, "vinactive"); 4084 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4085 4086 if ((vp->v_iflag & VI_OWEINACT) == 0) 4087 return (0); 4088 if (vp->v_iflag & VI_DOINGINACT) 4089 return (0); 4090 if (vp->v_usecount > 0) { 4091 vp->v_iflag &= ~VI_OWEINACT; 4092 return (0); 4093 } 4094 return (vinactivef(vp)); 4095 } 4096 4097 /* 4098 * Remove any vnodes in the vnode table belonging to mount point mp. 4099 * 4100 * If FORCECLOSE is not specified, there should not be any active ones, 4101 * return error if any are found (nb: this is a user error, not a 4102 * system error). If FORCECLOSE is specified, detach any active vnodes 4103 * that are found. 4104 * 4105 * If WRITECLOSE is set, only flush out regular file vnodes open for 4106 * writing. 4107 * 4108 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 4109 * 4110 * `rootrefs' specifies the base reference count for the root vnode 4111 * of this filesystem. The root vnode is considered busy if its 4112 * v_usecount exceeds this value. On a successful return, vflush(, td) 4113 * will call vrele() on the root vnode exactly rootrefs times. 4114 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 4115 * be zero. 4116 */ 4117 #ifdef DIAGNOSTIC 4118 static int busyprt = 0; /* print out busy vnodes */ 4119 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 4120 #endif 4121 4122 int 4123 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 4124 { 4125 struct vnode *vp, *mvp, *rootvp = NULL; 4126 struct vattr vattr; 4127 int busy = 0, error; 4128 4129 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 4130 rootrefs, flags); 4131 if (rootrefs > 0) { 4132 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 4133 ("vflush: bad args")); 4134 /* 4135 * Get the filesystem root vnode. We can vput() it 4136 * immediately, since with rootrefs > 0, it won't go away. 4137 */ 4138 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 4139 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 4140 __func__, error); 4141 return (error); 4142 } 4143 vput(rootvp); 4144 } 4145 loop: 4146 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 4147 vholdl(vp); 4148 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 4149 if (error) { 4150 vdrop(vp); 4151 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 4152 goto loop; 4153 } 4154 /* 4155 * Skip over a vnodes marked VV_SYSTEM. 4156 */ 4157 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 4158 VOP_UNLOCK(vp); 4159 vdrop(vp); 4160 continue; 4161 } 4162 /* 4163 * If WRITECLOSE is set, flush out unlinked but still open 4164 * files (even if open only for reading) and regular file 4165 * vnodes open for writing. 4166 */ 4167 if (flags & WRITECLOSE) { 4168 vnode_pager_clean_async(vp); 4169 do { 4170 error = VOP_FSYNC(vp, MNT_WAIT, td); 4171 } while (error == ERELOOKUP); 4172 if (error != 0) { 4173 VOP_UNLOCK(vp); 4174 vdrop(vp); 4175 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 4176 return (error); 4177 } 4178 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 4179 VI_LOCK(vp); 4180 4181 if ((vp->v_type == VNON || 4182 (error == 0 && vattr.va_nlink > 0)) && 4183 (vp->v_writecount <= 0 || vp->v_type != VREG)) { 4184 VOP_UNLOCK(vp); 4185 vdropl(vp); 4186 continue; 4187 } 4188 } else 4189 VI_LOCK(vp); 4190 /* 4191 * With v_usecount == 0, all we need to do is clear out the 4192 * vnode data structures and we are done. 4193 * 4194 * If FORCECLOSE is set, forcibly close the vnode. 4195 */ 4196 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 4197 vgonel(vp); 4198 } else { 4199 busy++; 4200 #ifdef DIAGNOSTIC 4201 if (busyprt) 4202 vn_printf(vp, "vflush: busy vnode "); 4203 #endif 4204 } 4205 VOP_UNLOCK(vp); 4206 vdropl(vp); 4207 } 4208 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 4209 /* 4210 * If just the root vnode is busy, and if its refcount 4211 * is equal to `rootrefs', then go ahead and kill it. 4212 */ 4213 VI_LOCK(rootvp); 4214 KASSERT(busy > 0, ("vflush: not busy")); 4215 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 4216 ("vflush: usecount %d < rootrefs %d", 4217 rootvp->v_usecount, rootrefs)); 4218 if (busy == 1 && rootvp->v_usecount == rootrefs) { 4219 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 4220 vgone(rootvp); 4221 VOP_UNLOCK(rootvp); 4222 busy = 0; 4223 } else 4224 VI_UNLOCK(rootvp); 4225 } 4226 if (busy) { 4227 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 4228 busy); 4229 return (EBUSY); 4230 } 4231 for (; rootrefs > 0; rootrefs--) 4232 vrele(rootvp); 4233 return (0); 4234 } 4235 4236 /* 4237 * Recycle an unused vnode. 4238 */ 4239 int 4240 vrecycle(struct vnode *vp) 4241 { 4242 int recycled; 4243 4244 VI_LOCK(vp); 4245 recycled = vrecyclel(vp); 4246 VI_UNLOCK(vp); 4247 return (recycled); 4248 } 4249 4250 /* 4251 * vrecycle, with the vp interlock held. 4252 */ 4253 int 4254 vrecyclel(struct vnode *vp) 4255 { 4256 int recycled; 4257 4258 ASSERT_VOP_ELOCKED(vp, __func__); 4259 ASSERT_VI_LOCKED(vp, __func__); 4260 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4261 recycled = 0; 4262 if (vp->v_usecount == 0) { 4263 recycled = 1; 4264 vgonel(vp); 4265 } 4266 return (recycled); 4267 } 4268 4269 /* 4270 * Eliminate all activity associated with a vnode 4271 * in preparation for reuse. 4272 */ 4273 void 4274 vgone(struct vnode *vp) 4275 { 4276 VI_LOCK(vp); 4277 vgonel(vp); 4278 VI_UNLOCK(vp); 4279 } 4280 4281 /* 4282 * Notify upper mounts about reclaimed or unlinked vnode. 4283 */ 4284 void 4285 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event) 4286 { 4287 struct mount *mp; 4288 struct mount_upper_node *ump; 4289 4290 mp = atomic_load_ptr(&vp->v_mount); 4291 if (mp == NULL) 4292 return; 4293 if (TAILQ_EMPTY(&mp->mnt_notify)) 4294 return; 4295 4296 MNT_ILOCK(mp); 4297 mp->mnt_upper_pending++; 4298 KASSERT(mp->mnt_upper_pending > 0, 4299 ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending)); 4300 TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) { 4301 MNT_IUNLOCK(mp); 4302 switch (event) { 4303 case VFS_NOTIFY_UPPER_RECLAIM: 4304 VFS_RECLAIM_LOWERVP(ump->mp, vp); 4305 break; 4306 case VFS_NOTIFY_UPPER_UNLINK: 4307 VFS_UNLINK_LOWERVP(ump->mp, vp); 4308 break; 4309 } 4310 MNT_ILOCK(mp); 4311 } 4312 mp->mnt_upper_pending--; 4313 if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 && 4314 mp->mnt_upper_pending == 0) { 4315 mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER; 4316 wakeup(&mp->mnt_uppers); 4317 } 4318 MNT_IUNLOCK(mp); 4319 } 4320 4321 /* 4322 * vgone, with the vp interlock held. 4323 */ 4324 static void 4325 vgonel(struct vnode *vp) 4326 { 4327 struct thread *td; 4328 struct mount *mp; 4329 vm_object_t object; 4330 bool active, doinginact, oweinact; 4331 4332 ASSERT_VOP_ELOCKED(vp, "vgonel"); 4333 ASSERT_VI_LOCKED(vp, "vgonel"); 4334 VNASSERT(vp->v_holdcnt, vp, 4335 ("vgonel: vp %p has no reference.", vp)); 4336 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4337 td = curthread; 4338 4339 /* 4340 * Don't vgonel if we're already doomed. 4341 */ 4342 if (VN_IS_DOOMED(vp)) { 4343 VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \ 4344 vn_get_state(vp) == VSTATE_DEAD, vp); 4345 return; 4346 } 4347 /* 4348 * Paired with freevnode. 4349 */ 4350 vn_seqc_write_begin_locked(vp); 4351 vunlazy_gone(vp); 4352 vn_irflag_set_locked(vp, VIRF_DOOMED); 4353 vn_set_state(vp, VSTATE_DESTROYING); 4354 4355 /* 4356 * Check to see if the vnode is in use. If so, we have to 4357 * call VOP_CLOSE() and VOP_INACTIVE(). 4358 * 4359 * It could be that VOP_INACTIVE() requested reclamation, in 4360 * which case we should avoid recursion, so check 4361 * VI_DOINGINACT. This is not precise but good enough. 4362 */ 4363 active = vp->v_usecount > 0; 4364 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 4365 doinginact = (vp->v_iflag & VI_DOINGINACT) != 0; 4366 4367 /* 4368 * If we need to do inactive VI_OWEINACT will be set. 4369 */ 4370 if (vp->v_iflag & VI_DEFINACT) { 4371 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 4372 vp->v_iflag &= ~VI_DEFINACT; 4373 vdropl(vp); 4374 } else { 4375 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); 4376 VI_UNLOCK(vp); 4377 } 4378 cache_purge_vgone(vp); 4379 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); 4380 4381 /* 4382 * If purging an active vnode, it must be closed and 4383 * deactivated before being reclaimed. 4384 */ 4385 if (active) 4386 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 4387 if (!doinginact) { 4388 do { 4389 if (oweinact || active) { 4390 VI_LOCK(vp); 4391 vinactivef(vp); 4392 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 4393 VI_UNLOCK(vp); 4394 } 4395 } while (oweinact); 4396 } 4397 if (vp->v_type == VSOCK) 4398 vfs_unp_reclaim(vp); 4399 4400 /* 4401 * Clean out any buffers associated with the vnode. 4402 * If the flush fails, just toss the buffers. 4403 */ 4404 mp = NULL; 4405 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 4406 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 4407 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { 4408 while (vinvalbuf(vp, 0, 0, 0) != 0) 4409 ; 4410 } 4411 4412 BO_LOCK(&vp->v_bufobj); 4413 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && 4414 vp->v_bufobj.bo_dirty.bv_cnt == 0 && 4415 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && 4416 vp->v_bufobj.bo_clean.bv_cnt == 0, 4417 ("vp %p bufobj not invalidated", vp)); 4418 4419 /* 4420 * For VMIO bufobj, BO_DEAD is set later, or in 4421 * vm_object_terminate() after the object's page queue is 4422 * flushed. 4423 */ 4424 object = vp->v_bufobj.bo_object; 4425 if (object == NULL) 4426 vp->v_bufobj.bo_flag |= BO_DEAD; 4427 BO_UNLOCK(&vp->v_bufobj); 4428 4429 /* 4430 * Handle the VM part. Tmpfs handles v_object on its own (the 4431 * OBJT_VNODE check). Nullfs or other bypassing filesystems 4432 * should not touch the object borrowed from the lower vnode 4433 * (the handle check). 4434 */ 4435 if (object != NULL && object->type == OBJT_VNODE && 4436 object->handle == vp) 4437 vnode_destroy_vobject(vp); 4438 4439 /* 4440 * Reclaim the vnode. 4441 */ 4442 if (VOP_RECLAIM(vp)) 4443 panic("vgone: cannot reclaim"); 4444 if (mp != NULL) 4445 vn_finished_secondary_write(mp); 4446 VNASSERT(vp->v_object == NULL, vp, 4447 ("vop_reclaim left v_object vp=%p", vp)); 4448 /* 4449 * Clear the advisory locks and wake up waiting threads. 4450 */ 4451 if (vp->v_lockf != NULL) { 4452 (void)VOP_ADVLOCKPURGE(vp); 4453 vp->v_lockf = NULL; 4454 } 4455 /* 4456 * Delete from old mount point vnode list. 4457 */ 4458 if (vp->v_mount == NULL) { 4459 VI_LOCK(vp); 4460 } else { 4461 delmntque(vp); 4462 ASSERT_VI_LOCKED(vp, "vgonel 2"); 4463 } 4464 /* 4465 * Done with purge, reset to the standard lock and invalidate 4466 * the vnode. 4467 */ 4468 vp->v_vnlock = &vp->v_lock; 4469 vp->v_op = &dead_vnodeops; 4470 vp->v_type = VBAD; 4471 vn_set_state(vp, VSTATE_DEAD); 4472 } 4473 4474 /* 4475 * Print out a description of a vnode. 4476 */ 4477 static const char *const vtypename[] = { 4478 [VNON] = "VNON", 4479 [VREG] = "VREG", 4480 [VDIR] = "VDIR", 4481 [VBLK] = "VBLK", 4482 [VCHR] = "VCHR", 4483 [VLNK] = "VLNK", 4484 [VSOCK] = "VSOCK", 4485 [VFIFO] = "VFIFO", 4486 [VBAD] = "VBAD", 4487 [VMARKER] = "VMARKER", 4488 }; 4489 _Static_assert(nitems(vtypename) == VLASTTYPE + 1, 4490 "vnode type name not added to vtypename"); 4491 4492 static const char *const vstatename[] = { 4493 [VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED", 4494 [VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED", 4495 [VSTATE_DESTROYING] = "VSTATE_DESTROYING", 4496 [VSTATE_DEAD] = "VSTATE_DEAD", 4497 }; 4498 _Static_assert(nitems(vstatename) == VLASTSTATE + 1, 4499 "vnode state name not added to vstatename"); 4500 4501 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0, 4502 "new hold count flag not added to vn_printf"); 4503 4504 void 4505 vn_printf(struct vnode *vp, const char *fmt, ...) 4506 { 4507 va_list ap; 4508 char buf[256], buf2[16]; 4509 u_long flags; 4510 u_int holdcnt; 4511 short irflag; 4512 4513 va_start(ap, fmt); 4514 vprintf(fmt, ap); 4515 va_end(ap); 4516 printf("%p: ", (void *)vp); 4517 printf("type %s state %s op %p\n", vtypename[vp->v_type], 4518 vstatename[vp->v_state], vp->v_op); 4519 holdcnt = atomic_load_int(&vp->v_holdcnt); 4520 printf(" usecount %d, writecount %d, refcount %d seqc users %d", 4521 vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS, 4522 vp->v_seqc_users); 4523 switch (vp->v_type) { 4524 case VDIR: 4525 printf(" mountedhere %p\n", vp->v_mountedhere); 4526 break; 4527 case VCHR: 4528 printf(" rdev %p\n", vp->v_rdev); 4529 break; 4530 case VSOCK: 4531 printf(" socket %p\n", vp->v_unpcb); 4532 break; 4533 case VFIFO: 4534 printf(" fifoinfo %p\n", vp->v_fifoinfo); 4535 break; 4536 default: 4537 printf("\n"); 4538 break; 4539 } 4540 buf[0] = '\0'; 4541 buf[1] = '\0'; 4542 if (holdcnt & VHOLD_NO_SMR) 4543 strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf)); 4544 printf(" hold count flags (%s)\n", buf + 1); 4545 4546 buf[0] = '\0'; 4547 buf[1] = '\0'; 4548 irflag = vn_irflag_read(vp); 4549 if (irflag & VIRF_DOOMED) 4550 strlcat(buf, "|VIRF_DOOMED", sizeof(buf)); 4551 if (irflag & VIRF_PGREAD) 4552 strlcat(buf, "|VIRF_PGREAD", sizeof(buf)); 4553 if (irflag & VIRF_MOUNTPOINT) 4554 strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf)); 4555 if (irflag & VIRF_TEXT_REF) 4556 strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf)); 4557 flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF); 4558 if (flags != 0) { 4559 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags); 4560 strlcat(buf, buf2, sizeof(buf)); 4561 } 4562 if (vp->v_vflag & VV_ROOT) 4563 strlcat(buf, "|VV_ROOT", sizeof(buf)); 4564 if (vp->v_vflag & VV_ISTTY) 4565 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 4566 if (vp->v_vflag & VV_NOSYNC) 4567 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 4568 if (vp->v_vflag & VV_ETERNALDEV) 4569 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 4570 if (vp->v_vflag & VV_CACHEDLABEL) 4571 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 4572 if (vp->v_vflag & VV_VMSIZEVNLOCK) 4573 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf)); 4574 if (vp->v_vflag & VV_COPYONWRITE) 4575 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 4576 if (vp->v_vflag & VV_SYSTEM) 4577 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 4578 if (vp->v_vflag & VV_PROCDEP) 4579 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 4580 if (vp->v_vflag & VV_DELETED) 4581 strlcat(buf, "|VV_DELETED", sizeof(buf)); 4582 if (vp->v_vflag & VV_MD) 4583 strlcat(buf, "|VV_MD", sizeof(buf)); 4584 if (vp->v_vflag & VV_FORCEINSMQ) 4585 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 4586 if (vp->v_vflag & VV_READLINK) 4587 strlcat(buf, "|VV_READLINK", sizeof(buf)); 4588 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 4589 VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM | 4590 VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK); 4591 if (flags != 0) { 4592 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 4593 strlcat(buf, buf2, sizeof(buf)); 4594 } 4595 if (vp->v_iflag & VI_MOUNT) 4596 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 4597 if (vp->v_iflag & VI_DOINGINACT) 4598 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 4599 if (vp->v_iflag & VI_OWEINACT) 4600 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 4601 if (vp->v_iflag & VI_DEFINACT) 4602 strlcat(buf, "|VI_DEFINACT", sizeof(buf)); 4603 if (vp->v_iflag & VI_FOPENING) 4604 strlcat(buf, "|VI_FOPENING", sizeof(buf)); 4605 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT | 4606 VI_OWEINACT | VI_DEFINACT | VI_FOPENING); 4607 if (flags != 0) { 4608 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 4609 strlcat(buf, buf2, sizeof(buf)); 4610 } 4611 if (vp->v_mflag & VMP_LAZYLIST) 4612 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf)); 4613 flags = vp->v_mflag & ~(VMP_LAZYLIST); 4614 if (flags != 0) { 4615 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags); 4616 strlcat(buf, buf2, sizeof(buf)); 4617 } 4618 printf(" flags (%s)", buf + 1); 4619 if (mtx_owned(VI_MTX(vp))) 4620 printf(" VI_LOCKed"); 4621 printf("\n"); 4622 if (vp->v_object != NULL) 4623 printf(" v_object %p ref %d pages %d " 4624 "cleanbuf %d dirtybuf %d\n", 4625 vp->v_object, vp->v_object->ref_count, 4626 vp->v_object->resident_page_count, 4627 vp->v_bufobj.bo_clean.bv_cnt, 4628 vp->v_bufobj.bo_dirty.bv_cnt); 4629 printf(" "); 4630 lockmgr_printinfo(vp->v_vnlock); 4631 if (vp->v_data != NULL) 4632 VOP_PRINT(vp); 4633 } 4634 4635 #ifdef DDB 4636 /* 4637 * List all of the locked vnodes in the system. 4638 * Called when debugging the kernel. 4639 */ 4640 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE) 4641 { 4642 struct mount *mp; 4643 struct vnode *vp; 4644 4645 /* 4646 * Note: because this is DDB, we can't obey the locking semantics 4647 * for these structures, which means we could catch an inconsistent 4648 * state and dereference a nasty pointer. Not much to be done 4649 * about that. 4650 */ 4651 db_printf("Locked vnodes\n"); 4652 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4653 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4654 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) 4655 vn_printf(vp, "vnode "); 4656 } 4657 } 4658 } 4659 4660 /* 4661 * Show details about the given vnode. 4662 */ 4663 DB_SHOW_COMMAND(vnode, db_show_vnode) 4664 { 4665 struct vnode *vp; 4666 4667 if (!have_addr) 4668 return; 4669 vp = (struct vnode *)addr; 4670 vn_printf(vp, "vnode "); 4671 } 4672 4673 /* 4674 * Show details about the given mount point. 4675 */ 4676 DB_SHOW_COMMAND(mount, db_show_mount) 4677 { 4678 struct mount *mp; 4679 struct vfsopt *opt; 4680 struct statfs *sp; 4681 struct vnode *vp; 4682 char buf[512]; 4683 uint64_t mflags; 4684 u_int flags; 4685 4686 if (!have_addr) { 4687 /* No address given, print short info about all mount points. */ 4688 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4689 db_printf("%p %s on %s (%s)\n", mp, 4690 mp->mnt_stat.f_mntfromname, 4691 mp->mnt_stat.f_mntonname, 4692 mp->mnt_stat.f_fstypename); 4693 if (db_pager_quit) 4694 break; 4695 } 4696 db_printf("\nMore info: show mount <addr>\n"); 4697 return; 4698 } 4699 4700 mp = (struct mount *)addr; 4701 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 4702 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 4703 4704 buf[0] = '\0'; 4705 mflags = mp->mnt_flag; 4706 #define MNT_FLAG(flag) do { \ 4707 if (mflags & (flag)) { \ 4708 if (buf[0] != '\0') \ 4709 strlcat(buf, ", ", sizeof(buf)); \ 4710 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 4711 mflags &= ~(flag); \ 4712 } \ 4713 } while (0) 4714 MNT_FLAG(MNT_RDONLY); 4715 MNT_FLAG(MNT_SYNCHRONOUS); 4716 MNT_FLAG(MNT_NOEXEC); 4717 MNT_FLAG(MNT_NOSUID); 4718 MNT_FLAG(MNT_NFS4ACLS); 4719 MNT_FLAG(MNT_UNION); 4720 MNT_FLAG(MNT_ASYNC); 4721 MNT_FLAG(MNT_SUIDDIR); 4722 MNT_FLAG(MNT_SOFTDEP); 4723 MNT_FLAG(MNT_NOSYMFOLLOW); 4724 MNT_FLAG(MNT_GJOURNAL); 4725 MNT_FLAG(MNT_MULTILABEL); 4726 MNT_FLAG(MNT_ACLS); 4727 MNT_FLAG(MNT_NOATIME); 4728 MNT_FLAG(MNT_NOCLUSTERR); 4729 MNT_FLAG(MNT_NOCLUSTERW); 4730 MNT_FLAG(MNT_SUJ); 4731 MNT_FLAG(MNT_EXRDONLY); 4732 MNT_FLAG(MNT_EXPORTED); 4733 MNT_FLAG(MNT_DEFEXPORTED); 4734 MNT_FLAG(MNT_EXPORTANON); 4735 MNT_FLAG(MNT_EXKERB); 4736 MNT_FLAG(MNT_EXPUBLIC); 4737 MNT_FLAG(MNT_LOCAL); 4738 MNT_FLAG(MNT_QUOTA); 4739 MNT_FLAG(MNT_ROOTFS); 4740 MNT_FLAG(MNT_USER); 4741 MNT_FLAG(MNT_IGNORE); 4742 MNT_FLAG(MNT_UPDATE); 4743 MNT_FLAG(MNT_DELEXPORT); 4744 MNT_FLAG(MNT_RELOAD); 4745 MNT_FLAG(MNT_FORCE); 4746 MNT_FLAG(MNT_SNAPSHOT); 4747 MNT_FLAG(MNT_BYFSID); 4748 #undef MNT_FLAG 4749 if (mflags != 0) { 4750 if (buf[0] != '\0') 4751 strlcat(buf, ", ", sizeof(buf)); 4752 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4753 "0x%016jx", mflags); 4754 } 4755 db_printf(" mnt_flag = %s\n", buf); 4756 4757 buf[0] = '\0'; 4758 flags = mp->mnt_kern_flag; 4759 #define MNT_KERN_FLAG(flag) do { \ 4760 if (flags & (flag)) { \ 4761 if (buf[0] != '\0') \ 4762 strlcat(buf, ", ", sizeof(buf)); \ 4763 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 4764 flags &= ~(flag); \ 4765 } \ 4766 } while (0) 4767 MNT_KERN_FLAG(MNTK_UNMOUNTF); 4768 MNT_KERN_FLAG(MNTK_ASYNC); 4769 MNT_KERN_FLAG(MNTK_SOFTDEP); 4770 MNT_KERN_FLAG(MNTK_NOMSYNC); 4771 MNT_KERN_FLAG(MNTK_DRAINING); 4772 MNT_KERN_FLAG(MNTK_REFEXPIRE); 4773 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 4774 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 4775 MNT_KERN_FLAG(MNTK_NO_IOPF); 4776 MNT_KERN_FLAG(MNTK_RECURSE); 4777 MNT_KERN_FLAG(MNTK_UPPER_WAITER); 4778 MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE); 4779 MNT_KERN_FLAG(MNTK_USES_BCACHE); 4780 MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG); 4781 MNT_KERN_FLAG(MNTK_FPLOOKUP); 4782 MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER); 4783 MNT_KERN_FLAG(MNTK_NOASYNC); 4784 MNT_KERN_FLAG(MNTK_UNMOUNT); 4785 MNT_KERN_FLAG(MNTK_MWAIT); 4786 MNT_KERN_FLAG(MNTK_SUSPEND); 4787 MNT_KERN_FLAG(MNTK_SUSPEND2); 4788 MNT_KERN_FLAG(MNTK_SUSPENDED); 4789 MNT_KERN_FLAG(MNTK_NULL_NOCACHE); 4790 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 4791 #undef MNT_KERN_FLAG 4792 if (flags != 0) { 4793 if (buf[0] != '\0') 4794 strlcat(buf, ", ", sizeof(buf)); 4795 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4796 "0x%08x", flags); 4797 } 4798 db_printf(" mnt_kern_flag = %s\n", buf); 4799 4800 db_printf(" mnt_opt = "); 4801 opt = TAILQ_FIRST(mp->mnt_opt); 4802 if (opt != NULL) { 4803 db_printf("%s", opt->name); 4804 opt = TAILQ_NEXT(opt, link); 4805 while (opt != NULL) { 4806 db_printf(", %s", opt->name); 4807 opt = TAILQ_NEXT(opt, link); 4808 } 4809 } 4810 db_printf("\n"); 4811 4812 sp = &mp->mnt_stat; 4813 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 4814 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 4815 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 4816 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 4817 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 4818 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 4819 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 4820 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 4821 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 4822 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 4823 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 4824 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 4825 4826 db_printf(" mnt_cred = { uid=%u ruid=%u", 4827 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 4828 if (jailed(mp->mnt_cred)) 4829 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 4830 db_printf(" }\n"); 4831 db_printf(" mnt_ref = %d (with %d in the struct)\n", 4832 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref); 4833 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 4834 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 4835 db_printf(" mnt_lazyvnodelistsize = %d\n", 4836 mp->mnt_lazyvnodelistsize); 4837 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n", 4838 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount); 4839 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 4840 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 4841 db_printf(" mnt_lockref = %d (with %d in the struct)\n", 4842 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref); 4843 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 4844 db_printf(" mnt_secondary_accwrites = %d\n", 4845 mp->mnt_secondary_accwrites); 4846 db_printf(" mnt_gjprovider = %s\n", 4847 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 4848 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops); 4849 4850 db_printf("\n\nList of active vnodes\n"); 4851 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4852 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) { 4853 vn_printf(vp, "vnode "); 4854 if (db_pager_quit) 4855 break; 4856 } 4857 } 4858 db_printf("\n\nList of inactive vnodes\n"); 4859 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4860 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) { 4861 vn_printf(vp, "vnode "); 4862 if (db_pager_quit) 4863 break; 4864 } 4865 } 4866 } 4867 #endif /* DDB */ 4868 4869 /* 4870 * Fill in a struct xvfsconf based on a struct vfsconf. 4871 */ 4872 static int 4873 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 4874 { 4875 struct xvfsconf xvfsp; 4876 4877 bzero(&xvfsp, sizeof(xvfsp)); 4878 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4879 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4880 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4881 xvfsp.vfc_flags = vfsp->vfc_flags; 4882 /* 4883 * These are unused in userland, we keep them 4884 * to not break binary compatibility. 4885 */ 4886 xvfsp.vfc_vfsops = NULL; 4887 xvfsp.vfc_next = NULL; 4888 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4889 } 4890 4891 #ifdef COMPAT_FREEBSD32 4892 struct xvfsconf32 { 4893 uint32_t vfc_vfsops; 4894 char vfc_name[MFSNAMELEN]; 4895 int32_t vfc_typenum; 4896 int32_t vfc_refcount; 4897 int32_t vfc_flags; 4898 uint32_t vfc_next; 4899 }; 4900 4901 static int 4902 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 4903 { 4904 struct xvfsconf32 xvfsp; 4905 4906 bzero(&xvfsp, sizeof(xvfsp)); 4907 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4908 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4909 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4910 xvfsp.vfc_flags = vfsp->vfc_flags; 4911 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4912 } 4913 #endif 4914 4915 /* 4916 * Top level filesystem related information gathering. 4917 */ 4918 static int 4919 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 4920 { 4921 struct vfsconf *vfsp; 4922 int error; 4923 4924 error = 0; 4925 vfsconf_slock(); 4926 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4927 #ifdef COMPAT_FREEBSD32 4928 if (req->flags & SCTL_MASK32) 4929 error = vfsconf2x32(req, vfsp); 4930 else 4931 #endif 4932 error = vfsconf2x(req, vfsp); 4933 if (error) 4934 break; 4935 } 4936 vfsconf_sunlock(); 4937 return (error); 4938 } 4939 4940 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | 4941 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, 4942 "S,xvfsconf", "List of all configured filesystems"); 4943 4944 #ifndef BURN_BRIDGES 4945 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 4946 4947 static int 4948 vfs_sysctl(SYSCTL_HANDLER_ARGS) 4949 { 4950 int *name = (int *)arg1 - 1; /* XXX */ 4951 u_int namelen = arg2 + 1; /* XXX */ 4952 struct vfsconf *vfsp; 4953 4954 log(LOG_WARNING, "userland calling deprecated sysctl, " 4955 "please rebuild world\n"); 4956 4957 #if 1 || defined(COMPAT_PRELITE2) 4958 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 4959 if (namelen == 1) 4960 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 4961 #endif 4962 4963 switch (name[1]) { 4964 case VFS_MAXTYPENUM: 4965 if (namelen != 2) 4966 return (ENOTDIR); 4967 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 4968 case VFS_CONF: 4969 if (namelen != 3) 4970 return (ENOTDIR); /* overloaded */ 4971 vfsconf_slock(); 4972 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4973 if (vfsp->vfc_typenum == name[2]) 4974 break; 4975 } 4976 vfsconf_sunlock(); 4977 if (vfsp == NULL) 4978 return (EOPNOTSUPP); 4979 #ifdef COMPAT_FREEBSD32 4980 if (req->flags & SCTL_MASK32) 4981 return (vfsconf2x32(req, vfsp)); 4982 else 4983 #endif 4984 return (vfsconf2x(req, vfsp)); 4985 } 4986 return (EOPNOTSUPP); 4987 } 4988 4989 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | 4990 CTLFLAG_MPSAFE, vfs_sysctl, 4991 "Generic filesystem"); 4992 4993 #if 1 || defined(COMPAT_PRELITE2) 4994 4995 static int 4996 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 4997 { 4998 int error; 4999 struct vfsconf *vfsp; 5000 struct ovfsconf ovfs; 5001 5002 vfsconf_slock(); 5003 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 5004 bzero(&ovfs, sizeof(ovfs)); 5005 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 5006 strcpy(ovfs.vfc_name, vfsp->vfc_name); 5007 ovfs.vfc_index = vfsp->vfc_typenum; 5008 ovfs.vfc_refcount = vfsp->vfc_refcount; 5009 ovfs.vfc_flags = vfsp->vfc_flags; 5010 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 5011 if (error != 0) { 5012 vfsconf_sunlock(); 5013 return (error); 5014 } 5015 } 5016 vfsconf_sunlock(); 5017 return (0); 5018 } 5019 5020 #endif /* 1 || COMPAT_PRELITE2 */ 5021 #endif /* !BURN_BRIDGES */ 5022 5023 static void 5024 unmount_or_warn(struct mount *mp) 5025 { 5026 int error; 5027 5028 error = dounmount(mp, MNT_FORCE, curthread); 5029 if (error != 0) { 5030 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); 5031 if (error == EBUSY) 5032 printf("BUSY)\n"); 5033 else 5034 printf("%d)\n", error); 5035 } 5036 } 5037 5038 /* 5039 * Unmount all filesystems. The list is traversed in reverse order 5040 * of mounting to avoid dependencies. 5041 */ 5042 void 5043 vfs_unmountall(void) 5044 { 5045 struct mount *mp, *tmp; 5046 5047 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 5048 5049 /* 5050 * Since this only runs when rebooting, it is not interlocked. 5051 */ 5052 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { 5053 vfs_ref(mp); 5054 5055 /* 5056 * Forcibly unmounting "/dev" before "/" would prevent clean 5057 * unmount of the latter. 5058 */ 5059 if (mp == rootdevmp) 5060 continue; 5061 5062 unmount_or_warn(mp); 5063 } 5064 5065 if (rootdevmp != NULL) 5066 unmount_or_warn(rootdevmp); 5067 } 5068 5069 static void 5070 vfs_deferred_inactive(struct vnode *vp, int lkflags) 5071 { 5072 5073 ASSERT_VI_LOCKED(vp, __func__); 5074 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp); 5075 if ((vp->v_iflag & VI_OWEINACT) == 0) { 5076 vdropl(vp); 5077 return; 5078 } 5079 if (vn_lock(vp, lkflags) == 0) { 5080 VI_LOCK(vp); 5081 vinactive(vp); 5082 VOP_UNLOCK(vp); 5083 vdropl(vp); 5084 return; 5085 } 5086 vdefer_inactive_unlocked(vp); 5087 } 5088 5089 static int 5090 vfs_periodic_inactive_filter(struct vnode *vp, void *arg) 5091 { 5092 5093 return (vp->v_iflag & VI_DEFINACT); 5094 } 5095 5096 static void __noinline 5097 vfs_periodic_inactive(struct mount *mp, int flags) 5098 { 5099 struct vnode *vp, *mvp; 5100 int lkflags; 5101 5102 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 5103 if (flags != MNT_WAIT) 5104 lkflags |= LK_NOWAIT; 5105 5106 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) { 5107 if ((vp->v_iflag & VI_DEFINACT) == 0) { 5108 VI_UNLOCK(vp); 5109 continue; 5110 } 5111 vp->v_iflag &= ~VI_DEFINACT; 5112 vfs_deferred_inactive(vp, lkflags); 5113 } 5114 } 5115 5116 static inline bool 5117 vfs_want_msync(struct vnode *vp) 5118 { 5119 struct vm_object *obj; 5120 5121 /* 5122 * This test may be performed without any locks held. 5123 * We rely on vm_object's type stability. 5124 */ 5125 if (vp->v_vflag & VV_NOSYNC) 5126 return (false); 5127 obj = vp->v_object; 5128 return (obj != NULL && vm_object_mightbedirty(obj)); 5129 } 5130 5131 static int 5132 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused) 5133 { 5134 5135 if (vp->v_vflag & VV_NOSYNC) 5136 return (false); 5137 if (vp->v_iflag & VI_DEFINACT) 5138 return (true); 5139 return (vfs_want_msync(vp)); 5140 } 5141 5142 static void __noinline 5143 vfs_periodic_msync_inactive(struct mount *mp, int flags) 5144 { 5145 struct vnode *vp, *mvp; 5146 int lkflags; 5147 bool seen_defer; 5148 5149 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 5150 if (flags != MNT_WAIT) 5151 lkflags |= LK_NOWAIT; 5152 5153 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) { 5154 seen_defer = false; 5155 if (vp->v_iflag & VI_DEFINACT) { 5156 vp->v_iflag &= ~VI_DEFINACT; 5157 seen_defer = true; 5158 } 5159 if (!vfs_want_msync(vp)) { 5160 if (seen_defer) 5161 vfs_deferred_inactive(vp, lkflags); 5162 else 5163 VI_UNLOCK(vp); 5164 continue; 5165 } 5166 if (vget(vp, lkflags) == 0) { 5167 if ((vp->v_vflag & VV_NOSYNC) == 0) { 5168 if (flags == MNT_WAIT) 5169 vnode_pager_clean_sync(vp); 5170 else 5171 vnode_pager_clean_async(vp); 5172 } 5173 vput(vp); 5174 if (seen_defer) 5175 vdrop(vp); 5176 } else { 5177 if (seen_defer) 5178 vdefer_inactive_unlocked(vp); 5179 } 5180 } 5181 } 5182 5183 void 5184 vfs_periodic(struct mount *mp, int flags) 5185 { 5186 5187 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 5188 5189 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0) 5190 vfs_periodic_inactive(mp, flags); 5191 else 5192 vfs_periodic_msync_inactive(mp, flags); 5193 } 5194 5195 static void 5196 destroy_vpollinfo_free(struct vpollinfo *vi) 5197 { 5198 5199 knlist_destroy(&vi->vpi_selinfo.si_note); 5200 mtx_destroy(&vi->vpi_lock); 5201 free(vi, M_VNODEPOLL); 5202 } 5203 5204 static void 5205 destroy_vpollinfo(struct vpollinfo *vi) 5206 { 5207 5208 knlist_clear(&vi->vpi_selinfo.si_note, 1); 5209 seldrain(&vi->vpi_selinfo); 5210 destroy_vpollinfo_free(vi); 5211 } 5212 5213 /* 5214 * Initialize per-vnode helper structure to hold poll-related state. 5215 */ 5216 void 5217 v_addpollinfo(struct vnode *vp) 5218 { 5219 struct vpollinfo *vi; 5220 5221 if (vp->v_pollinfo != NULL) 5222 return; 5223 vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO); 5224 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 5225 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 5226 vfs_knlunlock, vfs_knl_assert_lock); 5227 VI_LOCK(vp); 5228 if (vp->v_pollinfo != NULL) { 5229 VI_UNLOCK(vp); 5230 destroy_vpollinfo_free(vi); 5231 return; 5232 } 5233 vp->v_pollinfo = vi; 5234 VI_UNLOCK(vp); 5235 } 5236 5237 /* 5238 * Record a process's interest in events which might happen to 5239 * a vnode. Because poll uses the historic select-style interface 5240 * internally, this routine serves as both the ``check for any 5241 * pending events'' and the ``record my interest in future events'' 5242 * functions. (These are done together, while the lock is held, 5243 * to avoid race conditions.) 5244 */ 5245 int 5246 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 5247 { 5248 5249 v_addpollinfo(vp); 5250 mtx_lock(&vp->v_pollinfo->vpi_lock); 5251 if (vp->v_pollinfo->vpi_revents & events) { 5252 /* 5253 * This leaves events we are not interested 5254 * in available for the other process which 5255 * which presumably had requested them 5256 * (otherwise they would never have been 5257 * recorded). 5258 */ 5259 events &= vp->v_pollinfo->vpi_revents; 5260 vp->v_pollinfo->vpi_revents &= ~events; 5261 5262 mtx_unlock(&vp->v_pollinfo->vpi_lock); 5263 return (events); 5264 } 5265 vp->v_pollinfo->vpi_events |= events; 5266 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 5267 mtx_unlock(&vp->v_pollinfo->vpi_lock); 5268 return (0); 5269 } 5270 5271 /* 5272 * Routine to create and manage a filesystem syncer vnode. 5273 */ 5274 #define sync_close ((int (*)(struct vop_close_args *))nullop) 5275 static int sync_fsync(struct vop_fsync_args *); 5276 static int sync_inactive(struct vop_inactive_args *); 5277 static int sync_reclaim(struct vop_reclaim_args *); 5278 5279 static struct vop_vector sync_vnodeops = { 5280 .vop_bypass = VOP_EOPNOTSUPP, 5281 .vop_close = sync_close, 5282 .vop_fsync = sync_fsync, 5283 .vop_getwritemount = vop_stdgetwritemount, 5284 .vop_inactive = sync_inactive, 5285 .vop_need_inactive = vop_stdneed_inactive, 5286 .vop_reclaim = sync_reclaim, 5287 .vop_lock1 = vop_stdlock, 5288 .vop_unlock = vop_stdunlock, 5289 .vop_islocked = vop_stdislocked, 5290 .vop_fplookup_vexec = VOP_EAGAIN, 5291 .vop_fplookup_symlink = VOP_EAGAIN, 5292 }; 5293 VFS_VOP_VECTOR_REGISTER(sync_vnodeops); 5294 5295 /* 5296 * Create a new filesystem syncer vnode for the specified mount point. 5297 */ 5298 void 5299 vfs_allocate_syncvnode(struct mount *mp) 5300 { 5301 struct vnode *vp; 5302 struct bufobj *bo; 5303 static long start, incr, next; 5304 int error; 5305 5306 /* Allocate a new vnode */ 5307 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 5308 if (error != 0) 5309 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 5310 vp->v_type = VNON; 5311 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5312 vp->v_vflag |= VV_FORCEINSMQ; 5313 error = insmntque1(vp, mp); 5314 if (error != 0) 5315 panic("vfs_allocate_syncvnode: insmntque() failed"); 5316 vp->v_vflag &= ~VV_FORCEINSMQ; 5317 vn_set_state(vp, VSTATE_CONSTRUCTED); 5318 VOP_UNLOCK(vp); 5319 /* 5320 * Place the vnode onto the syncer worklist. We attempt to 5321 * scatter them about on the list so that they will go off 5322 * at evenly distributed times even if all the filesystems 5323 * are mounted at once. 5324 */ 5325 next += incr; 5326 if (next == 0 || next > syncer_maxdelay) { 5327 start /= 2; 5328 incr /= 2; 5329 if (start == 0) { 5330 start = syncer_maxdelay / 2; 5331 incr = syncer_maxdelay; 5332 } 5333 next = start; 5334 } 5335 bo = &vp->v_bufobj; 5336 BO_LOCK(bo); 5337 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 5338 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 5339 mtx_lock(&sync_mtx); 5340 sync_vnode_count++; 5341 if (mp->mnt_syncer == NULL) { 5342 mp->mnt_syncer = vp; 5343 vp = NULL; 5344 } 5345 mtx_unlock(&sync_mtx); 5346 BO_UNLOCK(bo); 5347 if (vp != NULL) { 5348 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5349 vgone(vp); 5350 vput(vp); 5351 } 5352 } 5353 5354 void 5355 vfs_deallocate_syncvnode(struct mount *mp) 5356 { 5357 struct vnode *vp; 5358 5359 mtx_lock(&sync_mtx); 5360 vp = mp->mnt_syncer; 5361 if (vp != NULL) 5362 mp->mnt_syncer = NULL; 5363 mtx_unlock(&sync_mtx); 5364 if (vp != NULL) 5365 vrele(vp); 5366 } 5367 5368 /* 5369 * Do a lazy sync of the filesystem. 5370 */ 5371 static int 5372 sync_fsync(struct vop_fsync_args *ap) 5373 { 5374 struct vnode *syncvp = ap->a_vp; 5375 struct mount *mp = syncvp->v_mount; 5376 int error, save; 5377 struct bufobj *bo; 5378 5379 /* 5380 * We only need to do something if this is a lazy evaluation. 5381 */ 5382 if (ap->a_waitfor != MNT_LAZY) 5383 return (0); 5384 5385 /* 5386 * Move ourselves to the back of the sync list. 5387 */ 5388 bo = &syncvp->v_bufobj; 5389 BO_LOCK(bo); 5390 vn_syncer_add_to_worklist(bo, syncdelay); 5391 BO_UNLOCK(bo); 5392 5393 /* 5394 * Walk the list of vnodes pushing all that are dirty and 5395 * not already on the sync list. 5396 */ 5397 if (vfs_busy(mp, MBF_NOWAIT) != 0) 5398 return (0); 5399 VOP_UNLOCK(syncvp); 5400 save = curthread_pflags_set(TDP_SYNCIO); 5401 /* 5402 * The filesystem at hand may be idle with free vnodes stored in the 5403 * batch. Return them instead of letting them stay there indefinitely. 5404 */ 5405 vfs_periodic(mp, MNT_NOWAIT); 5406 error = VFS_SYNC(mp, MNT_LAZY); 5407 curthread_pflags_restore(save); 5408 vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY); 5409 vfs_unbusy(mp); 5410 return (error); 5411 } 5412 5413 /* 5414 * The syncer vnode is no referenced. 5415 */ 5416 static int 5417 sync_inactive(struct vop_inactive_args *ap) 5418 { 5419 5420 vgone(ap->a_vp); 5421 return (0); 5422 } 5423 5424 /* 5425 * The syncer vnode is no longer needed and is being decommissioned. 5426 * 5427 * Modifications to the worklist must be protected by sync_mtx. 5428 */ 5429 static int 5430 sync_reclaim(struct vop_reclaim_args *ap) 5431 { 5432 struct vnode *vp = ap->a_vp; 5433 struct bufobj *bo; 5434 5435 bo = &vp->v_bufobj; 5436 BO_LOCK(bo); 5437 mtx_lock(&sync_mtx); 5438 if (vp->v_mount->mnt_syncer == vp) 5439 vp->v_mount->mnt_syncer = NULL; 5440 if (bo->bo_flag & BO_ONWORKLST) { 5441 LIST_REMOVE(bo, bo_synclist); 5442 syncer_worklist_len--; 5443 sync_vnode_count--; 5444 bo->bo_flag &= ~BO_ONWORKLST; 5445 } 5446 mtx_unlock(&sync_mtx); 5447 BO_UNLOCK(bo); 5448 5449 return (0); 5450 } 5451 5452 int 5453 vn_need_pageq_flush(struct vnode *vp) 5454 { 5455 struct vm_object *obj; 5456 5457 obj = vp->v_object; 5458 return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 5459 vm_object_mightbedirty(obj)); 5460 } 5461 5462 /* 5463 * Check if vnode represents a disk device 5464 */ 5465 bool 5466 vn_isdisk_error(struct vnode *vp, int *errp) 5467 { 5468 int error; 5469 5470 if (vp->v_type != VCHR) { 5471 error = ENOTBLK; 5472 goto out; 5473 } 5474 error = 0; 5475 dev_lock(); 5476 if (vp->v_rdev == NULL) 5477 error = ENXIO; 5478 else if (vp->v_rdev->si_devsw == NULL) 5479 error = ENXIO; 5480 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 5481 error = ENOTBLK; 5482 dev_unlock(); 5483 out: 5484 *errp = error; 5485 return (error == 0); 5486 } 5487 5488 bool 5489 vn_isdisk(struct vnode *vp) 5490 { 5491 int error; 5492 5493 return (vn_isdisk_error(vp, &error)); 5494 } 5495 5496 /* 5497 * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see 5498 * the comment above cache_fplookup for details. 5499 */ 5500 int 5501 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred) 5502 { 5503 int error; 5504 5505 VFS_SMR_ASSERT_ENTERED(); 5506 5507 /* Check the owner. */ 5508 if (cred->cr_uid == file_uid) { 5509 if (file_mode & S_IXUSR) 5510 return (0); 5511 goto out_error; 5512 } 5513 5514 /* Otherwise, check the groups (first match) */ 5515 if (groupmember(file_gid, cred)) { 5516 if (file_mode & S_IXGRP) 5517 return (0); 5518 goto out_error; 5519 } 5520 5521 /* Otherwise, check everyone else. */ 5522 if (file_mode & S_IXOTH) 5523 return (0); 5524 out_error: 5525 /* 5526 * Permission check failed, but it is possible denial will get overwritten 5527 * (e.g., when root is traversing through a 700 directory owned by someone 5528 * else). 5529 * 5530 * vaccess() calls priv_check_cred which in turn can descent into MAC 5531 * modules overriding this result. It's quite unclear what semantics 5532 * are allowed for them to operate, thus for safety we don't call them 5533 * from within the SMR section. This also means if any such modules 5534 * are present, we have to let the regular lookup decide. 5535 */ 5536 error = priv_check_cred_vfs_lookup_nomac(cred); 5537 switch (error) { 5538 case 0: 5539 return (0); 5540 case EAGAIN: 5541 /* 5542 * MAC modules present. 5543 */ 5544 return (EAGAIN); 5545 case EPERM: 5546 return (EACCES); 5547 default: 5548 return (error); 5549 } 5550 } 5551 5552 /* 5553 * Common filesystem object access control check routine. Accepts a 5554 * vnode's type, "mode", uid and gid, requested access mode, and credentials. 5555 * Returns 0 on success, or an errno on failure. 5556 */ 5557 int 5558 vaccess(__enum_uint8(vtype) type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 5559 accmode_t accmode, struct ucred *cred) 5560 { 5561 accmode_t dac_granted; 5562 accmode_t priv_granted; 5563 5564 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 5565 ("invalid bit in accmode")); 5566 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 5567 ("VAPPEND without VWRITE")); 5568 5569 /* 5570 * Look for a normal, non-privileged way to access the file/directory 5571 * as requested. If it exists, go with that. 5572 */ 5573 5574 dac_granted = 0; 5575 5576 /* Check the owner. */ 5577 if (cred->cr_uid == file_uid) { 5578 dac_granted |= VADMIN; 5579 if (file_mode & S_IXUSR) 5580 dac_granted |= VEXEC; 5581 if (file_mode & S_IRUSR) 5582 dac_granted |= VREAD; 5583 if (file_mode & S_IWUSR) 5584 dac_granted |= (VWRITE | VAPPEND); 5585 5586 if ((accmode & dac_granted) == accmode) 5587 return (0); 5588 5589 goto privcheck; 5590 } 5591 5592 /* Otherwise, check the groups (first match) */ 5593 if (groupmember(file_gid, cred)) { 5594 if (file_mode & S_IXGRP) 5595 dac_granted |= VEXEC; 5596 if (file_mode & S_IRGRP) 5597 dac_granted |= VREAD; 5598 if (file_mode & S_IWGRP) 5599 dac_granted |= (VWRITE | VAPPEND); 5600 5601 if ((accmode & dac_granted) == accmode) 5602 return (0); 5603 5604 goto privcheck; 5605 } 5606 5607 /* Otherwise, check everyone else. */ 5608 if (file_mode & S_IXOTH) 5609 dac_granted |= VEXEC; 5610 if (file_mode & S_IROTH) 5611 dac_granted |= VREAD; 5612 if (file_mode & S_IWOTH) 5613 dac_granted |= (VWRITE | VAPPEND); 5614 if ((accmode & dac_granted) == accmode) 5615 return (0); 5616 5617 privcheck: 5618 /* 5619 * Build a privilege mask to determine if the set of privileges 5620 * satisfies the requirements when combined with the granted mask 5621 * from above. For each privilege, if the privilege is required, 5622 * bitwise or the request type onto the priv_granted mask. 5623 */ 5624 priv_granted = 0; 5625 5626 if (type == VDIR) { 5627 /* 5628 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 5629 * requests, instead of PRIV_VFS_EXEC. 5630 */ 5631 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5632 !priv_check_cred(cred, PRIV_VFS_LOOKUP)) 5633 priv_granted |= VEXEC; 5634 } else { 5635 /* 5636 * Ensure that at least one execute bit is on. Otherwise, 5637 * a privileged user will always succeed, and we don't want 5638 * this to happen unless the file really is executable. 5639 */ 5640 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5641 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 5642 !priv_check_cred(cred, PRIV_VFS_EXEC)) 5643 priv_granted |= VEXEC; 5644 } 5645 5646 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 5647 !priv_check_cred(cred, PRIV_VFS_READ)) 5648 priv_granted |= VREAD; 5649 5650 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 5651 !priv_check_cred(cred, PRIV_VFS_WRITE)) 5652 priv_granted |= (VWRITE | VAPPEND); 5653 5654 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 5655 !priv_check_cred(cred, PRIV_VFS_ADMIN)) 5656 priv_granted |= VADMIN; 5657 5658 if ((accmode & (priv_granted | dac_granted)) == accmode) { 5659 return (0); 5660 } 5661 5662 return ((accmode & VADMIN) ? EPERM : EACCES); 5663 } 5664 5665 /* 5666 * Credential check based on process requesting service, and per-attribute 5667 * permissions. 5668 */ 5669 int 5670 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 5671 struct thread *td, accmode_t accmode) 5672 { 5673 5674 /* 5675 * Kernel-invoked always succeeds. 5676 */ 5677 if (cred == NOCRED) 5678 return (0); 5679 5680 /* 5681 * Do not allow privileged processes in jail to directly manipulate 5682 * system attributes. 5683 */ 5684 switch (attrnamespace) { 5685 case EXTATTR_NAMESPACE_SYSTEM: 5686 /* Potentially should be: return (EPERM); */ 5687 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM)); 5688 case EXTATTR_NAMESPACE_USER: 5689 return (VOP_ACCESS(vp, accmode, cred, td)); 5690 default: 5691 return (EPERM); 5692 } 5693 } 5694 5695 #ifdef DEBUG_VFS_LOCKS 5696 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 5697 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, 5698 "Drop into debugger on lock violation"); 5699 5700 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 5701 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 5702 0, "Check for interlock across VOPs"); 5703 5704 int vfs_badlock_print = 1; /* Print lock violations. */ 5705 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 5706 0, "Print lock violations"); 5707 5708 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */ 5709 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode, 5710 0, "Print vnode details on lock violations"); 5711 5712 #ifdef KDB 5713 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 5714 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, 5715 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations"); 5716 #endif 5717 5718 static void 5719 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 5720 { 5721 5722 #ifdef KDB 5723 if (vfs_badlock_backtrace) 5724 kdb_backtrace(); 5725 #endif 5726 if (vfs_badlock_vnode) 5727 vn_printf(vp, "vnode "); 5728 if (vfs_badlock_print) 5729 printf("%s: %p %s\n", str, (void *)vp, msg); 5730 if (vfs_badlock_ddb) 5731 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 5732 } 5733 5734 void 5735 assert_vi_locked(struct vnode *vp, const char *str) 5736 { 5737 5738 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 5739 vfs_badlock("interlock is not locked but should be", str, vp); 5740 } 5741 5742 void 5743 assert_vi_unlocked(struct vnode *vp, const char *str) 5744 { 5745 5746 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 5747 vfs_badlock("interlock is locked but should not be", str, vp); 5748 } 5749 5750 void 5751 assert_vop_locked(struct vnode *vp, const char *str) 5752 { 5753 if (KERNEL_PANICKED() || vp == NULL) 5754 return; 5755 5756 #ifdef WITNESS 5757 if ((vp->v_irflag & VIRF_CROSSMP) == 0 && 5758 witness_is_owned(&vp->v_vnlock->lock_object) == -1) 5759 #else 5760 int locked = VOP_ISLOCKED(vp); 5761 if (locked == 0 || locked == LK_EXCLOTHER) 5762 #endif 5763 vfs_badlock("is not locked but should be", str, vp); 5764 } 5765 5766 void 5767 assert_vop_unlocked(struct vnode *vp, const char *str) 5768 { 5769 if (KERNEL_PANICKED() || vp == NULL) 5770 return; 5771 5772 #ifdef WITNESS 5773 if ((vp->v_irflag & VIRF_CROSSMP) == 0 && 5774 witness_is_owned(&vp->v_vnlock->lock_object) == 1) 5775 #else 5776 if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE) 5777 #endif 5778 vfs_badlock("is locked but should not be", str, vp); 5779 } 5780 5781 void 5782 assert_vop_elocked(struct vnode *vp, const char *str) 5783 { 5784 if (KERNEL_PANICKED() || vp == NULL) 5785 return; 5786 5787 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 5788 vfs_badlock("is not exclusive locked but should be", str, vp); 5789 } 5790 #endif /* DEBUG_VFS_LOCKS */ 5791 5792 void 5793 vop_rename_fail(struct vop_rename_args *ap) 5794 { 5795 5796 if (ap->a_tvp != NULL) 5797 vput(ap->a_tvp); 5798 if (ap->a_tdvp == ap->a_tvp) 5799 vrele(ap->a_tdvp); 5800 else 5801 vput(ap->a_tdvp); 5802 vrele(ap->a_fdvp); 5803 vrele(ap->a_fvp); 5804 } 5805 5806 void 5807 vop_rename_pre(void *ap) 5808 { 5809 struct vop_rename_args *a = ap; 5810 5811 #ifdef DEBUG_VFS_LOCKS 5812 if (a->a_tvp) 5813 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 5814 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 5815 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 5816 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 5817 5818 /* Check the source (from). */ 5819 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 5820 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 5821 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 5822 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 5823 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 5824 5825 /* Check the target. */ 5826 if (a->a_tvp) 5827 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 5828 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 5829 #endif 5830 /* 5831 * It may be tempting to add vn_seqc_write_begin/end calls here and 5832 * in vop_rename_post but that's not going to work out since some 5833 * filesystems relookup vnodes mid-rename. This is probably a bug. 5834 * 5835 * For now filesystems are expected to do the relevant calls after they 5836 * decide what vnodes to operate on. 5837 */ 5838 if (a->a_tdvp != a->a_fdvp) 5839 vhold(a->a_fdvp); 5840 if (a->a_tvp != a->a_fvp) 5841 vhold(a->a_fvp); 5842 vhold(a->a_tdvp); 5843 if (a->a_tvp) 5844 vhold(a->a_tvp); 5845 } 5846 5847 #ifdef DEBUG_VFS_LOCKS 5848 void 5849 vop_fplookup_vexec_debugpre(void *ap __unused) 5850 { 5851 5852 VFS_SMR_ASSERT_ENTERED(); 5853 } 5854 5855 void 5856 vop_fplookup_vexec_debugpost(void *ap, int rc) 5857 { 5858 struct vop_fplookup_vexec_args *a; 5859 struct vnode *vp; 5860 5861 a = ap; 5862 vp = a->a_vp; 5863 5864 VFS_SMR_ASSERT_ENTERED(); 5865 if (rc == EOPNOTSUPP) 5866 VNPASS(VN_IS_DOOMED(vp), vp); 5867 } 5868 5869 void 5870 vop_fplookup_symlink_debugpre(void *ap __unused) 5871 { 5872 5873 VFS_SMR_ASSERT_ENTERED(); 5874 } 5875 5876 void 5877 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused) 5878 { 5879 5880 VFS_SMR_ASSERT_ENTERED(); 5881 } 5882 5883 static void 5884 vop_fsync_debugprepost(struct vnode *vp, const char *name) 5885 { 5886 if (vp->v_type == VCHR) 5887 ; 5888 /* 5889 * The shared vs. exclusive locking policy for fsync() 5890 * is actually determined by vp's write mount as indicated 5891 * by VOP_GETWRITEMOUNT(), which for stacked filesystems 5892 * may not be the same as vp->v_mount. However, if the 5893 * underlying filesystem which really handles the fsync() 5894 * supports shared locking, the stacked filesystem must also 5895 * be prepared for its VOP_FSYNC() operation to be called 5896 * with only a shared lock. On the other hand, if the 5897 * stacked filesystem claims support for shared write 5898 * locking but the underlying filesystem does not, and the 5899 * caller incorrectly uses a shared lock, this condition 5900 * should still be caught when the stacked filesystem 5901 * invokes VOP_FSYNC() on the underlying filesystem. 5902 */ 5903 else if (MNT_SHARED_WRITES(vp->v_mount)) 5904 ASSERT_VOP_LOCKED(vp, name); 5905 else 5906 ASSERT_VOP_ELOCKED(vp, name); 5907 } 5908 5909 void 5910 vop_fsync_debugpre(void *a) 5911 { 5912 struct vop_fsync_args *ap; 5913 5914 ap = a; 5915 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5916 } 5917 5918 void 5919 vop_fsync_debugpost(void *a, int rc __unused) 5920 { 5921 struct vop_fsync_args *ap; 5922 5923 ap = a; 5924 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5925 } 5926 5927 void 5928 vop_fdatasync_debugpre(void *a) 5929 { 5930 struct vop_fdatasync_args *ap; 5931 5932 ap = a; 5933 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5934 } 5935 5936 void 5937 vop_fdatasync_debugpost(void *a, int rc __unused) 5938 { 5939 struct vop_fdatasync_args *ap; 5940 5941 ap = a; 5942 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5943 } 5944 5945 void 5946 vop_strategy_debugpre(void *ap) 5947 { 5948 struct vop_strategy_args *a; 5949 struct buf *bp; 5950 5951 a = ap; 5952 bp = a->a_bp; 5953 5954 /* 5955 * Cluster ops lock their component buffers but not the IO container. 5956 */ 5957 if ((bp->b_flags & B_CLUSTER) != 0) 5958 return; 5959 5960 if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) { 5961 if (vfs_badlock_print) 5962 printf( 5963 "VOP_STRATEGY: bp is not locked but should be\n"); 5964 if (vfs_badlock_ddb) 5965 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 5966 } 5967 } 5968 5969 void 5970 vop_lock_debugpre(void *ap) 5971 { 5972 struct vop_lock1_args *a = ap; 5973 5974 if ((a->a_flags & LK_INTERLOCK) == 0) 5975 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5976 else 5977 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 5978 } 5979 5980 void 5981 vop_lock_debugpost(void *ap, int rc) 5982 { 5983 struct vop_lock1_args *a = ap; 5984 5985 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5986 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) 5987 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 5988 } 5989 5990 void 5991 vop_unlock_debugpre(void *ap) 5992 { 5993 struct vop_unlock_args *a = ap; 5994 struct vnode *vp = a->a_vp; 5995 5996 VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp); 5997 ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK"); 5998 } 5999 6000 void 6001 vop_need_inactive_debugpre(void *ap) 6002 { 6003 struct vop_need_inactive_args *a = ap; 6004 6005 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 6006 } 6007 6008 void 6009 vop_need_inactive_debugpost(void *ap, int rc) 6010 { 6011 struct vop_need_inactive_args *a = ap; 6012 6013 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 6014 } 6015 #endif 6016 6017 void 6018 vop_create_pre(void *ap) 6019 { 6020 struct vop_create_args *a; 6021 struct vnode *dvp; 6022 6023 a = ap; 6024 dvp = a->a_dvp; 6025 vn_seqc_write_begin(dvp); 6026 } 6027 6028 void 6029 vop_create_post(void *ap, int rc) 6030 { 6031 struct vop_create_args *a; 6032 struct vnode *dvp; 6033 6034 a = ap; 6035 dvp = a->a_dvp; 6036 vn_seqc_write_end(dvp); 6037 if (!rc) 6038 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6039 } 6040 6041 void 6042 vop_whiteout_pre(void *ap) 6043 { 6044 struct vop_whiteout_args *a; 6045 struct vnode *dvp; 6046 6047 a = ap; 6048 dvp = a->a_dvp; 6049 vn_seqc_write_begin(dvp); 6050 } 6051 6052 void 6053 vop_whiteout_post(void *ap, int rc) 6054 { 6055 struct vop_whiteout_args *a; 6056 struct vnode *dvp; 6057 6058 a = ap; 6059 dvp = a->a_dvp; 6060 vn_seqc_write_end(dvp); 6061 } 6062 6063 void 6064 vop_deleteextattr_pre(void *ap) 6065 { 6066 struct vop_deleteextattr_args *a; 6067 struct vnode *vp; 6068 6069 a = ap; 6070 vp = a->a_vp; 6071 vn_seqc_write_begin(vp); 6072 } 6073 6074 void 6075 vop_deleteextattr_post(void *ap, int rc) 6076 { 6077 struct vop_deleteextattr_args *a; 6078 struct vnode *vp; 6079 6080 a = ap; 6081 vp = a->a_vp; 6082 vn_seqc_write_end(vp); 6083 if (!rc) 6084 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 6085 } 6086 6087 void 6088 vop_link_pre(void *ap) 6089 { 6090 struct vop_link_args *a; 6091 struct vnode *vp, *tdvp; 6092 6093 a = ap; 6094 vp = a->a_vp; 6095 tdvp = a->a_tdvp; 6096 vn_seqc_write_begin(vp); 6097 vn_seqc_write_begin(tdvp); 6098 } 6099 6100 void 6101 vop_link_post(void *ap, int rc) 6102 { 6103 struct vop_link_args *a; 6104 struct vnode *vp, *tdvp; 6105 6106 a = ap; 6107 vp = a->a_vp; 6108 tdvp = a->a_tdvp; 6109 vn_seqc_write_end(vp); 6110 vn_seqc_write_end(tdvp); 6111 if (!rc) { 6112 VFS_KNOTE_LOCKED(vp, NOTE_LINK); 6113 VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE); 6114 } 6115 } 6116 6117 void 6118 vop_mkdir_pre(void *ap) 6119 { 6120 struct vop_mkdir_args *a; 6121 struct vnode *dvp; 6122 6123 a = ap; 6124 dvp = a->a_dvp; 6125 vn_seqc_write_begin(dvp); 6126 } 6127 6128 void 6129 vop_mkdir_post(void *ap, int rc) 6130 { 6131 struct vop_mkdir_args *a; 6132 struct vnode *dvp; 6133 6134 a = ap; 6135 dvp = a->a_dvp; 6136 vn_seqc_write_end(dvp); 6137 if (!rc) 6138 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 6139 } 6140 6141 #ifdef DEBUG_VFS_LOCKS 6142 void 6143 vop_mkdir_debugpost(void *ap, int rc) 6144 { 6145 struct vop_mkdir_args *a; 6146 6147 a = ap; 6148 if (!rc) 6149 cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp); 6150 } 6151 #endif 6152 6153 void 6154 vop_mknod_pre(void *ap) 6155 { 6156 struct vop_mknod_args *a; 6157 struct vnode *dvp; 6158 6159 a = ap; 6160 dvp = a->a_dvp; 6161 vn_seqc_write_begin(dvp); 6162 } 6163 6164 void 6165 vop_mknod_post(void *ap, int rc) 6166 { 6167 struct vop_mknod_args *a; 6168 struct vnode *dvp; 6169 6170 a = ap; 6171 dvp = a->a_dvp; 6172 vn_seqc_write_end(dvp); 6173 if (!rc) 6174 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6175 } 6176 6177 void 6178 vop_reclaim_post(void *ap, int rc) 6179 { 6180 struct vop_reclaim_args *a; 6181 struct vnode *vp; 6182 6183 a = ap; 6184 vp = a->a_vp; 6185 ASSERT_VOP_IN_SEQC(vp); 6186 if (!rc) 6187 VFS_KNOTE_LOCKED(vp, NOTE_REVOKE); 6188 } 6189 6190 void 6191 vop_remove_pre(void *ap) 6192 { 6193 struct vop_remove_args *a; 6194 struct vnode *dvp, *vp; 6195 6196 a = ap; 6197 dvp = a->a_dvp; 6198 vp = a->a_vp; 6199 vn_seqc_write_begin(dvp); 6200 vn_seqc_write_begin(vp); 6201 } 6202 6203 void 6204 vop_remove_post(void *ap, int rc) 6205 { 6206 struct vop_remove_args *a; 6207 struct vnode *dvp, *vp; 6208 6209 a = ap; 6210 dvp = a->a_dvp; 6211 vp = a->a_vp; 6212 vn_seqc_write_end(dvp); 6213 vn_seqc_write_end(vp); 6214 if (!rc) { 6215 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6216 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 6217 } 6218 } 6219 6220 void 6221 vop_rename_post(void *ap, int rc) 6222 { 6223 struct vop_rename_args *a = ap; 6224 long hint; 6225 6226 if (!rc) { 6227 hint = NOTE_WRITE; 6228 if (a->a_fdvp == a->a_tdvp) { 6229 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) 6230 hint |= NOTE_LINK; 6231 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 6232 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 6233 } else { 6234 hint |= NOTE_EXTEND; 6235 if (a->a_fvp->v_type == VDIR) 6236 hint |= NOTE_LINK; 6237 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 6238 6239 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && 6240 a->a_tvp->v_type == VDIR) 6241 hint &= ~NOTE_LINK; 6242 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 6243 } 6244 6245 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 6246 if (a->a_tvp) 6247 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 6248 } 6249 if (a->a_tdvp != a->a_fdvp) 6250 vdrop(a->a_fdvp); 6251 if (a->a_tvp != a->a_fvp) 6252 vdrop(a->a_fvp); 6253 vdrop(a->a_tdvp); 6254 if (a->a_tvp) 6255 vdrop(a->a_tvp); 6256 } 6257 6258 void 6259 vop_rmdir_pre(void *ap) 6260 { 6261 struct vop_rmdir_args *a; 6262 struct vnode *dvp, *vp; 6263 6264 a = ap; 6265 dvp = a->a_dvp; 6266 vp = a->a_vp; 6267 vn_seqc_write_begin(dvp); 6268 vn_seqc_write_begin(vp); 6269 } 6270 6271 void 6272 vop_rmdir_post(void *ap, int rc) 6273 { 6274 struct vop_rmdir_args *a; 6275 struct vnode *dvp, *vp; 6276 6277 a = ap; 6278 dvp = a->a_dvp; 6279 vp = a->a_vp; 6280 vn_seqc_write_end(dvp); 6281 vn_seqc_write_end(vp); 6282 if (!rc) { 6283 vp->v_vflag |= VV_UNLINKED; 6284 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 6285 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 6286 } 6287 } 6288 6289 void 6290 vop_setattr_pre(void *ap) 6291 { 6292 struct vop_setattr_args *a; 6293 struct vnode *vp; 6294 6295 a = ap; 6296 vp = a->a_vp; 6297 vn_seqc_write_begin(vp); 6298 } 6299 6300 void 6301 vop_setattr_post(void *ap, int rc) 6302 { 6303 struct vop_setattr_args *a; 6304 struct vnode *vp; 6305 6306 a = ap; 6307 vp = a->a_vp; 6308 vn_seqc_write_end(vp); 6309 if (!rc) 6310 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 6311 } 6312 6313 void 6314 vop_setacl_pre(void *ap) 6315 { 6316 struct vop_setacl_args *a; 6317 struct vnode *vp; 6318 6319 a = ap; 6320 vp = a->a_vp; 6321 vn_seqc_write_begin(vp); 6322 } 6323 6324 void 6325 vop_setacl_post(void *ap, int rc __unused) 6326 { 6327 struct vop_setacl_args *a; 6328 struct vnode *vp; 6329 6330 a = ap; 6331 vp = a->a_vp; 6332 vn_seqc_write_end(vp); 6333 } 6334 6335 void 6336 vop_setextattr_pre(void *ap) 6337 { 6338 struct vop_setextattr_args *a; 6339 struct vnode *vp; 6340 6341 a = ap; 6342 vp = a->a_vp; 6343 vn_seqc_write_begin(vp); 6344 } 6345 6346 void 6347 vop_setextattr_post(void *ap, int rc) 6348 { 6349 struct vop_setextattr_args *a; 6350 struct vnode *vp; 6351 6352 a = ap; 6353 vp = a->a_vp; 6354 vn_seqc_write_end(vp); 6355 if (!rc) 6356 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 6357 } 6358 6359 void 6360 vop_symlink_pre(void *ap) 6361 { 6362 struct vop_symlink_args *a; 6363 struct vnode *dvp; 6364 6365 a = ap; 6366 dvp = a->a_dvp; 6367 vn_seqc_write_begin(dvp); 6368 } 6369 6370 void 6371 vop_symlink_post(void *ap, int rc) 6372 { 6373 struct vop_symlink_args *a; 6374 struct vnode *dvp; 6375 6376 a = ap; 6377 dvp = a->a_dvp; 6378 vn_seqc_write_end(dvp); 6379 if (!rc) 6380 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6381 } 6382 6383 void 6384 vop_open_post(void *ap, int rc) 6385 { 6386 struct vop_open_args *a = ap; 6387 6388 if (!rc) 6389 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); 6390 } 6391 6392 void 6393 vop_close_post(void *ap, int rc) 6394 { 6395 struct vop_close_args *a = ap; 6396 6397 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ 6398 !VN_IS_DOOMED(a->a_vp))) { 6399 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 6400 NOTE_CLOSE_WRITE : NOTE_CLOSE); 6401 } 6402 } 6403 6404 void 6405 vop_read_post(void *ap, int rc) 6406 { 6407 struct vop_read_args *a = ap; 6408 6409 if (!rc) 6410 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 6411 } 6412 6413 void 6414 vop_read_pgcache_post(void *ap, int rc) 6415 { 6416 struct vop_read_pgcache_args *a = ap; 6417 6418 if (!rc) 6419 VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ); 6420 } 6421 6422 void 6423 vop_readdir_post(void *ap, int rc) 6424 { 6425 struct vop_readdir_args *a = ap; 6426 6427 if (!rc) 6428 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 6429 } 6430 6431 static struct knlist fs_knlist; 6432 6433 static void 6434 vfs_event_init(void *arg) 6435 { 6436 knlist_init_mtx(&fs_knlist, NULL); 6437 } 6438 /* XXX - correct order? */ 6439 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 6440 6441 void 6442 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 6443 { 6444 6445 KNOTE_UNLOCKED(&fs_knlist, event); 6446 } 6447 6448 static int filt_fsattach(struct knote *kn); 6449 static void filt_fsdetach(struct knote *kn); 6450 static int filt_fsevent(struct knote *kn, long hint); 6451 6452 struct filterops fs_filtops = { 6453 .f_isfd = 0, 6454 .f_attach = filt_fsattach, 6455 .f_detach = filt_fsdetach, 6456 .f_event = filt_fsevent 6457 }; 6458 6459 static int 6460 filt_fsattach(struct knote *kn) 6461 { 6462 6463 kn->kn_flags |= EV_CLEAR; 6464 knlist_add(&fs_knlist, kn, 0); 6465 return (0); 6466 } 6467 6468 static void 6469 filt_fsdetach(struct knote *kn) 6470 { 6471 6472 knlist_remove(&fs_knlist, kn, 0); 6473 } 6474 6475 static int 6476 filt_fsevent(struct knote *kn, long hint) 6477 { 6478 6479 kn->kn_fflags |= kn->kn_sfflags & hint; 6480 6481 return (kn->kn_fflags != 0); 6482 } 6483 6484 static int 6485 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 6486 { 6487 struct vfsidctl vc; 6488 int error; 6489 struct mount *mp; 6490 6491 error = SYSCTL_IN(req, &vc, sizeof(vc)); 6492 if (error) 6493 return (error); 6494 if (vc.vc_vers != VFS_CTL_VERS1) 6495 return (EINVAL); 6496 mp = vfs_getvfs(&vc.vc_fsid); 6497 if (mp == NULL) 6498 return (ENOENT); 6499 /* ensure that a specific sysctl goes to the right filesystem. */ 6500 if (strcmp(vc.vc_fstypename, "*") != 0 && 6501 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 6502 vfs_rel(mp); 6503 return (EINVAL); 6504 } 6505 VCTLTOREQ(&vc, req); 6506 error = VFS_SYSCTL(mp, vc.vc_op, req); 6507 vfs_rel(mp); 6508 return (error); 6509 } 6510 6511 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR, 6512 NULL, 0, sysctl_vfs_ctl, "", 6513 "Sysctl by fsid"); 6514 6515 /* 6516 * Function to initialize a va_filerev field sensibly. 6517 * XXX: Wouldn't a random number make a lot more sense ?? 6518 */ 6519 u_quad_t 6520 init_va_filerev(void) 6521 { 6522 struct bintime bt; 6523 6524 getbinuptime(&bt); 6525 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 6526 } 6527 6528 static int filt_vfsread(struct knote *kn, long hint); 6529 static int filt_vfswrite(struct knote *kn, long hint); 6530 static int filt_vfsvnode(struct knote *kn, long hint); 6531 static void filt_vfsdetach(struct knote *kn); 6532 static struct filterops vfsread_filtops = { 6533 .f_isfd = 1, 6534 .f_detach = filt_vfsdetach, 6535 .f_event = filt_vfsread 6536 }; 6537 static struct filterops vfswrite_filtops = { 6538 .f_isfd = 1, 6539 .f_detach = filt_vfsdetach, 6540 .f_event = filt_vfswrite 6541 }; 6542 static struct filterops vfsvnode_filtops = { 6543 .f_isfd = 1, 6544 .f_detach = filt_vfsdetach, 6545 .f_event = filt_vfsvnode 6546 }; 6547 6548 static void 6549 vfs_knllock(void *arg) 6550 { 6551 struct vnode *vp = arg; 6552 6553 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 6554 } 6555 6556 static void 6557 vfs_knlunlock(void *arg) 6558 { 6559 struct vnode *vp = arg; 6560 6561 VOP_UNLOCK(vp); 6562 } 6563 6564 static void 6565 vfs_knl_assert_lock(void *arg, int what) 6566 { 6567 #ifdef DEBUG_VFS_LOCKS 6568 struct vnode *vp = arg; 6569 6570 if (what == LA_LOCKED) 6571 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 6572 else 6573 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 6574 #endif 6575 } 6576 6577 int 6578 vfs_kqfilter(struct vop_kqfilter_args *ap) 6579 { 6580 struct vnode *vp = ap->a_vp; 6581 struct knote *kn = ap->a_kn; 6582 struct knlist *knl; 6583 6584 KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ && 6585 kn->kn_filter != EVFILT_WRITE), 6586 ("READ/WRITE filter on a FIFO leaked through")); 6587 switch (kn->kn_filter) { 6588 case EVFILT_READ: 6589 kn->kn_fop = &vfsread_filtops; 6590 break; 6591 case EVFILT_WRITE: 6592 kn->kn_fop = &vfswrite_filtops; 6593 break; 6594 case EVFILT_VNODE: 6595 kn->kn_fop = &vfsvnode_filtops; 6596 break; 6597 default: 6598 return (EINVAL); 6599 } 6600 6601 kn->kn_hook = (caddr_t)vp; 6602 6603 v_addpollinfo(vp); 6604 if (vp->v_pollinfo == NULL) 6605 return (ENOMEM); 6606 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 6607 vhold(vp); 6608 knlist_add(knl, kn, 0); 6609 6610 return (0); 6611 } 6612 6613 /* 6614 * Detach knote from vnode 6615 */ 6616 static void 6617 filt_vfsdetach(struct knote *kn) 6618 { 6619 struct vnode *vp = (struct vnode *)kn->kn_hook; 6620 6621 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 6622 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 6623 vdrop(vp); 6624 } 6625 6626 /*ARGSUSED*/ 6627 static int 6628 filt_vfsread(struct knote *kn, long hint) 6629 { 6630 struct vnode *vp = (struct vnode *)kn->kn_hook; 6631 off_t size; 6632 int res; 6633 6634 /* 6635 * filesystem is gone, so set the EOF flag and schedule 6636 * the knote for deletion. 6637 */ 6638 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6639 VI_LOCK(vp); 6640 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6641 VI_UNLOCK(vp); 6642 return (1); 6643 } 6644 6645 if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0) 6646 return (0); 6647 6648 VI_LOCK(vp); 6649 kn->kn_data = size - kn->kn_fp->f_offset; 6650 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; 6651 VI_UNLOCK(vp); 6652 return (res); 6653 } 6654 6655 /*ARGSUSED*/ 6656 static int 6657 filt_vfswrite(struct knote *kn, long hint) 6658 { 6659 struct vnode *vp = (struct vnode *)kn->kn_hook; 6660 6661 VI_LOCK(vp); 6662 6663 /* 6664 * filesystem is gone, so set the EOF flag and schedule 6665 * the knote for deletion. 6666 */ 6667 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) 6668 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6669 6670 kn->kn_data = 0; 6671 VI_UNLOCK(vp); 6672 return (1); 6673 } 6674 6675 static int 6676 filt_vfsvnode(struct knote *kn, long hint) 6677 { 6678 struct vnode *vp = (struct vnode *)kn->kn_hook; 6679 int res; 6680 6681 VI_LOCK(vp); 6682 if (kn->kn_sfflags & hint) 6683 kn->kn_fflags |= hint; 6684 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6685 kn->kn_flags |= EV_EOF; 6686 VI_UNLOCK(vp); 6687 return (1); 6688 } 6689 res = (kn->kn_fflags != 0); 6690 VI_UNLOCK(vp); 6691 return (res); 6692 } 6693 6694 int 6695 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 6696 { 6697 int error; 6698 6699 if (dp->d_reclen > ap->a_uio->uio_resid) 6700 return (ENAMETOOLONG); 6701 error = uiomove(dp, dp->d_reclen, ap->a_uio); 6702 if (error) { 6703 if (ap->a_ncookies != NULL) { 6704 if (ap->a_cookies != NULL) 6705 free(ap->a_cookies, M_TEMP); 6706 ap->a_cookies = NULL; 6707 *ap->a_ncookies = 0; 6708 } 6709 return (error); 6710 } 6711 if (ap->a_ncookies == NULL) 6712 return (0); 6713 6714 KASSERT(ap->a_cookies, 6715 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 6716 6717 *ap->a_cookies = realloc(*ap->a_cookies, 6718 (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO); 6719 (*ap->a_cookies)[*ap->a_ncookies] = off; 6720 *ap->a_ncookies += 1; 6721 return (0); 6722 } 6723 6724 /* 6725 * The purpose of this routine is to remove granularity from accmode_t, 6726 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 6727 * VADMIN and VAPPEND. 6728 * 6729 * If it returns 0, the caller is supposed to continue with the usual 6730 * access checks using 'accmode' as modified by this routine. If it 6731 * returns nonzero value, the caller is supposed to return that value 6732 * as errno. 6733 * 6734 * Note that after this routine runs, accmode may be zero. 6735 */ 6736 int 6737 vfs_unixify_accmode(accmode_t *accmode) 6738 { 6739 /* 6740 * There is no way to specify explicit "deny" rule using 6741 * file mode or POSIX.1e ACLs. 6742 */ 6743 if (*accmode & VEXPLICIT_DENY) { 6744 *accmode = 0; 6745 return (0); 6746 } 6747 6748 /* 6749 * None of these can be translated into usual access bits. 6750 * Also, the common case for NFSv4 ACLs is to not contain 6751 * either of these bits. Caller should check for VWRITE 6752 * on the containing directory instead. 6753 */ 6754 if (*accmode & (VDELETE_CHILD | VDELETE)) 6755 return (EPERM); 6756 6757 if (*accmode & VADMIN_PERMS) { 6758 *accmode &= ~VADMIN_PERMS; 6759 *accmode |= VADMIN; 6760 } 6761 6762 /* 6763 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 6764 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 6765 */ 6766 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 6767 6768 return (0); 6769 } 6770 6771 /* 6772 * Clear out a doomed vnode (if any) and replace it with a new one as long 6773 * as the fs is not being unmounted. Return the root vnode to the caller. 6774 */ 6775 static int __noinline 6776 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp) 6777 { 6778 struct vnode *vp; 6779 int error; 6780 6781 restart: 6782 if (mp->mnt_rootvnode != NULL) { 6783 MNT_ILOCK(mp); 6784 vp = mp->mnt_rootvnode; 6785 if (vp != NULL) { 6786 if (!VN_IS_DOOMED(vp)) { 6787 vrefact(vp); 6788 MNT_IUNLOCK(mp); 6789 error = vn_lock(vp, flags); 6790 if (error == 0) { 6791 *vpp = vp; 6792 return (0); 6793 } 6794 vrele(vp); 6795 goto restart; 6796 } 6797 /* 6798 * Clear the old one. 6799 */ 6800 mp->mnt_rootvnode = NULL; 6801 } 6802 MNT_IUNLOCK(mp); 6803 if (vp != NULL) { 6804 vfs_op_barrier_wait(mp); 6805 vrele(vp); 6806 } 6807 } 6808 error = VFS_CACHEDROOT(mp, flags, vpp); 6809 if (error != 0) 6810 return (error); 6811 if (mp->mnt_vfs_ops == 0) { 6812 MNT_ILOCK(mp); 6813 if (mp->mnt_vfs_ops != 0) { 6814 MNT_IUNLOCK(mp); 6815 return (0); 6816 } 6817 if (mp->mnt_rootvnode == NULL) { 6818 vrefact(*vpp); 6819 mp->mnt_rootvnode = *vpp; 6820 } else { 6821 if (mp->mnt_rootvnode != *vpp) { 6822 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) { 6823 panic("%s: mismatch between vnode returned " 6824 " by VFS_CACHEDROOT and the one cached " 6825 " (%p != %p)", 6826 __func__, *vpp, mp->mnt_rootvnode); 6827 } 6828 } 6829 } 6830 MNT_IUNLOCK(mp); 6831 } 6832 return (0); 6833 } 6834 6835 int 6836 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp) 6837 { 6838 struct mount_pcpu *mpcpu; 6839 struct vnode *vp; 6840 int error; 6841 6842 if (!vfs_op_thread_enter(mp, mpcpu)) 6843 return (vfs_cache_root_fallback(mp, flags, vpp)); 6844 vp = atomic_load_ptr(&mp->mnt_rootvnode); 6845 if (vp == NULL || VN_IS_DOOMED(vp)) { 6846 vfs_op_thread_exit(mp, mpcpu); 6847 return (vfs_cache_root_fallback(mp, flags, vpp)); 6848 } 6849 vrefact(vp); 6850 vfs_op_thread_exit(mp, mpcpu); 6851 error = vn_lock(vp, flags); 6852 if (error != 0) { 6853 vrele(vp); 6854 return (vfs_cache_root_fallback(mp, flags, vpp)); 6855 } 6856 *vpp = vp; 6857 return (0); 6858 } 6859 6860 struct vnode * 6861 vfs_cache_root_clear(struct mount *mp) 6862 { 6863 struct vnode *vp; 6864 6865 /* 6866 * ops > 0 guarantees there is nobody who can see this vnode 6867 */ 6868 MPASS(mp->mnt_vfs_ops > 0); 6869 vp = mp->mnt_rootvnode; 6870 if (vp != NULL) 6871 vn_seqc_write_begin(vp); 6872 mp->mnt_rootvnode = NULL; 6873 return (vp); 6874 } 6875 6876 void 6877 vfs_cache_root_set(struct mount *mp, struct vnode *vp) 6878 { 6879 6880 MPASS(mp->mnt_vfs_ops > 0); 6881 vrefact(vp); 6882 mp->mnt_rootvnode = vp; 6883 } 6884 6885 /* 6886 * These are helper functions for filesystems to traverse all 6887 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 6888 * 6889 * This interface replaces MNT_VNODE_FOREACH. 6890 */ 6891 6892 struct vnode * 6893 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 6894 { 6895 struct vnode *vp; 6896 6897 maybe_yield(); 6898 MNT_ILOCK(mp); 6899 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6900 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL; 6901 vp = TAILQ_NEXT(vp, v_nmntvnodes)) { 6902 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 6903 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 6904 continue; 6905 VI_LOCK(vp); 6906 if (VN_IS_DOOMED(vp)) { 6907 VI_UNLOCK(vp); 6908 continue; 6909 } 6910 break; 6911 } 6912 if (vp == NULL) { 6913 __mnt_vnode_markerfree_all(mvp, mp); 6914 /* MNT_IUNLOCK(mp); -- done in above function */ 6915 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 6916 return (NULL); 6917 } 6918 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 6919 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 6920 MNT_IUNLOCK(mp); 6921 return (vp); 6922 } 6923 6924 struct vnode * 6925 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 6926 { 6927 struct vnode *vp; 6928 6929 *mvp = vn_alloc_marker(mp); 6930 MNT_ILOCK(mp); 6931 MNT_REF(mp); 6932 6933 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 6934 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 6935 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 6936 continue; 6937 VI_LOCK(vp); 6938 if (VN_IS_DOOMED(vp)) { 6939 VI_UNLOCK(vp); 6940 continue; 6941 } 6942 break; 6943 } 6944 if (vp == NULL) { 6945 MNT_REL(mp); 6946 MNT_IUNLOCK(mp); 6947 vn_free_marker(*mvp); 6948 *mvp = NULL; 6949 return (NULL); 6950 } 6951 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 6952 MNT_IUNLOCK(mp); 6953 return (vp); 6954 } 6955 6956 void 6957 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 6958 { 6959 6960 if (*mvp == NULL) { 6961 MNT_IUNLOCK(mp); 6962 return; 6963 } 6964 6965 mtx_assert(MNT_MTX(mp), MA_OWNED); 6966 6967 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6968 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 6969 MNT_REL(mp); 6970 MNT_IUNLOCK(mp); 6971 vn_free_marker(*mvp); 6972 *mvp = NULL; 6973 } 6974 6975 /* 6976 * These are helper functions for filesystems to traverse their 6977 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h 6978 */ 6979 static void 6980 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 6981 { 6982 6983 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6984 6985 MNT_ILOCK(mp); 6986 MNT_REL(mp); 6987 MNT_IUNLOCK(mp); 6988 vn_free_marker(*mvp); 6989 *mvp = NULL; 6990 } 6991 6992 /* 6993 * Relock the mp mount vnode list lock with the vp vnode interlock in the 6994 * conventional lock order during mnt_vnode_next_lazy iteration. 6995 * 6996 * On entry, the mount vnode list lock is held and the vnode interlock is not. 6997 * The list lock is dropped and reacquired. On success, both locks are held. 6998 * On failure, the mount vnode list lock is held but the vnode interlock is 6999 * not, and the procedure may have yielded. 7000 */ 7001 static bool 7002 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp, 7003 struct vnode *vp) 7004 { 7005 7006 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER && 7007 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp, 7008 ("%s: bad marker", __func__)); 7009 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp, 7010 ("%s: inappropriate vnode", __func__)); 7011 ASSERT_VI_UNLOCKED(vp, __func__); 7012 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 7013 7014 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist); 7015 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist); 7016 7017 /* 7018 * Note we may be racing against vdrop which transitioned the hold 7019 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine, 7020 * if we are the only user after we get the interlock we will just 7021 * vdrop. 7022 */ 7023 vhold(vp); 7024 mtx_unlock(&mp->mnt_listmtx); 7025 VI_LOCK(vp); 7026 if (VN_IS_DOOMED(vp)) { 7027 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); 7028 goto out_lost; 7029 } 7030 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 7031 /* 7032 * There is nothing to do if we are the last user. 7033 */ 7034 if (!refcount_release_if_not_last(&vp->v_holdcnt)) 7035 goto out_lost; 7036 mtx_lock(&mp->mnt_listmtx); 7037 return (true); 7038 out_lost: 7039 vdropl(vp); 7040 maybe_yield(); 7041 mtx_lock(&mp->mnt_listmtx); 7042 return (false); 7043 } 7044 7045 static struct vnode * 7046 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 7047 void *cbarg) 7048 { 7049 struct vnode *vp; 7050 7051 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 7052 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 7053 restart: 7054 vp = TAILQ_NEXT(*mvp, v_lazylist); 7055 while (vp != NULL) { 7056 if (vp->v_type == VMARKER) { 7057 vp = TAILQ_NEXT(vp, v_lazylist); 7058 continue; 7059 } 7060 /* 7061 * See if we want to process the vnode. Note we may encounter a 7062 * long string of vnodes we don't care about and hog the list 7063 * as a result. Check for it and requeue the marker. 7064 */ 7065 VNPASS(!VN_IS_DOOMED(vp), vp); 7066 if (!cb(vp, cbarg)) { 7067 if (!should_yield()) { 7068 vp = TAILQ_NEXT(vp, v_lazylist); 7069 continue; 7070 } 7071 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, 7072 v_lazylist); 7073 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, 7074 v_lazylist); 7075 mtx_unlock(&mp->mnt_listmtx); 7076 kern_yield(PRI_USER); 7077 mtx_lock(&mp->mnt_listmtx); 7078 goto restart; 7079 } 7080 /* 7081 * Try-lock because this is the wrong lock order. 7082 */ 7083 if (!VI_TRYLOCK(vp) && 7084 !mnt_vnode_next_lazy_relock(*mvp, mp, vp)) 7085 goto restart; 7086 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 7087 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 7088 ("alien vnode on the lazy list %p %p", vp, mp)); 7089 VNPASS(vp->v_mount == mp, vp); 7090 VNPASS(!VN_IS_DOOMED(vp), vp); 7091 break; 7092 } 7093 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 7094 7095 /* Check if we are done */ 7096 if (vp == NULL) { 7097 mtx_unlock(&mp->mnt_listmtx); 7098 mnt_vnode_markerfree_lazy(mvp, mp); 7099 return (NULL); 7100 } 7101 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist); 7102 mtx_unlock(&mp->mnt_listmtx); 7103 ASSERT_VI_LOCKED(vp, "lazy iter"); 7104 return (vp); 7105 } 7106 7107 struct vnode * 7108 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 7109 void *cbarg) 7110 { 7111 7112 maybe_yield(); 7113 mtx_lock(&mp->mnt_listmtx); 7114 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 7115 } 7116 7117 struct vnode * 7118 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 7119 void *cbarg) 7120 { 7121 struct vnode *vp; 7122 7123 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist)) 7124 return (NULL); 7125 7126 *mvp = vn_alloc_marker(mp); 7127 MNT_ILOCK(mp); 7128 MNT_REF(mp); 7129 MNT_IUNLOCK(mp); 7130 7131 mtx_lock(&mp->mnt_listmtx); 7132 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist); 7133 if (vp == NULL) { 7134 mtx_unlock(&mp->mnt_listmtx); 7135 mnt_vnode_markerfree_lazy(mvp, mp); 7136 return (NULL); 7137 } 7138 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist); 7139 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 7140 } 7141 7142 void 7143 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 7144 { 7145 7146 if (*mvp == NULL) 7147 return; 7148 7149 mtx_lock(&mp->mnt_listmtx); 7150 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 7151 mtx_unlock(&mp->mnt_listmtx); 7152 mnt_vnode_markerfree_lazy(mvp, mp); 7153 } 7154 7155 int 7156 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp) 7157 { 7158 7159 if ((cnp->cn_flags & NOEXECCHECK) != 0) { 7160 cnp->cn_flags &= ~NOEXECCHECK; 7161 return (0); 7162 } 7163 7164 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread)); 7165 } 7166 7167 /* 7168 * Do not use this variant unless you have means other than the hold count 7169 * to prevent the vnode from getting freed. 7170 */ 7171 void 7172 vn_seqc_write_begin_locked(struct vnode *vp) 7173 { 7174 7175 ASSERT_VI_LOCKED(vp, __func__); 7176 VNPASS(vp->v_holdcnt > 0, vp); 7177 VNPASS(vp->v_seqc_users >= 0, vp); 7178 vp->v_seqc_users++; 7179 if (vp->v_seqc_users == 1) 7180 seqc_sleepable_write_begin(&vp->v_seqc); 7181 } 7182 7183 void 7184 vn_seqc_write_begin(struct vnode *vp) 7185 { 7186 7187 VI_LOCK(vp); 7188 vn_seqc_write_begin_locked(vp); 7189 VI_UNLOCK(vp); 7190 } 7191 7192 void 7193 vn_seqc_write_end_locked(struct vnode *vp) 7194 { 7195 7196 ASSERT_VI_LOCKED(vp, __func__); 7197 VNPASS(vp->v_seqc_users > 0, vp); 7198 vp->v_seqc_users--; 7199 if (vp->v_seqc_users == 0) 7200 seqc_sleepable_write_end(&vp->v_seqc); 7201 } 7202 7203 void 7204 vn_seqc_write_end(struct vnode *vp) 7205 { 7206 7207 VI_LOCK(vp); 7208 vn_seqc_write_end_locked(vp); 7209 VI_UNLOCK(vp); 7210 } 7211 7212 /* 7213 * Special case handling for allocating and freeing vnodes. 7214 * 7215 * The counter remains unchanged on free so that a doomed vnode will 7216 * keep testing as in modify as long as it is accessible with SMR. 7217 */ 7218 static void 7219 vn_seqc_init(struct vnode *vp) 7220 { 7221 7222 vp->v_seqc = 0; 7223 vp->v_seqc_users = 0; 7224 } 7225 7226 static void 7227 vn_seqc_write_end_free(struct vnode *vp) 7228 { 7229 7230 VNPASS(seqc_in_modify(vp->v_seqc), vp); 7231 VNPASS(vp->v_seqc_users == 1, vp); 7232 } 7233 7234 void 7235 vn_irflag_set_locked(struct vnode *vp, short toset) 7236 { 7237 short flags; 7238 7239 ASSERT_VI_LOCKED(vp, __func__); 7240 flags = vn_irflag_read(vp); 7241 VNASSERT((flags & toset) == 0, vp, 7242 ("%s: some of the passed flags already set (have %d, passed %d)\n", 7243 __func__, flags, toset)); 7244 atomic_store_short(&vp->v_irflag, flags | toset); 7245 } 7246 7247 void 7248 vn_irflag_set(struct vnode *vp, short toset) 7249 { 7250 7251 VI_LOCK(vp); 7252 vn_irflag_set_locked(vp, toset); 7253 VI_UNLOCK(vp); 7254 } 7255 7256 void 7257 vn_irflag_set_cond_locked(struct vnode *vp, short toset) 7258 { 7259 short flags; 7260 7261 ASSERT_VI_LOCKED(vp, __func__); 7262 flags = vn_irflag_read(vp); 7263 atomic_store_short(&vp->v_irflag, flags | toset); 7264 } 7265 7266 void 7267 vn_irflag_set_cond(struct vnode *vp, short toset) 7268 { 7269 7270 VI_LOCK(vp); 7271 vn_irflag_set_cond_locked(vp, toset); 7272 VI_UNLOCK(vp); 7273 } 7274 7275 void 7276 vn_irflag_unset_locked(struct vnode *vp, short tounset) 7277 { 7278 short flags; 7279 7280 ASSERT_VI_LOCKED(vp, __func__); 7281 flags = vn_irflag_read(vp); 7282 VNASSERT((flags & tounset) == tounset, vp, 7283 ("%s: some of the passed flags not set (have %d, passed %d)\n", 7284 __func__, flags, tounset)); 7285 atomic_store_short(&vp->v_irflag, flags & ~tounset); 7286 } 7287 7288 void 7289 vn_irflag_unset(struct vnode *vp, short tounset) 7290 { 7291 7292 VI_LOCK(vp); 7293 vn_irflag_unset_locked(vp, tounset); 7294 VI_UNLOCK(vp); 7295 } 7296 7297 int 7298 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred) 7299 { 7300 struct vattr vattr; 7301 int error; 7302 7303 ASSERT_VOP_LOCKED(vp, __func__); 7304 error = VOP_GETATTR(vp, &vattr, cred); 7305 if (__predict_true(error == 0)) { 7306 if (vattr.va_size <= OFF_MAX) 7307 *size = vattr.va_size; 7308 else 7309 error = EFBIG; 7310 } 7311 return (error); 7312 } 7313 7314 int 7315 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred) 7316 { 7317 int error; 7318 7319 VOP_LOCK(vp, LK_SHARED); 7320 error = vn_getsize_locked(vp, size, cred); 7321 VOP_UNLOCK(vp); 7322 return (error); 7323 } 7324 7325 #ifdef INVARIANTS 7326 void 7327 vn_set_state_validate(struct vnode *vp, __enum_uint8(vstate) state) 7328 { 7329 7330 switch (vp->v_state) { 7331 case VSTATE_UNINITIALIZED: 7332 switch (state) { 7333 case VSTATE_CONSTRUCTED: 7334 case VSTATE_DESTROYING: 7335 return; 7336 default: 7337 break; 7338 } 7339 break; 7340 case VSTATE_CONSTRUCTED: 7341 ASSERT_VOP_ELOCKED(vp, __func__); 7342 switch (state) { 7343 case VSTATE_DESTROYING: 7344 return; 7345 default: 7346 break; 7347 } 7348 break; 7349 case VSTATE_DESTROYING: 7350 ASSERT_VOP_ELOCKED(vp, __func__); 7351 switch (state) { 7352 case VSTATE_DEAD: 7353 return; 7354 default: 7355 break; 7356 } 7357 break; 7358 case VSTATE_DEAD: 7359 switch (state) { 7360 case VSTATE_UNINITIALIZED: 7361 return; 7362 default: 7363 break; 7364 } 7365 break; 7366 } 7367 7368 vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state); 7369 panic("invalid state transition %d -> %d\n", vp->v_state, state); 7370 } 7371 #endif 7372