1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1989, 1993 5 * The Regents of the University of California. All rights reserved. 6 * (c) UNIX System Laboratories, Inc. 7 * All or some portions of this file are derived from material licensed 8 * to the University of California by American Telephone and Telegraph 9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 10 * the permission of UNIX System Laboratories, Inc. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 37 */ 38 39 /* 40 * External virtual filesystem routines 41 */ 42 43 #include <sys/cdefs.h> 44 __FBSDID("$FreeBSD$"); 45 46 #include "opt_ddb.h" 47 #include "opt_watchdog.h" 48 49 #include <sys/param.h> 50 #include <sys/systm.h> 51 #include <sys/asan.h> 52 #include <sys/bio.h> 53 #include <sys/buf.h> 54 #include <sys/capsicum.h> 55 #include <sys/condvar.h> 56 #include <sys/conf.h> 57 #include <sys/counter.h> 58 #include <sys/dirent.h> 59 #include <sys/event.h> 60 #include <sys/eventhandler.h> 61 #include <sys/extattr.h> 62 #include <sys/file.h> 63 #include <sys/fcntl.h> 64 #include <sys/jail.h> 65 #include <sys/kdb.h> 66 #include <sys/kernel.h> 67 #include <sys/kthread.h> 68 #include <sys/ktr.h> 69 #include <sys/limits.h> 70 #include <sys/lockf.h> 71 #include <sys/malloc.h> 72 #include <sys/mount.h> 73 #include <sys/namei.h> 74 #include <sys/pctrie.h> 75 #include <sys/priv.h> 76 #include <sys/reboot.h> 77 #include <sys/refcount.h> 78 #include <sys/rwlock.h> 79 #include <sys/sched.h> 80 #include <sys/sleepqueue.h> 81 #include <sys/smr.h> 82 #include <sys/smp.h> 83 #include <sys/stat.h> 84 #include <sys/sysctl.h> 85 #include <sys/syslog.h> 86 #include <sys/vmmeter.h> 87 #include <sys/vnode.h> 88 #include <sys/watchdog.h> 89 90 #include <machine/stdarg.h> 91 92 #include <security/mac/mac_framework.h> 93 94 #include <vm/vm.h> 95 #include <vm/vm_object.h> 96 #include <vm/vm_extern.h> 97 #include <vm/pmap.h> 98 #include <vm/vm_map.h> 99 #include <vm/vm_page.h> 100 #include <vm/vm_kern.h> 101 #include <vm/uma.h> 102 103 #if defined(DEBUG_VFS_LOCKS) && (!defined(INVARIANTS) || !defined(WITNESS)) 104 #error DEBUG_VFS_LOCKS requires INVARIANTS and WITNESS 105 #endif 106 107 #ifdef DDB 108 #include <ddb/ddb.h> 109 #endif 110 111 static void delmntque(struct vnode *vp); 112 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, 113 int slpflag, int slptimeo); 114 static void syncer_shutdown(void *arg, int howto); 115 static int vtryrecycle(struct vnode *vp); 116 static void v_init_counters(struct vnode *); 117 static void vn_seqc_init(struct vnode *); 118 static void vn_seqc_write_end_free(struct vnode *vp); 119 static void vgonel(struct vnode *); 120 static bool vhold_recycle_free(struct vnode *); 121 static void vdropl_recycle(struct vnode *vp); 122 static void vdrop_recycle(struct vnode *vp); 123 static void vfs_knllock(void *arg); 124 static void vfs_knlunlock(void *arg); 125 static void vfs_knl_assert_lock(void *arg, int what); 126 static void destroy_vpollinfo(struct vpollinfo *vi); 127 static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, 128 daddr_t startlbn, daddr_t endlbn); 129 static void vnlru_recalc(void); 130 131 /* 132 * Number of vnodes in existence. Increased whenever getnewvnode() 133 * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode. 134 */ 135 static u_long __exclusive_cache_line numvnodes; 136 137 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, 138 "Number of vnodes in existence"); 139 140 static counter_u64_t vnodes_created; 141 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created, 142 "Number of vnodes created by getnewvnode"); 143 144 /* 145 * Conversion tables for conversion from vnode types to inode formats 146 * and back. 147 */ 148 enum vtype iftovt_tab[16] = { 149 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 150 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON 151 }; 152 int vttoif_tab[10] = { 153 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 154 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT 155 }; 156 157 /* 158 * List of allocates vnodes in the system. 159 */ 160 static TAILQ_HEAD(freelst, vnode) vnode_list; 161 static struct vnode *vnode_list_free_marker; 162 static struct vnode *vnode_list_reclaim_marker; 163 164 /* 165 * "Free" vnode target. Free vnodes are rarely completely free, but are 166 * just ones that are cheap to recycle. Usually they are for files which 167 * have been stat'd but not read; these usually have inode and namecache 168 * data attached to them. This target is the preferred minimum size of a 169 * sub-cache consisting mostly of such files. The system balances the size 170 * of this sub-cache with its complement to try to prevent either from 171 * thrashing while the other is relatively inactive. The targets express 172 * a preference for the best balance. 173 * 174 * "Above" this target there are 2 further targets (watermarks) related 175 * to recyling of free vnodes. In the best-operating case, the cache is 176 * exactly full, the free list has size between vlowat and vhiwat above the 177 * free target, and recycling from it and normal use maintains this state. 178 * Sometimes the free list is below vlowat or even empty, but this state 179 * is even better for immediate use provided the cache is not full. 180 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free 181 * ones) to reach one of these states. The watermarks are currently hard- 182 * coded as 4% and 9% of the available space higher. These and the default 183 * of 25% for wantfreevnodes are too large if the memory size is large. 184 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim 185 * whenever vnlru_proc() becomes active. 186 */ 187 static long wantfreevnodes; 188 static long __exclusive_cache_line freevnodes; 189 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, 190 &freevnodes, 0, "Number of \"free\" vnodes"); 191 static long freevnodes_old; 192 193 static counter_u64_t recycles_count; 194 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count, 195 "Number of vnodes recycled to meet vnode cache targets"); 196 197 static counter_u64_t recycles_free_count; 198 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count, 199 "Number of free vnodes recycled to meet vnode cache targets"); 200 201 static u_long deferred_inact; 202 SYSCTL_ULONG(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, 203 &deferred_inact, 0, "Number of times inactive processing was deferred"); 204 205 /* To keep more than one thread at a time from running vfs_getnewfsid */ 206 static struct mtx mntid_mtx; 207 208 /* 209 * Lock for any access to the following: 210 * vnode_list 211 * numvnodes 212 * freevnodes 213 */ 214 static struct mtx __exclusive_cache_line vnode_list_mtx; 215 216 /* Publicly exported FS */ 217 struct nfs_public nfs_pub; 218 219 static uma_zone_t buf_trie_zone; 220 static smr_t buf_trie_smr; 221 222 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 223 static uma_zone_t vnode_zone; 224 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll"); 225 226 __read_frequently smr_t vfs_smr; 227 228 /* 229 * The workitem queue. 230 * 231 * It is useful to delay writes of file data and filesystem metadata 232 * for tens of seconds so that quickly created and deleted files need 233 * not waste disk bandwidth being created and removed. To realize this, 234 * we append vnodes to a "workitem" queue. When running with a soft 235 * updates implementation, most pending metadata dependencies should 236 * not wait for more than a few seconds. Thus, mounted on block devices 237 * are delayed only about a half the time that file data is delayed. 238 * Similarly, directory updates are more critical, so are only delayed 239 * about a third the time that file data is delayed. Thus, there are 240 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 241 * one each second (driven off the filesystem syncer process). The 242 * syncer_delayno variable indicates the next queue that is to be processed. 243 * Items that need to be processed soon are placed in this queue: 244 * 245 * syncer_workitem_pending[syncer_delayno] 246 * 247 * A delay of fifteen seconds is done by placing the request fifteen 248 * entries later in the queue: 249 * 250 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 251 * 252 */ 253 static int syncer_delayno; 254 static long syncer_mask; 255 LIST_HEAD(synclist, bufobj); 256 static struct synclist *syncer_workitem_pending; 257 /* 258 * The sync_mtx protects: 259 * bo->bo_synclist 260 * sync_vnode_count 261 * syncer_delayno 262 * syncer_state 263 * syncer_workitem_pending 264 * syncer_worklist_len 265 * rushjob 266 */ 267 static struct mtx sync_mtx; 268 static struct cv sync_wakeup; 269 270 #define SYNCER_MAXDELAY 32 271 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 272 static int syncdelay = 30; /* max time to delay syncing data */ 273 static int filedelay = 30; /* time to delay syncing files */ 274 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, 275 "Time to delay syncing files (in seconds)"); 276 static int dirdelay = 29; /* time to delay syncing directories */ 277 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, 278 "Time to delay syncing directories (in seconds)"); 279 static int metadelay = 28; /* time to delay syncing metadata */ 280 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, 281 "Time to delay syncing metadata (in seconds)"); 282 static int rushjob; /* number of slots to run ASAP */ 283 static int stat_rush_requests; /* number of times I/O speeded up */ 284 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, 285 "Number of times I/O speeded up (rush requests)"); 286 287 #define VDBATCH_SIZE 8 288 struct vdbatch { 289 u_int index; 290 struct mtx lock; 291 struct vnode *tab[VDBATCH_SIZE]; 292 }; 293 DPCPU_DEFINE_STATIC(struct vdbatch, vd); 294 295 static void vdbatch_dequeue(struct vnode *vp); 296 297 /* 298 * When shutting down the syncer, run it at four times normal speed. 299 */ 300 #define SYNCER_SHUTDOWN_SPEEDUP 4 301 static int sync_vnode_count; 302 static int syncer_worklist_len; 303 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 304 syncer_state; 305 306 /* Target for maximum number of vnodes. */ 307 u_long desiredvnodes; 308 static u_long gapvnodes; /* gap between wanted and desired */ 309 static u_long vhiwat; /* enough extras after expansion */ 310 static u_long vlowat; /* minimal extras before expansion */ 311 static u_long vstir; /* nonzero to stir non-free vnodes */ 312 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */ 313 314 static u_long vnlru_read_freevnodes(void); 315 316 /* 317 * Note that no attempt is made to sanitize these parameters. 318 */ 319 static int 320 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS) 321 { 322 u_long val; 323 int error; 324 325 val = desiredvnodes; 326 error = sysctl_handle_long(oidp, &val, 0, req); 327 if (error != 0 || req->newptr == NULL) 328 return (error); 329 330 if (val == desiredvnodes) 331 return (0); 332 mtx_lock(&vnode_list_mtx); 333 desiredvnodes = val; 334 wantfreevnodes = desiredvnodes / 4; 335 vnlru_recalc(); 336 mtx_unlock(&vnode_list_mtx); 337 /* 338 * XXX There is no protection against multiple threads changing 339 * desiredvnodes at the same time. Locking above only helps vnlru and 340 * getnewvnode. 341 */ 342 vfs_hash_changesize(desiredvnodes); 343 cache_changesize(desiredvnodes); 344 return (0); 345 } 346 347 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes, 348 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes, 349 "LU", "Target for maximum number of vnodes"); 350 351 static int 352 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS) 353 { 354 u_long val; 355 int error; 356 357 val = wantfreevnodes; 358 error = sysctl_handle_long(oidp, &val, 0, req); 359 if (error != 0 || req->newptr == NULL) 360 return (error); 361 362 if (val == wantfreevnodes) 363 return (0); 364 mtx_lock(&vnode_list_mtx); 365 wantfreevnodes = val; 366 vnlru_recalc(); 367 mtx_unlock(&vnode_list_mtx); 368 return (0); 369 } 370 371 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes, 372 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes, 373 "LU", "Target for minimum number of \"free\" vnodes"); 374 375 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 376 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)"); 377 static int vnlru_nowhere; 378 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW | CTLFLAG_STATS, 379 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 380 381 static int 382 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS) 383 { 384 struct vnode *vp; 385 struct nameidata nd; 386 char *buf; 387 unsigned long ndflags; 388 int error; 389 390 if (req->newptr == NULL) 391 return (EINVAL); 392 if (req->newlen >= PATH_MAX) 393 return (E2BIG); 394 395 buf = malloc(PATH_MAX, M_TEMP, M_WAITOK); 396 error = SYSCTL_IN(req, buf, req->newlen); 397 if (error != 0) 398 goto out; 399 400 buf[req->newlen] = '\0'; 401 402 ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1; 403 NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf); 404 if ((error = namei(&nd)) != 0) 405 goto out; 406 vp = nd.ni_vp; 407 408 if (VN_IS_DOOMED(vp)) { 409 /* 410 * This vnode is being recycled. Return != 0 to let the caller 411 * know that the sysctl had no effect. Return EAGAIN because a 412 * subsequent call will likely succeed (since namei will create 413 * a new vnode if necessary) 414 */ 415 error = EAGAIN; 416 goto putvnode; 417 } 418 419 counter_u64_add(recycles_count, 1); 420 vgone(vp); 421 putvnode: 422 vput(vp); 423 NDFREE_PNBUF(&nd); 424 out: 425 free(buf, M_TEMP); 426 return (error); 427 } 428 429 static int 430 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS) 431 { 432 struct thread *td = curthread; 433 struct vnode *vp; 434 struct file *fp; 435 int error; 436 int fd; 437 438 if (req->newptr == NULL) 439 return (EBADF); 440 441 error = sysctl_handle_int(oidp, &fd, 0, req); 442 if (error != 0) 443 return (error); 444 error = getvnode(curthread, fd, &cap_fcntl_rights, &fp); 445 if (error != 0) 446 return (error); 447 vp = fp->f_vnode; 448 449 error = vn_lock(vp, LK_EXCLUSIVE); 450 if (error != 0) 451 goto drop; 452 453 counter_u64_add(recycles_count, 1); 454 vgone(vp); 455 VOP_UNLOCK(vp); 456 drop: 457 fdrop(fp, td); 458 return (error); 459 } 460 461 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode, 462 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, 463 sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname"); 464 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode, 465 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, 466 sysctl_ftry_reclaim_vnode, "I", 467 "Try to reclaim a vnode by its file descriptor"); 468 469 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */ 470 #define vnsz2log 8 471 #ifndef DEBUG_LOCKS 472 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log && 473 sizeof(struct vnode) < 1UL << (vnsz2log + 1), 474 "vnsz2log needs to be updated"); 475 #endif 476 477 /* 478 * Support for the bufobj clean & dirty pctrie. 479 */ 480 static void * 481 buf_trie_alloc(struct pctrie *ptree) 482 { 483 return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT)); 484 } 485 486 static void 487 buf_trie_free(struct pctrie *ptree, void *node) 488 { 489 uma_zfree_smr(buf_trie_zone, node); 490 } 491 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free, 492 buf_trie_smr); 493 494 /* 495 * Initialize the vnode management data structures. 496 * 497 * Reevaluate the following cap on the number of vnodes after the physical 498 * memory size exceeds 512GB. In the limit, as the physical memory size 499 * grows, the ratio of the memory size in KB to vnodes approaches 64:1. 500 */ 501 #ifndef MAXVNODES_MAX 502 #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */ 503 #endif 504 505 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker"); 506 507 static struct vnode * 508 vn_alloc_marker(struct mount *mp) 509 { 510 struct vnode *vp; 511 512 vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 513 vp->v_type = VMARKER; 514 vp->v_mount = mp; 515 516 return (vp); 517 } 518 519 static void 520 vn_free_marker(struct vnode *vp) 521 { 522 523 MPASS(vp->v_type == VMARKER); 524 free(vp, M_VNODE_MARKER); 525 } 526 527 #ifdef KASAN 528 static int 529 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused) 530 { 531 kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0); 532 return (0); 533 } 534 535 static void 536 vnode_dtor(void *mem, int size, void *arg __unused) 537 { 538 size_t end1, end2, off1, off2; 539 540 _Static_assert(offsetof(struct vnode, v_vnodelist) < 541 offsetof(struct vnode, v_dbatchcpu), 542 "KASAN marks require updating"); 543 544 off1 = offsetof(struct vnode, v_vnodelist); 545 off2 = offsetof(struct vnode, v_dbatchcpu); 546 end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist); 547 end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu); 548 549 /* 550 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even 551 * after the vnode has been freed. Try to get some KASAN coverage by 552 * marking everything except those two fields as invalid. Because 553 * KASAN's tracking is not byte-granular, any preceding fields sharing 554 * the same 8-byte aligned word must also be marked valid. 555 */ 556 557 /* Handle the area from the start until v_vnodelist... */ 558 off1 = rounddown2(off1, KASAN_SHADOW_SCALE); 559 kasan_mark(mem, off1, off1, KASAN_UMA_FREED); 560 561 /* ... then the area between v_vnodelist and v_dbatchcpu ... */ 562 off1 = roundup2(end1, KASAN_SHADOW_SCALE); 563 off2 = rounddown2(off2, KASAN_SHADOW_SCALE); 564 if (off2 > off1) 565 kasan_mark((void *)((char *)mem + off1), off2 - off1, 566 off2 - off1, KASAN_UMA_FREED); 567 568 /* ... and finally the area from v_dbatchcpu to the end. */ 569 off2 = roundup2(end2, KASAN_SHADOW_SCALE); 570 kasan_mark((void *)((char *)mem + off2), size - off2, size - off2, 571 KASAN_UMA_FREED); 572 } 573 #endif /* KASAN */ 574 575 /* 576 * Initialize a vnode as it first enters the zone. 577 */ 578 static int 579 vnode_init(void *mem, int size, int flags) 580 { 581 struct vnode *vp; 582 583 vp = mem; 584 bzero(vp, size); 585 /* 586 * Setup locks. 587 */ 588 vp->v_vnlock = &vp->v_lock; 589 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 590 /* 591 * By default, don't allow shared locks unless filesystems opt-in. 592 */ 593 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT, 594 LK_NOSHARE | LK_IS_VNODE); 595 /* 596 * Initialize bufobj. 597 */ 598 bufobj_init(&vp->v_bufobj, vp); 599 /* 600 * Initialize namecache. 601 */ 602 cache_vnode_init(vp); 603 /* 604 * Initialize rangelocks. 605 */ 606 rangelock_init(&vp->v_rl); 607 608 vp->v_dbatchcpu = NOCPU; 609 610 vp->v_state = VSTATE_DEAD; 611 612 /* 613 * Check vhold_recycle_free for an explanation. 614 */ 615 vp->v_holdcnt = VHOLD_NO_SMR; 616 vp->v_type = VNON; 617 mtx_lock(&vnode_list_mtx); 618 TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist); 619 mtx_unlock(&vnode_list_mtx); 620 return (0); 621 } 622 623 /* 624 * Free a vnode when it is cleared from the zone. 625 */ 626 static void 627 vnode_fini(void *mem, int size) 628 { 629 struct vnode *vp; 630 struct bufobj *bo; 631 632 vp = mem; 633 vdbatch_dequeue(vp); 634 mtx_lock(&vnode_list_mtx); 635 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 636 mtx_unlock(&vnode_list_mtx); 637 rangelock_destroy(&vp->v_rl); 638 lockdestroy(vp->v_vnlock); 639 mtx_destroy(&vp->v_interlock); 640 bo = &vp->v_bufobj; 641 rw_destroy(BO_LOCKPTR(bo)); 642 643 kasan_mark(mem, size, size, 0); 644 } 645 646 /* 647 * Provide the size of NFS nclnode and NFS fh for calculation of the 648 * vnode memory consumption. The size is specified directly to 649 * eliminate dependency on NFS-private header. 650 * 651 * Other filesystems may use bigger or smaller (like UFS and ZFS) 652 * private inode data, but the NFS-based estimation is ample enough. 653 * Still, we care about differences in the size between 64- and 32-bit 654 * platforms. 655 * 656 * Namecache structure size is heuristically 657 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1. 658 */ 659 #ifdef _LP64 660 #define NFS_NCLNODE_SZ (528 + 64) 661 #define NC_SZ 148 662 #else 663 #define NFS_NCLNODE_SZ (360 + 32) 664 #define NC_SZ 92 665 #endif 666 667 static void 668 vntblinit(void *dummy __unused) 669 { 670 struct vdbatch *vd; 671 uma_ctor ctor; 672 uma_dtor dtor; 673 int cpu, physvnodes, virtvnodes; 674 675 /* 676 * Desiredvnodes is a function of the physical memory size and the 677 * kernel's heap size. Generally speaking, it scales with the 678 * physical memory size. The ratio of desiredvnodes to the physical 679 * memory size is 1:16 until desiredvnodes exceeds 98,304. 680 * Thereafter, the 681 * marginal ratio of desiredvnodes to the physical memory size is 682 * 1:64. However, desiredvnodes is limited by the kernel's heap 683 * size. The memory required by desiredvnodes vnodes and vm objects 684 * must not exceed 1/10th of the kernel's heap size. 685 */ 686 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 + 687 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64; 688 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) + 689 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ)); 690 desiredvnodes = min(physvnodes, virtvnodes); 691 if (desiredvnodes > MAXVNODES_MAX) { 692 if (bootverbose) 693 printf("Reducing kern.maxvnodes %lu -> %lu\n", 694 desiredvnodes, MAXVNODES_MAX); 695 desiredvnodes = MAXVNODES_MAX; 696 } 697 wantfreevnodes = desiredvnodes / 4; 698 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 699 TAILQ_INIT(&vnode_list); 700 mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF); 701 /* 702 * The lock is taken to appease WITNESS. 703 */ 704 mtx_lock(&vnode_list_mtx); 705 vnlru_recalc(); 706 mtx_unlock(&vnode_list_mtx); 707 vnode_list_free_marker = vn_alloc_marker(NULL); 708 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist); 709 vnode_list_reclaim_marker = vn_alloc_marker(NULL); 710 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist); 711 712 #ifdef KASAN 713 ctor = vnode_ctor; 714 dtor = vnode_dtor; 715 #else 716 ctor = NULL; 717 dtor = NULL; 718 #endif 719 vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor, 720 vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN); 721 uma_zone_set_smr(vnode_zone, vfs_smr); 722 723 /* 724 * Preallocate enough nodes to support one-per buf so that 725 * we can not fail an insert. reassignbuf() callers can not 726 * tolerate the insertion failure. 727 */ 728 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(), 729 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, 730 UMA_ZONE_NOFREE | UMA_ZONE_SMR); 731 buf_trie_smr = uma_zone_get_smr(buf_trie_zone); 732 uma_prealloc(buf_trie_zone, nbuf); 733 734 vnodes_created = counter_u64_alloc(M_WAITOK); 735 recycles_count = counter_u64_alloc(M_WAITOK); 736 recycles_free_count = counter_u64_alloc(M_WAITOK); 737 738 /* 739 * Initialize the filesystem syncer. 740 */ 741 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 742 &syncer_mask); 743 syncer_maxdelay = syncer_mask + 1; 744 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 745 cv_init(&sync_wakeup, "syncer"); 746 747 CPU_FOREACH(cpu) { 748 vd = DPCPU_ID_PTR((cpu), vd); 749 bzero(vd, sizeof(*vd)); 750 mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF); 751 } 752 } 753 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL); 754 755 /* 756 * Mark a mount point as busy. Used to synchronize access and to delay 757 * unmounting. Eventually, mountlist_mtx is not released on failure. 758 * 759 * vfs_busy() is a custom lock, it can block the caller. 760 * vfs_busy() only sleeps if the unmount is active on the mount point. 761 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any 762 * vnode belonging to mp. 763 * 764 * Lookup uses vfs_busy() to traverse mount points. 765 * root fs var fs 766 * / vnode lock A / vnode lock (/var) D 767 * /var vnode lock B /log vnode lock(/var/log) E 768 * vfs_busy lock C vfs_busy lock F 769 * 770 * Within each file system, the lock order is C->A->B and F->D->E. 771 * 772 * When traversing across mounts, the system follows that lock order: 773 * 774 * C->A->B 775 * | 776 * +->F->D->E 777 * 778 * The lookup() process for namei("/var") illustrates the process: 779 * 1. VOP_LOOKUP() obtains B while A is held 780 * 2. vfs_busy() obtains a shared lock on F while A and B are held 781 * 3. vput() releases lock on B 782 * 4. vput() releases lock on A 783 * 5. VFS_ROOT() obtains lock on D while shared lock on F is held 784 * 6. vfs_unbusy() releases shared lock on F 785 * 7. vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A. 786 * Attempt to lock A (instead of vp_crossmp) while D is held would 787 * violate the global order, causing deadlocks. 788 * 789 * dounmount() locks B while F is drained. Note that for stacked 790 * filesystems, D and B in the example above may be the same lock, 791 * which introdues potential lock order reversal deadlock between 792 * dounmount() and step 5 above. These filesystems may avoid the LOR 793 * by setting VV_CROSSLOCK on the covered vnode so that lock B will 794 * remain held until after step 5. 795 */ 796 int 797 vfs_busy(struct mount *mp, int flags) 798 { 799 struct mount_pcpu *mpcpu; 800 801 MPASS((flags & ~MBF_MASK) == 0); 802 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags); 803 804 if (vfs_op_thread_enter(mp, mpcpu)) { 805 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 806 MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0); 807 MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0); 808 vfs_mp_count_add_pcpu(mpcpu, ref, 1); 809 vfs_mp_count_add_pcpu(mpcpu, lockref, 1); 810 vfs_op_thread_exit(mp, mpcpu); 811 if (flags & MBF_MNTLSTLOCK) 812 mtx_unlock(&mountlist_mtx); 813 return (0); 814 } 815 816 MNT_ILOCK(mp); 817 vfs_assert_mount_counters(mp); 818 MNT_REF(mp); 819 /* 820 * If mount point is currently being unmounted, sleep until the 821 * mount point fate is decided. If thread doing the unmounting fails, 822 * it will clear MNTK_UNMOUNT flag before waking us up, indicating 823 * that this mount point has survived the unmount attempt and vfs_busy 824 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE 825 * flag in addition to MNTK_UNMOUNT, indicating that mount point is 826 * about to be really destroyed. vfs_busy needs to release its 827 * reference on the mount point in this case and return with ENOENT, 828 * telling the caller the mount it tried to busy is no longer valid. 829 */ 830 while (mp->mnt_kern_flag & MNTK_UNMOUNT) { 831 KASSERT(TAILQ_EMPTY(&mp->mnt_uppers), 832 ("%s: non-empty upper mount list with pending unmount", 833 __func__)); 834 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) { 835 MNT_REL(mp); 836 MNT_IUNLOCK(mp); 837 CTR1(KTR_VFS, "%s: failed busying before sleeping", 838 __func__); 839 return (ENOENT); 840 } 841 if (flags & MBF_MNTLSTLOCK) 842 mtx_unlock(&mountlist_mtx); 843 mp->mnt_kern_flag |= MNTK_MWAIT; 844 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0); 845 if (flags & MBF_MNTLSTLOCK) 846 mtx_lock(&mountlist_mtx); 847 MNT_ILOCK(mp); 848 } 849 if (flags & MBF_MNTLSTLOCK) 850 mtx_unlock(&mountlist_mtx); 851 mp->mnt_lockref++; 852 MNT_IUNLOCK(mp); 853 return (0); 854 } 855 856 /* 857 * Free a busy filesystem. 858 */ 859 void 860 vfs_unbusy(struct mount *mp) 861 { 862 struct mount_pcpu *mpcpu; 863 int c; 864 865 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 866 867 if (vfs_op_thread_enter(mp, mpcpu)) { 868 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 869 vfs_mp_count_sub_pcpu(mpcpu, lockref, 1); 870 vfs_mp_count_sub_pcpu(mpcpu, ref, 1); 871 vfs_op_thread_exit(mp, mpcpu); 872 return; 873 } 874 875 MNT_ILOCK(mp); 876 vfs_assert_mount_counters(mp); 877 MNT_REL(mp); 878 c = --mp->mnt_lockref; 879 if (mp->mnt_vfs_ops == 0) { 880 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 881 MNT_IUNLOCK(mp); 882 return; 883 } 884 if (c < 0) 885 vfs_dump_mount_counters(mp); 886 if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) { 887 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT); 888 CTR1(KTR_VFS, "%s: waking up waiters", __func__); 889 mp->mnt_kern_flag &= ~MNTK_DRAINING; 890 wakeup(&mp->mnt_lockref); 891 } 892 MNT_IUNLOCK(mp); 893 } 894 895 /* 896 * Lookup a mount point by filesystem identifier. 897 */ 898 struct mount * 899 vfs_getvfs(fsid_t *fsid) 900 { 901 struct mount *mp; 902 903 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 904 mtx_lock(&mountlist_mtx); 905 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 906 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) { 907 vfs_ref(mp); 908 mtx_unlock(&mountlist_mtx); 909 return (mp); 910 } 911 } 912 mtx_unlock(&mountlist_mtx); 913 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 914 return ((struct mount *) 0); 915 } 916 917 /* 918 * Lookup a mount point by filesystem identifier, busying it before 919 * returning. 920 * 921 * To avoid congestion on mountlist_mtx, implement simple direct-mapped 922 * cache for popular filesystem identifiers. The cache is lockess, using 923 * the fact that struct mount's are never freed. In worst case we may 924 * get pointer to unmounted or even different filesystem, so we have to 925 * check what we got, and go slow way if so. 926 */ 927 struct mount * 928 vfs_busyfs(fsid_t *fsid) 929 { 930 #define FSID_CACHE_SIZE 256 931 typedef struct mount * volatile vmp_t; 932 static vmp_t cache[FSID_CACHE_SIZE]; 933 struct mount *mp; 934 int error; 935 uint32_t hash; 936 937 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 938 hash = fsid->val[0] ^ fsid->val[1]; 939 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1); 940 mp = cache[hash]; 941 if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0) 942 goto slow; 943 if (vfs_busy(mp, 0) != 0) { 944 cache[hash] = NULL; 945 goto slow; 946 } 947 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) 948 return (mp); 949 else 950 vfs_unbusy(mp); 951 952 slow: 953 mtx_lock(&mountlist_mtx); 954 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 955 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) { 956 error = vfs_busy(mp, MBF_MNTLSTLOCK); 957 if (error) { 958 cache[hash] = NULL; 959 mtx_unlock(&mountlist_mtx); 960 return (NULL); 961 } 962 cache[hash] = mp; 963 return (mp); 964 } 965 } 966 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 967 mtx_unlock(&mountlist_mtx); 968 return ((struct mount *) 0); 969 } 970 971 /* 972 * Check if a user can access privileged mount options. 973 */ 974 int 975 vfs_suser(struct mount *mp, struct thread *td) 976 { 977 int error; 978 979 if (jailed(td->td_ucred)) { 980 /* 981 * If the jail of the calling thread lacks permission for 982 * this type of file system, deny immediately. 983 */ 984 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag)) 985 return (EPERM); 986 987 /* 988 * If the file system was mounted outside the jail of the 989 * calling thread, deny immediately. 990 */ 991 if (prison_check(td->td_ucred, mp->mnt_cred) != 0) 992 return (EPERM); 993 } 994 995 /* 996 * If file system supports delegated administration, we don't check 997 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified 998 * by the file system itself. 999 * If this is not the user that did original mount, we check for 1000 * the PRIV_VFS_MOUNT_OWNER privilege. 1001 */ 1002 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) && 1003 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 1004 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) 1005 return (error); 1006 } 1007 return (0); 1008 } 1009 1010 /* 1011 * Get a new unique fsid. Try to make its val[0] unique, since this value 1012 * will be used to create fake device numbers for stat(). Also try (but 1013 * not so hard) make its val[0] unique mod 2^16, since some emulators only 1014 * support 16-bit device numbers. We end up with unique val[0]'s for the 1015 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 1016 * 1017 * Keep in mind that several mounts may be running in parallel. Starting 1018 * the search one past where the previous search terminated is both a 1019 * micro-optimization and a defense against returning the same fsid to 1020 * different mounts. 1021 */ 1022 void 1023 vfs_getnewfsid(struct mount *mp) 1024 { 1025 static uint16_t mntid_base; 1026 struct mount *nmp; 1027 fsid_t tfsid; 1028 int mtype; 1029 1030 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 1031 mtx_lock(&mntid_mtx); 1032 mtype = mp->mnt_vfc->vfc_typenum; 1033 tfsid.val[1] = mtype; 1034 mtype = (mtype & 0xFF) << 24; 1035 for (;;) { 1036 tfsid.val[0] = makedev(255, 1037 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 1038 mntid_base++; 1039 if ((nmp = vfs_getvfs(&tfsid)) == NULL) 1040 break; 1041 vfs_rel(nmp); 1042 } 1043 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 1044 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 1045 mtx_unlock(&mntid_mtx); 1046 } 1047 1048 /* 1049 * Knob to control the precision of file timestamps: 1050 * 1051 * 0 = seconds only; nanoseconds zeroed. 1052 * 1 = seconds and nanoseconds, accurate within 1/HZ. 1053 * 2 = seconds and nanoseconds, truncated to microseconds. 1054 * >=3 = seconds and nanoseconds, maximum precision. 1055 */ 1056 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 1057 1058 static int timestamp_precision = TSP_USEC; 1059 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 1060 ×tamp_precision, 0, "File timestamp precision (0: seconds, " 1061 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, " 1062 "3+: sec + ns (max. precision))"); 1063 1064 /* 1065 * Get a current timestamp. 1066 */ 1067 void 1068 vfs_timestamp(struct timespec *tsp) 1069 { 1070 struct timeval tv; 1071 1072 switch (timestamp_precision) { 1073 case TSP_SEC: 1074 tsp->tv_sec = time_second; 1075 tsp->tv_nsec = 0; 1076 break; 1077 case TSP_HZ: 1078 getnanotime(tsp); 1079 break; 1080 case TSP_USEC: 1081 microtime(&tv); 1082 TIMEVAL_TO_TIMESPEC(&tv, tsp); 1083 break; 1084 case TSP_NSEC: 1085 default: 1086 nanotime(tsp); 1087 break; 1088 } 1089 } 1090 1091 /* 1092 * Set vnode attributes to VNOVAL 1093 */ 1094 void 1095 vattr_null(struct vattr *vap) 1096 { 1097 1098 vap->va_type = VNON; 1099 vap->va_size = VNOVAL; 1100 vap->va_bytes = VNOVAL; 1101 vap->va_mode = VNOVAL; 1102 vap->va_nlink = VNOVAL; 1103 vap->va_uid = VNOVAL; 1104 vap->va_gid = VNOVAL; 1105 vap->va_fsid = VNOVAL; 1106 vap->va_fileid = VNOVAL; 1107 vap->va_blocksize = VNOVAL; 1108 vap->va_rdev = VNOVAL; 1109 vap->va_atime.tv_sec = VNOVAL; 1110 vap->va_atime.tv_nsec = VNOVAL; 1111 vap->va_mtime.tv_sec = VNOVAL; 1112 vap->va_mtime.tv_nsec = VNOVAL; 1113 vap->va_ctime.tv_sec = VNOVAL; 1114 vap->va_ctime.tv_nsec = VNOVAL; 1115 vap->va_birthtime.tv_sec = VNOVAL; 1116 vap->va_birthtime.tv_nsec = VNOVAL; 1117 vap->va_flags = VNOVAL; 1118 vap->va_gen = VNOVAL; 1119 vap->va_vaflags = 0; 1120 } 1121 1122 /* 1123 * Try to reduce the total number of vnodes. 1124 * 1125 * This routine (and its user) are buggy in at least the following ways: 1126 * - all parameters were picked years ago when RAM sizes were significantly 1127 * smaller 1128 * - it can pick vnodes based on pages used by the vm object, but filesystems 1129 * like ZFS don't use it making the pick broken 1130 * - since ZFS has its own aging policy it gets partially combated by this one 1131 * - a dedicated method should be provided for filesystems to let them decide 1132 * whether the vnode should be recycled 1133 * 1134 * This routine is called when we have too many vnodes. It attempts 1135 * to free <count> vnodes and will potentially free vnodes that still 1136 * have VM backing store (VM backing store is typically the cause 1137 * of a vnode blowout so we want to do this). Therefore, this operation 1138 * is not considered cheap. 1139 * 1140 * A number of conditions may prevent a vnode from being reclaimed. 1141 * the buffer cache may have references on the vnode, a directory 1142 * vnode may still have references due to the namei cache representing 1143 * underlying files, or the vnode may be in active use. It is not 1144 * desirable to reuse such vnodes. These conditions may cause the 1145 * number of vnodes to reach some minimum value regardless of what 1146 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 1147 * 1148 * @param reclaim_nc_src Only reclaim directories with outgoing namecache 1149 * entries if this argument is strue 1150 * @param trigger Only reclaim vnodes with fewer than this many resident 1151 * pages. 1152 * @param target How many vnodes to reclaim. 1153 * @return The number of vnodes that were reclaimed. 1154 */ 1155 static int 1156 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target) 1157 { 1158 struct vnode *vp, *mvp; 1159 struct mount *mp; 1160 struct vm_object *object; 1161 u_long done; 1162 bool retried; 1163 1164 mtx_assert(&vnode_list_mtx, MA_OWNED); 1165 1166 retried = false; 1167 done = 0; 1168 1169 mvp = vnode_list_reclaim_marker; 1170 restart: 1171 vp = mvp; 1172 while (done < target) { 1173 vp = TAILQ_NEXT(vp, v_vnodelist); 1174 if (__predict_false(vp == NULL)) 1175 break; 1176 1177 if (__predict_false(vp->v_type == VMARKER)) 1178 continue; 1179 1180 /* 1181 * If it's been deconstructed already, it's still 1182 * referenced, or it exceeds the trigger, skip it. 1183 * Also skip free vnodes. We are trying to make space 1184 * to expand the free list, not reduce it. 1185 */ 1186 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 || 1187 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src))) 1188 goto next_iter; 1189 1190 if (vp->v_type == VBAD || vp->v_type == VNON) 1191 goto next_iter; 1192 1193 object = atomic_load_ptr(&vp->v_object); 1194 if (object == NULL || object->resident_page_count > trigger) { 1195 goto next_iter; 1196 } 1197 1198 /* 1199 * Handle races against vnode allocation. Filesystems lock the 1200 * vnode some time after it gets returned from getnewvnode, 1201 * despite type and hold count being manipulated earlier. 1202 * Resorting to checking v_mount restores guarantees present 1203 * before the global list was reworked to contain all vnodes. 1204 */ 1205 if (!VI_TRYLOCK(vp)) 1206 goto next_iter; 1207 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) { 1208 VI_UNLOCK(vp); 1209 goto next_iter; 1210 } 1211 if (vp->v_mount == NULL) { 1212 VI_UNLOCK(vp); 1213 goto next_iter; 1214 } 1215 vholdl(vp); 1216 VI_UNLOCK(vp); 1217 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1218 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1219 mtx_unlock(&vnode_list_mtx); 1220 1221 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1222 vdrop_recycle(vp); 1223 goto next_iter_unlocked; 1224 } 1225 if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) { 1226 vdrop_recycle(vp); 1227 vn_finished_write(mp); 1228 goto next_iter_unlocked; 1229 } 1230 1231 VI_LOCK(vp); 1232 if (vp->v_usecount > 0 || 1233 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 1234 (vp->v_object != NULL && vp->v_object->handle == vp && 1235 vp->v_object->resident_page_count > trigger)) { 1236 VOP_UNLOCK(vp); 1237 vdropl_recycle(vp); 1238 vn_finished_write(mp); 1239 goto next_iter_unlocked; 1240 } 1241 counter_u64_add(recycles_count, 1); 1242 vgonel(vp); 1243 VOP_UNLOCK(vp); 1244 vdropl_recycle(vp); 1245 vn_finished_write(mp); 1246 done++; 1247 next_iter_unlocked: 1248 maybe_yield(); 1249 mtx_lock(&vnode_list_mtx); 1250 goto restart; 1251 next_iter: 1252 MPASS(vp->v_type != VMARKER); 1253 if (!should_yield()) 1254 continue; 1255 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1256 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1257 mtx_unlock(&vnode_list_mtx); 1258 kern_yield(PRI_USER); 1259 mtx_lock(&vnode_list_mtx); 1260 goto restart; 1261 } 1262 if (done == 0 && !retried) { 1263 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1264 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist); 1265 retried = true; 1266 goto restart; 1267 } 1268 return (done); 1269 } 1270 1271 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */ 1272 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free, 1273 0, 1274 "limit on vnode free requests per call to the vnlru_free routine"); 1275 1276 /* 1277 * Attempt to reduce the free list by the requested amount. 1278 */ 1279 static int 1280 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp) 1281 { 1282 struct vnode *vp; 1283 struct mount *mp; 1284 int ocount; 1285 1286 mtx_assert(&vnode_list_mtx, MA_OWNED); 1287 if (count > max_vnlru_free) 1288 count = max_vnlru_free; 1289 ocount = count; 1290 vp = mvp; 1291 for (;;) { 1292 if (count == 0) { 1293 break; 1294 } 1295 vp = TAILQ_NEXT(vp, v_vnodelist); 1296 if (__predict_false(vp == NULL)) { 1297 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1298 TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist); 1299 break; 1300 } 1301 if (__predict_false(vp->v_type == VMARKER)) 1302 continue; 1303 if (vp->v_holdcnt > 0) 1304 continue; 1305 /* 1306 * Don't recycle if our vnode is from different type 1307 * of mount point. Note that mp is type-safe, the 1308 * check does not reach unmapped address even if 1309 * vnode is reclaimed. 1310 */ 1311 if (mnt_op != NULL && (mp = vp->v_mount) != NULL && 1312 mp->mnt_op != mnt_op) { 1313 continue; 1314 } 1315 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) { 1316 continue; 1317 } 1318 if (!vhold_recycle_free(vp)) 1319 continue; 1320 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1321 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1322 mtx_unlock(&vnode_list_mtx); 1323 /* 1324 * FIXME: ignores the return value, meaning it may be nothing 1325 * got recycled but it claims otherwise to the caller. 1326 * 1327 * Originally the value started being ignored in 2005 with 1328 * 114a1006a8204aa156e1f9ad6476cdff89cada7f . 1329 * 1330 * Respecting the value can run into significant stalls if most 1331 * vnodes belong to one file system and it has writes 1332 * suspended. In presence of many threads and millions of 1333 * vnodes they keep contending on the vnode_list_mtx lock only 1334 * to find vnodes they can't recycle. 1335 * 1336 * The solution would be to pre-check if the vnode is likely to 1337 * be recycle-able, but it needs to happen with the 1338 * vnode_list_mtx lock held. This runs into a problem where 1339 * VOP_GETWRITEMOUNT (currently needed to find out about if 1340 * writes are frozen) can take locks which LOR against it. 1341 * 1342 * Check nullfs for one example (null_getwritemount). 1343 */ 1344 vtryrecycle(vp); 1345 count--; 1346 mtx_lock(&vnode_list_mtx); 1347 vp = mvp; 1348 } 1349 return (ocount - count); 1350 } 1351 1352 static int 1353 vnlru_free_locked(int count) 1354 { 1355 1356 mtx_assert(&vnode_list_mtx, MA_OWNED); 1357 return (vnlru_free_impl(count, NULL, vnode_list_free_marker)); 1358 } 1359 1360 void 1361 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp) 1362 { 1363 1364 MPASS(mnt_op != NULL); 1365 MPASS(mvp != NULL); 1366 VNPASS(mvp->v_type == VMARKER, mvp); 1367 mtx_lock(&vnode_list_mtx); 1368 vnlru_free_impl(count, mnt_op, mvp); 1369 mtx_unlock(&vnode_list_mtx); 1370 } 1371 1372 struct vnode * 1373 vnlru_alloc_marker(void) 1374 { 1375 struct vnode *mvp; 1376 1377 mvp = vn_alloc_marker(NULL); 1378 mtx_lock(&vnode_list_mtx); 1379 TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist); 1380 mtx_unlock(&vnode_list_mtx); 1381 return (mvp); 1382 } 1383 1384 void 1385 vnlru_free_marker(struct vnode *mvp) 1386 { 1387 mtx_lock(&vnode_list_mtx); 1388 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1389 mtx_unlock(&vnode_list_mtx); 1390 vn_free_marker(mvp); 1391 } 1392 1393 static void 1394 vnlru_recalc(void) 1395 { 1396 1397 mtx_assert(&vnode_list_mtx, MA_OWNED); 1398 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100); 1399 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */ 1400 vlowat = vhiwat / 2; 1401 } 1402 1403 /* 1404 * Attempt to recycle vnodes in a context that is always safe to block. 1405 * Calling vlrurecycle() from the bowels of filesystem code has some 1406 * interesting deadlock problems. 1407 */ 1408 static struct proc *vnlruproc; 1409 static int vnlruproc_sig; 1410 1411 /* 1412 * The main freevnodes counter is only updated when threads requeue their vnode 1413 * batches. CPUs are conditionally walked to compute a more accurate total. 1414 * 1415 * Limit how much of a slop are we willing to tolerate. Note: the actual value 1416 * at any given moment can still exceed slop, but it should not be by significant 1417 * margin in practice. 1418 */ 1419 #define VNLRU_FREEVNODES_SLOP 126 1420 1421 static void __noinline 1422 vfs_freevnodes_rollup(int8_t *lfreevnodes) 1423 { 1424 1425 atomic_add_long(&freevnodes, *lfreevnodes); 1426 *lfreevnodes = 0; 1427 critical_exit(); 1428 } 1429 1430 static __inline void 1431 vfs_freevnodes_inc(void) 1432 { 1433 int8_t *lfreevnodes; 1434 1435 critical_enter(); 1436 lfreevnodes = PCPU_PTR(vfs_freevnodes); 1437 (*lfreevnodes)++; 1438 if (__predict_false(*lfreevnodes == VNLRU_FREEVNODES_SLOP)) 1439 vfs_freevnodes_rollup(lfreevnodes); 1440 else 1441 critical_exit(); 1442 } 1443 1444 static __inline void 1445 vfs_freevnodes_dec(void) 1446 { 1447 int8_t *lfreevnodes; 1448 1449 critical_enter(); 1450 lfreevnodes = PCPU_PTR(vfs_freevnodes); 1451 (*lfreevnodes)--; 1452 if (__predict_false(*lfreevnodes == -VNLRU_FREEVNODES_SLOP)) 1453 vfs_freevnodes_rollup(lfreevnodes); 1454 else 1455 critical_exit(); 1456 } 1457 1458 static u_long 1459 vnlru_read_freevnodes(void) 1460 { 1461 long slop, rfreevnodes; 1462 int cpu; 1463 1464 rfreevnodes = atomic_load_long(&freevnodes); 1465 1466 if (rfreevnodes > freevnodes_old) 1467 slop = rfreevnodes - freevnodes_old; 1468 else 1469 slop = freevnodes_old - rfreevnodes; 1470 if (slop < VNLRU_FREEVNODES_SLOP) 1471 return (rfreevnodes >= 0 ? rfreevnodes : 0); 1472 freevnodes_old = rfreevnodes; 1473 CPU_FOREACH(cpu) { 1474 freevnodes_old += cpuid_to_pcpu[cpu]->pc_vfs_freevnodes; 1475 } 1476 return (freevnodes_old >= 0 ? freevnodes_old : 0); 1477 } 1478 1479 static bool 1480 vnlru_under(u_long rnumvnodes, u_long limit) 1481 { 1482 u_long rfreevnodes, space; 1483 1484 if (__predict_false(rnumvnodes > desiredvnodes)) 1485 return (true); 1486 1487 space = desiredvnodes - rnumvnodes; 1488 if (space < limit) { 1489 rfreevnodes = vnlru_read_freevnodes(); 1490 if (rfreevnodes > wantfreevnodes) 1491 space += rfreevnodes - wantfreevnodes; 1492 } 1493 return (space < limit); 1494 } 1495 1496 static bool 1497 vnlru_under_unlocked(u_long rnumvnodes, u_long limit) 1498 { 1499 long rfreevnodes, space; 1500 1501 if (__predict_false(rnumvnodes > desiredvnodes)) 1502 return (true); 1503 1504 space = desiredvnodes - rnumvnodes; 1505 if (space < limit) { 1506 rfreevnodes = atomic_load_long(&freevnodes); 1507 if (rfreevnodes > wantfreevnodes) 1508 space += rfreevnodes - wantfreevnodes; 1509 } 1510 return (space < limit); 1511 } 1512 1513 static void 1514 vnlru_kick(void) 1515 { 1516 1517 mtx_assert(&vnode_list_mtx, MA_OWNED); 1518 if (vnlruproc_sig == 0) { 1519 vnlruproc_sig = 1; 1520 wakeup(vnlruproc); 1521 } 1522 } 1523 1524 static void 1525 vnlru_proc(void) 1526 { 1527 u_long rnumvnodes, rfreevnodes, target; 1528 unsigned long onumvnodes; 1529 int done, force, trigger, usevnodes; 1530 bool reclaim_nc_src, want_reread; 1531 1532 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc, 1533 SHUTDOWN_PRI_FIRST); 1534 1535 force = 0; 1536 want_reread = false; 1537 for (;;) { 1538 kproc_suspend_check(vnlruproc); 1539 mtx_lock(&vnode_list_mtx); 1540 rnumvnodes = atomic_load_long(&numvnodes); 1541 1542 if (want_reread) { 1543 force = vnlru_under(numvnodes, vhiwat) ? 1 : 0; 1544 want_reread = false; 1545 } 1546 1547 /* 1548 * If numvnodes is too large (due to desiredvnodes being 1549 * adjusted using its sysctl, or emergency growth), first 1550 * try to reduce it by discarding from the free list. 1551 */ 1552 if (rnumvnodes > desiredvnodes) { 1553 vnlru_free_locked(rnumvnodes - desiredvnodes); 1554 rnumvnodes = atomic_load_long(&numvnodes); 1555 } 1556 /* 1557 * Sleep if the vnode cache is in a good state. This is 1558 * when it is not over-full and has space for about a 4% 1559 * or 9% expansion (by growing its size or inexcessively 1560 * reducing its free list). Otherwise, try to reclaim 1561 * space for a 10% expansion. 1562 */ 1563 if (vstir && force == 0) { 1564 force = 1; 1565 vstir = 0; 1566 } 1567 if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) { 1568 vnlruproc_sig = 0; 1569 wakeup(&vnlruproc_sig); 1570 msleep(vnlruproc, &vnode_list_mtx, 1571 PVFS|PDROP, "vlruwt", hz); 1572 continue; 1573 } 1574 rfreevnodes = vnlru_read_freevnodes(); 1575 1576 onumvnodes = rnumvnodes; 1577 /* 1578 * Calculate parameters for recycling. These are the same 1579 * throughout the loop to give some semblance of fairness. 1580 * The trigger point is to avoid recycling vnodes with lots 1581 * of resident pages. We aren't trying to free memory; we 1582 * are trying to recycle or at least free vnodes. 1583 */ 1584 if (rnumvnodes <= desiredvnodes) 1585 usevnodes = rnumvnodes - rfreevnodes; 1586 else 1587 usevnodes = rnumvnodes; 1588 if (usevnodes <= 0) 1589 usevnodes = 1; 1590 /* 1591 * The trigger value is chosen to give a conservatively 1592 * large value to ensure that it alone doesn't prevent 1593 * making progress. The value can easily be so large that 1594 * it is effectively infinite in some congested and 1595 * misconfigured cases, and this is necessary. Normally 1596 * it is about 8 to 100 (pages), which is quite large. 1597 */ 1598 trigger = vm_cnt.v_page_count * 2 / usevnodes; 1599 if (force < 2) 1600 trigger = vsmalltrigger; 1601 reclaim_nc_src = force >= 3; 1602 target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1); 1603 target = target / 10 + 1; 1604 done = vlrureclaim(reclaim_nc_src, trigger, target); 1605 mtx_unlock(&vnode_list_mtx); 1606 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes) 1607 uma_reclaim(UMA_RECLAIM_DRAIN); 1608 if (done == 0) { 1609 if (force == 0 || force == 1) { 1610 force = 2; 1611 continue; 1612 } 1613 if (force == 2) { 1614 force = 3; 1615 continue; 1616 } 1617 want_reread = true; 1618 force = 0; 1619 vnlru_nowhere++; 1620 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 1621 } else { 1622 want_reread = true; 1623 kern_yield(PRI_USER); 1624 } 1625 } 1626 } 1627 1628 static struct kproc_desc vnlru_kp = { 1629 "vnlru", 1630 vnlru_proc, 1631 &vnlruproc 1632 }; 1633 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, 1634 &vnlru_kp); 1635 1636 /* 1637 * Routines having to do with the management of the vnode table. 1638 */ 1639 1640 /* 1641 * Try to recycle a freed vnode. We abort if anyone picks up a reference 1642 * before we actually vgone(). This function must be called with the vnode 1643 * held to prevent the vnode from being returned to the free list midway 1644 * through vgone(). 1645 */ 1646 static int 1647 vtryrecycle(struct vnode *vp) 1648 { 1649 struct mount *vnmp; 1650 1651 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 1652 VNPASS(vp->v_holdcnt > 0, vp); 1653 /* 1654 * This vnode may found and locked via some other list, if so we 1655 * can't recycle it yet. 1656 */ 1657 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 1658 CTR2(KTR_VFS, 1659 "%s: impossible to recycle, vp %p lock is already held", 1660 __func__, vp); 1661 vdrop_recycle(vp); 1662 return (EWOULDBLOCK); 1663 } 1664 /* 1665 * Don't recycle if its filesystem is being suspended. 1666 */ 1667 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 1668 VOP_UNLOCK(vp); 1669 CTR2(KTR_VFS, 1670 "%s: impossible to recycle, cannot start the write for %p", 1671 __func__, vp); 1672 vdrop_recycle(vp); 1673 return (EBUSY); 1674 } 1675 /* 1676 * If we got this far, we need to acquire the interlock and see if 1677 * anyone picked up this vnode from another list. If not, we will 1678 * mark it with DOOMED via vgonel() so that anyone who does find it 1679 * will skip over it. 1680 */ 1681 VI_LOCK(vp); 1682 if (vp->v_usecount) { 1683 VOP_UNLOCK(vp); 1684 vdropl_recycle(vp); 1685 vn_finished_write(vnmp); 1686 CTR2(KTR_VFS, 1687 "%s: impossible to recycle, %p is already referenced", 1688 __func__, vp); 1689 return (EBUSY); 1690 } 1691 if (!VN_IS_DOOMED(vp)) { 1692 counter_u64_add(recycles_free_count, 1); 1693 vgonel(vp); 1694 } 1695 VOP_UNLOCK(vp); 1696 vdropl_recycle(vp); 1697 vn_finished_write(vnmp); 1698 return (0); 1699 } 1700 1701 /* 1702 * Allocate a new vnode. 1703 * 1704 * The operation never returns an error. Returning an error was disabled 1705 * in r145385 (dated 2005) with the following comment: 1706 * 1707 * XXX Not all VFS_VGET/ffs_vget callers check returns. 1708 * 1709 * Given the age of this commit (almost 15 years at the time of writing this 1710 * comment) restoring the ability to fail requires a significant audit of 1711 * all codepaths. 1712 * 1713 * The routine can try to free a vnode or stall for up to 1 second waiting for 1714 * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation. 1715 */ 1716 static u_long vn_alloc_cyclecount; 1717 1718 static struct vnode * __noinline 1719 vn_alloc_hard(struct mount *mp) 1720 { 1721 u_long rnumvnodes, rfreevnodes; 1722 1723 mtx_lock(&vnode_list_mtx); 1724 rnumvnodes = atomic_load_long(&numvnodes); 1725 if (rnumvnodes + 1 < desiredvnodes) { 1726 vn_alloc_cyclecount = 0; 1727 goto alloc; 1728 } 1729 rfreevnodes = vnlru_read_freevnodes(); 1730 if (vn_alloc_cyclecount++ >= rfreevnodes) { 1731 vn_alloc_cyclecount = 0; 1732 vstir = 1; 1733 } 1734 /* 1735 * Grow the vnode cache if it will not be above its target max 1736 * after growing. Otherwise, if the free list is nonempty, try 1737 * to reclaim 1 item from it before growing the cache (possibly 1738 * above its target max if the reclamation failed or is delayed). 1739 * Otherwise, wait for some space. In all cases, schedule 1740 * vnlru_proc() if we are getting short of space. The watermarks 1741 * should be chosen so that we never wait or even reclaim from 1742 * the free list to below its target minimum. 1743 */ 1744 if (vnlru_free_locked(1) > 0) 1745 goto alloc; 1746 if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 1747 /* 1748 * Wait for space for a new vnode. 1749 */ 1750 vnlru_kick(); 1751 msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz); 1752 if (atomic_load_long(&numvnodes) + 1 > desiredvnodes && 1753 vnlru_read_freevnodes() > 1) 1754 vnlru_free_locked(1); 1755 } 1756 alloc: 1757 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; 1758 if (vnlru_under(rnumvnodes, vlowat)) 1759 vnlru_kick(); 1760 mtx_unlock(&vnode_list_mtx); 1761 return (uma_zalloc_smr(vnode_zone, M_WAITOK)); 1762 } 1763 1764 static struct vnode * 1765 vn_alloc(struct mount *mp) 1766 { 1767 u_long rnumvnodes; 1768 1769 if (__predict_false(vn_alloc_cyclecount != 0)) 1770 return (vn_alloc_hard(mp)); 1771 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; 1772 if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) { 1773 atomic_subtract_long(&numvnodes, 1); 1774 return (vn_alloc_hard(mp)); 1775 } 1776 1777 return (uma_zalloc_smr(vnode_zone, M_WAITOK)); 1778 } 1779 1780 static void 1781 vn_free(struct vnode *vp) 1782 { 1783 1784 atomic_subtract_long(&numvnodes, 1); 1785 uma_zfree_smr(vnode_zone, vp); 1786 } 1787 1788 /* 1789 * Return the next vnode from the free list. 1790 */ 1791 int 1792 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, 1793 struct vnode **vpp) 1794 { 1795 struct vnode *vp; 1796 struct thread *td; 1797 struct lock_object *lo; 1798 1799 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag); 1800 1801 KASSERT(vops->registered, 1802 ("%s: not registered vector op %p\n", __func__, vops)); 1803 cache_validate_vop_vector(mp, vops); 1804 1805 td = curthread; 1806 if (td->td_vp_reserved != NULL) { 1807 vp = td->td_vp_reserved; 1808 td->td_vp_reserved = NULL; 1809 } else { 1810 vp = vn_alloc(mp); 1811 } 1812 counter_u64_add(vnodes_created, 1); 1813 1814 vn_set_state(vp, VSTATE_UNINITIALIZED); 1815 1816 /* 1817 * Locks are given the generic name "vnode" when created. 1818 * Follow the historic practice of using the filesystem 1819 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc. 1820 * 1821 * Locks live in a witness group keyed on their name. Thus, 1822 * when a lock is renamed, it must also move from the witness 1823 * group of its old name to the witness group of its new name. 1824 * 1825 * The change only needs to be made when the vnode moves 1826 * from one filesystem type to another. We ensure that each 1827 * filesystem use a single static name pointer for its tag so 1828 * that we can compare pointers rather than doing a strcmp(). 1829 */ 1830 lo = &vp->v_vnlock->lock_object; 1831 #ifdef WITNESS 1832 if (lo->lo_name != tag) { 1833 #endif 1834 lo->lo_name = tag; 1835 #ifdef WITNESS 1836 WITNESS_DESTROY(lo); 1837 WITNESS_INIT(lo, tag); 1838 } 1839 #endif 1840 /* 1841 * By default, don't allow shared locks unless filesystems opt-in. 1842 */ 1843 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE; 1844 /* 1845 * Finalize various vnode identity bits. 1846 */ 1847 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp)); 1848 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp)); 1849 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp)); 1850 vp->v_type = VNON; 1851 vp->v_op = vops; 1852 vp->v_irflag = 0; 1853 v_init_counters(vp); 1854 vn_seqc_init(vp); 1855 vp->v_bufobj.bo_ops = &buf_ops_bio; 1856 #ifdef DIAGNOSTIC 1857 if (mp == NULL && vops != &dead_vnodeops) 1858 printf("NULL mp in getnewvnode(9), tag %s\n", tag); 1859 #endif 1860 #ifdef MAC 1861 mac_vnode_init(vp); 1862 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 1863 mac_vnode_associate_singlelabel(mp, vp); 1864 #endif 1865 if (mp != NULL) { 1866 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize; 1867 } 1868 1869 /* 1870 * For the filesystems which do not use vfs_hash_insert(), 1871 * still initialize v_hash to have vfs_hash_index() useful. 1872 * E.g., nullfs uses vfs_hash_index() on the lower vnode for 1873 * its own hashing. 1874 */ 1875 vp->v_hash = (uintptr_t)vp >> vnsz2log; 1876 1877 *vpp = vp; 1878 return (0); 1879 } 1880 1881 void 1882 getnewvnode_reserve(void) 1883 { 1884 struct thread *td; 1885 1886 td = curthread; 1887 MPASS(td->td_vp_reserved == NULL); 1888 td->td_vp_reserved = vn_alloc(NULL); 1889 } 1890 1891 void 1892 getnewvnode_drop_reserve(void) 1893 { 1894 struct thread *td; 1895 1896 td = curthread; 1897 if (td->td_vp_reserved != NULL) { 1898 vn_free(td->td_vp_reserved); 1899 td->td_vp_reserved = NULL; 1900 } 1901 } 1902 1903 static void __noinline 1904 freevnode(struct vnode *vp) 1905 { 1906 struct bufobj *bo; 1907 1908 /* 1909 * The vnode has been marked for destruction, so free it. 1910 * 1911 * The vnode will be returned to the zone where it will 1912 * normally remain until it is needed for another vnode. We 1913 * need to cleanup (or verify that the cleanup has already 1914 * been done) any residual data left from its current use 1915 * so as not to contaminate the freshly allocated vnode. 1916 */ 1917 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp); 1918 /* 1919 * Paired with vgone. 1920 */ 1921 vn_seqc_write_end_free(vp); 1922 1923 bo = &vp->v_bufobj; 1924 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 1925 VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp); 1926 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 1927 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 1928 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 1929 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 1930 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp, 1931 ("clean blk trie not empty")); 1932 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 1933 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp, 1934 ("dirty blk trie not empty")); 1935 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp, 1936 ("Dangling rangelock waiters")); 1937 VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp, 1938 ("Leaked inactivation")); 1939 VI_UNLOCK(vp); 1940 cache_assert_no_entries(vp); 1941 1942 #ifdef MAC 1943 mac_vnode_destroy(vp); 1944 #endif 1945 if (vp->v_pollinfo != NULL) { 1946 /* 1947 * Use LK_NOWAIT to shut up witness about the lock. We may get 1948 * here while having another vnode locked when trying to 1949 * satisfy a lookup and needing to recycle. 1950 */ 1951 VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT); 1952 destroy_vpollinfo(vp->v_pollinfo); 1953 VOP_UNLOCK(vp); 1954 vp->v_pollinfo = NULL; 1955 } 1956 vp->v_mountedhere = NULL; 1957 vp->v_unpcb = NULL; 1958 vp->v_rdev = NULL; 1959 vp->v_fifoinfo = NULL; 1960 vp->v_iflag = 0; 1961 vp->v_vflag = 0; 1962 bo->bo_flag = 0; 1963 vn_free(vp); 1964 } 1965 1966 /* 1967 * Delete from old mount point vnode list, if on one. 1968 */ 1969 static void 1970 delmntque(struct vnode *vp) 1971 { 1972 struct mount *mp; 1973 1974 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); 1975 1976 mp = vp->v_mount; 1977 MNT_ILOCK(mp); 1978 VI_LOCK(vp); 1979 vp->v_mount = NULL; 1980 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 1981 ("bad mount point vnode list size")); 1982 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1983 mp->mnt_nvnodelistsize--; 1984 MNT_REL(mp); 1985 MNT_IUNLOCK(mp); 1986 /* 1987 * The caller expects the interlock to be still held. 1988 */ 1989 ASSERT_VI_LOCKED(vp, __func__); 1990 } 1991 1992 static int 1993 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr) 1994 { 1995 1996 KASSERT(vp->v_mount == NULL, 1997 ("insmntque: vnode already on per mount vnode list")); 1998 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 1999 if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) { 2000 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); 2001 } else { 2002 KASSERT(!dtr, 2003 ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup", 2004 __func__)); 2005 } 2006 2007 /* 2008 * We acquire the vnode interlock early to ensure that the 2009 * vnode cannot be recycled by another process releasing a 2010 * holdcnt on it before we get it on both the vnode list 2011 * and the active vnode list. The mount mutex protects only 2012 * manipulation of the vnode list and the vnode freelist 2013 * mutex protects only manipulation of the active vnode list. 2014 * Hence the need to hold the vnode interlock throughout. 2015 */ 2016 MNT_ILOCK(mp); 2017 VI_LOCK(vp); 2018 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 && 2019 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || 2020 mp->mnt_nvnodelistsize == 0)) && 2021 (vp->v_vflag & VV_FORCEINSMQ) == 0) { 2022 VI_UNLOCK(vp); 2023 MNT_IUNLOCK(mp); 2024 if (dtr) { 2025 vp->v_data = NULL; 2026 vp->v_op = &dead_vnodeops; 2027 vgone(vp); 2028 vput(vp); 2029 } 2030 return (EBUSY); 2031 } 2032 vp->v_mount = mp; 2033 MNT_REF(mp); 2034 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2035 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 2036 ("neg mount point vnode list size")); 2037 mp->mnt_nvnodelistsize++; 2038 VI_UNLOCK(vp); 2039 MNT_IUNLOCK(mp); 2040 return (0); 2041 } 2042 2043 /* 2044 * Insert into list of vnodes for the new mount point, if available. 2045 * insmntque() reclaims the vnode on insertion failure, insmntque1() 2046 * leaves handling of the vnode to the caller. 2047 */ 2048 int 2049 insmntque(struct vnode *vp, struct mount *mp) 2050 { 2051 return (insmntque1_int(vp, mp, true)); 2052 } 2053 2054 int 2055 insmntque1(struct vnode *vp, struct mount *mp) 2056 { 2057 return (insmntque1_int(vp, mp, false)); 2058 } 2059 2060 /* 2061 * Flush out and invalidate all buffers associated with a bufobj 2062 * Called with the underlying object locked. 2063 */ 2064 int 2065 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) 2066 { 2067 int error; 2068 2069 BO_LOCK(bo); 2070 if (flags & V_SAVE) { 2071 error = bufobj_wwait(bo, slpflag, slptimeo); 2072 if (error) { 2073 BO_UNLOCK(bo); 2074 return (error); 2075 } 2076 if (bo->bo_dirty.bv_cnt > 0) { 2077 BO_UNLOCK(bo); 2078 do { 2079 error = BO_SYNC(bo, MNT_WAIT); 2080 } while (error == ERELOOKUP); 2081 if (error != 0) 2082 return (error); 2083 BO_LOCK(bo); 2084 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) { 2085 BO_UNLOCK(bo); 2086 return (EBUSY); 2087 } 2088 } 2089 } 2090 /* 2091 * If you alter this loop please notice that interlock is dropped and 2092 * reacquired in flushbuflist. Special care is needed to ensure that 2093 * no race conditions occur from this. 2094 */ 2095 do { 2096 error = flushbuflist(&bo->bo_clean, 2097 flags, bo, slpflag, slptimeo); 2098 if (error == 0 && !(flags & V_CLEANONLY)) 2099 error = flushbuflist(&bo->bo_dirty, 2100 flags, bo, slpflag, slptimeo); 2101 if (error != 0 && error != EAGAIN) { 2102 BO_UNLOCK(bo); 2103 return (error); 2104 } 2105 } while (error != 0); 2106 2107 /* 2108 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 2109 * have write I/O in-progress but if there is a VM object then the 2110 * VM object can also have read-I/O in-progress. 2111 */ 2112 do { 2113 bufobj_wwait(bo, 0, 0); 2114 if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) { 2115 BO_UNLOCK(bo); 2116 vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx"); 2117 BO_LOCK(bo); 2118 } 2119 } while (bo->bo_numoutput > 0); 2120 BO_UNLOCK(bo); 2121 2122 /* 2123 * Destroy the copy in the VM cache, too. 2124 */ 2125 if (bo->bo_object != NULL && 2126 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) { 2127 VM_OBJECT_WLOCK(bo->bo_object); 2128 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? 2129 OBJPR_CLEANONLY : 0); 2130 VM_OBJECT_WUNLOCK(bo->bo_object); 2131 } 2132 2133 #ifdef INVARIANTS 2134 BO_LOCK(bo); 2135 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO | 2136 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 || 2137 bo->bo_clean.bv_cnt > 0)) 2138 panic("vinvalbuf: flush failed"); 2139 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 && 2140 bo->bo_dirty.bv_cnt > 0) 2141 panic("vinvalbuf: flush dirty failed"); 2142 BO_UNLOCK(bo); 2143 #endif 2144 return (0); 2145 } 2146 2147 /* 2148 * Flush out and invalidate all buffers associated with a vnode. 2149 * Called with the underlying object locked. 2150 */ 2151 int 2152 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 2153 { 2154 2155 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 2156 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 2157 if (vp->v_object != NULL && vp->v_object->handle != vp) 2158 return (0); 2159 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); 2160 } 2161 2162 /* 2163 * Flush out buffers on the specified list. 2164 * 2165 */ 2166 static int 2167 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 2168 int slptimeo) 2169 { 2170 struct buf *bp, *nbp; 2171 int retval, error; 2172 daddr_t lblkno; 2173 b_xflags_t xflags; 2174 2175 ASSERT_BO_WLOCKED(bo); 2176 2177 retval = 0; 2178 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 2179 /* 2180 * If we are flushing both V_NORMAL and V_ALT buffers then 2181 * do not skip any buffers. If we are flushing only V_NORMAL 2182 * buffers then skip buffers marked as BX_ALTDATA. If we are 2183 * flushing only V_ALT buffers then skip buffers not marked 2184 * as BX_ALTDATA. 2185 */ 2186 if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) && 2187 (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) || 2188 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) { 2189 continue; 2190 } 2191 if (nbp != NULL) { 2192 lblkno = nbp->b_lblkno; 2193 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN); 2194 } 2195 retval = EAGAIN; 2196 error = BUF_TIMELOCK(bp, 2197 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), 2198 "flushbuf", slpflag, slptimeo); 2199 if (error) { 2200 BO_LOCK(bo); 2201 return (error != ENOLCK ? error : EAGAIN); 2202 } 2203 KASSERT(bp->b_bufobj == bo, 2204 ("bp %p wrong b_bufobj %p should be %p", 2205 bp, bp->b_bufobj, bo)); 2206 /* 2207 * XXX Since there are no node locks for NFS, I 2208 * believe there is a slight chance that a delayed 2209 * write will occur while sleeping just above, so 2210 * check for it. 2211 */ 2212 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 2213 (flags & V_SAVE)) { 2214 bremfree(bp); 2215 bp->b_flags |= B_ASYNC; 2216 bwrite(bp); 2217 BO_LOCK(bo); 2218 return (EAGAIN); /* XXX: why not loop ? */ 2219 } 2220 bremfree(bp); 2221 bp->b_flags |= (B_INVAL | B_RELBUF); 2222 bp->b_flags &= ~B_ASYNC; 2223 brelse(bp); 2224 BO_LOCK(bo); 2225 if (nbp == NULL) 2226 break; 2227 nbp = gbincore(bo, lblkno); 2228 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2229 != xflags) 2230 break; /* nbp invalid */ 2231 } 2232 return (retval); 2233 } 2234 2235 int 2236 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn) 2237 { 2238 struct buf *bp; 2239 int error; 2240 daddr_t lblkno; 2241 2242 ASSERT_BO_LOCKED(bo); 2243 2244 for (lblkno = startn;;) { 2245 again: 2246 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno); 2247 if (bp == NULL || bp->b_lblkno >= endn || 2248 bp->b_lblkno < startn) 2249 break; 2250 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 2251 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0); 2252 if (error != 0) { 2253 BO_RLOCK(bo); 2254 if (error == ENOLCK) 2255 goto again; 2256 return (error); 2257 } 2258 KASSERT(bp->b_bufobj == bo, 2259 ("bp %p wrong b_bufobj %p should be %p", 2260 bp, bp->b_bufobj, bo)); 2261 lblkno = bp->b_lblkno + 1; 2262 if ((bp->b_flags & B_MANAGED) == 0) 2263 bremfree(bp); 2264 bp->b_flags |= B_RELBUF; 2265 /* 2266 * In the VMIO case, use the B_NOREUSE flag to hint that the 2267 * pages backing each buffer in the range are unlikely to be 2268 * reused. Dirty buffers will have the hint applied once 2269 * they've been written. 2270 */ 2271 if ((bp->b_flags & B_VMIO) != 0) 2272 bp->b_flags |= B_NOREUSE; 2273 brelse(bp); 2274 BO_RLOCK(bo); 2275 } 2276 return (0); 2277 } 2278 2279 /* 2280 * Truncate a file's buffer and pages to a specified length. This 2281 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 2282 * sync activity. 2283 */ 2284 int 2285 vtruncbuf(struct vnode *vp, off_t length, int blksize) 2286 { 2287 struct buf *bp, *nbp; 2288 struct bufobj *bo; 2289 daddr_t startlbn; 2290 2291 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__, 2292 vp, blksize, (uintmax_t)length); 2293 2294 /* 2295 * Round up to the *next* lbn. 2296 */ 2297 startlbn = howmany(length, blksize); 2298 2299 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 2300 2301 bo = &vp->v_bufobj; 2302 restart_unlocked: 2303 BO_LOCK(bo); 2304 2305 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN) 2306 ; 2307 2308 if (length > 0) { 2309 restartsync: 2310 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2311 if (bp->b_lblkno > 0) 2312 continue; 2313 /* 2314 * Since we hold the vnode lock this should only 2315 * fail if we're racing with the buf daemon. 2316 */ 2317 if (BUF_LOCK(bp, 2318 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2319 BO_LOCKPTR(bo)) == ENOLCK) 2320 goto restart_unlocked; 2321 2322 VNASSERT((bp->b_flags & B_DELWRI), vp, 2323 ("buf(%p) on dirty queue without DELWRI", bp)); 2324 2325 bremfree(bp); 2326 bawrite(bp); 2327 BO_LOCK(bo); 2328 goto restartsync; 2329 } 2330 } 2331 2332 bufobj_wwait(bo, 0, 0); 2333 BO_UNLOCK(bo); 2334 vnode_pager_setsize(vp, length); 2335 2336 return (0); 2337 } 2338 2339 /* 2340 * Invalidate the cached pages of a file's buffer within the range of block 2341 * numbers [startlbn, endlbn). 2342 */ 2343 void 2344 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn, 2345 int blksize) 2346 { 2347 struct bufobj *bo; 2348 off_t start, end; 2349 2350 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range"); 2351 2352 start = blksize * startlbn; 2353 end = blksize * endlbn; 2354 2355 bo = &vp->v_bufobj; 2356 BO_LOCK(bo); 2357 MPASS(blksize == bo->bo_bsize); 2358 2359 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN) 2360 ; 2361 2362 BO_UNLOCK(bo); 2363 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1)); 2364 } 2365 2366 static int 2367 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, 2368 daddr_t startlbn, daddr_t endlbn) 2369 { 2370 struct buf *bp, *nbp; 2371 bool anyfreed; 2372 2373 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked"); 2374 ASSERT_BO_LOCKED(bo); 2375 2376 do { 2377 anyfreed = false; 2378 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 2379 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) 2380 continue; 2381 if (BUF_LOCK(bp, 2382 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2383 BO_LOCKPTR(bo)) == ENOLCK) { 2384 BO_LOCK(bo); 2385 return (EAGAIN); 2386 } 2387 2388 bremfree(bp); 2389 bp->b_flags |= B_INVAL | B_RELBUF; 2390 bp->b_flags &= ~B_ASYNC; 2391 brelse(bp); 2392 anyfreed = true; 2393 2394 BO_LOCK(bo); 2395 if (nbp != NULL && 2396 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 2397 nbp->b_vp != vp || 2398 (nbp->b_flags & B_DELWRI) != 0)) 2399 return (EAGAIN); 2400 } 2401 2402 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2403 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) 2404 continue; 2405 if (BUF_LOCK(bp, 2406 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2407 BO_LOCKPTR(bo)) == ENOLCK) { 2408 BO_LOCK(bo); 2409 return (EAGAIN); 2410 } 2411 bremfree(bp); 2412 bp->b_flags |= B_INVAL | B_RELBUF; 2413 bp->b_flags &= ~B_ASYNC; 2414 brelse(bp); 2415 anyfreed = true; 2416 2417 BO_LOCK(bo); 2418 if (nbp != NULL && 2419 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 2420 (nbp->b_vp != vp) || 2421 (nbp->b_flags & B_DELWRI) == 0)) 2422 return (EAGAIN); 2423 } 2424 } while (anyfreed); 2425 return (0); 2426 } 2427 2428 static void 2429 buf_vlist_remove(struct buf *bp) 2430 { 2431 struct bufv *bv; 2432 b_xflags_t flags; 2433 2434 flags = bp->b_xflags; 2435 2436 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2437 ASSERT_BO_WLOCKED(bp->b_bufobj); 2438 KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 && 2439 (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN), 2440 ("%s: buffer %p has invalid queue state", __func__, bp)); 2441 2442 if ((flags & BX_VNDIRTY) != 0) 2443 bv = &bp->b_bufobj->bo_dirty; 2444 else 2445 bv = &bp->b_bufobj->bo_clean; 2446 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno); 2447 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 2448 bv->bv_cnt--; 2449 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 2450 } 2451 2452 /* 2453 * Add the buffer to the sorted clean or dirty block list. 2454 * 2455 * NOTE: xflags is passed as a constant, optimizing this inline function! 2456 */ 2457 static void 2458 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 2459 { 2460 struct bufv *bv; 2461 struct buf *n; 2462 int error; 2463 2464 ASSERT_BO_WLOCKED(bo); 2465 KASSERT((bo->bo_flag & BO_NOBUFS) == 0, 2466 ("buf_vlist_add: bo %p does not allow bufs", bo)); 2467 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0, 2468 ("dead bo %p", bo)); 2469 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 2470 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 2471 bp->b_xflags |= xflags; 2472 if (xflags & BX_VNDIRTY) 2473 bv = &bo->bo_dirty; 2474 else 2475 bv = &bo->bo_clean; 2476 2477 /* 2478 * Keep the list ordered. Optimize empty list insertion. Assume 2479 * we tend to grow at the tail so lookup_le should usually be cheaper 2480 * than _ge. 2481 */ 2482 if (bv->bv_cnt == 0 || 2483 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno) 2484 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 2485 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL) 2486 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs); 2487 else 2488 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs); 2489 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp); 2490 if (error) 2491 panic("buf_vlist_add: Preallocated nodes insufficient."); 2492 bv->bv_cnt++; 2493 } 2494 2495 /* 2496 * Look up a buffer using the buffer tries. 2497 */ 2498 struct buf * 2499 gbincore(struct bufobj *bo, daddr_t lblkno) 2500 { 2501 struct buf *bp; 2502 2503 ASSERT_BO_LOCKED(bo); 2504 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno); 2505 if (bp != NULL) 2506 return (bp); 2507 return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno)); 2508 } 2509 2510 /* 2511 * Look up a buf using the buffer tries, without the bufobj lock. This relies 2512 * on SMR for safe lookup, and bufs being in a no-free zone to provide type 2513 * stability of the result. Like other lockless lookups, the found buf may 2514 * already be invalid by the time this function returns. 2515 */ 2516 struct buf * 2517 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno) 2518 { 2519 struct buf *bp; 2520 2521 ASSERT_BO_UNLOCKED(bo); 2522 bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno); 2523 if (bp != NULL) 2524 return (bp); 2525 return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno)); 2526 } 2527 2528 /* 2529 * Associate a buffer with a vnode. 2530 */ 2531 void 2532 bgetvp(struct vnode *vp, struct buf *bp) 2533 { 2534 struct bufobj *bo; 2535 2536 bo = &vp->v_bufobj; 2537 ASSERT_BO_WLOCKED(bo); 2538 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 2539 2540 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 2541 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 2542 ("bgetvp: bp already attached! %p", bp)); 2543 2544 vhold(vp); 2545 bp->b_vp = vp; 2546 bp->b_bufobj = bo; 2547 /* 2548 * Insert onto list for new vnode. 2549 */ 2550 buf_vlist_add(bp, bo, BX_VNCLEAN); 2551 } 2552 2553 /* 2554 * Disassociate a buffer from a vnode. 2555 */ 2556 void 2557 brelvp(struct buf *bp) 2558 { 2559 struct bufobj *bo; 2560 struct vnode *vp; 2561 2562 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2563 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 2564 2565 /* 2566 * Delete from old vnode list, if on one. 2567 */ 2568 vp = bp->b_vp; /* XXX */ 2569 bo = bp->b_bufobj; 2570 BO_LOCK(bo); 2571 buf_vlist_remove(bp); 2572 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2573 bo->bo_flag &= ~BO_ONWORKLST; 2574 mtx_lock(&sync_mtx); 2575 LIST_REMOVE(bo, bo_synclist); 2576 syncer_worklist_len--; 2577 mtx_unlock(&sync_mtx); 2578 } 2579 bp->b_vp = NULL; 2580 bp->b_bufobj = NULL; 2581 BO_UNLOCK(bo); 2582 vdrop(vp); 2583 } 2584 2585 /* 2586 * Add an item to the syncer work queue. 2587 */ 2588 static void 2589 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 2590 { 2591 int slot; 2592 2593 ASSERT_BO_WLOCKED(bo); 2594 2595 mtx_lock(&sync_mtx); 2596 if (bo->bo_flag & BO_ONWORKLST) 2597 LIST_REMOVE(bo, bo_synclist); 2598 else { 2599 bo->bo_flag |= BO_ONWORKLST; 2600 syncer_worklist_len++; 2601 } 2602 2603 if (delay > syncer_maxdelay - 2) 2604 delay = syncer_maxdelay - 2; 2605 slot = (syncer_delayno + delay) & syncer_mask; 2606 2607 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 2608 mtx_unlock(&sync_mtx); 2609 } 2610 2611 static int 2612 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 2613 { 2614 int error, len; 2615 2616 mtx_lock(&sync_mtx); 2617 len = syncer_worklist_len - sync_vnode_count; 2618 mtx_unlock(&sync_mtx); 2619 error = SYSCTL_OUT(req, &len, sizeof(len)); 2620 return (error); 2621 } 2622 2623 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, 2624 CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0, 2625 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 2626 2627 static struct proc *updateproc; 2628 static void sched_sync(void); 2629 static struct kproc_desc up_kp = { 2630 "syncer", 2631 sched_sync, 2632 &updateproc 2633 }; 2634 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); 2635 2636 static int 2637 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) 2638 { 2639 struct vnode *vp; 2640 struct mount *mp; 2641 2642 *bo = LIST_FIRST(slp); 2643 if (*bo == NULL) 2644 return (0); 2645 vp = bo2vnode(*bo); 2646 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) 2647 return (1); 2648 /* 2649 * We use vhold in case the vnode does not 2650 * successfully sync. vhold prevents the vnode from 2651 * going away when we unlock the sync_mtx so that 2652 * we can acquire the vnode interlock. 2653 */ 2654 vholdl(vp); 2655 mtx_unlock(&sync_mtx); 2656 VI_UNLOCK(vp); 2657 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2658 vdrop(vp); 2659 mtx_lock(&sync_mtx); 2660 return (*bo == LIST_FIRST(slp)); 2661 } 2662 MPASSERT(mp == NULL || (curthread->td_pflags & TDP_IGNSUSP) != 0 || 2663 (mp->mnt_kern_flag & MNTK_SUSPENDED) == 0, mp, 2664 ("suspended mp syncing vp %p", vp)); 2665 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2666 (void) VOP_FSYNC(vp, MNT_LAZY, td); 2667 VOP_UNLOCK(vp); 2668 vn_finished_write(mp); 2669 BO_LOCK(*bo); 2670 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { 2671 /* 2672 * Put us back on the worklist. The worklist 2673 * routine will remove us from our current 2674 * position and then add us back in at a later 2675 * position. 2676 */ 2677 vn_syncer_add_to_worklist(*bo, syncdelay); 2678 } 2679 BO_UNLOCK(*bo); 2680 vdrop(vp); 2681 mtx_lock(&sync_mtx); 2682 return (0); 2683 } 2684 2685 static int first_printf = 1; 2686 2687 /* 2688 * System filesystem synchronizer daemon. 2689 */ 2690 static void 2691 sched_sync(void) 2692 { 2693 struct synclist *next, *slp; 2694 struct bufobj *bo; 2695 long starttime; 2696 struct thread *td = curthread; 2697 int last_work_seen; 2698 int net_worklist_len; 2699 int syncer_final_iter; 2700 int error; 2701 2702 last_work_seen = 0; 2703 syncer_final_iter = 0; 2704 syncer_state = SYNCER_RUNNING; 2705 starttime = time_uptime; 2706 td->td_pflags |= TDP_NORUNNINGBUF; 2707 2708 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 2709 SHUTDOWN_PRI_LAST); 2710 2711 mtx_lock(&sync_mtx); 2712 for (;;) { 2713 if (syncer_state == SYNCER_FINAL_DELAY && 2714 syncer_final_iter == 0) { 2715 mtx_unlock(&sync_mtx); 2716 kproc_suspend_check(td->td_proc); 2717 mtx_lock(&sync_mtx); 2718 } 2719 net_worklist_len = syncer_worklist_len - sync_vnode_count; 2720 if (syncer_state != SYNCER_RUNNING && 2721 starttime != time_uptime) { 2722 if (first_printf) { 2723 printf("\nSyncing disks, vnodes remaining... "); 2724 first_printf = 0; 2725 } 2726 printf("%d ", net_worklist_len); 2727 } 2728 starttime = time_uptime; 2729 2730 /* 2731 * Push files whose dirty time has expired. Be careful 2732 * of interrupt race on slp queue. 2733 * 2734 * Skip over empty worklist slots when shutting down. 2735 */ 2736 do { 2737 slp = &syncer_workitem_pending[syncer_delayno]; 2738 syncer_delayno += 1; 2739 if (syncer_delayno == syncer_maxdelay) 2740 syncer_delayno = 0; 2741 next = &syncer_workitem_pending[syncer_delayno]; 2742 /* 2743 * If the worklist has wrapped since the 2744 * it was emptied of all but syncer vnodes, 2745 * switch to the FINAL_DELAY state and run 2746 * for one more second. 2747 */ 2748 if (syncer_state == SYNCER_SHUTTING_DOWN && 2749 net_worklist_len == 0 && 2750 last_work_seen == syncer_delayno) { 2751 syncer_state = SYNCER_FINAL_DELAY; 2752 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 2753 } 2754 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 2755 syncer_worklist_len > 0); 2756 2757 /* 2758 * Keep track of the last time there was anything 2759 * on the worklist other than syncer vnodes. 2760 * Return to the SHUTTING_DOWN state if any 2761 * new work appears. 2762 */ 2763 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 2764 last_work_seen = syncer_delayno; 2765 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 2766 syncer_state = SYNCER_SHUTTING_DOWN; 2767 while (!LIST_EMPTY(slp)) { 2768 error = sync_vnode(slp, &bo, td); 2769 if (error == 1) { 2770 LIST_REMOVE(bo, bo_synclist); 2771 LIST_INSERT_HEAD(next, bo, bo_synclist); 2772 continue; 2773 } 2774 2775 if (first_printf == 0) { 2776 /* 2777 * Drop the sync mutex, because some watchdog 2778 * drivers need to sleep while patting 2779 */ 2780 mtx_unlock(&sync_mtx); 2781 wdog_kern_pat(WD_LASTVAL); 2782 mtx_lock(&sync_mtx); 2783 } 2784 } 2785 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 2786 syncer_final_iter--; 2787 /* 2788 * The variable rushjob allows the kernel to speed up the 2789 * processing of the filesystem syncer process. A rushjob 2790 * value of N tells the filesystem syncer to process the next 2791 * N seconds worth of work on its queue ASAP. Currently rushjob 2792 * is used by the soft update code to speed up the filesystem 2793 * syncer process when the incore state is getting so far 2794 * ahead of the disk that the kernel memory pool is being 2795 * threatened with exhaustion. 2796 */ 2797 if (rushjob > 0) { 2798 rushjob -= 1; 2799 continue; 2800 } 2801 /* 2802 * Just sleep for a short period of time between 2803 * iterations when shutting down to allow some I/O 2804 * to happen. 2805 * 2806 * If it has taken us less than a second to process the 2807 * current work, then wait. Otherwise start right over 2808 * again. We can still lose time if any single round 2809 * takes more than two seconds, but it does not really 2810 * matter as we are just trying to generally pace the 2811 * filesystem activity. 2812 */ 2813 if (syncer_state != SYNCER_RUNNING || 2814 time_uptime == starttime) { 2815 thread_lock(td); 2816 sched_prio(td, PPAUSE); 2817 thread_unlock(td); 2818 } 2819 if (syncer_state != SYNCER_RUNNING) 2820 cv_timedwait(&sync_wakeup, &sync_mtx, 2821 hz / SYNCER_SHUTDOWN_SPEEDUP); 2822 else if (time_uptime == starttime) 2823 cv_timedwait(&sync_wakeup, &sync_mtx, hz); 2824 } 2825 } 2826 2827 /* 2828 * Request the syncer daemon to speed up its work. 2829 * We never push it to speed up more than half of its 2830 * normal turn time, otherwise it could take over the cpu. 2831 */ 2832 int 2833 speedup_syncer(void) 2834 { 2835 int ret = 0; 2836 2837 mtx_lock(&sync_mtx); 2838 if (rushjob < syncdelay / 2) { 2839 rushjob += 1; 2840 stat_rush_requests += 1; 2841 ret = 1; 2842 } 2843 mtx_unlock(&sync_mtx); 2844 cv_broadcast(&sync_wakeup); 2845 return (ret); 2846 } 2847 2848 /* 2849 * Tell the syncer to speed up its work and run though its work 2850 * list several times, then tell it to shut down. 2851 */ 2852 static void 2853 syncer_shutdown(void *arg, int howto) 2854 { 2855 2856 if (howto & RB_NOSYNC) 2857 return; 2858 mtx_lock(&sync_mtx); 2859 syncer_state = SYNCER_SHUTTING_DOWN; 2860 rushjob = 0; 2861 mtx_unlock(&sync_mtx); 2862 cv_broadcast(&sync_wakeup); 2863 kproc_shutdown(arg, howto); 2864 } 2865 2866 void 2867 syncer_suspend(void) 2868 { 2869 2870 syncer_shutdown(updateproc, 0); 2871 } 2872 2873 void 2874 syncer_resume(void) 2875 { 2876 2877 mtx_lock(&sync_mtx); 2878 first_printf = 1; 2879 syncer_state = SYNCER_RUNNING; 2880 mtx_unlock(&sync_mtx); 2881 cv_broadcast(&sync_wakeup); 2882 kproc_resume(updateproc); 2883 } 2884 2885 /* 2886 * Move the buffer between the clean and dirty lists of its vnode. 2887 */ 2888 void 2889 reassignbuf(struct buf *bp) 2890 { 2891 struct vnode *vp; 2892 struct bufobj *bo; 2893 int delay; 2894 #ifdef INVARIANTS 2895 struct bufv *bv; 2896 #endif 2897 2898 vp = bp->b_vp; 2899 bo = bp->b_bufobj; 2900 2901 KASSERT((bp->b_flags & B_PAGING) == 0, 2902 ("%s: cannot reassign paging buffer %p", __func__, bp)); 2903 2904 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 2905 bp, bp->b_vp, bp->b_flags); 2906 2907 BO_LOCK(bo); 2908 buf_vlist_remove(bp); 2909 2910 /* 2911 * If dirty, put on list of dirty buffers; otherwise insert onto list 2912 * of clean buffers. 2913 */ 2914 if (bp->b_flags & B_DELWRI) { 2915 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 2916 switch (vp->v_type) { 2917 case VDIR: 2918 delay = dirdelay; 2919 break; 2920 case VCHR: 2921 delay = metadelay; 2922 break; 2923 default: 2924 delay = filedelay; 2925 } 2926 vn_syncer_add_to_worklist(bo, delay); 2927 } 2928 buf_vlist_add(bp, bo, BX_VNDIRTY); 2929 } else { 2930 buf_vlist_add(bp, bo, BX_VNCLEAN); 2931 2932 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2933 mtx_lock(&sync_mtx); 2934 LIST_REMOVE(bo, bo_synclist); 2935 syncer_worklist_len--; 2936 mtx_unlock(&sync_mtx); 2937 bo->bo_flag &= ~BO_ONWORKLST; 2938 } 2939 } 2940 #ifdef INVARIANTS 2941 bv = &bo->bo_clean; 2942 bp = TAILQ_FIRST(&bv->bv_hd); 2943 KASSERT(bp == NULL || bp->b_bufobj == bo, 2944 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2945 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2946 KASSERT(bp == NULL || bp->b_bufobj == bo, 2947 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2948 bv = &bo->bo_dirty; 2949 bp = TAILQ_FIRST(&bv->bv_hd); 2950 KASSERT(bp == NULL || bp->b_bufobj == bo, 2951 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2952 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2953 KASSERT(bp == NULL || bp->b_bufobj == bo, 2954 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2955 #endif 2956 BO_UNLOCK(bo); 2957 } 2958 2959 static void 2960 v_init_counters(struct vnode *vp) 2961 { 2962 2963 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0, 2964 vp, ("%s called for an initialized vnode", __FUNCTION__)); 2965 ASSERT_VI_UNLOCKED(vp, __FUNCTION__); 2966 2967 refcount_init(&vp->v_holdcnt, 1); 2968 refcount_init(&vp->v_usecount, 1); 2969 } 2970 2971 /* 2972 * Grab a particular vnode from the free list, increment its 2973 * reference count and lock it. VIRF_DOOMED is set if the vnode 2974 * is being destroyed. Only callers who specify LK_RETRY will 2975 * see doomed vnodes. If inactive processing was delayed in 2976 * vput try to do it here. 2977 * 2978 * usecount is manipulated using atomics without holding any locks. 2979 * 2980 * holdcnt can be manipulated using atomics without holding any locks, 2981 * except when transitioning 1<->0, in which case the interlock is held. 2982 * 2983 * Consumers which don't guarantee liveness of the vnode can use SMR to 2984 * try to get a reference. Note this operation can fail since the vnode 2985 * may be awaiting getting freed by the time they get to it. 2986 */ 2987 enum vgetstate 2988 vget_prep_smr(struct vnode *vp) 2989 { 2990 enum vgetstate vs; 2991 2992 VFS_SMR_ASSERT_ENTERED(); 2993 2994 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2995 vs = VGET_USECOUNT; 2996 } else { 2997 if (vhold_smr(vp)) 2998 vs = VGET_HOLDCNT; 2999 else 3000 vs = VGET_NONE; 3001 } 3002 return (vs); 3003 } 3004 3005 enum vgetstate 3006 vget_prep(struct vnode *vp) 3007 { 3008 enum vgetstate vs; 3009 3010 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 3011 vs = VGET_USECOUNT; 3012 } else { 3013 vhold(vp); 3014 vs = VGET_HOLDCNT; 3015 } 3016 return (vs); 3017 } 3018 3019 void 3020 vget_abort(struct vnode *vp, enum vgetstate vs) 3021 { 3022 3023 switch (vs) { 3024 case VGET_USECOUNT: 3025 vrele(vp); 3026 break; 3027 case VGET_HOLDCNT: 3028 vdrop(vp); 3029 break; 3030 default: 3031 __assert_unreachable(); 3032 } 3033 } 3034 3035 int 3036 vget(struct vnode *vp, int flags) 3037 { 3038 enum vgetstate vs; 3039 3040 vs = vget_prep(vp); 3041 return (vget_finish(vp, flags, vs)); 3042 } 3043 3044 int 3045 vget_finish(struct vnode *vp, int flags, enum vgetstate vs) 3046 { 3047 int error; 3048 3049 if ((flags & LK_INTERLOCK) != 0) 3050 ASSERT_VI_LOCKED(vp, __func__); 3051 else 3052 ASSERT_VI_UNLOCKED(vp, __func__); 3053 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); 3054 VNPASS(vp->v_holdcnt > 0, vp); 3055 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); 3056 3057 error = vn_lock(vp, flags); 3058 if (__predict_false(error != 0)) { 3059 vget_abort(vp, vs); 3060 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, 3061 vp); 3062 return (error); 3063 } 3064 3065 vget_finish_ref(vp, vs); 3066 return (0); 3067 } 3068 3069 void 3070 vget_finish_ref(struct vnode *vp, enum vgetstate vs) 3071 { 3072 int old; 3073 3074 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); 3075 VNPASS(vp->v_holdcnt > 0, vp); 3076 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); 3077 3078 if (vs == VGET_USECOUNT) 3079 return; 3080 3081 /* 3082 * We hold the vnode. If the usecount is 0 it will be utilized to keep 3083 * the vnode around. Otherwise someone else lended their hold count and 3084 * we have to drop ours. 3085 */ 3086 old = atomic_fetchadd_int(&vp->v_usecount, 1); 3087 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old)); 3088 if (old != 0) { 3089 #ifdef INVARIANTS 3090 old = atomic_fetchadd_int(&vp->v_holdcnt, -1); 3091 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); 3092 #else 3093 refcount_release(&vp->v_holdcnt); 3094 #endif 3095 } 3096 } 3097 3098 void 3099 vref(struct vnode *vp) 3100 { 3101 enum vgetstate vs; 3102 3103 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3104 vs = vget_prep(vp); 3105 vget_finish_ref(vp, vs); 3106 } 3107 3108 void 3109 vrefact(struct vnode *vp) 3110 { 3111 3112 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3113 #ifdef INVARIANTS 3114 int old = atomic_fetchadd_int(&vp->v_usecount, 1); 3115 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); 3116 #else 3117 refcount_acquire(&vp->v_usecount); 3118 #endif 3119 } 3120 3121 void 3122 vlazy(struct vnode *vp) 3123 { 3124 struct mount *mp; 3125 3126 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__)); 3127 3128 if ((vp->v_mflag & VMP_LAZYLIST) != 0) 3129 return; 3130 /* 3131 * We may get here for inactive routines after the vnode got doomed. 3132 */ 3133 if (VN_IS_DOOMED(vp)) 3134 return; 3135 mp = vp->v_mount; 3136 mtx_lock(&mp->mnt_listmtx); 3137 if ((vp->v_mflag & VMP_LAZYLIST) == 0) { 3138 vp->v_mflag |= VMP_LAZYLIST; 3139 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3140 mp->mnt_lazyvnodelistsize++; 3141 } 3142 mtx_unlock(&mp->mnt_listmtx); 3143 } 3144 3145 static void 3146 vunlazy(struct vnode *vp) 3147 { 3148 struct mount *mp; 3149 3150 ASSERT_VI_LOCKED(vp, __func__); 3151 VNPASS(!VN_IS_DOOMED(vp), vp); 3152 3153 mp = vp->v_mount; 3154 mtx_lock(&mp->mnt_listmtx); 3155 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 3156 /* 3157 * Don't remove the vnode from the lazy list if another thread 3158 * has increased the hold count. It may have re-enqueued the 3159 * vnode to the lazy list and is now responsible for its 3160 * removal. 3161 */ 3162 if (vp->v_holdcnt == 0) { 3163 vp->v_mflag &= ~VMP_LAZYLIST; 3164 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3165 mp->mnt_lazyvnodelistsize--; 3166 } 3167 mtx_unlock(&mp->mnt_listmtx); 3168 } 3169 3170 /* 3171 * This routine is only meant to be called from vgonel prior to dooming 3172 * the vnode. 3173 */ 3174 static void 3175 vunlazy_gone(struct vnode *vp) 3176 { 3177 struct mount *mp; 3178 3179 ASSERT_VOP_ELOCKED(vp, __func__); 3180 ASSERT_VI_LOCKED(vp, __func__); 3181 VNPASS(!VN_IS_DOOMED(vp), vp); 3182 3183 if (vp->v_mflag & VMP_LAZYLIST) { 3184 mp = vp->v_mount; 3185 mtx_lock(&mp->mnt_listmtx); 3186 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 3187 vp->v_mflag &= ~VMP_LAZYLIST; 3188 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3189 mp->mnt_lazyvnodelistsize--; 3190 mtx_unlock(&mp->mnt_listmtx); 3191 } 3192 } 3193 3194 static void 3195 vdefer_inactive(struct vnode *vp) 3196 { 3197 3198 ASSERT_VI_LOCKED(vp, __func__); 3199 VNPASS(vp->v_holdcnt > 0, vp); 3200 if (VN_IS_DOOMED(vp)) { 3201 vdropl(vp); 3202 return; 3203 } 3204 if (vp->v_iflag & VI_DEFINACT) { 3205 VNPASS(vp->v_holdcnt > 1, vp); 3206 vdropl(vp); 3207 return; 3208 } 3209 if (vp->v_usecount > 0) { 3210 vp->v_iflag &= ~VI_OWEINACT; 3211 vdropl(vp); 3212 return; 3213 } 3214 vlazy(vp); 3215 vp->v_iflag |= VI_DEFINACT; 3216 VI_UNLOCK(vp); 3217 atomic_add_long(&deferred_inact, 1); 3218 } 3219 3220 static void 3221 vdefer_inactive_unlocked(struct vnode *vp) 3222 { 3223 3224 VI_LOCK(vp); 3225 if ((vp->v_iflag & VI_OWEINACT) == 0) { 3226 vdropl(vp); 3227 return; 3228 } 3229 vdefer_inactive(vp); 3230 } 3231 3232 enum vput_op { VRELE, VPUT, VUNREF }; 3233 3234 /* 3235 * Handle ->v_usecount transitioning to 0. 3236 * 3237 * By releasing the last usecount we take ownership of the hold count which 3238 * provides liveness of the vnode, meaning we have to vdrop. 3239 * 3240 * For all vnodes we may need to perform inactive processing. It requires an 3241 * exclusive lock on the vnode, while it is legal to call here with only a 3242 * shared lock (or no locks). If locking the vnode in an expected manner fails, 3243 * inactive processing gets deferred to the syncer. 3244 * 3245 * XXX Some filesystems pass in an exclusively locked vnode and strongly depend 3246 * on the lock being held all the way until VOP_INACTIVE. This in particular 3247 * happens with UFS which adds half-constructed vnodes to the hash, where they 3248 * can be found by other code. 3249 */ 3250 static void 3251 vput_final(struct vnode *vp, enum vput_op func) 3252 { 3253 int error; 3254 bool want_unlock; 3255 3256 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3257 VNPASS(vp->v_holdcnt > 0, vp); 3258 3259 VI_LOCK(vp); 3260 3261 /* 3262 * By the time we got here someone else might have transitioned 3263 * the count back to > 0. 3264 */ 3265 if (vp->v_usecount > 0) 3266 goto out; 3267 3268 /* 3269 * If the vnode is doomed vgone already performed inactive processing 3270 * (if needed). 3271 */ 3272 if (VN_IS_DOOMED(vp)) 3273 goto out; 3274 3275 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0)) 3276 goto out; 3277 3278 if (vp->v_iflag & VI_DOINGINACT) 3279 goto out; 3280 3281 /* 3282 * Locking operations here will drop the interlock and possibly the 3283 * vnode lock, opening a window where the vnode can get doomed all the 3284 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to 3285 * perform inactive. 3286 */ 3287 vp->v_iflag |= VI_OWEINACT; 3288 want_unlock = false; 3289 error = 0; 3290 switch (func) { 3291 case VRELE: 3292 switch (VOP_ISLOCKED(vp)) { 3293 case LK_EXCLUSIVE: 3294 break; 3295 case LK_EXCLOTHER: 3296 case 0: 3297 want_unlock = true; 3298 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 3299 VI_LOCK(vp); 3300 break; 3301 default: 3302 /* 3303 * The lock has at least one sharer, but we have no way 3304 * to conclude whether this is us. Play it safe and 3305 * defer processing. 3306 */ 3307 error = EAGAIN; 3308 break; 3309 } 3310 break; 3311 case VPUT: 3312 want_unlock = true; 3313 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3314 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | 3315 LK_NOWAIT); 3316 VI_LOCK(vp); 3317 } 3318 break; 3319 case VUNREF: 3320 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3321 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); 3322 VI_LOCK(vp); 3323 } 3324 break; 3325 } 3326 if (error == 0) { 3327 if (func == VUNREF) { 3328 VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp, 3329 ("recursive vunref")); 3330 vp->v_vflag |= VV_UNREF; 3331 } 3332 for (;;) { 3333 error = vinactive(vp); 3334 if (want_unlock) 3335 VOP_UNLOCK(vp); 3336 if (error != ERELOOKUP || !want_unlock) 3337 break; 3338 VOP_LOCK(vp, LK_EXCLUSIVE); 3339 } 3340 if (func == VUNREF) 3341 vp->v_vflag &= ~VV_UNREF; 3342 vdropl(vp); 3343 } else { 3344 vdefer_inactive(vp); 3345 } 3346 return; 3347 out: 3348 if (func == VPUT) 3349 VOP_UNLOCK(vp); 3350 vdropl(vp); 3351 } 3352 3353 /* 3354 * Decrement ->v_usecount for a vnode. 3355 * 3356 * Releasing the last use count requires additional processing, see vput_final 3357 * above for details. 3358 * 3359 * Comment above each variant denotes lock state on entry and exit. 3360 */ 3361 3362 /* 3363 * in: any 3364 * out: same as passed in 3365 */ 3366 void 3367 vrele(struct vnode *vp) 3368 { 3369 3370 ASSERT_VI_UNLOCKED(vp, __func__); 3371 if (!refcount_release(&vp->v_usecount)) 3372 return; 3373 vput_final(vp, VRELE); 3374 } 3375 3376 /* 3377 * in: locked 3378 * out: unlocked 3379 */ 3380 void 3381 vput(struct vnode *vp) 3382 { 3383 3384 ASSERT_VOP_LOCKED(vp, __func__); 3385 ASSERT_VI_UNLOCKED(vp, __func__); 3386 if (!refcount_release(&vp->v_usecount)) { 3387 VOP_UNLOCK(vp); 3388 return; 3389 } 3390 vput_final(vp, VPUT); 3391 } 3392 3393 /* 3394 * in: locked 3395 * out: locked 3396 */ 3397 void 3398 vunref(struct vnode *vp) 3399 { 3400 3401 ASSERT_VOP_LOCKED(vp, __func__); 3402 ASSERT_VI_UNLOCKED(vp, __func__); 3403 if (!refcount_release(&vp->v_usecount)) 3404 return; 3405 vput_final(vp, VUNREF); 3406 } 3407 3408 void 3409 vhold(struct vnode *vp) 3410 { 3411 int old; 3412 3413 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3414 old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3415 VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3416 ("%s: wrong hold count %d", __func__, old)); 3417 if (old == 0) 3418 vfs_freevnodes_dec(); 3419 } 3420 3421 void 3422 vholdnz(struct vnode *vp) 3423 { 3424 3425 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3426 #ifdef INVARIANTS 3427 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3428 VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3429 ("%s: wrong hold count %d", __func__, old)); 3430 #else 3431 atomic_add_int(&vp->v_holdcnt, 1); 3432 #endif 3433 } 3434 3435 /* 3436 * Grab a hold count unless the vnode is freed. 3437 * 3438 * Only use this routine if vfs smr is the only protection you have against 3439 * freeing the vnode. 3440 * 3441 * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag 3442 * is not set. After the flag is set the vnode becomes immutable to anyone but 3443 * the thread which managed to set the flag. 3444 * 3445 * It may be tempting to replace the loop with: 3446 * count = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3447 * if (count & VHOLD_NO_SMR) { 3448 * backpedal and error out; 3449 * } 3450 * 3451 * However, while this is more performant, it hinders debugging by eliminating 3452 * the previously mentioned invariant. 3453 */ 3454 bool 3455 vhold_smr(struct vnode *vp) 3456 { 3457 int count; 3458 3459 VFS_SMR_ASSERT_ENTERED(); 3460 3461 count = atomic_load_int(&vp->v_holdcnt); 3462 for (;;) { 3463 if (count & VHOLD_NO_SMR) { 3464 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, 3465 ("non-zero hold count with flags %d\n", count)); 3466 return (false); 3467 } 3468 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); 3469 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) { 3470 if (count == 0) 3471 vfs_freevnodes_dec(); 3472 return (true); 3473 } 3474 } 3475 } 3476 3477 /* 3478 * Hold a free vnode for recycling. 3479 * 3480 * Note: vnode_init references this comment. 3481 * 3482 * Attempts to recycle only need the global vnode list lock and have no use for 3483 * SMR. 3484 * 3485 * However, vnodes get inserted into the global list before they get fully 3486 * initialized and stay there until UMA decides to free the memory. This in 3487 * particular means the target can be found before it becomes usable and after 3488 * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to 3489 * VHOLD_NO_SMR. 3490 * 3491 * Note: the vnode may gain more references after we transition the count 0->1. 3492 */ 3493 static bool 3494 vhold_recycle_free(struct vnode *vp) 3495 { 3496 int count; 3497 3498 mtx_assert(&vnode_list_mtx, MA_OWNED); 3499 3500 count = atomic_load_int(&vp->v_holdcnt); 3501 for (;;) { 3502 if (count & VHOLD_NO_SMR) { 3503 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, 3504 ("non-zero hold count with flags %d\n", count)); 3505 return (false); 3506 } 3507 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); 3508 if (count > 0) { 3509 return (false); 3510 } 3511 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) { 3512 vfs_freevnodes_dec(); 3513 return (true); 3514 } 3515 } 3516 } 3517 3518 static void __noinline 3519 vdbatch_process(struct vdbatch *vd) 3520 { 3521 struct vnode *vp; 3522 int i; 3523 3524 mtx_assert(&vd->lock, MA_OWNED); 3525 MPASS(curthread->td_pinned > 0); 3526 MPASS(vd->index == VDBATCH_SIZE); 3527 3528 critical_enter(); 3529 if (mtx_trylock(&vnode_list_mtx)) { 3530 for (i = 0; i < VDBATCH_SIZE; i++) { 3531 vp = vd->tab[i]; 3532 vd->tab[i] = NULL; 3533 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 3534 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist); 3535 MPASS(vp->v_dbatchcpu != NOCPU); 3536 vp->v_dbatchcpu = NOCPU; 3537 } 3538 mtx_unlock(&vnode_list_mtx); 3539 } else { 3540 for (i = 0; i < VDBATCH_SIZE; i++) { 3541 vp = vd->tab[i]; 3542 vd->tab[i] = NULL; 3543 MPASS(vp->v_dbatchcpu != NOCPU); 3544 vp->v_dbatchcpu = NOCPU; 3545 } 3546 } 3547 vd->index = 0; 3548 critical_exit(); 3549 } 3550 3551 static void 3552 vdbatch_enqueue(struct vnode *vp) 3553 { 3554 struct vdbatch *vd; 3555 3556 ASSERT_VI_LOCKED(vp, __func__); 3557 VNPASS(!VN_IS_DOOMED(vp), vp); 3558 3559 if (vp->v_dbatchcpu != NOCPU) { 3560 VI_UNLOCK(vp); 3561 return; 3562 } 3563 3564 sched_pin(); 3565 vd = DPCPU_PTR(vd); 3566 mtx_lock(&vd->lock); 3567 MPASS(vd->index < VDBATCH_SIZE); 3568 MPASS(vd->tab[vd->index] == NULL); 3569 /* 3570 * A hack: we depend on being pinned so that we know what to put in 3571 * ->v_dbatchcpu. 3572 */ 3573 vp->v_dbatchcpu = curcpu; 3574 vd->tab[vd->index] = vp; 3575 vd->index++; 3576 VI_UNLOCK(vp); 3577 if (vd->index == VDBATCH_SIZE) 3578 vdbatch_process(vd); 3579 mtx_unlock(&vd->lock); 3580 sched_unpin(); 3581 } 3582 3583 /* 3584 * This routine must only be called for vnodes which are about to be 3585 * deallocated. Supporting dequeue for arbitrary vndoes would require 3586 * validating that the locked batch matches. 3587 */ 3588 static void 3589 vdbatch_dequeue(struct vnode *vp) 3590 { 3591 struct vdbatch *vd; 3592 int i; 3593 short cpu; 3594 3595 VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp); 3596 3597 cpu = vp->v_dbatchcpu; 3598 if (cpu == NOCPU) 3599 return; 3600 3601 vd = DPCPU_ID_PTR(cpu, vd); 3602 mtx_lock(&vd->lock); 3603 for (i = 0; i < vd->index; i++) { 3604 if (vd->tab[i] != vp) 3605 continue; 3606 vp->v_dbatchcpu = NOCPU; 3607 vd->index--; 3608 vd->tab[i] = vd->tab[vd->index]; 3609 vd->tab[vd->index] = NULL; 3610 break; 3611 } 3612 mtx_unlock(&vd->lock); 3613 /* 3614 * Either we dequeued the vnode above or the target CPU beat us to it. 3615 */ 3616 MPASS(vp->v_dbatchcpu == NOCPU); 3617 } 3618 3619 /* 3620 * Drop the hold count of the vnode. If this is the last reference to 3621 * the vnode we place it on the free list unless it has been vgone'd 3622 * (marked VIRF_DOOMED) in which case we will free it. 3623 * 3624 * Because the vnode vm object keeps a hold reference on the vnode if 3625 * there is at least one resident non-cached page, the vnode cannot 3626 * leave the active list without the page cleanup done. 3627 */ 3628 static void __noinline 3629 vdropl_final(struct vnode *vp) 3630 { 3631 3632 ASSERT_VI_LOCKED(vp, __func__); 3633 VNPASS(VN_IS_DOOMED(vp), vp); 3634 /* 3635 * Set the VHOLD_NO_SMR flag. 3636 * 3637 * We may be racing against vhold_smr. If they win we can just pretend 3638 * we never got this far, they will vdrop later. 3639 */ 3640 if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) { 3641 vfs_freevnodes_inc(); 3642 VI_UNLOCK(vp); 3643 /* 3644 * We lost the aforementioned race. Any subsequent access is 3645 * invalid as they might have managed to vdropl on their own. 3646 */ 3647 return; 3648 } 3649 /* 3650 * Don't bump freevnodes as this one is going away. 3651 */ 3652 freevnode(vp); 3653 } 3654 3655 void 3656 vdrop(struct vnode *vp) 3657 { 3658 3659 ASSERT_VI_UNLOCKED(vp, __func__); 3660 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3661 if (refcount_release_if_not_last(&vp->v_holdcnt)) 3662 return; 3663 VI_LOCK(vp); 3664 vdropl(vp); 3665 } 3666 3667 static void __always_inline 3668 vdropl_impl(struct vnode *vp, bool enqueue) 3669 { 3670 3671 ASSERT_VI_LOCKED(vp, __func__); 3672 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3673 if (!refcount_release(&vp->v_holdcnt)) { 3674 VI_UNLOCK(vp); 3675 return; 3676 } 3677 VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp); 3678 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp); 3679 if (VN_IS_DOOMED(vp)) { 3680 vdropl_final(vp); 3681 return; 3682 } 3683 3684 vfs_freevnodes_inc(); 3685 if (vp->v_mflag & VMP_LAZYLIST) { 3686 vunlazy(vp); 3687 } 3688 3689 if (!enqueue) { 3690 VI_UNLOCK(vp); 3691 return; 3692 } 3693 3694 /* 3695 * Also unlocks the interlock. We can't assert on it as we 3696 * released our hold and by now the vnode might have been 3697 * freed. 3698 */ 3699 vdbatch_enqueue(vp); 3700 } 3701 3702 void 3703 vdropl(struct vnode *vp) 3704 { 3705 3706 vdropl_impl(vp, true); 3707 } 3708 3709 /* 3710 * vdrop a vnode when recycling 3711 * 3712 * This is a special case routine only to be used when recycling, differs from 3713 * regular vdrop by not requeieing the vnode on LRU. 3714 * 3715 * Consider a case where vtryrecycle continuously fails with all vnodes (due to 3716 * e.g., frozen writes on the filesystem), filling the batch and causing it to 3717 * be requeued. Then vnlru will end up revisiting the same vnodes. This is a 3718 * loop which can last for as long as writes are frozen. 3719 */ 3720 static void 3721 vdropl_recycle(struct vnode *vp) 3722 { 3723 3724 vdropl_impl(vp, false); 3725 } 3726 3727 static void 3728 vdrop_recycle(struct vnode *vp) 3729 { 3730 3731 VI_LOCK(vp); 3732 vdropl_recycle(vp); 3733 } 3734 3735 /* 3736 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 3737 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 3738 */ 3739 static int 3740 vinactivef(struct vnode *vp) 3741 { 3742 struct vm_object *obj; 3743 int error; 3744 3745 ASSERT_VOP_ELOCKED(vp, "vinactive"); 3746 ASSERT_VI_LOCKED(vp, "vinactive"); 3747 VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp); 3748 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3749 vp->v_iflag |= VI_DOINGINACT; 3750 vp->v_iflag &= ~VI_OWEINACT; 3751 VI_UNLOCK(vp); 3752 /* 3753 * Before moving off the active list, we must be sure that any 3754 * modified pages are converted into the vnode's dirty 3755 * buffers, since these will no longer be checked once the 3756 * vnode is on the inactive list. 3757 * 3758 * The write-out of the dirty pages is asynchronous. At the 3759 * point that VOP_INACTIVE() is called, there could still be 3760 * pending I/O and dirty pages in the object. 3761 */ 3762 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 3763 vm_object_mightbedirty(obj)) { 3764 VM_OBJECT_WLOCK(obj); 3765 vm_object_page_clean(obj, 0, 0, 0); 3766 VM_OBJECT_WUNLOCK(obj); 3767 } 3768 error = VOP_INACTIVE(vp); 3769 VI_LOCK(vp); 3770 VNPASS(vp->v_iflag & VI_DOINGINACT, vp); 3771 vp->v_iflag &= ~VI_DOINGINACT; 3772 return (error); 3773 } 3774 3775 int 3776 vinactive(struct vnode *vp) 3777 { 3778 3779 ASSERT_VOP_ELOCKED(vp, "vinactive"); 3780 ASSERT_VI_LOCKED(vp, "vinactive"); 3781 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3782 3783 if ((vp->v_iflag & VI_OWEINACT) == 0) 3784 return (0); 3785 if (vp->v_iflag & VI_DOINGINACT) 3786 return (0); 3787 if (vp->v_usecount > 0) { 3788 vp->v_iflag &= ~VI_OWEINACT; 3789 return (0); 3790 } 3791 return (vinactivef(vp)); 3792 } 3793 3794 /* 3795 * Remove any vnodes in the vnode table belonging to mount point mp. 3796 * 3797 * If FORCECLOSE is not specified, there should not be any active ones, 3798 * return error if any are found (nb: this is a user error, not a 3799 * system error). If FORCECLOSE is specified, detach any active vnodes 3800 * that are found. 3801 * 3802 * If WRITECLOSE is set, only flush out regular file vnodes open for 3803 * writing. 3804 * 3805 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 3806 * 3807 * `rootrefs' specifies the base reference count for the root vnode 3808 * of this filesystem. The root vnode is considered busy if its 3809 * v_usecount exceeds this value. On a successful return, vflush(, td) 3810 * will call vrele() on the root vnode exactly rootrefs times. 3811 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 3812 * be zero. 3813 */ 3814 #ifdef DIAGNOSTIC 3815 static int busyprt = 0; /* print out busy vnodes */ 3816 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 3817 #endif 3818 3819 int 3820 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 3821 { 3822 struct vnode *vp, *mvp, *rootvp = NULL; 3823 struct vattr vattr; 3824 int busy = 0, error; 3825 3826 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 3827 rootrefs, flags); 3828 if (rootrefs > 0) { 3829 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 3830 ("vflush: bad args")); 3831 /* 3832 * Get the filesystem root vnode. We can vput() it 3833 * immediately, since with rootrefs > 0, it won't go away. 3834 */ 3835 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 3836 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 3837 __func__, error); 3838 return (error); 3839 } 3840 vput(rootvp); 3841 } 3842 loop: 3843 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 3844 vholdl(vp); 3845 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 3846 if (error) { 3847 vdrop(vp); 3848 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3849 goto loop; 3850 } 3851 /* 3852 * Skip over a vnodes marked VV_SYSTEM. 3853 */ 3854 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 3855 VOP_UNLOCK(vp); 3856 vdrop(vp); 3857 continue; 3858 } 3859 /* 3860 * If WRITECLOSE is set, flush out unlinked but still open 3861 * files (even if open only for reading) and regular file 3862 * vnodes open for writing. 3863 */ 3864 if (flags & WRITECLOSE) { 3865 if (vp->v_object != NULL) { 3866 VM_OBJECT_WLOCK(vp->v_object); 3867 vm_object_page_clean(vp->v_object, 0, 0, 0); 3868 VM_OBJECT_WUNLOCK(vp->v_object); 3869 } 3870 do { 3871 error = VOP_FSYNC(vp, MNT_WAIT, td); 3872 } while (error == ERELOOKUP); 3873 if (error != 0) { 3874 VOP_UNLOCK(vp); 3875 vdrop(vp); 3876 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3877 return (error); 3878 } 3879 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 3880 VI_LOCK(vp); 3881 3882 if ((vp->v_type == VNON || 3883 (error == 0 && vattr.va_nlink > 0)) && 3884 (vp->v_writecount <= 0 || vp->v_type != VREG)) { 3885 VOP_UNLOCK(vp); 3886 vdropl(vp); 3887 continue; 3888 } 3889 } else 3890 VI_LOCK(vp); 3891 /* 3892 * With v_usecount == 0, all we need to do is clear out the 3893 * vnode data structures and we are done. 3894 * 3895 * If FORCECLOSE is set, forcibly close the vnode. 3896 */ 3897 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 3898 vgonel(vp); 3899 } else { 3900 busy++; 3901 #ifdef DIAGNOSTIC 3902 if (busyprt) 3903 vn_printf(vp, "vflush: busy vnode "); 3904 #endif 3905 } 3906 VOP_UNLOCK(vp); 3907 vdropl(vp); 3908 } 3909 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 3910 /* 3911 * If just the root vnode is busy, and if its refcount 3912 * is equal to `rootrefs', then go ahead and kill it. 3913 */ 3914 VI_LOCK(rootvp); 3915 KASSERT(busy > 0, ("vflush: not busy")); 3916 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 3917 ("vflush: usecount %d < rootrefs %d", 3918 rootvp->v_usecount, rootrefs)); 3919 if (busy == 1 && rootvp->v_usecount == rootrefs) { 3920 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 3921 vgone(rootvp); 3922 VOP_UNLOCK(rootvp); 3923 busy = 0; 3924 } else 3925 VI_UNLOCK(rootvp); 3926 } 3927 if (busy) { 3928 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 3929 busy); 3930 return (EBUSY); 3931 } 3932 for (; rootrefs > 0; rootrefs--) 3933 vrele(rootvp); 3934 return (0); 3935 } 3936 3937 /* 3938 * Recycle an unused vnode to the front of the free list. 3939 */ 3940 int 3941 vrecycle(struct vnode *vp) 3942 { 3943 int recycled; 3944 3945 VI_LOCK(vp); 3946 recycled = vrecyclel(vp); 3947 VI_UNLOCK(vp); 3948 return (recycled); 3949 } 3950 3951 /* 3952 * vrecycle, with the vp interlock held. 3953 */ 3954 int 3955 vrecyclel(struct vnode *vp) 3956 { 3957 int recycled; 3958 3959 ASSERT_VOP_ELOCKED(vp, __func__); 3960 ASSERT_VI_LOCKED(vp, __func__); 3961 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3962 recycled = 0; 3963 if (vp->v_usecount == 0) { 3964 recycled = 1; 3965 vgonel(vp); 3966 } 3967 return (recycled); 3968 } 3969 3970 /* 3971 * Eliminate all activity associated with a vnode 3972 * in preparation for reuse. 3973 */ 3974 void 3975 vgone(struct vnode *vp) 3976 { 3977 VI_LOCK(vp); 3978 vgonel(vp); 3979 VI_UNLOCK(vp); 3980 } 3981 3982 /* 3983 * Notify upper mounts about reclaimed or unlinked vnode. 3984 */ 3985 void 3986 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event) 3987 { 3988 struct mount *mp; 3989 struct mount_upper_node *ump; 3990 3991 mp = atomic_load_ptr(&vp->v_mount); 3992 if (mp == NULL) 3993 return; 3994 if (TAILQ_EMPTY(&mp->mnt_notify)) 3995 return; 3996 3997 MNT_ILOCK(mp); 3998 mp->mnt_upper_pending++; 3999 KASSERT(mp->mnt_upper_pending > 0, 4000 ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending)); 4001 TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) { 4002 MNT_IUNLOCK(mp); 4003 switch (event) { 4004 case VFS_NOTIFY_UPPER_RECLAIM: 4005 VFS_RECLAIM_LOWERVP(ump->mp, vp); 4006 break; 4007 case VFS_NOTIFY_UPPER_UNLINK: 4008 VFS_UNLINK_LOWERVP(ump->mp, vp); 4009 break; 4010 } 4011 MNT_ILOCK(mp); 4012 } 4013 mp->mnt_upper_pending--; 4014 if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 && 4015 mp->mnt_upper_pending == 0) { 4016 mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER; 4017 wakeup(&mp->mnt_uppers); 4018 } 4019 MNT_IUNLOCK(mp); 4020 } 4021 4022 /* 4023 * vgone, with the vp interlock held. 4024 */ 4025 static void 4026 vgonel(struct vnode *vp) 4027 { 4028 struct thread *td; 4029 struct mount *mp; 4030 vm_object_t object; 4031 bool active, doinginact, oweinact; 4032 4033 ASSERT_VOP_ELOCKED(vp, "vgonel"); 4034 ASSERT_VI_LOCKED(vp, "vgonel"); 4035 VNASSERT(vp->v_holdcnt, vp, 4036 ("vgonel: vp %p has no reference.", vp)); 4037 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 4038 td = curthread; 4039 4040 /* 4041 * Don't vgonel if we're already doomed. 4042 */ 4043 if (VN_IS_DOOMED(vp)) { 4044 VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \ 4045 vn_get_state(vp) == VSTATE_DEAD, vp); 4046 return; 4047 } 4048 /* 4049 * Paired with freevnode. 4050 */ 4051 vn_seqc_write_begin_locked(vp); 4052 vunlazy_gone(vp); 4053 vn_irflag_set_locked(vp, VIRF_DOOMED); 4054 vn_set_state(vp, VSTATE_DESTROYING); 4055 4056 /* 4057 * Check to see if the vnode is in use. If so, we have to 4058 * call VOP_CLOSE() and VOP_INACTIVE(). 4059 * 4060 * It could be that VOP_INACTIVE() requested reclamation, in 4061 * which case we should avoid recursion, so check 4062 * VI_DOINGINACT. This is not precise but good enough. 4063 */ 4064 active = vp->v_usecount > 0; 4065 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 4066 doinginact = (vp->v_iflag & VI_DOINGINACT) != 0; 4067 4068 /* 4069 * If we need to do inactive VI_OWEINACT will be set. 4070 */ 4071 if (vp->v_iflag & VI_DEFINACT) { 4072 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 4073 vp->v_iflag &= ~VI_DEFINACT; 4074 vdropl(vp); 4075 } else { 4076 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); 4077 VI_UNLOCK(vp); 4078 } 4079 cache_purge_vgone(vp); 4080 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); 4081 4082 /* 4083 * If purging an active vnode, it must be closed and 4084 * deactivated before being reclaimed. 4085 */ 4086 if (active) 4087 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 4088 if (!doinginact) { 4089 do { 4090 if (oweinact || active) { 4091 VI_LOCK(vp); 4092 vinactivef(vp); 4093 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 4094 VI_UNLOCK(vp); 4095 } 4096 } while (oweinact); 4097 } 4098 if (vp->v_type == VSOCK) 4099 vfs_unp_reclaim(vp); 4100 4101 /* 4102 * Clean out any buffers associated with the vnode. 4103 * If the flush fails, just toss the buffers. 4104 */ 4105 mp = NULL; 4106 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 4107 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 4108 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { 4109 while (vinvalbuf(vp, 0, 0, 0) != 0) 4110 ; 4111 } 4112 4113 BO_LOCK(&vp->v_bufobj); 4114 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && 4115 vp->v_bufobj.bo_dirty.bv_cnt == 0 && 4116 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && 4117 vp->v_bufobj.bo_clean.bv_cnt == 0, 4118 ("vp %p bufobj not invalidated", vp)); 4119 4120 /* 4121 * For VMIO bufobj, BO_DEAD is set later, or in 4122 * vm_object_terminate() after the object's page queue is 4123 * flushed. 4124 */ 4125 object = vp->v_bufobj.bo_object; 4126 if (object == NULL) 4127 vp->v_bufobj.bo_flag |= BO_DEAD; 4128 BO_UNLOCK(&vp->v_bufobj); 4129 4130 /* 4131 * Handle the VM part. Tmpfs handles v_object on its own (the 4132 * OBJT_VNODE check). Nullfs or other bypassing filesystems 4133 * should not touch the object borrowed from the lower vnode 4134 * (the handle check). 4135 */ 4136 if (object != NULL && object->type == OBJT_VNODE && 4137 object->handle == vp) 4138 vnode_destroy_vobject(vp); 4139 4140 /* 4141 * Reclaim the vnode. 4142 */ 4143 if (VOP_RECLAIM(vp)) 4144 panic("vgone: cannot reclaim"); 4145 if (mp != NULL) 4146 vn_finished_secondary_write(mp); 4147 VNASSERT(vp->v_object == NULL, vp, 4148 ("vop_reclaim left v_object vp=%p", vp)); 4149 /* 4150 * Clear the advisory locks and wake up waiting threads. 4151 */ 4152 if (vp->v_lockf != NULL) { 4153 (void)VOP_ADVLOCKPURGE(vp); 4154 vp->v_lockf = NULL; 4155 } 4156 /* 4157 * Delete from old mount point vnode list. 4158 */ 4159 if (vp->v_mount == NULL) { 4160 VI_LOCK(vp); 4161 } else { 4162 delmntque(vp); 4163 ASSERT_VI_LOCKED(vp, "vgonel 2"); 4164 } 4165 /* 4166 * Done with purge, reset to the standard lock and invalidate 4167 * the vnode. 4168 */ 4169 vp->v_vnlock = &vp->v_lock; 4170 vp->v_op = &dead_vnodeops; 4171 vp->v_type = VBAD; 4172 vn_set_state(vp, VSTATE_DEAD); 4173 } 4174 4175 /* 4176 * Print out a description of a vnode. 4177 */ 4178 static const char *const vtypename[] = { 4179 [VNON] = "VNON", 4180 [VREG] = "VREG", 4181 [VDIR] = "VDIR", 4182 [VBLK] = "VBLK", 4183 [VCHR] = "VCHR", 4184 [VLNK] = "VLNK", 4185 [VSOCK] = "VSOCK", 4186 [VFIFO] = "VFIFO", 4187 [VBAD] = "VBAD", 4188 [VMARKER] = "VMARKER", 4189 }; 4190 _Static_assert(nitems(vtypename) == VLASTTYPE + 1, 4191 "vnode type name not added to vtypename"); 4192 4193 static const char *const vstatename[] = { 4194 [VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED", 4195 [VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED", 4196 [VSTATE_DESTROYING] = "VSTATE_DESTROYING", 4197 [VSTATE_DEAD] = "VSTATE_DEAD", 4198 }; 4199 _Static_assert(nitems(vstatename) == VLASTSTATE + 1, 4200 "vnode state name not added to vstatename"); 4201 4202 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0, 4203 "new hold count flag not added to vn_printf"); 4204 4205 void 4206 vn_printf(struct vnode *vp, const char *fmt, ...) 4207 { 4208 va_list ap; 4209 char buf[256], buf2[16]; 4210 u_long flags; 4211 u_int holdcnt; 4212 short irflag; 4213 4214 va_start(ap, fmt); 4215 vprintf(fmt, ap); 4216 va_end(ap); 4217 printf("%p: ", (void *)vp); 4218 printf("type %s state %s op %p\n", vtypename[vp->v_type], 4219 vstatename[vp->v_state], vp->v_op); 4220 holdcnt = atomic_load_int(&vp->v_holdcnt); 4221 printf(" usecount %d, writecount %d, refcount %d seqc users %d", 4222 vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS, 4223 vp->v_seqc_users); 4224 switch (vp->v_type) { 4225 case VDIR: 4226 printf(" mountedhere %p\n", vp->v_mountedhere); 4227 break; 4228 case VCHR: 4229 printf(" rdev %p\n", vp->v_rdev); 4230 break; 4231 case VSOCK: 4232 printf(" socket %p\n", vp->v_unpcb); 4233 break; 4234 case VFIFO: 4235 printf(" fifoinfo %p\n", vp->v_fifoinfo); 4236 break; 4237 default: 4238 printf("\n"); 4239 break; 4240 } 4241 buf[0] = '\0'; 4242 buf[1] = '\0'; 4243 if (holdcnt & VHOLD_NO_SMR) 4244 strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf)); 4245 printf(" hold count flags (%s)\n", buf + 1); 4246 4247 buf[0] = '\0'; 4248 buf[1] = '\0'; 4249 irflag = vn_irflag_read(vp); 4250 if (irflag & VIRF_DOOMED) 4251 strlcat(buf, "|VIRF_DOOMED", sizeof(buf)); 4252 if (irflag & VIRF_PGREAD) 4253 strlcat(buf, "|VIRF_PGREAD", sizeof(buf)); 4254 if (irflag & VIRF_MOUNTPOINT) 4255 strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf)); 4256 if (irflag & VIRF_TEXT_REF) 4257 strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf)); 4258 flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF); 4259 if (flags != 0) { 4260 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags); 4261 strlcat(buf, buf2, sizeof(buf)); 4262 } 4263 if (vp->v_vflag & VV_ROOT) 4264 strlcat(buf, "|VV_ROOT", sizeof(buf)); 4265 if (vp->v_vflag & VV_ISTTY) 4266 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 4267 if (vp->v_vflag & VV_NOSYNC) 4268 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 4269 if (vp->v_vflag & VV_ETERNALDEV) 4270 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 4271 if (vp->v_vflag & VV_CACHEDLABEL) 4272 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 4273 if (vp->v_vflag & VV_VMSIZEVNLOCK) 4274 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf)); 4275 if (vp->v_vflag & VV_COPYONWRITE) 4276 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 4277 if (vp->v_vflag & VV_SYSTEM) 4278 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 4279 if (vp->v_vflag & VV_PROCDEP) 4280 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 4281 if (vp->v_vflag & VV_DELETED) 4282 strlcat(buf, "|VV_DELETED", sizeof(buf)); 4283 if (vp->v_vflag & VV_MD) 4284 strlcat(buf, "|VV_MD", sizeof(buf)); 4285 if (vp->v_vflag & VV_FORCEINSMQ) 4286 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 4287 if (vp->v_vflag & VV_READLINK) 4288 strlcat(buf, "|VV_READLINK", sizeof(buf)); 4289 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 4290 VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM | 4291 VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK); 4292 if (flags != 0) { 4293 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 4294 strlcat(buf, buf2, sizeof(buf)); 4295 } 4296 if (vp->v_iflag & VI_MOUNT) 4297 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 4298 if (vp->v_iflag & VI_DOINGINACT) 4299 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 4300 if (vp->v_iflag & VI_OWEINACT) 4301 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 4302 if (vp->v_iflag & VI_DEFINACT) 4303 strlcat(buf, "|VI_DEFINACT", sizeof(buf)); 4304 if (vp->v_iflag & VI_FOPENING) 4305 strlcat(buf, "|VI_FOPENING", sizeof(buf)); 4306 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT | 4307 VI_OWEINACT | VI_DEFINACT | VI_FOPENING); 4308 if (flags != 0) { 4309 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 4310 strlcat(buf, buf2, sizeof(buf)); 4311 } 4312 if (vp->v_mflag & VMP_LAZYLIST) 4313 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf)); 4314 flags = vp->v_mflag & ~(VMP_LAZYLIST); 4315 if (flags != 0) { 4316 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags); 4317 strlcat(buf, buf2, sizeof(buf)); 4318 } 4319 printf(" flags (%s)", buf + 1); 4320 if (mtx_owned(VI_MTX(vp))) 4321 printf(" VI_LOCKed"); 4322 printf("\n"); 4323 if (vp->v_object != NULL) 4324 printf(" v_object %p ref %d pages %d " 4325 "cleanbuf %d dirtybuf %d\n", 4326 vp->v_object, vp->v_object->ref_count, 4327 vp->v_object->resident_page_count, 4328 vp->v_bufobj.bo_clean.bv_cnt, 4329 vp->v_bufobj.bo_dirty.bv_cnt); 4330 printf(" "); 4331 lockmgr_printinfo(vp->v_vnlock); 4332 if (vp->v_data != NULL) 4333 VOP_PRINT(vp); 4334 } 4335 4336 #ifdef DDB 4337 /* 4338 * List all of the locked vnodes in the system. 4339 * Called when debugging the kernel. 4340 */ 4341 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE) 4342 { 4343 struct mount *mp; 4344 struct vnode *vp; 4345 4346 /* 4347 * Note: because this is DDB, we can't obey the locking semantics 4348 * for these structures, which means we could catch an inconsistent 4349 * state and dereference a nasty pointer. Not much to be done 4350 * about that. 4351 */ 4352 db_printf("Locked vnodes\n"); 4353 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4354 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4355 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) 4356 vn_printf(vp, "vnode "); 4357 } 4358 } 4359 } 4360 4361 /* 4362 * Show details about the given vnode. 4363 */ 4364 DB_SHOW_COMMAND(vnode, db_show_vnode) 4365 { 4366 struct vnode *vp; 4367 4368 if (!have_addr) 4369 return; 4370 vp = (struct vnode *)addr; 4371 vn_printf(vp, "vnode "); 4372 } 4373 4374 /* 4375 * Show details about the given mount point. 4376 */ 4377 DB_SHOW_COMMAND(mount, db_show_mount) 4378 { 4379 struct mount *mp; 4380 struct vfsopt *opt; 4381 struct statfs *sp; 4382 struct vnode *vp; 4383 char buf[512]; 4384 uint64_t mflags; 4385 u_int flags; 4386 4387 if (!have_addr) { 4388 /* No address given, print short info about all mount points. */ 4389 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4390 db_printf("%p %s on %s (%s)\n", mp, 4391 mp->mnt_stat.f_mntfromname, 4392 mp->mnt_stat.f_mntonname, 4393 mp->mnt_stat.f_fstypename); 4394 if (db_pager_quit) 4395 break; 4396 } 4397 db_printf("\nMore info: show mount <addr>\n"); 4398 return; 4399 } 4400 4401 mp = (struct mount *)addr; 4402 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 4403 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 4404 4405 buf[0] = '\0'; 4406 mflags = mp->mnt_flag; 4407 #define MNT_FLAG(flag) do { \ 4408 if (mflags & (flag)) { \ 4409 if (buf[0] != '\0') \ 4410 strlcat(buf, ", ", sizeof(buf)); \ 4411 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 4412 mflags &= ~(flag); \ 4413 } \ 4414 } while (0) 4415 MNT_FLAG(MNT_RDONLY); 4416 MNT_FLAG(MNT_SYNCHRONOUS); 4417 MNT_FLAG(MNT_NOEXEC); 4418 MNT_FLAG(MNT_NOSUID); 4419 MNT_FLAG(MNT_NFS4ACLS); 4420 MNT_FLAG(MNT_UNION); 4421 MNT_FLAG(MNT_ASYNC); 4422 MNT_FLAG(MNT_SUIDDIR); 4423 MNT_FLAG(MNT_SOFTDEP); 4424 MNT_FLAG(MNT_NOSYMFOLLOW); 4425 MNT_FLAG(MNT_GJOURNAL); 4426 MNT_FLAG(MNT_MULTILABEL); 4427 MNT_FLAG(MNT_ACLS); 4428 MNT_FLAG(MNT_NOATIME); 4429 MNT_FLAG(MNT_NOCLUSTERR); 4430 MNT_FLAG(MNT_NOCLUSTERW); 4431 MNT_FLAG(MNT_SUJ); 4432 MNT_FLAG(MNT_EXRDONLY); 4433 MNT_FLAG(MNT_EXPORTED); 4434 MNT_FLAG(MNT_DEFEXPORTED); 4435 MNT_FLAG(MNT_EXPORTANON); 4436 MNT_FLAG(MNT_EXKERB); 4437 MNT_FLAG(MNT_EXPUBLIC); 4438 MNT_FLAG(MNT_LOCAL); 4439 MNT_FLAG(MNT_QUOTA); 4440 MNT_FLAG(MNT_ROOTFS); 4441 MNT_FLAG(MNT_USER); 4442 MNT_FLAG(MNT_IGNORE); 4443 MNT_FLAG(MNT_UPDATE); 4444 MNT_FLAG(MNT_DELEXPORT); 4445 MNT_FLAG(MNT_RELOAD); 4446 MNT_FLAG(MNT_FORCE); 4447 MNT_FLAG(MNT_SNAPSHOT); 4448 MNT_FLAG(MNT_BYFSID); 4449 #undef MNT_FLAG 4450 if (mflags != 0) { 4451 if (buf[0] != '\0') 4452 strlcat(buf, ", ", sizeof(buf)); 4453 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4454 "0x%016jx", mflags); 4455 } 4456 db_printf(" mnt_flag = %s\n", buf); 4457 4458 buf[0] = '\0'; 4459 flags = mp->mnt_kern_flag; 4460 #define MNT_KERN_FLAG(flag) do { \ 4461 if (flags & (flag)) { \ 4462 if (buf[0] != '\0') \ 4463 strlcat(buf, ", ", sizeof(buf)); \ 4464 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 4465 flags &= ~(flag); \ 4466 } \ 4467 } while (0) 4468 MNT_KERN_FLAG(MNTK_UNMOUNTF); 4469 MNT_KERN_FLAG(MNTK_ASYNC); 4470 MNT_KERN_FLAG(MNTK_SOFTDEP); 4471 MNT_KERN_FLAG(MNTK_NOMSYNC); 4472 MNT_KERN_FLAG(MNTK_DRAINING); 4473 MNT_KERN_FLAG(MNTK_REFEXPIRE); 4474 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 4475 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 4476 MNT_KERN_FLAG(MNTK_NO_IOPF); 4477 MNT_KERN_FLAG(MNTK_RECURSE); 4478 MNT_KERN_FLAG(MNTK_UPPER_WAITER); 4479 MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE); 4480 MNT_KERN_FLAG(MNTK_USES_BCACHE); 4481 MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG); 4482 MNT_KERN_FLAG(MNTK_FPLOOKUP); 4483 MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER); 4484 MNT_KERN_FLAG(MNTK_NOASYNC); 4485 MNT_KERN_FLAG(MNTK_UNMOUNT); 4486 MNT_KERN_FLAG(MNTK_MWAIT); 4487 MNT_KERN_FLAG(MNTK_SUSPEND); 4488 MNT_KERN_FLAG(MNTK_SUSPEND2); 4489 MNT_KERN_FLAG(MNTK_SUSPENDED); 4490 MNT_KERN_FLAG(MNTK_NULL_NOCACHE); 4491 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 4492 #undef MNT_KERN_FLAG 4493 if (flags != 0) { 4494 if (buf[0] != '\0') 4495 strlcat(buf, ", ", sizeof(buf)); 4496 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4497 "0x%08x", flags); 4498 } 4499 db_printf(" mnt_kern_flag = %s\n", buf); 4500 4501 db_printf(" mnt_opt = "); 4502 opt = TAILQ_FIRST(mp->mnt_opt); 4503 if (opt != NULL) { 4504 db_printf("%s", opt->name); 4505 opt = TAILQ_NEXT(opt, link); 4506 while (opt != NULL) { 4507 db_printf(", %s", opt->name); 4508 opt = TAILQ_NEXT(opt, link); 4509 } 4510 } 4511 db_printf("\n"); 4512 4513 sp = &mp->mnt_stat; 4514 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 4515 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 4516 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 4517 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 4518 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 4519 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 4520 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 4521 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 4522 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 4523 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 4524 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 4525 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 4526 4527 db_printf(" mnt_cred = { uid=%u ruid=%u", 4528 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 4529 if (jailed(mp->mnt_cred)) 4530 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 4531 db_printf(" }\n"); 4532 db_printf(" mnt_ref = %d (with %d in the struct)\n", 4533 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref); 4534 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 4535 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 4536 db_printf(" mnt_lazyvnodelistsize = %d\n", 4537 mp->mnt_lazyvnodelistsize); 4538 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n", 4539 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount); 4540 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 4541 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 4542 db_printf(" mnt_lockref = %d (with %d in the struct)\n", 4543 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref); 4544 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 4545 db_printf(" mnt_secondary_accwrites = %d\n", 4546 mp->mnt_secondary_accwrites); 4547 db_printf(" mnt_gjprovider = %s\n", 4548 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 4549 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops); 4550 4551 db_printf("\n\nList of active vnodes\n"); 4552 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4553 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) { 4554 vn_printf(vp, "vnode "); 4555 if (db_pager_quit) 4556 break; 4557 } 4558 } 4559 db_printf("\n\nList of inactive vnodes\n"); 4560 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4561 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) { 4562 vn_printf(vp, "vnode "); 4563 if (db_pager_quit) 4564 break; 4565 } 4566 } 4567 } 4568 #endif /* DDB */ 4569 4570 /* 4571 * Fill in a struct xvfsconf based on a struct vfsconf. 4572 */ 4573 static int 4574 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 4575 { 4576 struct xvfsconf xvfsp; 4577 4578 bzero(&xvfsp, sizeof(xvfsp)); 4579 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4580 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4581 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4582 xvfsp.vfc_flags = vfsp->vfc_flags; 4583 /* 4584 * These are unused in userland, we keep them 4585 * to not break binary compatibility. 4586 */ 4587 xvfsp.vfc_vfsops = NULL; 4588 xvfsp.vfc_next = NULL; 4589 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4590 } 4591 4592 #ifdef COMPAT_FREEBSD32 4593 struct xvfsconf32 { 4594 uint32_t vfc_vfsops; 4595 char vfc_name[MFSNAMELEN]; 4596 int32_t vfc_typenum; 4597 int32_t vfc_refcount; 4598 int32_t vfc_flags; 4599 uint32_t vfc_next; 4600 }; 4601 4602 static int 4603 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 4604 { 4605 struct xvfsconf32 xvfsp; 4606 4607 bzero(&xvfsp, sizeof(xvfsp)); 4608 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4609 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4610 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4611 xvfsp.vfc_flags = vfsp->vfc_flags; 4612 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4613 } 4614 #endif 4615 4616 /* 4617 * Top level filesystem related information gathering. 4618 */ 4619 static int 4620 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 4621 { 4622 struct vfsconf *vfsp; 4623 int error; 4624 4625 error = 0; 4626 vfsconf_slock(); 4627 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4628 #ifdef COMPAT_FREEBSD32 4629 if (req->flags & SCTL_MASK32) 4630 error = vfsconf2x32(req, vfsp); 4631 else 4632 #endif 4633 error = vfsconf2x(req, vfsp); 4634 if (error) 4635 break; 4636 } 4637 vfsconf_sunlock(); 4638 return (error); 4639 } 4640 4641 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | 4642 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, 4643 "S,xvfsconf", "List of all configured filesystems"); 4644 4645 #ifndef BURN_BRIDGES 4646 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 4647 4648 static int 4649 vfs_sysctl(SYSCTL_HANDLER_ARGS) 4650 { 4651 int *name = (int *)arg1 - 1; /* XXX */ 4652 u_int namelen = arg2 + 1; /* XXX */ 4653 struct vfsconf *vfsp; 4654 4655 log(LOG_WARNING, "userland calling deprecated sysctl, " 4656 "please rebuild world\n"); 4657 4658 #if 1 || defined(COMPAT_PRELITE2) 4659 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 4660 if (namelen == 1) 4661 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 4662 #endif 4663 4664 switch (name[1]) { 4665 case VFS_MAXTYPENUM: 4666 if (namelen != 2) 4667 return (ENOTDIR); 4668 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 4669 case VFS_CONF: 4670 if (namelen != 3) 4671 return (ENOTDIR); /* overloaded */ 4672 vfsconf_slock(); 4673 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4674 if (vfsp->vfc_typenum == name[2]) 4675 break; 4676 } 4677 vfsconf_sunlock(); 4678 if (vfsp == NULL) 4679 return (EOPNOTSUPP); 4680 #ifdef COMPAT_FREEBSD32 4681 if (req->flags & SCTL_MASK32) 4682 return (vfsconf2x32(req, vfsp)); 4683 else 4684 #endif 4685 return (vfsconf2x(req, vfsp)); 4686 } 4687 return (EOPNOTSUPP); 4688 } 4689 4690 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | 4691 CTLFLAG_MPSAFE, vfs_sysctl, 4692 "Generic filesystem"); 4693 4694 #if 1 || defined(COMPAT_PRELITE2) 4695 4696 static int 4697 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 4698 { 4699 int error; 4700 struct vfsconf *vfsp; 4701 struct ovfsconf ovfs; 4702 4703 vfsconf_slock(); 4704 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4705 bzero(&ovfs, sizeof(ovfs)); 4706 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 4707 strcpy(ovfs.vfc_name, vfsp->vfc_name); 4708 ovfs.vfc_index = vfsp->vfc_typenum; 4709 ovfs.vfc_refcount = vfsp->vfc_refcount; 4710 ovfs.vfc_flags = vfsp->vfc_flags; 4711 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 4712 if (error != 0) { 4713 vfsconf_sunlock(); 4714 return (error); 4715 } 4716 } 4717 vfsconf_sunlock(); 4718 return (0); 4719 } 4720 4721 #endif /* 1 || COMPAT_PRELITE2 */ 4722 #endif /* !BURN_BRIDGES */ 4723 4724 static void 4725 unmount_or_warn(struct mount *mp) 4726 { 4727 int error; 4728 4729 error = dounmount(mp, MNT_FORCE, curthread); 4730 if (error != 0) { 4731 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); 4732 if (error == EBUSY) 4733 printf("BUSY)\n"); 4734 else 4735 printf("%d)\n", error); 4736 } 4737 } 4738 4739 /* 4740 * Unmount all filesystems. The list is traversed in reverse order 4741 * of mounting to avoid dependencies. 4742 */ 4743 void 4744 vfs_unmountall(void) 4745 { 4746 struct mount *mp, *tmp; 4747 4748 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 4749 4750 /* 4751 * Since this only runs when rebooting, it is not interlocked. 4752 */ 4753 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { 4754 vfs_ref(mp); 4755 4756 /* 4757 * Forcibly unmounting "/dev" before "/" would prevent clean 4758 * unmount of the latter. 4759 */ 4760 if (mp == rootdevmp) 4761 continue; 4762 4763 unmount_or_warn(mp); 4764 } 4765 4766 if (rootdevmp != NULL) 4767 unmount_or_warn(rootdevmp); 4768 } 4769 4770 static void 4771 vfs_deferred_inactive(struct vnode *vp, int lkflags) 4772 { 4773 4774 ASSERT_VI_LOCKED(vp, __func__); 4775 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp); 4776 if ((vp->v_iflag & VI_OWEINACT) == 0) { 4777 vdropl(vp); 4778 return; 4779 } 4780 if (vn_lock(vp, lkflags) == 0) { 4781 VI_LOCK(vp); 4782 vinactive(vp); 4783 VOP_UNLOCK(vp); 4784 vdropl(vp); 4785 return; 4786 } 4787 vdefer_inactive_unlocked(vp); 4788 } 4789 4790 static int 4791 vfs_periodic_inactive_filter(struct vnode *vp, void *arg) 4792 { 4793 4794 return (vp->v_iflag & VI_DEFINACT); 4795 } 4796 4797 static void __noinline 4798 vfs_periodic_inactive(struct mount *mp, int flags) 4799 { 4800 struct vnode *vp, *mvp; 4801 int lkflags; 4802 4803 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 4804 if (flags != MNT_WAIT) 4805 lkflags |= LK_NOWAIT; 4806 4807 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) { 4808 if ((vp->v_iflag & VI_DEFINACT) == 0) { 4809 VI_UNLOCK(vp); 4810 continue; 4811 } 4812 vp->v_iflag &= ~VI_DEFINACT; 4813 vfs_deferred_inactive(vp, lkflags); 4814 } 4815 } 4816 4817 static inline bool 4818 vfs_want_msync(struct vnode *vp) 4819 { 4820 struct vm_object *obj; 4821 4822 /* 4823 * This test may be performed without any locks held. 4824 * We rely on vm_object's type stability. 4825 */ 4826 if (vp->v_vflag & VV_NOSYNC) 4827 return (false); 4828 obj = vp->v_object; 4829 return (obj != NULL && vm_object_mightbedirty(obj)); 4830 } 4831 4832 static int 4833 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused) 4834 { 4835 4836 if (vp->v_vflag & VV_NOSYNC) 4837 return (false); 4838 if (vp->v_iflag & VI_DEFINACT) 4839 return (true); 4840 return (vfs_want_msync(vp)); 4841 } 4842 4843 static void __noinline 4844 vfs_periodic_msync_inactive(struct mount *mp, int flags) 4845 { 4846 struct vnode *vp, *mvp; 4847 struct vm_object *obj; 4848 int lkflags, objflags; 4849 bool seen_defer; 4850 4851 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 4852 if (flags != MNT_WAIT) { 4853 lkflags |= LK_NOWAIT; 4854 objflags = OBJPC_NOSYNC; 4855 } else { 4856 objflags = OBJPC_SYNC; 4857 } 4858 4859 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) { 4860 seen_defer = false; 4861 if (vp->v_iflag & VI_DEFINACT) { 4862 vp->v_iflag &= ~VI_DEFINACT; 4863 seen_defer = true; 4864 } 4865 if (!vfs_want_msync(vp)) { 4866 if (seen_defer) 4867 vfs_deferred_inactive(vp, lkflags); 4868 else 4869 VI_UNLOCK(vp); 4870 continue; 4871 } 4872 if (vget(vp, lkflags) == 0) { 4873 obj = vp->v_object; 4874 if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) { 4875 VM_OBJECT_WLOCK(obj); 4876 vm_object_page_clean(obj, 0, 0, objflags); 4877 VM_OBJECT_WUNLOCK(obj); 4878 } 4879 vput(vp); 4880 if (seen_defer) 4881 vdrop(vp); 4882 } else { 4883 if (seen_defer) 4884 vdefer_inactive_unlocked(vp); 4885 } 4886 } 4887 } 4888 4889 void 4890 vfs_periodic(struct mount *mp, int flags) 4891 { 4892 4893 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 4894 4895 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0) 4896 vfs_periodic_inactive(mp, flags); 4897 else 4898 vfs_periodic_msync_inactive(mp, flags); 4899 } 4900 4901 static void 4902 destroy_vpollinfo_free(struct vpollinfo *vi) 4903 { 4904 4905 knlist_destroy(&vi->vpi_selinfo.si_note); 4906 mtx_destroy(&vi->vpi_lock); 4907 free(vi, M_VNODEPOLL); 4908 } 4909 4910 static void 4911 destroy_vpollinfo(struct vpollinfo *vi) 4912 { 4913 4914 knlist_clear(&vi->vpi_selinfo.si_note, 1); 4915 seldrain(&vi->vpi_selinfo); 4916 destroy_vpollinfo_free(vi); 4917 } 4918 4919 /* 4920 * Initialize per-vnode helper structure to hold poll-related state. 4921 */ 4922 void 4923 v_addpollinfo(struct vnode *vp) 4924 { 4925 struct vpollinfo *vi; 4926 4927 if (vp->v_pollinfo != NULL) 4928 return; 4929 vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO); 4930 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 4931 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 4932 vfs_knlunlock, vfs_knl_assert_lock); 4933 VI_LOCK(vp); 4934 if (vp->v_pollinfo != NULL) { 4935 VI_UNLOCK(vp); 4936 destroy_vpollinfo_free(vi); 4937 return; 4938 } 4939 vp->v_pollinfo = vi; 4940 VI_UNLOCK(vp); 4941 } 4942 4943 /* 4944 * Record a process's interest in events which might happen to 4945 * a vnode. Because poll uses the historic select-style interface 4946 * internally, this routine serves as both the ``check for any 4947 * pending events'' and the ``record my interest in future events'' 4948 * functions. (These are done together, while the lock is held, 4949 * to avoid race conditions.) 4950 */ 4951 int 4952 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 4953 { 4954 4955 v_addpollinfo(vp); 4956 mtx_lock(&vp->v_pollinfo->vpi_lock); 4957 if (vp->v_pollinfo->vpi_revents & events) { 4958 /* 4959 * This leaves events we are not interested 4960 * in available for the other process which 4961 * which presumably had requested them 4962 * (otherwise they would never have been 4963 * recorded). 4964 */ 4965 events &= vp->v_pollinfo->vpi_revents; 4966 vp->v_pollinfo->vpi_revents &= ~events; 4967 4968 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4969 return (events); 4970 } 4971 vp->v_pollinfo->vpi_events |= events; 4972 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 4973 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4974 return (0); 4975 } 4976 4977 /* 4978 * Routine to create and manage a filesystem syncer vnode. 4979 */ 4980 #define sync_close ((int (*)(struct vop_close_args *))nullop) 4981 static int sync_fsync(struct vop_fsync_args *); 4982 static int sync_inactive(struct vop_inactive_args *); 4983 static int sync_reclaim(struct vop_reclaim_args *); 4984 4985 static struct vop_vector sync_vnodeops = { 4986 .vop_bypass = VOP_EOPNOTSUPP, 4987 .vop_close = sync_close, 4988 .vop_fsync = sync_fsync, 4989 .vop_inactive = sync_inactive, 4990 .vop_need_inactive = vop_stdneed_inactive, 4991 .vop_reclaim = sync_reclaim, 4992 .vop_lock1 = vop_stdlock, 4993 .vop_unlock = vop_stdunlock, 4994 .vop_islocked = vop_stdislocked, 4995 .vop_fplookup_vexec = VOP_EAGAIN, 4996 .vop_fplookup_symlink = VOP_EAGAIN, 4997 }; 4998 VFS_VOP_VECTOR_REGISTER(sync_vnodeops); 4999 5000 /* 5001 * Create a new filesystem syncer vnode for the specified mount point. 5002 */ 5003 void 5004 vfs_allocate_syncvnode(struct mount *mp) 5005 { 5006 struct vnode *vp; 5007 struct bufobj *bo; 5008 static long start, incr, next; 5009 int error; 5010 5011 /* Allocate a new vnode */ 5012 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 5013 if (error != 0) 5014 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 5015 vp->v_type = VNON; 5016 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5017 vp->v_vflag |= VV_FORCEINSMQ; 5018 error = insmntque1(vp, mp); 5019 if (error != 0) 5020 panic("vfs_allocate_syncvnode: insmntque() failed"); 5021 vp->v_vflag &= ~VV_FORCEINSMQ; 5022 vn_set_state(vp, VSTATE_CONSTRUCTED); 5023 VOP_UNLOCK(vp); 5024 /* 5025 * Place the vnode onto the syncer worklist. We attempt to 5026 * scatter them about on the list so that they will go off 5027 * at evenly distributed times even if all the filesystems 5028 * are mounted at once. 5029 */ 5030 next += incr; 5031 if (next == 0 || next > syncer_maxdelay) { 5032 start /= 2; 5033 incr /= 2; 5034 if (start == 0) { 5035 start = syncer_maxdelay / 2; 5036 incr = syncer_maxdelay; 5037 } 5038 next = start; 5039 } 5040 bo = &vp->v_bufobj; 5041 BO_LOCK(bo); 5042 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 5043 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 5044 mtx_lock(&sync_mtx); 5045 sync_vnode_count++; 5046 if (mp->mnt_syncer == NULL) { 5047 mp->mnt_syncer = vp; 5048 vp = NULL; 5049 } 5050 mtx_unlock(&sync_mtx); 5051 BO_UNLOCK(bo); 5052 if (vp != NULL) { 5053 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5054 vgone(vp); 5055 vput(vp); 5056 } 5057 } 5058 5059 void 5060 vfs_deallocate_syncvnode(struct mount *mp) 5061 { 5062 struct vnode *vp; 5063 5064 mtx_lock(&sync_mtx); 5065 vp = mp->mnt_syncer; 5066 if (vp != NULL) 5067 mp->mnt_syncer = NULL; 5068 mtx_unlock(&sync_mtx); 5069 if (vp != NULL) 5070 vrele(vp); 5071 } 5072 5073 /* 5074 * Do a lazy sync of the filesystem. 5075 */ 5076 static int 5077 sync_fsync(struct vop_fsync_args *ap) 5078 { 5079 struct vnode *syncvp = ap->a_vp; 5080 struct mount *mp = syncvp->v_mount; 5081 int error, save; 5082 struct bufobj *bo; 5083 5084 /* 5085 * We only need to do something if this is a lazy evaluation. 5086 */ 5087 if (ap->a_waitfor != MNT_LAZY) 5088 return (0); 5089 5090 /* 5091 * Move ourselves to the back of the sync list. 5092 */ 5093 bo = &syncvp->v_bufobj; 5094 BO_LOCK(bo); 5095 vn_syncer_add_to_worklist(bo, syncdelay); 5096 BO_UNLOCK(bo); 5097 5098 /* 5099 * Walk the list of vnodes pushing all that are dirty and 5100 * not already on the sync list. 5101 */ 5102 if (vfs_busy(mp, MBF_NOWAIT) != 0) 5103 return (0); 5104 VOP_UNLOCK(syncvp); 5105 save = curthread_pflags_set(TDP_SYNCIO); 5106 /* 5107 * The filesystem at hand may be idle with free vnodes stored in the 5108 * batch. Return them instead of letting them stay there indefinitely. 5109 */ 5110 vfs_periodic(mp, MNT_NOWAIT); 5111 error = VFS_SYNC(mp, MNT_LAZY); 5112 curthread_pflags_restore(save); 5113 vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY); 5114 vfs_unbusy(mp); 5115 return (error); 5116 } 5117 5118 /* 5119 * The syncer vnode is no referenced. 5120 */ 5121 static int 5122 sync_inactive(struct vop_inactive_args *ap) 5123 { 5124 5125 vgone(ap->a_vp); 5126 return (0); 5127 } 5128 5129 /* 5130 * The syncer vnode is no longer needed and is being decommissioned. 5131 * 5132 * Modifications to the worklist must be protected by sync_mtx. 5133 */ 5134 static int 5135 sync_reclaim(struct vop_reclaim_args *ap) 5136 { 5137 struct vnode *vp = ap->a_vp; 5138 struct bufobj *bo; 5139 5140 bo = &vp->v_bufobj; 5141 BO_LOCK(bo); 5142 mtx_lock(&sync_mtx); 5143 if (vp->v_mount->mnt_syncer == vp) 5144 vp->v_mount->mnt_syncer = NULL; 5145 if (bo->bo_flag & BO_ONWORKLST) { 5146 LIST_REMOVE(bo, bo_synclist); 5147 syncer_worklist_len--; 5148 sync_vnode_count--; 5149 bo->bo_flag &= ~BO_ONWORKLST; 5150 } 5151 mtx_unlock(&sync_mtx); 5152 BO_UNLOCK(bo); 5153 5154 return (0); 5155 } 5156 5157 int 5158 vn_need_pageq_flush(struct vnode *vp) 5159 { 5160 struct vm_object *obj; 5161 5162 obj = vp->v_object; 5163 return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 5164 vm_object_mightbedirty(obj)); 5165 } 5166 5167 /* 5168 * Check if vnode represents a disk device 5169 */ 5170 bool 5171 vn_isdisk_error(struct vnode *vp, int *errp) 5172 { 5173 int error; 5174 5175 if (vp->v_type != VCHR) { 5176 error = ENOTBLK; 5177 goto out; 5178 } 5179 error = 0; 5180 dev_lock(); 5181 if (vp->v_rdev == NULL) 5182 error = ENXIO; 5183 else if (vp->v_rdev->si_devsw == NULL) 5184 error = ENXIO; 5185 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 5186 error = ENOTBLK; 5187 dev_unlock(); 5188 out: 5189 *errp = error; 5190 return (error == 0); 5191 } 5192 5193 bool 5194 vn_isdisk(struct vnode *vp) 5195 { 5196 int error; 5197 5198 return (vn_isdisk_error(vp, &error)); 5199 } 5200 5201 /* 5202 * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see 5203 * the comment above cache_fplookup for details. 5204 */ 5205 int 5206 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred) 5207 { 5208 int error; 5209 5210 VFS_SMR_ASSERT_ENTERED(); 5211 5212 /* Check the owner. */ 5213 if (cred->cr_uid == file_uid) { 5214 if (file_mode & S_IXUSR) 5215 return (0); 5216 goto out_error; 5217 } 5218 5219 /* Otherwise, check the groups (first match) */ 5220 if (groupmember(file_gid, cred)) { 5221 if (file_mode & S_IXGRP) 5222 return (0); 5223 goto out_error; 5224 } 5225 5226 /* Otherwise, check everyone else. */ 5227 if (file_mode & S_IXOTH) 5228 return (0); 5229 out_error: 5230 /* 5231 * Permission check failed, but it is possible denial will get overwritten 5232 * (e.g., when root is traversing through a 700 directory owned by someone 5233 * else). 5234 * 5235 * vaccess() calls priv_check_cred which in turn can descent into MAC 5236 * modules overriding this result. It's quite unclear what semantics 5237 * are allowed for them to operate, thus for safety we don't call them 5238 * from within the SMR section. This also means if any such modules 5239 * are present, we have to let the regular lookup decide. 5240 */ 5241 error = priv_check_cred_vfs_lookup_nomac(cred); 5242 switch (error) { 5243 case 0: 5244 return (0); 5245 case EAGAIN: 5246 /* 5247 * MAC modules present. 5248 */ 5249 return (EAGAIN); 5250 case EPERM: 5251 return (EACCES); 5252 default: 5253 return (error); 5254 } 5255 } 5256 5257 /* 5258 * Common filesystem object access control check routine. Accepts a 5259 * vnode's type, "mode", uid and gid, requested access mode, and credentials. 5260 * Returns 0 on success, or an errno on failure. 5261 */ 5262 int 5263 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 5264 accmode_t accmode, struct ucred *cred) 5265 { 5266 accmode_t dac_granted; 5267 accmode_t priv_granted; 5268 5269 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 5270 ("invalid bit in accmode")); 5271 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 5272 ("VAPPEND without VWRITE")); 5273 5274 /* 5275 * Look for a normal, non-privileged way to access the file/directory 5276 * as requested. If it exists, go with that. 5277 */ 5278 5279 dac_granted = 0; 5280 5281 /* Check the owner. */ 5282 if (cred->cr_uid == file_uid) { 5283 dac_granted |= VADMIN; 5284 if (file_mode & S_IXUSR) 5285 dac_granted |= VEXEC; 5286 if (file_mode & S_IRUSR) 5287 dac_granted |= VREAD; 5288 if (file_mode & S_IWUSR) 5289 dac_granted |= (VWRITE | VAPPEND); 5290 5291 if ((accmode & dac_granted) == accmode) 5292 return (0); 5293 5294 goto privcheck; 5295 } 5296 5297 /* Otherwise, check the groups (first match) */ 5298 if (groupmember(file_gid, cred)) { 5299 if (file_mode & S_IXGRP) 5300 dac_granted |= VEXEC; 5301 if (file_mode & S_IRGRP) 5302 dac_granted |= VREAD; 5303 if (file_mode & S_IWGRP) 5304 dac_granted |= (VWRITE | VAPPEND); 5305 5306 if ((accmode & dac_granted) == accmode) 5307 return (0); 5308 5309 goto privcheck; 5310 } 5311 5312 /* Otherwise, check everyone else. */ 5313 if (file_mode & S_IXOTH) 5314 dac_granted |= VEXEC; 5315 if (file_mode & S_IROTH) 5316 dac_granted |= VREAD; 5317 if (file_mode & S_IWOTH) 5318 dac_granted |= (VWRITE | VAPPEND); 5319 if ((accmode & dac_granted) == accmode) 5320 return (0); 5321 5322 privcheck: 5323 /* 5324 * Build a privilege mask to determine if the set of privileges 5325 * satisfies the requirements when combined with the granted mask 5326 * from above. For each privilege, if the privilege is required, 5327 * bitwise or the request type onto the priv_granted mask. 5328 */ 5329 priv_granted = 0; 5330 5331 if (type == VDIR) { 5332 /* 5333 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 5334 * requests, instead of PRIV_VFS_EXEC. 5335 */ 5336 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5337 !priv_check_cred(cred, PRIV_VFS_LOOKUP)) 5338 priv_granted |= VEXEC; 5339 } else { 5340 /* 5341 * Ensure that at least one execute bit is on. Otherwise, 5342 * a privileged user will always succeed, and we don't want 5343 * this to happen unless the file really is executable. 5344 */ 5345 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5346 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 5347 !priv_check_cred(cred, PRIV_VFS_EXEC)) 5348 priv_granted |= VEXEC; 5349 } 5350 5351 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 5352 !priv_check_cred(cred, PRIV_VFS_READ)) 5353 priv_granted |= VREAD; 5354 5355 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 5356 !priv_check_cred(cred, PRIV_VFS_WRITE)) 5357 priv_granted |= (VWRITE | VAPPEND); 5358 5359 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 5360 !priv_check_cred(cred, PRIV_VFS_ADMIN)) 5361 priv_granted |= VADMIN; 5362 5363 if ((accmode & (priv_granted | dac_granted)) == accmode) { 5364 return (0); 5365 } 5366 5367 return ((accmode & VADMIN) ? EPERM : EACCES); 5368 } 5369 5370 /* 5371 * Credential check based on process requesting service, and per-attribute 5372 * permissions. 5373 */ 5374 int 5375 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 5376 struct thread *td, accmode_t accmode) 5377 { 5378 5379 /* 5380 * Kernel-invoked always succeeds. 5381 */ 5382 if (cred == NOCRED) 5383 return (0); 5384 5385 /* 5386 * Do not allow privileged processes in jail to directly manipulate 5387 * system attributes. 5388 */ 5389 switch (attrnamespace) { 5390 case EXTATTR_NAMESPACE_SYSTEM: 5391 /* Potentially should be: return (EPERM); */ 5392 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM)); 5393 case EXTATTR_NAMESPACE_USER: 5394 return (VOP_ACCESS(vp, accmode, cred, td)); 5395 default: 5396 return (EPERM); 5397 } 5398 } 5399 5400 #ifdef DEBUG_VFS_LOCKS 5401 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 5402 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, 5403 "Drop into debugger on lock violation"); 5404 5405 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 5406 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 5407 0, "Check for interlock across VOPs"); 5408 5409 int vfs_badlock_print = 1; /* Print lock violations. */ 5410 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 5411 0, "Print lock violations"); 5412 5413 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */ 5414 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode, 5415 0, "Print vnode details on lock violations"); 5416 5417 #ifdef KDB 5418 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 5419 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, 5420 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations"); 5421 #endif 5422 5423 static void 5424 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 5425 { 5426 5427 #ifdef KDB 5428 if (vfs_badlock_backtrace) 5429 kdb_backtrace(); 5430 #endif 5431 if (vfs_badlock_vnode) 5432 vn_printf(vp, "vnode "); 5433 if (vfs_badlock_print) 5434 printf("%s: %p %s\n", str, (void *)vp, msg); 5435 if (vfs_badlock_ddb) 5436 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 5437 } 5438 5439 void 5440 assert_vi_locked(struct vnode *vp, const char *str) 5441 { 5442 5443 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 5444 vfs_badlock("interlock is not locked but should be", str, vp); 5445 } 5446 5447 void 5448 assert_vi_unlocked(struct vnode *vp, const char *str) 5449 { 5450 5451 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 5452 vfs_badlock("interlock is locked but should not be", str, vp); 5453 } 5454 5455 void 5456 assert_vop_locked(struct vnode *vp, const char *str) 5457 { 5458 if (KERNEL_PANICKED() || vp == NULL) 5459 return; 5460 5461 #ifdef WITNESS 5462 if ((vp->v_irflag & VIRF_CROSSMP) == 0 && 5463 witness_is_owned(&vp->v_vnlock->lock_object) == -1) 5464 #else 5465 int locked = VOP_ISLOCKED(vp); 5466 if (locked == 0 || locked == LK_EXCLOTHER) 5467 #endif 5468 vfs_badlock("is not locked but should be", str, vp); 5469 } 5470 5471 void 5472 assert_vop_unlocked(struct vnode *vp, const char *str) 5473 { 5474 if (KERNEL_PANICKED() || vp == NULL) 5475 return; 5476 5477 #ifdef WITNESS 5478 if ((vp->v_irflag & VIRF_CROSSMP) == 0 && 5479 witness_is_owned(&vp->v_vnlock->lock_object) == 1) 5480 #else 5481 if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE) 5482 #endif 5483 vfs_badlock("is locked but should not be", str, vp); 5484 } 5485 5486 void 5487 assert_vop_elocked(struct vnode *vp, const char *str) 5488 { 5489 if (KERNEL_PANICKED() || vp == NULL) 5490 return; 5491 5492 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 5493 vfs_badlock("is not exclusive locked but should be", str, vp); 5494 } 5495 #endif /* DEBUG_VFS_LOCKS */ 5496 5497 void 5498 vop_rename_fail(struct vop_rename_args *ap) 5499 { 5500 5501 if (ap->a_tvp != NULL) 5502 vput(ap->a_tvp); 5503 if (ap->a_tdvp == ap->a_tvp) 5504 vrele(ap->a_tdvp); 5505 else 5506 vput(ap->a_tdvp); 5507 vrele(ap->a_fdvp); 5508 vrele(ap->a_fvp); 5509 } 5510 5511 void 5512 vop_rename_pre(void *ap) 5513 { 5514 struct vop_rename_args *a = ap; 5515 5516 #ifdef DEBUG_VFS_LOCKS 5517 if (a->a_tvp) 5518 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 5519 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 5520 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 5521 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 5522 5523 /* Check the source (from). */ 5524 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 5525 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 5526 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 5527 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 5528 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 5529 5530 /* Check the target. */ 5531 if (a->a_tvp) 5532 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 5533 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 5534 #endif 5535 /* 5536 * It may be tempting to add vn_seqc_write_begin/end calls here and 5537 * in vop_rename_post but that's not going to work out since some 5538 * filesystems relookup vnodes mid-rename. This is probably a bug. 5539 * 5540 * For now filesystems are expected to do the relevant calls after they 5541 * decide what vnodes to operate on. 5542 */ 5543 if (a->a_tdvp != a->a_fdvp) 5544 vhold(a->a_fdvp); 5545 if (a->a_tvp != a->a_fvp) 5546 vhold(a->a_fvp); 5547 vhold(a->a_tdvp); 5548 if (a->a_tvp) 5549 vhold(a->a_tvp); 5550 } 5551 5552 #ifdef DEBUG_VFS_LOCKS 5553 void 5554 vop_fplookup_vexec_debugpre(void *ap __unused) 5555 { 5556 5557 VFS_SMR_ASSERT_ENTERED(); 5558 } 5559 5560 void 5561 vop_fplookup_vexec_debugpost(void *ap, int rc) 5562 { 5563 struct vop_fplookup_vexec_args *a; 5564 struct vnode *vp; 5565 5566 a = ap; 5567 vp = a->a_vp; 5568 5569 VFS_SMR_ASSERT_ENTERED(); 5570 if (rc == EOPNOTSUPP) 5571 VNPASS(VN_IS_DOOMED(vp), vp); 5572 } 5573 5574 void 5575 vop_fplookup_symlink_debugpre(void *ap __unused) 5576 { 5577 5578 VFS_SMR_ASSERT_ENTERED(); 5579 } 5580 5581 void 5582 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused) 5583 { 5584 5585 VFS_SMR_ASSERT_ENTERED(); 5586 } 5587 5588 static void 5589 vop_fsync_debugprepost(struct vnode *vp, const char *name) 5590 { 5591 if (vp->v_type == VCHR) 5592 ; 5593 else if (MNT_EXTENDED_SHARED(vp->v_mount)) 5594 ASSERT_VOP_LOCKED(vp, name); 5595 else 5596 ASSERT_VOP_ELOCKED(vp, name); 5597 } 5598 5599 void 5600 vop_fsync_debugpre(void *a) 5601 { 5602 struct vop_fsync_args *ap; 5603 5604 ap = a; 5605 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5606 } 5607 5608 void 5609 vop_fsync_debugpost(void *a, int rc __unused) 5610 { 5611 struct vop_fsync_args *ap; 5612 5613 ap = a; 5614 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5615 } 5616 5617 void 5618 vop_fdatasync_debugpre(void *a) 5619 { 5620 struct vop_fdatasync_args *ap; 5621 5622 ap = a; 5623 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5624 } 5625 5626 void 5627 vop_fdatasync_debugpost(void *a, int rc __unused) 5628 { 5629 struct vop_fdatasync_args *ap; 5630 5631 ap = a; 5632 vop_fsync_debugprepost(ap->a_vp, "fsync"); 5633 } 5634 5635 void 5636 vop_strategy_debugpre(void *ap) 5637 { 5638 struct vop_strategy_args *a; 5639 struct buf *bp; 5640 5641 a = ap; 5642 bp = a->a_bp; 5643 5644 /* 5645 * Cluster ops lock their component buffers but not the IO container. 5646 */ 5647 if ((bp->b_flags & B_CLUSTER) != 0) 5648 return; 5649 5650 if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) { 5651 if (vfs_badlock_print) 5652 printf( 5653 "VOP_STRATEGY: bp is not locked but should be\n"); 5654 if (vfs_badlock_ddb) 5655 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 5656 } 5657 } 5658 5659 void 5660 vop_lock_debugpre(void *ap) 5661 { 5662 struct vop_lock1_args *a = ap; 5663 5664 if ((a->a_flags & LK_INTERLOCK) == 0) 5665 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5666 else 5667 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 5668 } 5669 5670 void 5671 vop_lock_debugpost(void *ap, int rc) 5672 { 5673 struct vop_lock1_args *a = ap; 5674 5675 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5676 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) 5677 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 5678 } 5679 5680 void 5681 vop_unlock_debugpre(void *ap) 5682 { 5683 struct vop_unlock_args *a = ap; 5684 struct vnode *vp = a->a_vp; 5685 5686 VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp); 5687 ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK"); 5688 } 5689 5690 void 5691 vop_need_inactive_debugpre(void *ap) 5692 { 5693 struct vop_need_inactive_args *a = ap; 5694 5695 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 5696 } 5697 5698 void 5699 vop_need_inactive_debugpost(void *ap, int rc) 5700 { 5701 struct vop_need_inactive_args *a = ap; 5702 5703 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 5704 } 5705 #endif 5706 5707 void 5708 vop_create_pre(void *ap) 5709 { 5710 struct vop_create_args *a; 5711 struct vnode *dvp; 5712 5713 a = ap; 5714 dvp = a->a_dvp; 5715 vn_seqc_write_begin(dvp); 5716 } 5717 5718 void 5719 vop_create_post(void *ap, int rc) 5720 { 5721 struct vop_create_args *a; 5722 struct vnode *dvp; 5723 5724 a = ap; 5725 dvp = a->a_dvp; 5726 vn_seqc_write_end(dvp); 5727 if (!rc) 5728 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 5729 } 5730 5731 void 5732 vop_whiteout_pre(void *ap) 5733 { 5734 struct vop_whiteout_args *a; 5735 struct vnode *dvp; 5736 5737 a = ap; 5738 dvp = a->a_dvp; 5739 vn_seqc_write_begin(dvp); 5740 } 5741 5742 void 5743 vop_whiteout_post(void *ap, int rc) 5744 { 5745 struct vop_whiteout_args *a; 5746 struct vnode *dvp; 5747 5748 a = ap; 5749 dvp = a->a_dvp; 5750 vn_seqc_write_end(dvp); 5751 } 5752 5753 void 5754 vop_deleteextattr_pre(void *ap) 5755 { 5756 struct vop_deleteextattr_args *a; 5757 struct vnode *vp; 5758 5759 a = ap; 5760 vp = a->a_vp; 5761 vn_seqc_write_begin(vp); 5762 } 5763 5764 void 5765 vop_deleteextattr_post(void *ap, int rc) 5766 { 5767 struct vop_deleteextattr_args *a; 5768 struct vnode *vp; 5769 5770 a = ap; 5771 vp = a->a_vp; 5772 vn_seqc_write_end(vp); 5773 if (!rc) 5774 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 5775 } 5776 5777 void 5778 vop_link_pre(void *ap) 5779 { 5780 struct vop_link_args *a; 5781 struct vnode *vp, *tdvp; 5782 5783 a = ap; 5784 vp = a->a_vp; 5785 tdvp = a->a_tdvp; 5786 vn_seqc_write_begin(vp); 5787 vn_seqc_write_begin(tdvp); 5788 } 5789 5790 void 5791 vop_link_post(void *ap, int rc) 5792 { 5793 struct vop_link_args *a; 5794 struct vnode *vp, *tdvp; 5795 5796 a = ap; 5797 vp = a->a_vp; 5798 tdvp = a->a_tdvp; 5799 vn_seqc_write_end(vp); 5800 vn_seqc_write_end(tdvp); 5801 if (!rc) { 5802 VFS_KNOTE_LOCKED(vp, NOTE_LINK); 5803 VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE); 5804 } 5805 } 5806 5807 void 5808 vop_mkdir_pre(void *ap) 5809 { 5810 struct vop_mkdir_args *a; 5811 struct vnode *dvp; 5812 5813 a = ap; 5814 dvp = a->a_dvp; 5815 vn_seqc_write_begin(dvp); 5816 } 5817 5818 void 5819 vop_mkdir_post(void *ap, int rc) 5820 { 5821 struct vop_mkdir_args *a; 5822 struct vnode *dvp; 5823 5824 a = ap; 5825 dvp = a->a_dvp; 5826 vn_seqc_write_end(dvp); 5827 if (!rc) 5828 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 5829 } 5830 5831 #ifdef DEBUG_VFS_LOCKS 5832 void 5833 vop_mkdir_debugpost(void *ap, int rc) 5834 { 5835 struct vop_mkdir_args *a; 5836 5837 a = ap; 5838 if (!rc) 5839 cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp); 5840 } 5841 #endif 5842 5843 void 5844 vop_mknod_pre(void *ap) 5845 { 5846 struct vop_mknod_args *a; 5847 struct vnode *dvp; 5848 5849 a = ap; 5850 dvp = a->a_dvp; 5851 vn_seqc_write_begin(dvp); 5852 } 5853 5854 void 5855 vop_mknod_post(void *ap, int rc) 5856 { 5857 struct vop_mknod_args *a; 5858 struct vnode *dvp; 5859 5860 a = ap; 5861 dvp = a->a_dvp; 5862 vn_seqc_write_end(dvp); 5863 if (!rc) 5864 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 5865 } 5866 5867 void 5868 vop_reclaim_post(void *ap, int rc) 5869 { 5870 struct vop_reclaim_args *a; 5871 struct vnode *vp; 5872 5873 a = ap; 5874 vp = a->a_vp; 5875 ASSERT_VOP_IN_SEQC(vp); 5876 if (!rc) 5877 VFS_KNOTE_LOCKED(vp, NOTE_REVOKE); 5878 } 5879 5880 void 5881 vop_remove_pre(void *ap) 5882 { 5883 struct vop_remove_args *a; 5884 struct vnode *dvp, *vp; 5885 5886 a = ap; 5887 dvp = a->a_dvp; 5888 vp = a->a_vp; 5889 vn_seqc_write_begin(dvp); 5890 vn_seqc_write_begin(vp); 5891 } 5892 5893 void 5894 vop_remove_post(void *ap, int rc) 5895 { 5896 struct vop_remove_args *a; 5897 struct vnode *dvp, *vp; 5898 5899 a = ap; 5900 dvp = a->a_dvp; 5901 vp = a->a_vp; 5902 vn_seqc_write_end(dvp); 5903 vn_seqc_write_end(vp); 5904 if (!rc) { 5905 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 5906 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 5907 } 5908 } 5909 5910 void 5911 vop_rename_post(void *ap, int rc) 5912 { 5913 struct vop_rename_args *a = ap; 5914 long hint; 5915 5916 if (!rc) { 5917 hint = NOTE_WRITE; 5918 if (a->a_fdvp == a->a_tdvp) { 5919 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) 5920 hint |= NOTE_LINK; 5921 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 5922 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 5923 } else { 5924 hint |= NOTE_EXTEND; 5925 if (a->a_fvp->v_type == VDIR) 5926 hint |= NOTE_LINK; 5927 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 5928 5929 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && 5930 a->a_tvp->v_type == VDIR) 5931 hint &= ~NOTE_LINK; 5932 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 5933 } 5934 5935 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 5936 if (a->a_tvp) 5937 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 5938 } 5939 if (a->a_tdvp != a->a_fdvp) 5940 vdrop(a->a_fdvp); 5941 if (a->a_tvp != a->a_fvp) 5942 vdrop(a->a_fvp); 5943 vdrop(a->a_tdvp); 5944 if (a->a_tvp) 5945 vdrop(a->a_tvp); 5946 } 5947 5948 void 5949 vop_rmdir_pre(void *ap) 5950 { 5951 struct vop_rmdir_args *a; 5952 struct vnode *dvp, *vp; 5953 5954 a = ap; 5955 dvp = a->a_dvp; 5956 vp = a->a_vp; 5957 vn_seqc_write_begin(dvp); 5958 vn_seqc_write_begin(vp); 5959 } 5960 5961 void 5962 vop_rmdir_post(void *ap, int rc) 5963 { 5964 struct vop_rmdir_args *a; 5965 struct vnode *dvp, *vp; 5966 5967 a = ap; 5968 dvp = a->a_dvp; 5969 vp = a->a_vp; 5970 vn_seqc_write_end(dvp); 5971 vn_seqc_write_end(vp); 5972 if (!rc) { 5973 vp->v_vflag |= VV_UNLINKED; 5974 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 5975 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 5976 } 5977 } 5978 5979 void 5980 vop_setattr_pre(void *ap) 5981 { 5982 struct vop_setattr_args *a; 5983 struct vnode *vp; 5984 5985 a = ap; 5986 vp = a->a_vp; 5987 vn_seqc_write_begin(vp); 5988 } 5989 5990 void 5991 vop_setattr_post(void *ap, int rc) 5992 { 5993 struct vop_setattr_args *a; 5994 struct vnode *vp; 5995 5996 a = ap; 5997 vp = a->a_vp; 5998 vn_seqc_write_end(vp); 5999 if (!rc) 6000 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 6001 } 6002 6003 void 6004 vop_setacl_pre(void *ap) 6005 { 6006 struct vop_setacl_args *a; 6007 struct vnode *vp; 6008 6009 a = ap; 6010 vp = a->a_vp; 6011 vn_seqc_write_begin(vp); 6012 } 6013 6014 void 6015 vop_setacl_post(void *ap, int rc __unused) 6016 { 6017 struct vop_setacl_args *a; 6018 struct vnode *vp; 6019 6020 a = ap; 6021 vp = a->a_vp; 6022 vn_seqc_write_end(vp); 6023 } 6024 6025 void 6026 vop_setextattr_pre(void *ap) 6027 { 6028 struct vop_setextattr_args *a; 6029 struct vnode *vp; 6030 6031 a = ap; 6032 vp = a->a_vp; 6033 vn_seqc_write_begin(vp); 6034 } 6035 6036 void 6037 vop_setextattr_post(void *ap, int rc) 6038 { 6039 struct vop_setextattr_args *a; 6040 struct vnode *vp; 6041 6042 a = ap; 6043 vp = a->a_vp; 6044 vn_seqc_write_end(vp); 6045 if (!rc) 6046 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 6047 } 6048 6049 void 6050 vop_symlink_pre(void *ap) 6051 { 6052 struct vop_symlink_args *a; 6053 struct vnode *dvp; 6054 6055 a = ap; 6056 dvp = a->a_dvp; 6057 vn_seqc_write_begin(dvp); 6058 } 6059 6060 void 6061 vop_symlink_post(void *ap, int rc) 6062 { 6063 struct vop_symlink_args *a; 6064 struct vnode *dvp; 6065 6066 a = ap; 6067 dvp = a->a_dvp; 6068 vn_seqc_write_end(dvp); 6069 if (!rc) 6070 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 6071 } 6072 6073 void 6074 vop_open_post(void *ap, int rc) 6075 { 6076 struct vop_open_args *a = ap; 6077 6078 if (!rc) 6079 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); 6080 } 6081 6082 void 6083 vop_close_post(void *ap, int rc) 6084 { 6085 struct vop_close_args *a = ap; 6086 6087 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ 6088 !VN_IS_DOOMED(a->a_vp))) { 6089 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 6090 NOTE_CLOSE_WRITE : NOTE_CLOSE); 6091 } 6092 } 6093 6094 void 6095 vop_read_post(void *ap, int rc) 6096 { 6097 struct vop_read_args *a = ap; 6098 6099 if (!rc) 6100 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 6101 } 6102 6103 void 6104 vop_read_pgcache_post(void *ap, int rc) 6105 { 6106 struct vop_read_pgcache_args *a = ap; 6107 6108 if (!rc) 6109 VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ); 6110 } 6111 6112 void 6113 vop_readdir_post(void *ap, int rc) 6114 { 6115 struct vop_readdir_args *a = ap; 6116 6117 if (!rc) 6118 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 6119 } 6120 6121 static struct knlist fs_knlist; 6122 6123 static void 6124 vfs_event_init(void *arg) 6125 { 6126 knlist_init_mtx(&fs_knlist, NULL); 6127 } 6128 /* XXX - correct order? */ 6129 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 6130 6131 void 6132 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 6133 { 6134 6135 KNOTE_UNLOCKED(&fs_knlist, event); 6136 } 6137 6138 static int filt_fsattach(struct knote *kn); 6139 static void filt_fsdetach(struct knote *kn); 6140 static int filt_fsevent(struct knote *kn, long hint); 6141 6142 struct filterops fs_filtops = { 6143 .f_isfd = 0, 6144 .f_attach = filt_fsattach, 6145 .f_detach = filt_fsdetach, 6146 .f_event = filt_fsevent 6147 }; 6148 6149 static int 6150 filt_fsattach(struct knote *kn) 6151 { 6152 6153 kn->kn_flags |= EV_CLEAR; 6154 knlist_add(&fs_knlist, kn, 0); 6155 return (0); 6156 } 6157 6158 static void 6159 filt_fsdetach(struct knote *kn) 6160 { 6161 6162 knlist_remove(&fs_knlist, kn, 0); 6163 } 6164 6165 static int 6166 filt_fsevent(struct knote *kn, long hint) 6167 { 6168 6169 kn->kn_fflags |= kn->kn_sfflags & hint; 6170 6171 return (kn->kn_fflags != 0); 6172 } 6173 6174 static int 6175 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 6176 { 6177 struct vfsidctl vc; 6178 int error; 6179 struct mount *mp; 6180 6181 error = SYSCTL_IN(req, &vc, sizeof(vc)); 6182 if (error) 6183 return (error); 6184 if (vc.vc_vers != VFS_CTL_VERS1) 6185 return (EINVAL); 6186 mp = vfs_getvfs(&vc.vc_fsid); 6187 if (mp == NULL) 6188 return (ENOENT); 6189 /* ensure that a specific sysctl goes to the right filesystem. */ 6190 if (strcmp(vc.vc_fstypename, "*") != 0 && 6191 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 6192 vfs_rel(mp); 6193 return (EINVAL); 6194 } 6195 VCTLTOREQ(&vc, req); 6196 error = VFS_SYSCTL(mp, vc.vc_op, req); 6197 vfs_rel(mp); 6198 return (error); 6199 } 6200 6201 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR, 6202 NULL, 0, sysctl_vfs_ctl, "", 6203 "Sysctl by fsid"); 6204 6205 /* 6206 * Function to initialize a va_filerev field sensibly. 6207 * XXX: Wouldn't a random number make a lot more sense ?? 6208 */ 6209 u_quad_t 6210 init_va_filerev(void) 6211 { 6212 struct bintime bt; 6213 6214 getbinuptime(&bt); 6215 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 6216 } 6217 6218 static int filt_vfsread(struct knote *kn, long hint); 6219 static int filt_vfswrite(struct knote *kn, long hint); 6220 static int filt_vfsvnode(struct knote *kn, long hint); 6221 static void filt_vfsdetach(struct knote *kn); 6222 static struct filterops vfsread_filtops = { 6223 .f_isfd = 1, 6224 .f_detach = filt_vfsdetach, 6225 .f_event = filt_vfsread 6226 }; 6227 static struct filterops vfswrite_filtops = { 6228 .f_isfd = 1, 6229 .f_detach = filt_vfsdetach, 6230 .f_event = filt_vfswrite 6231 }; 6232 static struct filterops vfsvnode_filtops = { 6233 .f_isfd = 1, 6234 .f_detach = filt_vfsdetach, 6235 .f_event = filt_vfsvnode 6236 }; 6237 6238 static void 6239 vfs_knllock(void *arg) 6240 { 6241 struct vnode *vp = arg; 6242 6243 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 6244 } 6245 6246 static void 6247 vfs_knlunlock(void *arg) 6248 { 6249 struct vnode *vp = arg; 6250 6251 VOP_UNLOCK(vp); 6252 } 6253 6254 static void 6255 vfs_knl_assert_lock(void *arg, int what) 6256 { 6257 #ifdef DEBUG_VFS_LOCKS 6258 struct vnode *vp = arg; 6259 6260 if (what == LA_LOCKED) 6261 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 6262 else 6263 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 6264 #endif 6265 } 6266 6267 int 6268 vfs_kqfilter(struct vop_kqfilter_args *ap) 6269 { 6270 struct vnode *vp = ap->a_vp; 6271 struct knote *kn = ap->a_kn; 6272 struct knlist *knl; 6273 6274 KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ && 6275 kn->kn_filter != EVFILT_WRITE), 6276 ("READ/WRITE filter on a FIFO leaked through")); 6277 switch (kn->kn_filter) { 6278 case EVFILT_READ: 6279 kn->kn_fop = &vfsread_filtops; 6280 break; 6281 case EVFILT_WRITE: 6282 kn->kn_fop = &vfswrite_filtops; 6283 break; 6284 case EVFILT_VNODE: 6285 kn->kn_fop = &vfsvnode_filtops; 6286 break; 6287 default: 6288 return (EINVAL); 6289 } 6290 6291 kn->kn_hook = (caddr_t)vp; 6292 6293 v_addpollinfo(vp); 6294 if (vp->v_pollinfo == NULL) 6295 return (ENOMEM); 6296 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 6297 vhold(vp); 6298 knlist_add(knl, kn, 0); 6299 6300 return (0); 6301 } 6302 6303 /* 6304 * Detach knote from vnode 6305 */ 6306 static void 6307 filt_vfsdetach(struct knote *kn) 6308 { 6309 struct vnode *vp = (struct vnode *)kn->kn_hook; 6310 6311 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 6312 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 6313 vdrop(vp); 6314 } 6315 6316 /*ARGSUSED*/ 6317 static int 6318 filt_vfsread(struct knote *kn, long hint) 6319 { 6320 struct vnode *vp = (struct vnode *)kn->kn_hook; 6321 off_t size; 6322 int res; 6323 6324 /* 6325 * filesystem is gone, so set the EOF flag and schedule 6326 * the knote for deletion. 6327 */ 6328 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6329 VI_LOCK(vp); 6330 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6331 VI_UNLOCK(vp); 6332 return (1); 6333 } 6334 6335 if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0) 6336 return (0); 6337 6338 VI_LOCK(vp); 6339 kn->kn_data = size - kn->kn_fp->f_offset; 6340 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; 6341 VI_UNLOCK(vp); 6342 return (res); 6343 } 6344 6345 /*ARGSUSED*/ 6346 static int 6347 filt_vfswrite(struct knote *kn, long hint) 6348 { 6349 struct vnode *vp = (struct vnode *)kn->kn_hook; 6350 6351 VI_LOCK(vp); 6352 6353 /* 6354 * filesystem is gone, so set the EOF flag and schedule 6355 * the knote for deletion. 6356 */ 6357 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) 6358 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6359 6360 kn->kn_data = 0; 6361 VI_UNLOCK(vp); 6362 return (1); 6363 } 6364 6365 static int 6366 filt_vfsvnode(struct knote *kn, long hint) 6367 { 6368 struct vnode *vp = (struct vnode *)kn->kn_hook; 6369 int res; 6370 6371 VI_LOCK(vp); 6372 if (kn->kn_sfflags & hint) 6373 kn->kn_fflags |= hint; 6374 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6375 kn->kn_flags |= EV_EOF; 6376 VI_UNLOCK(vp); 6377 return (1); 6378 } 6379 res = (kn->kn_fflags != 0); 6380 VI_UNLOCK(vp); 6381 return (res); 6382 } 6383 6384 /* 6385 * Returns whether the directory is empty or not. 6386 * If it is empty, the return value is 0; otherwise 6387 * the return value is an error value (which may 6388 * be ENOTEMPTY). 6389 */ 6390 int 6391 vfs_emptydir(struct vnode *vp) 6392 { 6393 struct uio uio; 6394 struct iovec iov; 6395 struct dirent *dirent, *dp, *endp; 6396 int error, eof; 6397 6398 error = 0; 6399 eof = 0; 6400 6401 ASSERT_VOP_LOCKED(vp, "vfs_emptydir"); 6402 VNPASS(vp->v_type == VDIR, vp); 6403 6404 dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK); 6405 iov.iov_base = dirent; 6406 iov.iov_len = sizeof(struct dirent); 6407 6408 uio.uio_iov = &iov; 6409 uio.uio_iovcnt = 1; 6410 uio.uio_offset = 0; 6411 uio.uio_resid = sizeof(struct dirent); 6412 uio.uio_segflg = UIO_SYSSPACE; 6413 uio.uio_rw = UIO_READ; 6414 uio.uio_td = curthread; 6415 6416 while (eof == 0 && error == 0) { 6417 error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof, 6418 NULL, NULL); 6419 if (error != 0) 6420 break; 6421 endp = (void *)((uint8_t *)dirent + 6422 sizeof(struct dirent) - uio.uio_resid); 6423 for (dp = dirent; dp < endp; 6424 dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) { 6425 if (dp->d_type == DT_WHT) 6426 continue; 6427 if (dp->d_namlen == 0) 6428 continue; 6429 if (dp->d_type != DT_DIR && 6430 dp->d_type != DT_UNKNOWN) { 6431 error = ENOTEMPTY; 6432 break; 6433 } 6434 if (dp->d_namlen > 2) { 6435 error = ENOTEMPTY; 6436 break; 6437 } 6438 if (dp->d_namlen == 1 && 6439 dp->d_name[0] != '.') { 6440 error = ENOTEMPTY; 6441 break; 6442 } 6443 if (dp->d_namlen == 2 && 6444 dp->d_name[1] != '.') { 6445 error = ENOTEMPTY; 6446 break; 6447 } 6448 uio.uio_resid = sizeof(struct dirent); 6449 } 6450 } 6451 free(dirent, M_TEMP); 6452 return (error); 6453 } 6454 6455 int 6456 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 6457 { 6458 int error; 6459 6460 if (dp->d_reclen > ap->a_uio->uio_resid) 6461 return (ENAMETOOLONG); 6462 error = uiomove(dp, dp->d_reclen, ap->a_uio); 6463 if (error) { 6464 if (ap->a_ncookies != NULL) { 6465 if (ap->a_cookies != NULL) 6466 free(ap->a_cookies, M_TEMP); 6467 ap->a_cookies = NULL; 6468 *ap->a_ncookies = 0; 6469 } 6470 return (error); 6471 } 6472 if (ap->a_ncookies == NULL) 6473 return (0); 6474 6475 KASSERT(ap->a_cookies, 6476 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 6477 6478 *ap->a_cookies = realloc(*ap->a_cookies, 6479 (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO); 6480 (*ap->a_cookies)[*ap->a_ncookies] = off; 6481 *ap->a_ncookies += 1; 6482 return (0); 6483 } 6484 6485 /* 6486 * The purpose of this routine is to remove granularity from accmode_t, 6487 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 6488 * VADMIN and VAPPEND. 6489 * 6490 * If it returns 0, the caller is supposed to continue with the usual 6491 * access checks using 'accmode' as modified by this routine. If it 6492 * returns nonzero value, the caller is supposed to return that value 6493 * as errno. 6494 * 6495 * Note that after this routine runs, accmode may be zero. 6496 */ 6497 int 6498 vfs_unixify_accmode(accmode_t *accmode) 6499 { 6500 /* 6501 * There is no way to specify explicit "deny" rule using 6502 * file mode or POSIX.1e ACLs. 6503 */ 6504 if (*accmode & VEXPLICIT_DENY) { 6505 *accmode = 0; 6506 return (0); 6507 } 6508 6509 /* 6510 * None of these can be translated into usual access bits. 6511 * Also, the common case for NFSv4 ACLs is to not contain 6512 * either of these bits. Caller should check for VWRITE 6513 * on the containing directory instead. 6514 */ 6515 if (*accmode & (VDELETE_CHILD | VDELETE)) 6516 return (EPERM); 6517 6518 if (*accmode & VADMIN_PERMS) { 6519 *accmode &= ~VADMIN_PERMS; 6520 *accmode |= VADMIN; 6521 } 6522 6523 /* 6524 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 6525 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 6526 */ 6527 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 6528 6529 return (0); 6530 } 6531 6532 /* 6533 * Clear out a doomed vnode (if any) and replace it with a new one as long 6534 * as the fs is not being unmounted. Return the root vnode to the caller. 6535 */ 6536 static int __noinline 6537 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp) 6538 { 6539 struct vnode *vp; 6540 int error; 6541 6542 restart: 6543 if (mp->mnt_rootvnode != NULL) { 6544 MNT_ILOCK(mp); 6545 vp = mp->mnt_rootvnode; 6546 if (vp != NULL) { 6547 if (!VN_IS_DOOMED(vp)) { 6548 vrefact(vp); 6549 MNT_IUNLOCK(mp); 6550 error = vn_lock(vp, flags); 6551 if (error == 0) { 6552 *vpp = vp; 6553 return (0); 6554 } 6555 vrele(vp); 6556 goto restart; 6557 } 6558 /* 6559 * Clear the old one. 6560 */ 6561 mp->mnt_rootvnode = NULL; 6562 } 6563 MNT_IUNLOCK(mp); 6564 if (vp != NULL) { 6565 vfs_op_barrier_wait(mp); 6566 vrele(vp); 6567 } 6568 } 6569 error = VFS_CACHEDROOT(mp, flags, vpp); 6570 if (error != 0) 6571 return (error); 6572 if (mp->mnt_vfs_ops == 0) { 6573 MNT_ILOCK(mp); 6574 if (mp->mnt_vfs_ops != 0) { 6575 MNT_IUNLOCK(mp); 6576 return (0); 6577 } 6578 if (mp->mnt_rootvnode == NULL) { 6579 vrefact(*vpp); 6580 mp->mnt_rootvnode = *vpp; 6581 } else { 6582 if (mp->mnt_rootvnode != *vpp) { 6583 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) { 6584 panic("%s: mismatch between vnode returned " 6585 " by VFS_CACHEDROOT and the one cached " 6586 " (%p != %p)", 6587 __func__, *vpp, mp->mnt_rootvnode); 6588 } 6589 } 6590 } 6591 MNT_IUNLOCK(mp); 6592 } 6593 return (0); 6594 } 6595 6596 int 6597 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp) 6598 { 6599 struct mount_pcpu *mpcpu; 6600 struct vnode *vp; 6601 int error; 6602 6603 if (!vfs_op_thread_enter(mp, mpcpu)) 6604 return (vfs_cache_root_fallback(mp, flags, vpp)); 6605 vp = atomic_load_ptr(&mp->mnt_rootvnode); 6606 if (vp == NULL || VN_IS_DOOMED(vp)) { 6607 vfs_op_thread_exit(mp, mpcpu); 6608 return (vfs_cache_root_fallback(mp, flags, vpp)); 6609 } 6610 vrefact(vp); 6611 vfs_op_thread_exit(mp, mpcpu); 6612 error = vn_lock(vp, flags); 6613 if (error != 0) { 6614 vrele(vp); 6615 return (vfs_cache_root_fallback(mp, flags, vpp)); 6616 } 6617 *vpp = vp; 6618 return (0); 6619 } 6620 6621 struct vnode * 6622 vfs_cache_root_clear(struct mount *mp) 6623 { 6624 struct vnode *vp; 6625 6626 /* 6627 * ops > 0 guarantees there is nobody who can see this vnode 6628 */ 6629 MPASS(mp->mnt_vfs_ops > 0); 6630 vp = mp->mnt_rootvnode; 6631 if (vp != NULL) 6632 vn_seqc_write_begin(vp); 6633 mp->mnt_rootvnode = NULL; 6634 return (vp); 6635 } 6636 6637 void 6638 vfs_cache_root_set(struct mount *mp, struct vnode *vp) 6639 { 6640 6641 MPASS(mp->mnt_vfs_ops > 0); 6642 vrefact(vp); 6643 mp->mnt_rootvnode = vp; 6644 } 6645 6646 /* 6647 * These are helper functions for filesystems to traverse all 6648 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 6649 * 6650 * This interface replaces MNT_VNODE_FOREACH. 6651 */ 6652 6653 struct vnode * 6654 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 6655 { 6656 struct vnode *vp; 6657 6658 maybe_yield(); 6659 MNT_ILOCK(mp); 6660 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6661 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL; 6662 vp = TAILQ_NEXT(vp, v_nmntvnodes)) { 6663 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 6664 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 6665 continue; 6666 VI_LOCK(vp); 6667 if (VN_IS_DOOMED(vp)) { 6668 VI_UNLOCK(vp); 6669 continue; 6670 } 6671 break; 6672 } 6673 if (vp == NULL) { 6674 __mnt_vnode_markerfree_all(mvp, mp); 6675 /* MNT_IUNLOCK(mp); -- done in above function */ 6676 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 6677 return (NULL); 6678 } 6679 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 6680 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 6681 MNT_IUNLOCK(mp); 6682 return (vp); 6683 } 6684 6685 struct vnode * 6686 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 6687 { 6688 struct vnode *vp; 6689 6690 *mvp = vn_alloc_marker(mp); 6691 MNT_ILOCK(mp); 6692 MNT_REF(mp); 6693 6694 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 6695 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 6696 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 6697 continue; 6698 VI_LOCK(vp); 6699 if (VN_IS_DOOMED(vp)) { 6700 VI_UNLOCK(vp); 6701 continue; 6702 } 6703 break; 6704 } 6705 if (vp == NULL) { 6706 MNT_REL(mp); 6707 MNT_IUNLOCK(mp); 6708 vn_free_marker(*mvp); 6709 *mvp = NULL; 6710 return (NULL); 6711 } 6712 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 6713 MNT_IUNLOCK(mp); 6714 return (vp); 6715 } 6716 6717 void 6718 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 6719 { 6720 6721 if (*mvp == NULL) { 6722 MNT_IUNLOCK(mp); 6723 return; 6724 } 6725 6726 mtx_assert(MNT_MTX(mp), MA_OWNED); 6727 6728 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6729 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 6730 MNT_REL(mp); 6731 MNT_IUNLOCK(mp); 6732 vn_free_marker(*mvp); 6733 *mvp = NULL; 6734 } 6735 6736 /* 6737 * These are helper functions for filesystems to traverse their 6738 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h 6739 */ 6740 static void 6741 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 6742 { 6743 6744 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6745 6746 MNT_ILOCK(mp); 6747 MNT_REL(mp); 6748 MNT_IUNLOCK(mp); 6749 vn_free_marker(*mvp); 6750 *mvp = NULL; 6751 } 6752 6753 /* 6754 * Relock the mp mount vnode list lock with the vp vnode interlock in the 6755 * conventional lock order during mnt_vnode_next_lazy iteration. 6756 * 6757 * On entry, the mount vnode list lock is held and the vnode interlock is not. 6758 * The list lock is dropped and reacquired. On success, both locks are held. 6759 * On failure, the mount vnode list lock is held but the vnode interlock is 6760 * not, and the procedure may have yielded. 6761 */ 6762 static bool 6763 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp, 6764 struct vnode *vp) 6765 { 6766 6767 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER && 6768 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp, 6769 ("%s: bad marker", __func__)); 6770 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp, 6771 ("%s: inappropriate vnode", __func__)); 6772 ASSERT_VI_UNLOCKED(vp, __func__); 6773 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 6774 6775 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist); 6776 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist); 6777 6778 /* 6779 * Note we may be racing against vdrop which transitioned the hold 6780 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine, 6781 * if we are the only user after we get the interlock we will just 6782 * vdrop. 6783 */ 6784 vhold(vp); 6785 mtx_unlock(&mp->mnt_listmtx); 6786 VI_LOCK(vp); 6787 if (VN_IS_DOOMED(vp)) { 6788 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); 6789 goto out_lost; 6790 } 6791 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 6792 /* 6793 * There is nothing to do if we are the last user. 6794 */ 6795 if (!refcount_release_if_not_last(&vp->v_holdcnt)) 6796 goto out_lost; 6797 mtx_lock(&mp->mnt_listmtx); 6798 return (true); 6799 out_lost: 6800 vdropl(vp); 6801 maybe_yield(); 6802 mtx_lock(&mp->mnt_listmtx); 6803 return (false); 6804 } 6805 6806 static struct vnode * 6807 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 6808 void *cbarg) 6809 { 6810 struct vnode *vp; 6811 6812 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 6813 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6814 restart: 6815 vp = TAILQ_NEXT(*mvp, v_lazylist); 6816 while (vp != NULL) { 6817 if (vp->v_type == VMARKER) { 6818 vp = TAILQ_NEXT(vp, v_lazylist); 6819 continue; 6820 } 6821 /* 6822 * See if we want to process the vnode. Note we may encounter a 6823 * long string of vnodes we don't care about and hog the list 6824 * as a result. Check for it and requeue the marker. 6825 */ 6826 VNPASS(!VN_IS_DOOMED(vp), vp); 6827 if (!cb(vp, cbarg)) { 6828 if (!should_yield()) { 6829 vp = TAILQ_NEXT(vp, v_lazylist); 6830 continue; 6831 } 6832 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, 6833 v_lazylist); 6834 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, 6835 v_lazylist); 6836 mtx_unlock(&mp->mnt_listmtx); 6837 kern_yield(PRI_USER); 6838 mtx_lock(&mp->mnt_listmtx); 6839 goto restart; 6840 } 6841 /* 6842 * Try-lock because this is the wrong lock order. 6843 */ 6844 if (!VI_TRYLOCK(vp) && 6845 !mnt_vnode_next_lazy_relock(*mvp, mp, vp)) 6846 goto restart; 6847 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 6848 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 6849 ("alien vnode on the lazy list %p %p", vp, mp)); 6850 VNPASS(vp->v_mount == mp, vp); 6851 VNPASS(!VN_IS_DOOMED(vp), vp); 6852 break; 6853 } 6854 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 6855 6856 /* Check if we are done */ 6857 if (vp == NULL) { 6858 mtx_unlock(&mp->mnt_listmtx); 6859 mnt_vnode_markerfree_lazy(mvp, mp); 6860 return (NULL); 6861 } 6862 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist); 6863 mtx_unlock(&mp->mnt_listmtx); 6864 ASSERT_VI_LOCKED(vp, "lazy iter"); 6865 return (vp); 6866 } 6867 6868 struct vnode * 6869 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 6870 void *cbarg) 6871 { 6872 6873 maybe_yield(); 6874 mtx_lock(&mp->mnt_listmtx); 6875 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 6876 } 6877 6878 struct vnode * 6879 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 6880 void *cbarg) 6881 { 6882 struct vnode *vp; 6883 6884 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist)) 6885 return (NULL); 6886 6887 *mvp = vn_alloc_marker(mp); 6888 MNT_ILOCK(mp); 6889 MNT_REF(mp); 6890 MNT_IUNLOCK(mp); 6891 6892 mtx_lock(&mp->mnt_listmtx); 6893 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist); 6894 if (vp == NULL) { 6895 mtx_unlock(&mp->mnt_listmtx); 6896 mnt_vnode_markerfree_lazy(mvp, mp); 6897 return (NULL); 6898 } 6899 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist); 6900 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 6901 } 6902 6903 void 6904 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 6905 { 6906 6907 if (*mvp == NULL) 6908 return; 6909 6910 mtx_lock(&mp->mnt_listmtx); 6911 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 6912 mtx_unlock(&mp->mnt_listmtx); 6913 mnt_vnode_markerfree_lazy(mvp, mp); 6914 } 6915 6916 int 6917 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp) 6918 { 6919 6920 if ((cnp->cn_flags & NOEXECCHECK) != 0) { 6921 cnp->cn_flags &= ~NOEXECCHECK; 6922 return (0); 6923 } 6924 6925 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread)); 6926 } 6927 6928 /* 6929 * Do not use this variant unless you have means other than the hold count 6930 * to prevent the vnode from getting freed. 6931 */ 6932 void 6933 vn_seqc_write_begin_locked(struct vnode *vp) 6934 { 6935 6936 ASSERT_VI_LOCKED(vp, __func__); 6937 VNPASS(vp->v_holdcnt > 0, vp); 6938 VNPASS(vp->v_seqc_users >= 0, vp); 6939 vp->v_seqc_users++; 6940 if (vp->v_seqc_users == 1) 6941 seqc_sleepable_write_begin(&vp->v_seqc); 6942 } 6943 6944 void 6945 vn_seqc_write_begin(struct vnode *vp) 6946 { 6947 6948 VI_LOCK(vp); 6949 vn_seqc_write_begin_locked(vp); 6950 VI_UNLOCK(vp); 6951 } 6952 6953 void 6954 vn_seqc_write_end_locked(struct vnode *vp) 6955 { 6956 6957 ASSERT_VI_LOCKED(vp, __func__); 6958 VNPASS(vp->v_seqc_users > 0, vp); 6959 vp->v_seqc_users--; 6960 if (vp->v_seqc_users == 0) 6961 seqc_sleepable_write_end(&vp->v_seqc); 6962 } 6963 6964 void 6965 vn_seqc_write_end(struct vnode *vp) 6966 { 6967 6968 VI_LOCK(vp); 6969 vn_seqc_write_end_locked(vp); 6970 VI_UNLOCK(vp); 6971 } 6972 6973 /* 6974 * Special case handling for allocating and freeing vnodes. 6975 * 6976 * The counter remains unchanged on free so that a doomed vnode will 6977 * keep testing as in modify as long as it is accessible with SMR. 6978 */ 6979 static void 6980 vn_seqc_init(struct vnode *vp) 6981 { 6982 6983 vp->v_seqc = 0; 6984 vp->v_seqc_users = 0; 6985 } 6986 6987 static void 6988 vn_seqc_write_end_free(struct vnode *vp) 6989 { 6990 6991 VNPASS(seqc_in_modify(vp->v_seqc), vp); 6992 VNPASS(vp->v_seqc_users == 1, vp); 6993 } 6994 6995 void 6996 vn_irflag_set_locked(struct vnode *vp, short toset) 6997 { 6998 short flags; 6999 7000 ASSERT_VI_LOCKED(vp, __func__); 7001 flags = vn_irflag_read(vp); 7002 VNASSERT((flags & toset) == 0, vp, 7003 ("%s: some of the passed flags already set (have %d, passed %d)\n", 7004 __func__, flags, toset)); 7005 atomic_store_short(&vp->v_irflag, flags | toset); 7006 } 7007 7008 void 7009 vn_irflag_set(struct vnode *vp, short toset) 7010 { 7011 7012 VI_LOCK(vp); 7013 vn_irflag_set_locked(vp, toset); 7014 VI_UNLOCK(vp); 7015 } 7016 7017 void 7018 vn_irflag_set_cond_locked(struct vnode *vp, short toset) 7019 { 7020 short flags; 7021 7022 ASSERT_VI_LOCKED(vp, __func__); 7023 flags = vn_irflag_read(vp); 7024 atomic_store_short(&vp->v_irflag, flags | toset); 7025 } 7026 7027 void 7028 vn_irflag_set_cond(struct vnode *vp, short toset) 7029 { 7030 7031 VI_LOCK(vp); 7032 vn_irflag_set_cond_locked(vp, toset); 7033 VI_UNLOCK(vp); 7034 } 7035 7036 void 7037 vn_irflag_unset_locked(struct vnode *vp, short tounset) 7038 { 7039 short flags; 7040 7041 ASSERT_VI_LOCKED(vp, __func__); 7042 flags = vn_irflag_read(vp); 7043 VNASSERT((flags & tounset) == tounset, vp, 7044 ("%s: some of the passed flags not set (have %d, passed %d)\n", 7045 __func__, flags, tounset)); 7046 atomic_store_short(&vp->v_irflag, flags & ~tounset); 7047 } 7048 7049 void 7050 vn_irflag_unset(struct vnode *vp, short tounset) 7051 { 7052 7053 VI_LOCK(vp); 7054 vn_irflag_unset_locked(vp, tounset); 7055 VI_UNLOCK(vp); 7056 } 7057 7058 int 7059 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred) 7060 { 7061 struct vattr vattr; 7062 int error; 7063 7064 ASSERT_VOP_LOCKED(vp, __func__); 7065 error = VOP_GETATTR(vp, &vattr, cred); 7066 if (__predict_true(error == 0)) { 7067 if (vattr.va_size <= OFF_MAX) 7068 *size = vattr.va_size; 7069 else 7070 error = EFBIG; 7071 } 7072 return (error); 7073 } 7074 7075 int 7076 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred) 7077 { 7078 int error; 7079 7080 VOP_LOCK(vp, LK_SHARED); 7081 error = vn_getsize_locked(vp, size, cred); 7082 VOP_UNLOCK(vp); 7083 return (error); 7084 } 7085 7086 #ifdef INVARIANTS 7087 void 7088 vn_set_state_validate(struct vnode *vp, enum vstate state) 7089 { 7090 7091 switch (vp->v_state) { 7092 case VSTATE_UNINITIALIZED: 7093 switch (state) { 7094 case VSTATE_CONSTRUCTED: 7095 case VSTATE_DESTROYING: 7096 return; 7097 default: 7098 break; 7099 } 7100 break; 7101 case VSTATE_CONSTRUCTED: 7102 ASSERT_VOP_ELOCKED(vp, __func__); 7103 switch (state) { 7104 case VSTATE_DESTROYING: 7105 return; 7106 default: 7107 break; 7108 } 7109 break; 7110 case VSTATE_DESTROYING: 7111 ASSERT_VOP_ELOCKED(vp, __func__); 7112 switch (state) { 7113 case VSTATE_DEAD: 7114 return; 7115 default: 7116 break; 7117 } 7118 break; 7119 case VSTATE_DEAD: 7120 switch (state) { 7121 case VSTATE_UNINITIALIZED: 7122 return; 7123 default: 7124 break; 7125 } 7126 break; 7127 } 7128 7129 vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state); 7130 panic("invalid state transition %d -> %d\n", vp->v_state, state); 7131 } 7132 #endif 7133