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/bio.h> 52 #include <sys/buf.h> 53 #include <sys/capsicum.h> 54 #include <sys/condvar.h> 55 #include <sys/conf.h> 56 #include <sys/counter.h> 57 #include <sys/dirent.h> 58 #include <sys/event.h> 59 #include <sys/eventhandler.h> 60 #include <sys/extattr.h> 61 #include <sys/file.h> 62 #include <sys/fcntl.h> 63 #include <sys/jail.h> 64 #include <sys/kdb.h> 65 #include <sys/kernel.h> 66 #include <sys/kthread.h> 67 #include <sys/ktr.h> 68 #include <sys/lockf.h> 69 #include <sys/malloc.h> 70 #include <sys/mount.h> 71 #include <sys/namei.h> 72 #include <sys/pctrie.h> 73 #include <sys/priv.h> 74 #include <sys/reboot.h> 75 #include <sys/refcount.h> 76 #include <sys/rwlock.h> 77 #include <sys/sched.h> 78 #include <sys/sleepqueue.h> 79 #include <sys/smr.h> 80 #include <sys/smp.h> 81 #include <sys/stat.h> 82 #include <sys/sysctl.h> 83 #include <sys/syslog.h> 84 #include <sys/vmmeter.h> 85 #include <sys/vnode.h> 86 #include <sys/watchdog.h> 87 88 #include <machine/stdarg.h> 89 90 #include <security/mac/mac_framework.h> 91 92 #include <vm/vm.h> 93 #include <vm/vm_object.h> 94 #include <vm/vm_extern.h> 95 #include <vm/pmap.h> 96 #include <vm/vm_map.h> 97 #include <vm/vm_page.h> 98 #include <vm/vm_kern.h> 99 #include <vm/uma.h> 100 101 #ifdef DDB 102 #include <ddb/ddb.h> 103 #endif 104 105 static void delmntque(struct vnode *vp); 106 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, 107 int slpflag, int slptimeo); 108 static void syncer_shutdown(void *arg, int howto); 109 static int vtryrecycle(struct vnode *vp); 110 static void v_init_counters(struct vnode *); 111 static void vgonel(struct vnode *); 112 static void vfs_knllock(void *arg); 113 static void vfs_knlunlock(void *arg); 114 static void vfs_knl_assert_locked(void *arg); 115 static void vfs_knl_assert_unlocked(void *arg); 116 static void destroy_vpollinfo(struct vpollinfo *vi); 117 static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, 118 daddr_t startlbn, daddr_t endlbn); 119 static void vnlru_recalc(void); 120 121 /* 122 * These fences are intended for cases where some synchronization is 123 * needed between access of v_iflags and lockless vnode refcount (v_holdcnt 124 * and v_usecount) updates. Access to v_iflags is generally synchronized 125 * by the interlock, but we have some internal assertions that check vnode 126 * flags without acquiring the lock. Thus, these fences are INVARIANTS-only 127 * for now. 128 */ 129 #ifdef INVARIANTS 130 #define VNODE_REFCOUNT_FENCE_ACQ() atomic_thread_fence_acq() 131 #define VNODE_REFCOUNT_FENCE_REL() atomic_thread_fence_rel() 132 #else 133 #define VNODE_REFCOUNT_FENCE_ACQ() 134 #define VNODE_REFCOUNT_FENCE_REL() 135 #endif 136 137 /* 138 * Number of vnodes in existence. Increased whenever getnewvnode() 139 * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode. 140 */ 141 static u_long __exclusive_cache_line numvnodes; 142 143 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, 144 "Number of vnodes in existence"); 145 146 static counter_u64_t vnodes_created; 147 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created, 148 "Number of vnodes created by getnewvnode"); 149 150 /* 151 * Conversion tables for conversion from vnode types to inode formats 152 * and back. 153 */ 154 enum vtype iftovt_tab[16] = { 155 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 156 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON 157 }; 158 int vttoif_tab[10] = { 159 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 160 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT 161 }; 162 163 /* 164 * List of allocates vnodes in the system. 165 */ 166 static TAILQ_HEAD(freelst, vnode) vnode_list; 167 static struct vnode *vnode_list_free_marker; 168 static struct vnode *vnode_list_reclaim_marker; 169 170 /* 171 * "Free" vnode target. Free vnodes are rarely completely free, but are 172 * just ones that are cheap to recycle. Usually they are for files which 173 * have been stat'd but not read; these usually have inode and namecache 174 * data attached to them. This target is the preferred minimum size of a 175 * sub-cache consisting mostly of such files. The system balances the size 176 * of this sub-cache with its complement to try to prevent either from 177 * thrashing while the other is relatively inactive. The targets express 178 * a preference for the best balance. 179 * 180 * "Above" this target there are 2 further targets (watermarks) related 181 * to recyling of free vnodes. In the best-operating case, the cache is 182 * exactly full, the free list has size between vlowat and vhiwat above the 183 * free target, and recycling from it and normal use maintains this state. 184 * Sometimes the free list is below vlowat or even empty, but this state 185 * is even better for immediate use provided the cache is not full. 186 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free 187 * ones) to reach one of these states. The watermarks are currently hard- 188 * coded as 4% and 9% of the available space higher. These and the default 189 * of 25% for wantfreevnodes are too large if the memory size is large. 190 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim 191 * whenever vnlru_proc() becomes active. 192 */ 193 static long wantfreevnodes; 194 static long __exclusive_cache_line freevnodes; 195 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, 196 &freevnodes, 0, "Number of \"free\" vnodes"); 197 static long freevnodes_old; 198 199 static counter_u64_t recycles_count; 200 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count, 201 "Number of vnodes recycled to meet vnode cache targets"); 202 203 static counter_u64_t recycles_free_count; 204 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count, 205 "Number of free vnodes recycled to meet vnode cache targets"); 206 207 static counter_u64_t deferred_inact; 208 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact, 209 "Number of times inactive processing was deferred"); 210 211 /* To keep more than one thread at a time from running vfs_getnewfsid */ 212 static struct mtx mntid_mtx; 213 214 /* 215 * Lock for any access to the following: 216 * vnode_list 217 * numvnodes 218 * freevnodes 219 */ 220 static struct mtx __exclusive_cache_line vnode_list_mtx; 221 222 /* Publicly exported FS */ 223 struct nfs_public nfs_pub; 224 225 static uma_zone_t buf_trie_zone; 226 static smr_t buf_trie_smr; 227 228 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 229 static uma_zone_t vnode_zone; 230 static uma_zone_t vnodepoll_zone; 231 232 __read_frequently smr_t vfs_smr; 233 234 /* 235 * The workitem queue. 236 * 237 * It is useful to delay writes of file data and filesystem metadata 238 * for tens of seconds so that quickly created and deleted files need 239 * not waste disk bandwidth being created and removed. To realize this, 240 * we append vnodes to a "workitem" queue. When running with a soft 241 * updates implementation, most pending metadata dependencies should 242 * not wait for more than a few seconds. Thus, mounted on block devices 243 * are delayed only about a half the time that file data is delayed. 244 * Similarly, directory updates are more critical, so are only delayed 245 * about a third the time that file data is delayed. Thus, there are 246 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 247 * one each second (driven off the filesystem syncer process). The 248 * syncer_delayno variable indicates the next queue that is to be processed. 249 * Items that need to be processed soon are placed in this queue: 250 * 251 * syncer_workitem_pending[syncer_delayno] 252 * 253 * A delay of fifteen seconds is done by placing the request fifteen 254 * entries later in the queue: 255 * 256 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 257 * 258 */ 259 static int syncer_delayno; 260 static long syncer_mask; 261 LIST_HEAD(synclist, bufobj); 262 static struct synclist *syncer_workitem_pending; 263 /* 264 * The sync_mtx protects: 265 * bo->bo_synclist 266 * sync_vnode_count 267 * syncer_delayno 268 * syncer_state 269 * syncer_workitem_pending 270 * syncer_worklist_len 271 * rushjob 272 */ 273 static struct mtx sync_mtx; 274 static struct cv sync_wakeup; 275 276 #define SYNCER_MAXDELAY 32 277 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 278 static int syncdelay = 30; /* max time to delay syncing data */ 279 static int filedelay = 30; /* time to delay syncing files */ 280 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, 281 "Time to delay syncing files (in seconds)"); 282 static int dirdelay = 29; /* time to delay syncing directories */ 283 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, 284 "Time to delay syncing directories (in seconds)"); 285 static int metadelay = 28; /* time to delay syncing metadata */ 286 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, 287 "Time to delay syncing metadata (in seconds)"); 288 static int rushjob; /* number of slots to run ASAP */ 289 static int stat_rush_requests; /* number of times I/O speeded up */ 290 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, 291 "Number of times I/O speeded up (rush requests)"); 292 293 #define VDBATCH_SIZE 8 294 struct vdbatch { 295 u_int index; 296 long freevnodes; 297 struct mtx lock; 298 struct vnode *tab[VDBATCH_SIZE]; 299 }; 300 DPCPU_DEFINE_STATIC(struct vdbatch, vd); 301 302 static void vdbatch_dequeue(struct vnode *vp); 303 304 /* 305 * When shutting down the syncer, run it at four times normal speed. 306 */ 307 #define SYNCER_SHUTDOWN_SPEEDUP 4 308 static int sync_vnode_count; 309 static int syncer_worklist_len; 310 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 311 syncer_state; 312 313 /* Target for maximum number of vnodes. */ 314 u_long desiredvnodes; 315 static u_long gapvnodes; /* gap between wanted and desired */ 316 static u_long vhiwat; /* enough extras after expansion */ 317 static u_long vlowat; /* minimal extras before expansion */ 318 static u_long vstir; /* nonzero to stir non-free vnodes */ 319 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */ 320 321 static u_long vnlru_read_freevnodes(void); 322 323 /* 324 * Note that no attempt is made to sanitize these parameters. 325 */ 326 static int 327 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS) 328 { 329 u_long val; 330 int error; 331 332 val = desiredvnodes; 333 error = sysctl_handle_long(oidp, &val, 0, req); 334 if (error != 0 || req->newptr == NULL) 335 return (error); 336 337 if (val == desiredvnodes) 338 return (0); 339 mtx_lock(&vnode_list_mtx); 340 desiredvnodes = val; 341 wantfreevnodes = desiredvnodes / 4; 342 vnlru_recalc(); 343 mtx_unlock(&vnode_list_mtx); 344 /* 345 * XXX There is no protection against multiple threads changing 346 * desiredvnodes at the same time. Locking above only helps vnlru and 347 * getnewvnode. 348 */ 349 vfs_hash_changesize(desiredvnodes); 350 cache_changesize(desiredvnodes); 351 return (0); 352 } 353 354 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes, 355 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes, 356 "LU", "Target for maximum number of vnodes"); 357 358 static int 359 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS) 360 { 361 u_long val; 362 int error; 363 364 val = wantfreevnodes; 365 error = sysctl_handle_long(oidp, &val, 0, req); 366 if (error != 0 || req->newptr == NULL) 367 return (error); 368 369 if (val == wantfreevnodes) 370 return (0); 371 mtx_lock(&vnode_list_mtx); 372 wantfreevnodes = val; 373 vnlru_recalc(); 374 mtx_unlock(&vnode_list_mtx); 375 return (0); 376 } 377 378 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes, 379 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes, 380 "LU", "Target for minimum number of \"free\" vnodes"); 381 382 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 383 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)"); 384 static int vnlru_nowhere; 385 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 386 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 387 388 static int 389 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS) 390 { 391 struct vnode *vp; 392 struct nameidata nd; 393 char *buf; 394 unsigned long ndflags; 395 int error; 396 397 if (req->newptr == NULL) 398 return (EINVAL); 399 if (req->newlen >= PATH_MAX) 400 return (E2BIG); 401 402 buf = malloc(PATH_MAX, M_TEMP, M_WAITOK); 403 error = SYSCTL_IN(req, buf, req->newlen); 404 if (error != 0) 405 goto out; 406 407 buf[req->newlen] = '\0'; 408 409 ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | SAVENAME; 410 NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread); 411 if ((error = namei(&nd)) != 0) 412 goto out; 413 vp = nd.ni_vp; 414 415 if (VN_IS_DOOMED(vp)) { 416 /* 417 * This vnode is being recycled. Return != 0 to let the caller 418 * know that the sysctl had no effect. Return EAGAIN because a 419 * subsequent call will likely succeed (since namei will create 420 * a new vnode if necessary) 421 */ 422 error = EAGAIN; 423 goto putvnode; 424 } 425 426 counter_u64_add(recycles_count, 1); 427 vgone(vp); 428 putvnode: 429 NDFREE(&nd, 0); 430 out: 431 free(buf, M_TEMP); 432 return (error); 433 } 434 435 static int 436 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS) 437 { 438 struct thread *td = curthread; 439 struct vnode *vp; 440 struct file *fp; 441 int error; 442 int fd; 443 444 if (req->newptr == NULL) 445 return (EBADF); 446 447 error = sysctl_handle_int(oidp, &fd, 0, req); 448 if (error != 0) 449 return (error); 450 error = getvnode(curthread, fd, &cap_fcntl_rights, &fp); 451 if (error != 0) 452 return (error); 453 vp = fp->f_vnode; 454 455 error = vn_lock(vp, LK_EXCLUSIVE); 456 if (error != 0) 457 goto drop; 458 459 counter_u64_add(recycles_count, 1); 460 vgone(vp); 461 VOP_UNLOCK(vp); 462 drop: 463 fdrop(fp, td); 464 return (error); 465 } 466 467 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode, 468 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, 469 sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname"); 470 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode, 471 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, 472 sysctl_ftry_reclaim_vnode, "I", 473 "Try to reclaim a vnode by its file descriptor"); 474 475 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */ 476 static int vnsz2log; 477 478 /* 479 * Support for the bufobj clean & dirty pctrie. 480 */ 481 static void * 482 buf_trie_alloc(struct pctrie *ptree) 483 { 484 return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT)); 485 } 486 487 static void 488 buf_trie_free(struct pctrie *ptree, void *node) 489 { 490 uma_zfree_smr(buf_trie_zone, node); 491 } 492 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free, 493 buf_trie_smr); 494 495 /* 496 * Initialize the vnode management data structures. 497 * 498 * Reevaluate the following cap on the number of vnodes after the physical 499 * memory size exceeds 512GB. In the limit, as the physical memory size 500 * grows, the ratio of the memory size in KB to vnodes approaches 64:1. 501 */ 502 #ifndef MAXVNODES_MAX 503 #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */ 504 #endif 505 506 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker"); 507 508 static struct vnode * 509 vn_alloc_marker(struct mount *mp) 510 { 511 struct vnode *vp; 512 513 vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 514 vp->v_type = VMARKER; 515 vp->v_mount = mp; 516 517 return (vp); 518 } 519 520 static void 521 vn_free_marker(struct vnode *vp) 522 { 523 524 MPASS(vp->v_type == VMARKER); 525 free(vp, M_VNODE_MARKER); 526 } 527 528 /* 529 * Initialize a vnode as it first enters the zone. 530 */ 531 static int 532 vnode_init(void *mem, int size, int flags) 533 { 534 struct vnode *vp; 535 536 vp = mem; 537 bzero(vp, size); 538 /* 539 * Setup locks. 540 */ 541 vp->v_vnlock = &vp->v_lock; 542 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 543 /* 544 * By default, don't allow shared locks unless filesystems opt-in. 545 */ 546 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT, 547 LK_NOSHARE | LK_IS_VNODE); 548 /* 549 * Initialize bufobj. 550 */ 551 bufobj_init(&vp->v_bufobj, vp); 552 /* 553 * Initialize namecache. 554 */ 555 cache_vnode_init(vp); 556 /* 557 * Initialize rangelocks. 558 */ 559 rangelock_init(&vp->v_rl); 560 561 vp->v_dbatchcpu = NOCPU; 562 563 mtx_lock(&vnode_list_mtx); 564 TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist); 565 mtx_unlock(&vnode_list_mtx); 566 return (0); 567 } 568 569 /* 570 * Free a vnode when it is cleared from the zone. 571 */ 572 static void 573 vnode_fini(void *mem, int size) 574 { 575 struct vnode *vp; 576 struct bufobj *bo; 577 578 vp = mem; 579 vdbatch_dequeue(vp); 580 mtx_lock(&vnode_list_mtx); 581 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 582 mtx_unlock(&vnode_list_mtx); 583 rangelock_destroy(&vp->v_rl); 584 lockdestroy(vp->v_vnlock); 585 mtx_destroy(&vp->v_interlock); 586 bo = &vp->v_bufobj; 587 rw_destroy(BO_LOCKPTR(bo)); 588 } 589 590 /* 591 * Provide the size of NFS nclnode and NFS fh for calculation of the 592 * vnode memory consumption. The size is specified directly to 593 * eliminate dependency on NFS-private header. 594 * 595 * Other filesystems may use bigger or smaller (like UFS and ZFS) 596 * private inode data, but the NFS-based estimation is ample enough. 597 * Still, we care about differences in the size between 64- and 32-bit 598 * platforms. 599 * 600 * Namecache structure size is heuristically 601 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1. 602 */ 603 #ifdef _LP64 604 #define NFS_NCLNODE_SZ (528 + 64) 605 #define NC_SZ 148 606 #else 607 #define NFS_NCLNODE_SZ (360 + 32) 608 #define NC_SZ 92 609 #endif 610 611 static void 612 vntblinit(void *dummy __unused) 613 { 614 struct vdbatch *vd; 615 int cpu, physvnodes, virtvnodes; 616 u_int i; 617 618 /* 619 * Desiredvnodes is a function of the physical memory size and the 620 * kernel's heap size. Generally speaking, it scales with the 621 * physical memory size. The ratio of desiredvnodes to the physical 622 * memory size is 1:16 until desiredvnodes exceeds 98,304. 623 * Thereafter, the 624 * marginal ratio of desiredvnodes to the physical memory size is 625 * 1:64. However, desiredvnodes is limited by the kernel's heap 626 * size. The memory required by desiredvnodes vnodes and vm objects 627 * must not exceed 1/10th of the kernel's heap size. 628 */ 629 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 + 630 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64; 631 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) + 632 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ)); 633 desiredvnodes = min(physvnodes, virtvnodes); 634 if (desiredvnodes > MAXVNODES_MAX) { 635 if (bootverbose) 636 printf("Reducing kern.maxvnodes %lu -> %lu\n", 637 desiredvnodes, MAXVNODES_MAX); 638 desiredvnodes = MAXVNODES_MAX; 639 } 640 wantfreevnodes = desiredvnodes / 4; 641 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 642 TAILQ_INIT(&vnode_list); 643 mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF); 644 /* 645 * The lock is taken to appease WITNESS. 646 */ 647 mtx_lock(&vnode_list_mtx); 648 vnlru_recalc(); 649 mtx_unlock(&vnode_list_mtx); 650 vnode_list_free_marker = vn_alloc_marker(NULL); 651 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist); 652 vnode_list_reclaim_marker = vn_alloc_marker(NULL); 653 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist); 654 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 655 vnode_init, vnode_fini, UMA_ALIGN_PTR, 0); 656 uma_zone_set_smr(vnode_zone, vfs_smr); 657 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 658 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 659 /* 660 * Preallocate enough nodes to support one-per buf so that 661 * we can not fail an insert. reassignbuf() callers can not 662 * tolerate the insertion failure. 663 */ 664 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(), 665 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, 666 UMA_ZONE_NOFREE | UMA_ZONE_SMR); 667 buf_trie_smr = uma_zone_get_smr(buf_trie_zone); 668 uma_prealloc(buf_trie_zone, nbuf); 669 670 vnodes_created = counter_u64_alloc(M_WAITOK); 671 recycles_count = counter_u64_alloc(M_WAITOK); 672 recycles_free_count = counter_u64_alloc(M_WAITOK); 673 deferred_inact = counter_u64_alloc(M_WAITOK); 674 675 /* 676 * Initialize the filesystem syncer. 677 */ 678 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 679 &syncer_mask); 680 syncer_maxdelay = syncer_mask + 1; 681 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 682 cv_init(&sync_wakeup, "syncer"); 683 for (i = 1; i <= sizeof(struct vnode); i <<= 1) 684 vnsz2log++; 685 vnsz2log--; 686 687 CPU_FOREACH(cpu) { 688 vd = DPCPU_ID_PTR((cpu), vd); 689 bzero(vd, sizeof(*vd)); 690 mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF); 691 } 692 } 693 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL); 694 695 /* 696 * Mark a mount point as busy. Used to synchronize access and to delay 697 * unmounting. Eventually, mountlist_mtx is not released on failure. 698 * 699 * vfs_busy() is a custom lock, it can block the caller. 700 * vfs_busy() only sleeps if the unmount is active on the mount point. 701 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any 702 * vnode belonging to mp. 703 * 704 * Lookup uses vfs_busy() to traverse mount points. 705 * root fs var fs 706 * / vnode lock A / vnode lock (/var) D 707 * /var vnode lock B /log vnode lock(/var/log) E 708 * vfs_busy lock C vfs_busy lock F 709 * 710 * Within each file system, the lock order is C->A->B and F->D->E. 711 * 712 * When traversing across mounts, the system follows that lock order: 713 * 714 * C->A->B 715 * | 716 * +->F->D->E 717 * 718 * The lookup() process for namei("/var") illustrates the process: 719 * VOP_LOOKUP() obtains B while A is held 720 * vfs_busy() obtains a shared lock on F while A and B are held 721 * vput() releases lock on B 722 * vput() releases lock on A 723 * VFS_ROOT() obtains lock on D while shared lock on F is held 724 * vfs_unbusy() releases shared lock on F 725 * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A. 726 * Attempt to lock A (instead of vp_crossmp) while D is held would 727 * violate the global order, causing deadlocks. 728 * 729 * dounmount() locks B while F is drained. 730 */ 731 int 732 vfs_busy(struct mount *mp, int flags) 733 { 734 735 MPASS((flags & ~MBF_MASK) == 0); 736 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags); 737 738 if (vfs_op_thread_enter(mp)) { 739 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 740 MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0); 741 MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0); 742 vfs_mp_count_add_pcpu(mp, ref, 1); 743 vfs_mp_count_add_pcpu(mp, lockref, 1); 744 vfs_op_thread_exit(mp); 745 if (flags & MBF_MNTLSTLOCK) 746 mtx_unlock(&mountlist_mtx); 747 return (0); 748 } 749 750 MNT_ILOCK(mp); 751 vfs_assert_mount_counters(mp); 752 MNT_REF(mp); 753 /* 754 * If mount point is currently being unmounted, sleep until the 755 * mount point fate is decided. If thread doing the unmounting fails, 756 * it will clear MNTK_UNMOUNT flag before waking us up, indicating 757 * that this mount point has survived the unmount attempt and vfs_busy 758 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE 759 * flag in addition to MNTK_UNMOUNT, indicating that mount point is 760 * about to be really destroyed. vfs_busy needs to release its 761 * reference on the mount point in this case and return with ENOENT, 762 * telling the caller that mount mount it tried to busy is no longer 763 * valid. 764 */ 765 while (mp->mnt_kern_flag & MNTK_UNMOUNT) { 766 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) { 767 MNT_REL(mp); 768 MNT_IUNLOCK(mp); 769 CTR1(KTR_VFS, "%s: failed busying before sleeping", 770 __func__); 771 return (ENOENT); 772 } 773 if (flags & MBF_MNTLSTLOCK) 774 mtx_unlock(&mountlist_mtx); 775 mp->mnt_kern_flag |= MNTK_MWAIT; 776 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0); 777 if (flags & MBF_MNTLSTLOCK) 778 mtx_lock(&mountlist_mtx); 779 MNT_ILOCK(mp); 780 } 781 if (flags & MBF_MNTLSTLOCK) 782 mtx_unlock(&mountlist_mtx); 783 mp->mnt_lockref++; 784 MNT_IUNLOCK(mp); 785 return (0); 786 } 787 788 /* 789 * Free a busy filesystem. 790 */ 791 void 792 vfs_unbusy(struct mount *mp) 793 { 794 int c; 795 796 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 797 798 if (vfs_op_thread_enter(mp)) { 799 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 800 vfs_mp_count_sub_pcpu(mp, lockref, 1); 801 vfs_mp_count_sub_pcpu(mp, ref, 1); 802 vfs_op_thread_exit(mp); 803 return; 804 } 805 806 MNT_ILOCK(mp); 807 vfs_assert_mount_counters(mp); 808 MNT_REL(mp); 809 c = --mp->mnt_lockref; 810 if (mp->mnt_vfs_ops == 0) { 811 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 812 MNT_IUNLOCK(mp); 813 return; 814 } 815 if (c < 0) 816 vfs_dump_mount_counters(mp); 817 if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) { 818 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT); 819 CTR1(KTR_VFS, "%s: waking up waiters", __func__); 820 mp->mnt_kern_flag &= ~MNTK_DRAINING; 821 wakeup(&mp->mnt_lockref); 822 } 823 MNT_IUNLOCK(mp); 824 } 825 826 /* 827 * Lookup a mount point by filesystem identifier. 828 */ 829 struct mount * 830 vfs_getvfs(fsid_t *fsid) 831 { 832 struct mount *mp; 833 834 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 835 mtx_lock(&mountlist_mtx); 836 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 837 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) { 838 vfs_ref(mp); 839 mtx_unlock(&mountlist_mtx); 840 return (mp); 841 } 842 } 843 mtx_unlock(&mountlist_mtx); 844 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 845 return ((struct mount *) 0); 846 } 847 848 /* 849 * Lookup a mount point by filesystem identifier, busying it before 850 * returning. 851 * 852 * To avoid congestion on mountlist_mtx, implement simple direct-mapped 853 * cache for popular filesystem identifiers. The cache is lockess, using 854 * the fact that struct mount's are never freed. In worst case we may 855 * get pointer to unmounted or even different filesystem, so we have to 856 * check what we got, and go slow way if so. 857 */ 858 struct mount * 859 vfs_busyfs(fsid_t *fsid) 860 { 861 #define FSID_CACHE_SIZE 256 862 typedef struct mount * volatile vmp_t; 863 static vmp_t cache[FSID_CACHE_SIZE]; 864 struct mount *mp; 865 int error; 866 uint32_t hash; 867 868 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 869 hash = fsid->val[0] ^ fsid->val[1]; 870 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1); 871 mp = cache[hash]; 872 if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0) 873 goto slow; 874 if (vfs_busy(mp, 0) != 0) { 875 cache[hash] = NULL; 876 goto slow; 877 } 878 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) 879 return (mp); 880 else 881 vfs_unbusy(mp); 882 883 slow: 884 mtx_lock(&mountlist_mtx); 885 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 886 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) { 887 error = vfs_busy(mp, MBF_MNTLSTLOCK); 888 if (error) { 889 cache[hash] = NULL; 890 mtx_unlock(&mountlist_mtx); 891 return (NULL); 892 } 893 cache[hash] = mp; 894 return (mp); 895 } 896 } 897 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 898 mtx_unlock(&mountlist_mtx); 899 return ((struct mount *) 0); 900 } 901 902 /* 903 * Check if a user can access privileged mount options. 904 */ 905 int 906 vfs_suser(struct mount *mp, struct thread *td) 907 { 908 int error; 909 910 if (jailed(td->td_ucred)) { 911 /* 912 * If the jail of the calling thread lacks permission for 913 * this type of file system, deny immediately. 914 */ 915 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag)) 916 return (EPERM); 917 918 /* 919 * If the file system was mounted outside the jail of the 920 * calling thread, deny immediately. 921 */ 922 if (prison_check(td->td_ucred, mp->mnt_cred) != 0) 923 return (EPERM); 924 } 925 926 /* 927 * If file system supports delegated administration, we don't check 928 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified 929 * by the file system itself. 930 * If this is not the user that did original mount, we check for 931 * the PRIV_VFS_MOUNT_OWNER privilege. 932 */ 933 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) && 934 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 935 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) 936 return (error); 937 } 938 return (0); 939 } 940 941 /* 942 * Get a new unique fsid. Try to make its val[0] unique, since this value 943 * will be used to create fake device numbers for stat(). Also try (but 944 * not so hard) make its val[0] unique mod 2^16, since some emulators only 945 * support 16-bit device numbers. We end up with unique val[0]'s for the 946 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 947 * 948 * Keep in mind that several mounts may be running in parallel. Starting 949 * the search one past where the previous search terminated is both a 950 * micro-optimization and a defense against returning the same fsid to 951 * different mounts. 952 */ 953 void 954 vfs_getnewfsid(struct mount *mp) 955 { 956 static uint16_t mntid_base; 957 struct mount *nmp; 958 fsid_t tfsid; 959 int mtype; 960 961 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 962 mtx_lock(&mntid_mtx); 963 mtype = mp->mnt_vfc->vfc_typenum; 964 tfsid.val[1] = mtype; 965 mtype = (mtype & 0xFF) << 24; 966 for (;;) { 967 tfsid.val[0] = makedev(255, 968 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 969 mntid_base++; 970 if ((nmp = vfs_getvfs(&tfsid)) == NULL) 971 break; 972 vfs_rel(nmp); 973 } 974 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 975 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 976 mtx_unlock(&mntid_mtx); 977 } 978 979 /* 980 * Knob to control the precision of file timestamps: 981 * 982 * 0 = seconds only; nanoseconds zeroed. 983 * 1 = seconds and nanoseconds, accurate within 1/HZ. 984 * 2 = seconds and nanoseconds, truncated to microseconds. 985 * >=3 = seconds and nanoseconds, maximum precision. 986 */ 987 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 988 989 static int timestamp_precision = TSP_USEC; 990 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 991 ×tamp_precision, 0, "File timestamp precision (0: seconds, " 992 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, " 993 "3+: sec + ns (max. precision))"); 994 995 /* 996 * Get a current timestamp. 997 */ 998 void 999 vfs_timestamp(struct timespec *tsp) 1000 { 1001 struct timeval tv; 1002 1003 switch (timestamp_precision) { 1004 case TSP_SEC: 1005 tsp->tv_sec = time_second; 1006 tsp->tv_nsec = 0; 1007 break; 1008 case TSP_HZ: 1009 getnanotime(tsp); 1010 break; 1011 case TSP_USEC: 1012 microtime(&tv); 1013 TIMEVAL_TO_TIMESPEC(&tv, tsp); 1014 break; 1015 case TSP_NSEC: 1016 default: 1017 nanotime(tsp); 1018 break; 1019 } 1020 } 1021 1022 /* 1023 * Set vnode attributes to VNOVAL 1024 */ 1025 void 1026 vattr_null(struct vattr *vap) 1027 { 1028 1029 vap->va_type = VNON; 1030 vap->va_size = VNOVAL; 1031 vap->va_bytes = VNOVAL; 1032 vap->va_mode = VNOVAL; 1033 vap->va_nlink = VNOVAL; 1034 vap->va_uid = VNOVAL; 1035 vap->va_gid = VNOVAL; 1036 vap->va_fsid = VNOVAL; 1037 vap->va_fileid = VNOVAL; 1038 vap->va_blocksize = VNOVAL; 1039 vap->va_rdev = VNOVAL; 1040 vap->va_atime.tv_sec = VNOVAL; 1041 vap->va_atime.tv_nsec = VNOVAL; 1042 vap->va_mtime.tv_sec = VNOVAL; 1043 vap->va_mtime.tv_nsec = VNOVAL; 1044 vap->va_ctime.tv_sec = VNOVAL; 1045 vap->va_ctime.tv_nsec = VNOVAL; 1046 vap->va_birthtime.tv_sec = VNOVAL; 1047 vap->va_birthtime.tv_nsec = VNOVAL; 1048 vap->va_flags = VNOVAL; 1049 vap->va_gen = VNOVAL; 1050 vap->va_vaflags = 0; 1051 } 1052 1053 /* 1054 * Try to reduce the total number of vnodes. 1055 * 1056 * This routine (and its user) are buggy in at least the following ways: 1057 * - all parameters were picked years ago when RAM sizes were significantly 1058 * smaller 1059 * - it can pick vnodes based on pages used by the vm object, but filesystems 1060 * like ZFS don't use it making the pick broken 1061 * - since ZFS has its own aging policy it gets partially combated by this one 1062 * - a dedicated method should be provided for filesystems to let them decide 1063 * whether the vnode should be recycled 1064 * 1065 * This routine is called when we have too many vnodes. It attempts 1066 * to free <count> vnodes and will potentially free vnodes that still 1067 * have VM backing store (VM backing store is typically the cause 1068 * of a vnode blowout so we want to do this). Therefore, this operation 1069 * is not considered cheap. 1070 * 1071 * A number of conditions may prevent a vnode from being reclaimed. 1072 * the buffer cache may have references on the vnode, a directory 1073 * vnode may still have references due to the namei cache representing 1074 * underlying files, or the vnode may be in active use. It is not 1075 * desirable to reuse such vnodes. These conditions may cause the 1076 * number of vnodes to reach some minimum value regardless of what 1077 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 1078 * 1079 * @param reclaim_nc_src Only reclaim directories with outgoing namecache 1080 * entries if this argument is strue 1081 * @param trigger Only reclaim vnodes with fewer than this many resident 1082 * pages. 1083 * @param target How many vnodes to reclaim. 1084 * @return The number of vnodes that were reclaimed. 1085 */ 1086 static int 1087 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target) 1088 { 1089 struct vnode *vp, *mvp; 1090 struct mount *mp; 1091 struct vm_object *object; 1092 u_long done; 1093 bool retried; 1094 1095 mtx_assert(&vnode_list_mtx, MA_OWNED); 1096 1097 retried = false; 1098 done = 0; 1099 1100 mvp = vnode_list_reclaim_marker; 1101 restart: 1102 vp = mvp; 1103 while (done < target) { 1104 vp = TAILQ_NEXT(vp, v_vnodelist); 1105 if (__predict_false(vp == NULL)) 1106 break; 1107 1108 if (__predict_false(vp->v_type == VMARKER)) 1109 continue; 1110 1111 /* 1112 * If it's been deconstructed already, it's still 1113 * referenced, or it exceeds the trigger, skip it. 1114 * Also skip free vnodes. We are trying to make space 1115 * to expand the free list, not reduce it. 1116 */ 1117 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 || 1118 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src))) 1119 goto next_iter; 1120 1121 if (vp->v_type == VBAD || vp->v_type == VNON) 1122 goto next_iter; 1123 1124 if (!VI_TRYLOCK(vp)) 1125 goto next_iter; 1126 1127 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 || 1128 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 1129 VN_IS_DOOMED(vp) || vp->v_type == VNON) { 1130 VI_UNLOCK(vp); 1131 goto next_iter; 1132 } 1133 1134 object = atomic_load_ptr(&vp->v_object); 1135 if (object == NULL || object->resident_page_count > trigger) { 1136 VI_UNLOCK(vp); 1137 goto next_iter; 1138 } 1139 1140 vholdl(vp); 1141 VI_UNLOCK(vp); 1142 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1143 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1144 mtx_unlock(&vnode_list_mtx); 1145 1146 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1147 vdrop(vp); 1148 goto next_iter_unlocked; 1149 } 1150 if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) { 1151 vdrop(vp); 1152 vn_finished_write(mp); 1153 goto next_iter_unlocked; 1154 } 1155 1156 VI_LOCK(vp); 1157 if (vp->v_usecount > 0 || 1158 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || 1159 (vp->v_object != NULL && 1160 vp->v_object->resident_page_count > trigger)) { 1161 VOP_UNLOCK(vp); 1162 vdropl(vp); 1163 vn_finished_write(mp); 1164 goto next_iter_unlocked; 1165 } 1166 counter_u64_add(recycles_count, 1); 1167 vgonel(vp); 1168 VOP_UNLOCK(vp); 1169 vdropl(vp); 1170 vn_finished_write(mp); 1171 done++; 1172 next_iter_unlocked: 1173 if (should_yield()) 1174 kern_yield(PRI_USER); 1175 mtx_lock(&vnode_list_mtx); 1176 goto restart; 1177 next_iter: 1178 MPASS(vp->v_type != VMARKER); 1179 if (!should_yield()) 1180 continue; 1181 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1182 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1183 mtx_unlock(&vnode_list_mtx); 1184 kern_yield(PRI_USER); 1185 mtx_lock(&vnode_list_mtx); 1186 goto restart; 1187 } 1188 if (done == 0 && !retried) { 1189 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1190 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist); 1191 retried = true; 1192 goto restart; 1193 } 1194 return (done); 1195 } 1196 1197 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */ 1198 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free, 1199 0, 1200 "limit on vnode free requests per call to the vnlru_free routine"); 1201 1202 /* 1203 * Attempt to reduce the free list by the requested amount. 1204 */ 1205 static int 1206 vnlru_free_locked(int count, struct vfsops *mnt_op) 1207 { 1208 struct vnode *vp, *mvp; 1209 struct mount *mp; 1210 int ocount; 1211 1212 mtx_assert(&vnode_list_mtx, MA_OWNED); 1213 if (count > max_vnlru_free) 1214 count = max_vnlru_free; 1215 ocount = count; 1216 mvp = vnode_list_free_marker; 1217 restart: 1218 vp = mvp; 1219 while (count > 0) { 1220 vp = TAILQ_NEXT(vp, v_vnodelist); 1221 if (__predict_false(vp == NULL)) { 1222 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1223 TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist); 1224 break; 1225 } 1226 if (__predict_false(vp->v_type == VMARKER)) 1227 continue; 1228 1229 /* 1230 * Don't recycle if our vnode is from different type 1231 * of mount point. Note that mp is type-safe, the 1232 * check does not reach unmapped address even if 1233 * vnode is reclaimed. 1234 * Don't recycle if we can't get the interlock without 1235 * blocking. 1236 */ 1237 if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL && 1238 mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) { 1239 continue; 1240 } 1241 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); 1242 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); 1243 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) { 1244 VI_UNLOCK(vp); 1245 continue; 1246 } 1247 vholdl(vp); 1248 count--; 1249 mtx_unlock(&vnode_list_mtx); 1250 VI_UNLOCK(vp); 1251 vtryrecycle(vp); 1252 vdrop(vp); 1253 mtx_lock(&vnode_list_mtx); 1254 goto restart; 1255 } 1256 return (ocount - count); 1257 } 1258 1259 void 1260 vnlru_free(int count, struct vfsops *mnt_op) 1261 { 1262 1263 mtx_lock(&vnode_list_mtx); 1264 vnlru_free_locked(count, mnt_op); 1265 mtx_unlock(&vnode_list_mtx); 1266 } 1267 1268 static void 1269 vnlru_recalc(void) 1270 { 1271 1272 mtx_assert(&vnode_list_mtx, MA_OWNED); 1273 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100); 1274 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */ 1275 vlowat = vhiwat / 2; 1276 } 1277 1278 /* 1279 * Attempt to recycle vnodes in a context that is always safe to block. 1280 * Calling vlrurecycle() from the bowels of filesystem code has some 1281 * interesting deadlock problems. 1282 */ 1283 static struct proc *vnlruproc; 1284 static int vnlruproc_sig; 1285 1286 /* 1287 * The main freevnodes counter is only updated when threads requeue their vnode 1288 * batches. CPUs are conditionally walked to compute a more accurate total. 1289 * 1290 * Limit how much of a slop are we willing to tolerate. Note: the actual value 1291 * at any given moment can still exceed slop, but it should not be by significant 1292 * margin in practice. 1293 */ 1294 #define VNLRU_FREEVNODES_SLOP 128 1295 1296 static u_long 1297 vnlru_read_freevnodes(void) 1298 { 1299 struct vdbatch *vd; 1300 long slop; 1301 int cpu; 1302 1303 mtx_assert(&vnode_list_mtx, MA_OWNED); 1304 if (freevnodes > freevnodes_old) 1305 slop = freevnodes - freevnodes_old; 1306 else 1307 slop = freevnodes_old - freevnodes; 1308 if (slop < VNLRU_FREEVNODES_SLOP) 1309 return (freevnodes >= 0 ? freevnodes : 0); 1310 freevnodes_old = freevnodes; 1311 CPU_FOREACH(cpu) { 1312 vd = DPCPU_ID_PTR((cpu), vd); 1313 freevnodes_old += vd->freevnodes; 1314 } 1315 return (freevnodes_old >= 0 ? freevnodes_old : 0); 1316 } 1317 1318 static bool 1319 vnlru_under(u_long rnumvnodes, u_long limit) 1320 { 1321 u_long rfreevnodes, space; 1322 1323 if (__predict_false(rnumvnodes > desiredvnodes)) 1324 return (true); 1325 1326 space = desiredvnodes - rnumvnodes; 1327 if (space < limit) { 1328 rfreevnodes = vnlru_read_freevnodes(); 1329 if (rfreevnodes > wantfreevnodes) 1330 space += rfreevnodes - wantfreevnodes; 1331 } 1332 return (space < limit); 1333 } 1334 1335 static bool 1336 vnlru_under_unlocked(u_long rnumvnodes, u_long limit) 1337 { 1338 long rfreevnodes, space; 1339 1340 if (__predict_false(rnumvnodes > desiredvnodes)) 1341 return (true); 1342 1343 space = desiredvnodes - rnumvnodes; 1344 if (space < limit) { 1345 rfreevnodes = atomic_load_long(&freevnodes); 1346 if (rfreevnodes > wantfreevnodes) 1347 space += rfreevnodes - wantfreevnodes; 1348 } 1349 return (space < limit); 1350 } 1351 1352 static void 1353 vnlru_kick(void) 1354 { 1355 1356 mtx_assert(&vnode_list_mtx, MA_OWNED); 1357 if (vnlruproc_sig == 0) { 1358 vnlruproc_sig = 1; 1359 wakeup(vnlruproc); 1360 } 1361 } 1362 1363 static void 1364 vnlru_proc(void) 1365 { 1366 u_long rnumvnodes, rfreevnodes, target; 1367 unsigned long onumvnodes; 1368 int done, force, trigger, usevnodes; 1369 bool reclaim_nc_src, want_reread; 1370 1371 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc, 1372 SHUTDOWN_PRI_FIRST); 1373 1374 force = 0; 1375 want_reread = false; 1376 for (;;) { 1377 kproc_suspend_check(vnlruproc); 1378 mtx_lock(&vnode_list_mtx); 1379 rnumvnodes = atomic_load_long(&numvnodes); 1380 1381 if (want_reread) { 1382 force = vnlru_under(numvnodes, vhiwat) ? 1 : 0; 1383 want_reread = false; 1384 } 1385 1386 /* 1387 * If numvnodes is too large (due to desiredvnodes being 1388 * adjusted using its sysctl, or emergency growth), first 1389 * try to reduce it by discarding from the free list. 1390 */ 1391 if (rnumvnodes > desiredvnodes) { 1392 vnlru_free_locked(rnumvnodes - desiredvnodes, NULL); 1393 rnumvnodes = atomic_load_long(&numvnodes); 1394 } 1395 /* 1396 * Sleep if the vnode cache is in a good state. This is 1397 * when it is not over-full and has space for about a 4% 1398 * or 9% expansion (by growing its size or inexcessively 1399 * reducing its free list). Otherwise, try to reclaim 1400 * space for a 10% expansion. 1401 */ 1402 if (vstir && force == 0) { 1403 force = 1; 1404 vstir = 0; 1405 } 1406 if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) { 1407 vnlruproc_sig = 0; 1408 wakeup(&vnlruproc_sig); 1409 msleep(vnlruproc, &vnode_list_mtx, 1410 PVFS|PDROP, "vlruwt", hz); 1411 continue; 1412 } 1413 rfreevnodes = vnlru_read_freevnodes(); 1414 1415 onumvnodes = rnumvnodes; 1416 /* 1417 * Calculate parameters for recycling. These are the same 1418 * throughout the loop to give some semblance of fairness. 1419 * The trigger point is to avoid recycling vnodes with lots 1420 * of resident pages. We aren't trying to free memory; we 1421 * are trying to recycle or at least free vnodes. 1422 */ 1423 if (rnumvnodes <= desiredvnodes) 1424 usevnodes = rnumvnodes - rfreevnodes; 1425 else 1426 usevnodes = rnumvnodes; 1427 if (usevnodes <= 0) 1428 usevnodes = 1; 1429 /* 1430 * The trigger value is is chosen to give a conservatively 1431 * large value to ensure that it alone doesn't prevent 1432 * making progress. The value can easily be so large that 1433 * it is effectively infinite in some congested and 1434 * misconfigured cases, and this is necessary. Normally 1435 * it is about 8 to 100 (pages), which is quite large. 1436 */ 1437 trigger = vm_cnt.v_page_count * 2 / usevnodes; 1438 if (force < 2) 1439 trigger = vsmalltrigger; 1440 reclaim_nc_src = force >= 3; 1441 target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1); 1442 target = target / 10 + 1; 1443 done = vlrureclaim(reclaim_nc_src, trigger, target); 1444 mtx_unlock(&vnode_list_mtx); 1445 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes) 1446 uma_reclaim(UMA_RECLAIM_DRAIN); 1447 if (done == 0) { 1448 if (force == 0 || force == 1) { 1449 force = 2; 1450 continue; 1451 } 1452 if (force == 2) { 1453 force = 3; 1454 continue; 1455 } 1456 want_reread = true; 1457 force = 0; 1458 vnlru_nowhere++; 1459 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 1460 } else { 1461 want_reread = true; 1462 kern_yield(PRI_USER); 1463 } 1464 } 1465 } 1466 1467 static struct kproc_desc vnlru_kp = { 1468 "vnlru", 1469 vnlru_proc, 1470 &vnlruproc 1471 }; 1472 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, 1473 &vnlru_kp); 1474 1475 /* 1476 * Routines having to do with the management of the vnode table. 1477 */ 1478 1479 /* 1480 * Try to recycle a freed vnode. We abort if anyone picks up a reference 1481 * before we actually vgone(). This function must be called with the vnode 1482 * held to prevent the vnode from being returned to the free list midway 1483 * through vgone(). 1484 */ 1485 static int 1486 vtryrecycle(struct vnode *vp) 1487 { 1488 struct mount *vnmp; 1489 1490 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 1491 VNASSERT(vp->v_holdcnt, vp, 1492 ("vtryrecycle: Recycling vp %p without a reference.", vp)); 1493 /* 1494 * This vnode may found and locked via some other list, if so we 1495 * can't recycle it yet. 1496 */ 1497 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 1498 CTR2(KTR_VFS, 1499 "%s: impossible to recycle, vp %p lock is already held", 1500 __func__, vp); 1501 return (EWOULDBLOCK); 1502 } 1503 /* 1504 * Don't recycle if its filesystem is being suspended. 1505 */ 1506 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 1507 VOP_UNLOCK(vp); 1508 CTR2(KTR_VFS, 1509 "%s: impossible to recycle, cannot start the write for %p", 1510 __func__, vp); 1511 return (EBUSY); 1512 } 1513 /* 1514 * If we got this far, we need to acquire the interlock and see if 1515 * anyone picked up this vnode from another list. If not, we will 1516 * mark it with DOOMED via vgonel() so that anyone who does find it 1517 * will skip over it. 1518 */ 1519 VI_LOCK(vp); 1520 if (vp->v_usecount) { 1521 VOP_UNLOCK(vp); 1522 VI_UNLOCK(vp); 1523 vn_finished_write(vnmp); 1524 CTR2(KTR_VFS, 1525 "%s: impossible to recycle, %p is already referenced", 1526 __func__, vp); 1527 return (EBUSY); 1528 } 1529 if (!VN_IS_DOOMED(vp)) { 1530 counter_u64_add(recycles_free_count, 1); 1531 vgonel(vp); 1532 } 1533 VOP_UNLOCK(vp); 1534 VI_UNLOCK(vp); 1535 vn_finished_write(vnmp); 1536 return (0); 1537 } 1538 1539 /* 1540 * Allocate a new vnode. 1541 * 1542 * The operation never returns an error. Returning an error was disabled 1543 * in r145385 (dated 2005) with the following comment: 1544 * 1545 * XXX Not all VFS_VGET/ffs_vget callers check returns. 1546 * 1547 * Given the age of this commit (almost 15 years at the time of writing this 1548 * comment) restoring the ability to fail requires a significant audit of 1549 * all codepaths. 1550 * 1551 * The routine can try to free a vnode or stall for up to 1 second waiting for 1552 * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation. 1553 */ 1554 static u_long vn_alloc_cyclecount; 1555 1556 static struct vnode * __noinline 1557 vn_alloc_hard(struct mount *mp) 1558 { 1559 u_long rnumvnodes, rfreevnodes; 1560 1561 mtx_lock(&vnode_list_mtx); 1562 rnumvnodes = atomic_load_long(&numvnodes); 1563 if (rnumvnodes + 1 < desiredvnodes) { 1564 vn_alloc_cyclecount = 0; 1565 goto alloc; 1566 } 1567 rfreevnodes = vnlru_read_freevnodes(); 1568 if (vn_alloc_cyclecount++ >= rfreevnodes) { 1569 vn_alloc_cyclecount = 0; 1570 vstir = 1; 1571 } 1572 /* 1573 * Grow the vnode cache if it will not be above its target max 1574 * after growing. Otherwise, if the free list is nonempty, try 1575 * to reclaim 1 item from it before growing the cache (possibly 1576 * above its target max if the reclamation failed or is delayed). 1577 * Otherwise, wait for some space. In all cases, schedule 1578 * vnlru_proc() if we are getting short of space. The watermarks 1579 * should be chosen so that we never wait or even reclaim from 1580 * the free list to below its target minimum. 1581 */ 1582 if (vnlru_free_locked(1, NULL) > 0) 1583 goto alloc; 1584 if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 1585 /* 1586 * Wait for space for a new vnode. 1587 */ 1588 vnlru_kick(); 1589 msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz); 1590 if (atomic_load_long(&numvnodes) + 1 > desiredvnodes && 1591 vnlru_read_freevnodes() > 1) 1592 vnlru_free_locked(1, NULL); 1593 } 1594 alloc: 1595 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; 1596 if (vnlru_under(rnumvnodes, vlowat)) 1597 vnlru_kick(); 1598 mtx_unlock(&vnode_list_mtx); 1599 return (uma_zalloc_smr(vnode_zone, M_WAITOK)); 1600 } 1601 1602 static struct vnode * 1603 vn_alloc(struct mount *mp) 1604 { 1605 u_long rnumvnodes; 1606 1607 if (__predict_false(vn_alloc_cyclecount != 0)) 1608 return (vn_alloc_hard(mp)); 1609 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; 1610 if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) { 1611 atomic_subtract_long(&numvnodes, 1); 1612 return (vn_alloc_hard(mp)); 1613 } 1614 1615 return (uma_zalloc_smr(vnode_zone, M_WAITOK)); 1616 } 1617 1618 static void 1619 vn_free(struct vnode *vp) 1620 { 1621 1622 atomic_subtract_long(&numvnodes, 1); 1623 uma_zfree_smr(vnode_zone, vp); 1624 } 1625 1626 /* 1627 * Return the next vnode from the free list. 1628 */ 1629 int 1630 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, 1631 struct vnode **vpp) 1632 { 1633 struct vnode *vp; 1634 struct thread *td; 1635 struct lock_object *lo; 1636 1637 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag); 1638 1639 KASSERT(vops->registered, 1640 ("%s: not registered vector op %p\n", __func__, vops)); 1641 1642 td = curthread; 1643 if (td->td_vp_reserved != NULL) { 1644 vp = td->td_vp_reserved; 1645 td->td_vp_reserved = NULL; 1646 } else { 1647 vp = vn_alloc(mp); 1648 } 1649 counter_u64_add(vnodes_created, 1); 1650 /* 1651 * Locks are given the generic name "vnode" when created. 1652 * Follow the historic practice of using the filesystem 1653 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc. 1654 * 1655 * Locks live in a witness group keyed on their name. Thus, 1656 * when a lock is renamed, it must also move from the witness 1657 * group of its old name to the witness group of its new name. 1658 * 1659 * The change only needs to be made when the vnode moves 1660 * from one filesystem type to another. We ensure that each 1661 * filesystem use a single static name pointer for its tag so 1662 * that we can compare pointers rather than doing a strcmp(). 1663 */ 1664 lo = &vp->v_vnlock->lock_object; 1665 #ifdef WITNESS 1666 if (lo->lo_name != tag) { 1667 #endif 1668 lo->lo_name = tag; 1669 #ifdef WITNESS 1670 WITNESS_DESTROY(lo); 1671 WITNESS_INIT(lo, tag); 1672 } 1673 #endif 1674 /* 1675 * By default, don't allow shared locks unless filesystems opt-in. 1676 */ 1677 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE; 1678 /* 1679 * Finalize various vnode identity bits. 1680 */ 1681 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp)); 1682 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp)); 1683 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp)); 1684 vp->v_type = VNON; 1685 vp->v_op = vops; 1686 v_init_counters(vp); 1687 vp->v_bufobj.bo_ops = &buf_ops_bio; 1688 #ifdef DIAGNOSTIC 1689 if (mp == NULL && vops != &dead_vnodeops) 1690 printf("NULL mp in getnewvnode(9), tag %s\n", tag); 1691 #endif 1692 #ifdef MAC 1693 mac_vnode_init(vp); 1694 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 1695 mac_vnode_associate_singlelabel(mp, vp); 1696 #endif 1697 if (mp != NULL) { 1698 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize; 1699 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) 1700 vp->v_vflag |= VV_NOKNOTE; 1701 } 1702 1703 /* 1704 * For the filesystems which do not use vfs_hash_insert(), 1705 * still initialize v_hash to have vfs_hash_index() useful. 1706 * E.g., nullfs uses vfs_hash_index() on the lower vnode for 1707 * its own hashing. 1708 */ 1709 vp->v_hash = (uintptr_t)vp >> vnsz2log; 1710 1711 *vpp = vp; 1712 return (0); 1713 } 1714 1715 void 1716 getnewvnode_reserve(void) 1717 { 1718 struct thread *td; 1719 1720 td = curthread; 1721 MPASS(td->td_vp_reserved == NULL); 1722 td->td_vp_reserved = vn_alloc(NULL); 1723 } 1724 1725 void 1726 getnewvnode_drop_reserve(void) 1727 { 1728 struct thread *td; 1729 1730 td = curthread; 1731 if (td->td_vp_reserved != NULL) { 1732 vn_free(td->td_vp_reserved); 1733 td->td_vp_reserved = NULL; 1734 } 1735 } 1736 1737 static void 1738 freevnode(struct vnode *vp) 1739 { 1740 struct bufobj *bo; 1741 1742 /* 1743 * The vnode has been marked for destruction, so free it. 1744 * 1745 * The vnode will be returned to the zone where it will 1746 * normally remain until it is needed for another vnode. We 1747 * need to cleanup (or verify that the cleanup has already 1748 * been done) any residual data left from its current use 1749 * so as not to contaminate the freshly allocated vnode. 1750 */ 1751 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp); 1752 /* 1753 * Paired with vgone. 1754 */ 1755 vn_seqc_write_end_locked(vp); 1756 VNPASS(vp->v_seqc_users == 0, vp); 1757 1758 bo = &vp->v_bufobj; 1759 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 1760 VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp); 1761 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 1762 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 1763 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 1764 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 1765 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp, 1766 ("clean blk trie not empty")); 1767 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 1768 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp, 1769 ("dirty blk trie not empty")); 1770 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); 1771 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); 1772 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for ..")); 1773 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp, 1774 ("Dangling rangelock waiters")); 1775 VI_UNLOCK(vp); 1776 #ifdef MAC 1777 mac_vnode_destroy(vp); 1778 #endif 1779 if (vp->v_pollinfo != NULL) { 1780 destroy_vpollinfo(vp->v_pollinfo); 1781 vp->v_pollinfo = NULL; 1782 } 1783 #ifdef INVARIANTS 1784 /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */ 1785 vp->v_op = NULL; 1786 #endif 1787 vp->v_mountedhere = NULL; 1788 vp->v_unpcb = NULL; 1789 vp->v_rdev = NULL; 1790 vp->v_fifoinfo = NULL; 1791 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 1792 vp->v_irflag = 0; 1793 vp->v_iflag = 0; 1794 vp->v_vflag = 0; 1795 bo->bo_flag = 0; 1796 vn_free(vp); 1797 } 1798 1799 /* 1800 * Delete from old mount point vnode list, if on one. 1801 */ 1802 static void 1803 delmntque(struct vnode *vp) 1804 { 1805 struct mount *mp; 1806 1807 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); 1808 1809 mp = vp->v_mount; 1810 if (mp == NULL) 1811 return; 1812 MNT_ILOCK(mp); 1813 VI_LOCK(vp); 1814 vp->v_mount = NULL; 1815 VI_UNLOCK(vp); 1816 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 1817 ("bad mount point vnode list size")); 1818 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1819 mp->mnt_nvnodelistsize--; 1820 MNT_REL(mp); 1821 MNT_IUNLOCK(mp); 1822 } 1823 1824 static void 1825 insmntque_stddtr(struct vnode *vp, void *dtr_arg) 1826 { 1827 1828 vp->v_data = NULL; 1829 vp->v_op = &dead_vnodeops; 1830 vgone(vp); 1831 vput(vp); 1832 } 1833 1834 /* 1835 * Insert into list of vnodes for the new mount point, if available. 1836 */ 1837 int 1838 insmntque1(struct vnode *vp, struct mount *mp, 1839 void (*dtr)(struct vnode *, void *), void *dtr_arg) 1840 { 1841 1842 KASSERT(vp->v_mount == NULL, 1843 ("insmntque: vnode already on per mount vnode list")); 1844 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 1845 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); 1846 1847 /* 1848 * We acquire the vnode interlock early to ensure that the 1849 * vnode cannot be recycled by another process releasing a 1850 * holdcnt on it before we get it on both the vnode list 1851 * and the active vnode list. The mount mutex protects only 1852 * manipulation of the vnode list and the vnode freelist 1853 * mutex protects only manipulation of the active vnode list. 1854 * Hence the need to hold the vnode interlock throughout. 1855 */ 1856 MNT_ILOCK(mp); 1857 VI_LOCK(vp); 1858 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 && 1859 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || 1860 mp->mnt_nvnodelistsize == 0)) && 1861 (vp->v_vflag & VV_FORCEINSMQ) == 0) { 1862 VI_UNLOCK(vp); 1863 MNT_IUNLOCK(mp); 1864 if (dtr != NULL) 1865 dtr(vp, dtr_arg); 1866 return (EBUSY); 1867 } 1868 vp->v_mount = mp; 1869 MNT_REF(mp); 1870 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1871 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 1872 ("neg mount point vnode list size")); 1873 mp->mnt_nvnodelistsize++; 1874 VI_UNLOCK(vp); 1875 MNT_IUNLOCK(mp); 1876 return (0); 1877 } 1878 1879 int 1880 insmntque(struct vnode *vp, struct mount *mp) 1881 { 1882 1883 return (insmntque1(vp, mp, insmntque_stddtr, NULL)); 1884 } 1885 1886 /* 1887 * Flush out and invalidate all buffers associated with a bufobj 1888 * Called with the underlying object locked. 1889 */ 1890 int 1891 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) 1892 { 1893 int error; 1894 1895 BO_LOCK(bo); 1896 if (flags & V_SAVE) { 1897 error = bufobj_wwait(bo, slpflag, slptimeo); 1898 if (error) { 1899 BO_UNLOCK(bo); 1900 return (error); 1901 } 1902 if (bo->bo_dirty.bv_cnt > 0) { 1903 BO_UNLOCK(bo); 1904 if ((error = BO_SYNC(bo, MNT_WAIT)) != 0) 1905 return (error); 1906 /* 1907 * XXX We could save a lock/unlock if this was only 1908 * enabled under INVARIANTS 1909 */ 1910 BO_LOCK(bo); 1911 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 1912 panic("vinvalbuf: dirty bufs"); 1913 } 1914 } 1915 /* 1916 * If you alter this loop please notice that interlock is dropped and 1917 * reacquired in flushbuflist. Special care is needed to ensure that 1918 * no race conditions occur from this. 1919 */ 1920 do { 1921 error = flushbuflist(&bo->bo_clean, 1922 flags, bo, slpflag, slptimeo); 1923 if (error == 0 && !(flags & V_CLEANONLY)) 1924 error = flushbuflist(&bo->bo_dirty, 1925 flags, bo, slpflag, slptimeo); 1926 if (error != 0 && error != EAGAIN) { 1927 BO_UNLOCK(bo); 1928 return (error); 1929 } 1930 } while (error != 0); 1931 1932 /* 1933 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1934 * have write I/O in-progress but if there is a VM object then the 1935 * VM object can also have read-I/O in-progress. 1936 */ 1937 do { 1938 bufobj_wwait(bo, 0, 0); 1939 if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) { 1940 BO_UNLOCK(bo); 1941 vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx"); 1942 BO_LOCK(bo); 1943 } 1944 } while (bo->bo_numoutput > 0); 1945 BO_UNLOCK(bo); 1946 1947 /* 1948 * Destroy the copy in the VM cache, too. 1949 */ 1950 if (bo->bo_object != NULL && 1951 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) { 1952 VM_OBJECT_WLOCK(bo->bo_object); 1953 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? 1954 OBJPR_CLEANONLY : 0); 1955 VM_OBJECT_WUNLOCK(bo->bo_object); 1956 } 1957 1958 #ifdef INVARIANTS 1959 BO_LOCK(bo); 1960 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO | 1961 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 || 1962 bo->bo_clean.bv_cnt > 0)) 1963 panic("vinvalbuf: flush failed"); 1964 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 && 1965 bo->bo_dirty.bv_cnt > 0) 1966 panic("vinvalbuf: flush dirty failed"); 1967 BO_UNLOCK(bo); 1968 #endif 1969 return (0); 1970 } 1971 1972 /* 1973 * Flush out and invalidate all buffers associated with a vnode. 1974 * Called with the underlying object locked. 1975 */ 1976 int 1977 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 1978 { 1979 1980 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 1981 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1982 if (vp->v_object != NULL && vp->v_object->handle != vp) 1983 return (0); 1984 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); 1985 } 1986 1987 /* 1988 * Flush out buffers on the specified list. 1989 * 1990 */ 1991 static int 1992 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 1993 int slptimeo) 1994 { 1995 struct buf *bp, *nbp; 1996 int retval, error; 1997 daddr_t lblkno; 1998 b_xflags_t xflags; 1999 2000 ASSERT_BO_WLOCKED(bo); 2001 2002 retval = 0; 2003 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 2004 /* 2005 * If we are flushing both V_NORMAL and V_ALT buffers then 2006 * do not skip any buffers. If we are flushing only V_NORMAL 2007 * buffers then skip buffers marked as BX_ALTDATA. If we are 2008 * flushing only V_ALT buffers then skip buffers not marked 2009 * as BX_ALTDATA. 2010 */ 2011 if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) && 2012 (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) || 2013 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) { 2014 continue; 2015 } 2016 if (nbp != NULL) { 2017 lblkno = nbp->b_lblkno; 2018 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN); 2019 } 2020 retval = EAGAIN; 2021 error = BUF_TIMELOCK(bp, 2022 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), 2023 "flushbuf", slpflag, slptimeo); 2024 if (error) { 2025 BO_LOCK(bo); 2026 return (error != ENOLCK ? error : EAGAIN); 2027 } 2028 KASSERT(bp->b_bufobj == bo, 2029 ("bp %p wrong b_bufobj %p should be %p", 2030 bp, bp->b_bufobj, bo)); 2031 /* 2032 * XXX Since there are no node locks for NFS, I 2033 * believe there is a slight chance that a delayed 2034 * write will occur while sleeping just above, so 2035 * check for it. 2036 */ 2037 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 2038 (flags & V_SAVE)) { 2039 bremfree(bp); 2040 bp->b_flags |= B_ASYNC; 2041 bwrite(bp); 2042 BO_LOCK(bo); 2043 return (EAGAIN); /* XXX: why not loop ? */ 2044 } 2045 bremfree(bp); 2046 bp->b_flags |= (B_INVAL | B_RELBUF); 2047 bp->b_flags &= ~B_ASYNC; 2048 brelse(bp); 2049 BO_LOCK(bo); 2050 if (nbp == NULL) 2051 break; 2052 nbp = gbincore(bo, lblkno); 2053 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2054 != xflags) 2055 break; /* nbp invalid */ 2056 } 2057 return (retval); 2058 } 2059 2060 int 2061 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn) 2062 { 2063 struct buf *bp; 2064 int error; 2065 daddr_t lblkno; 2066 2067 ASSERT_BO_LOCKED(bo); 2068 2069 for (lblkno = startn;;) { 2070 again: 2071 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno); 2072 if (bp == NULL || bp->b_lblkno >= endn || 2073 bp->b_lblkno < startn) 2074 break; 2075 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 2076 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0); 2077 if (error != 0) { 2078 BO_RLOCK(bo); 2079 if (error == ENOLCK) 2080 goto again; 2081 return (error); 2082 } 2083 KASSERT(bp->b_bufobj == bo, 2084 ("bp %p wrong b_bufobj %p should be %p", 2085 bp, bp->b_bufobj, bo)); 2086 lblkno = bp->b_lblkno + 1; 2087 if ((bp->b_flags & B_MANAGED) == 0) 2088 bremfree(bp); 2089 bp->b_flags |= B_RELBUF; 2090 /* 2091 * In the VMIO case, use the B_NOREUSE flag to hint that the 2092 * pages backing each buffer in the range are unlikely to be 2093 * reused. Dirty buffers will have the hint applied once 2094 * they've been written. 2095 */ 2096 if ((bp->b_flags & B_VMIO) != 0) 2097 bp->b_flags |= B_NOREUSE; 2098 brelse(bp); 2099 BO_RLOCK(bo); 2100 } 2101 return (0); 2102 } 2103 2104 /* 2105 * Truncate a file's buffer and pages to a specified length. This 2106 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 2107 * sync activity. 2108 */ 2109 int 2110 vtruncbuf(struct vnode *vp, off_t length, int blksize) 2111 { 2112 struct buf *bp, *nbp; 2113 struct bufobj *bo; 2114 daddr_t startlbn; 2115 2116 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__, 2117 vp, blksize, (uintmax_t)length); 2118 2119 /* 2120 * Round up to the *next* lbn. 2121 */ 2122 startlbn = howmany(length, blksize); 2123 2124 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 2125 2126 bo = &vp->v_bufobj; 2127 restart_unlocked: 2128 BO_LOCK(bo); 2129 2130 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN) 2131 ; 2132 2133 if (length > 0) { 2134 restartsync: 2135 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2136 if (bp->b_lblkno > 0) 2137 continue; 2138 /* 2139 * Since we hold the vnode lock this should only 2140 * fail if we're racing with the buf daemon. 2141 */ 2142 if (BUF_LOCK(bp, 2143 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2144 BO_LOCKPTR(bo)) == ENOLCK) 2145 goto restart_unlocked; 2146 2147 VNASSERT((bp->b_flags & B_DELWRI), vp, 2148 ("buf(%p) on dirty queue without DELWRI", bp)); 2149 2150 bremfree(bp); 2151 bawrite(bp); 2152 BO_LOCK(bo); 2153 goto restartsync; 2154 } 2155 } 2156 2157 bufobj_wwait(bo, 0, 0); 2158 BO_UNLOCK(bo); 2159 vnode_pager_setsize(vp, length); 2160 2161 return (0); 2162 } 2163 2164 /* 2165 * Invalidate the cached pages of a file's buffer within the range of block 2166 * numbers [startlbn, endlbn). 2167 */ 2168 void 2169 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn, 2170 int blksize) 2171 { 2172 struct bufobj *bo; 2173 off_t start, end; 2174 2175 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range"); 2176 2177 start = blksize * startlbn; 2178 end = blksize * endlbn; 2179 2180 bo = &vp->v_bufobj; 2181 BO_LOCK(bo); 2182 MPASS(blksize == bo->bo_bsize); 2183 2184 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN) 2185 ; 2186 2187 BO_UNLOCK(bo); 2188 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1)); 2189 } 2190 2191 static int 2192 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, 2193 daddr_t startlbn, daddr_t endlbn) 2194 { 2195 struct buf *bp, *nbp; 2196 bool anyfreed; 2197 2198 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked"); 2199 ASSERT_BO_LOCKED(bo); 2200 2201 do { 2202 anyfreed = false; 2203 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 2204 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) 2205 continue; 2206 if (BUF_LOCK(bp, 2207 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2208 BO_LOCKPTR(bo)) == ENOLCK) { 2209 BO_LOCK(bo); 2210 return (EAGAIN); 2211 } 2212 2213 bremfree(bp); 2214 bp->b_flags |= B_INVAL | B_RELBUF; 2215 bp->b_flags &= ~B_ASYNC; 2216 brelse(bp); 2217 anyfreed = true; 2218 2219 BO_LOCK(bo); 2220 if (nbp != NULL && 2221 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 2222 nbp->b_vp != vp || 2223 (nbp->b_flags & B_DELWRI) != 0)) 2224 return (EAGAIN); 2225 } 2226 2227 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 2228 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) 2229 continue; 2230 if (BUF_LOCK(bp, 2231 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2232 BO_LOCKPTR(bo)) == ENOLCK) { 2233 BO_LOCK(bo); 2234 return (EAGAIN); 2235 } 2236 bremfree(bp); 2237 bp->b_flags |= B_INVAL | B_RELBUF; 2238 bp->b_flags &= ~B_ASYNC; 2239 brelse(bp); 2240 anyfreed = true; 2241 2242 BO_LOCK(bo); 2243 if (nbp != NULL && 2244 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 2245 (nbp->b_vp != vp) || 2246 (nbp->b_flags & B_DELWRI) == 0)) 2247 return (EAGAIN); 2248 } 2249 } while (anyfreed); 2250 return (0); 2251 } 2252 2253 static void 2254 buf_vlist_remove(struct buf *bp) 2255 { 2256 struct bufv *bv; 2257 b_xflags_t flags; 2258 2259 flags = bp->b_xflags; 2260 2261 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2262 ASSERT_BO_WLOCKED(bp->b_bufobj); 2263 KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 && 2264 (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN), 2265 ("%s: buffer %p has invalid queue state", __func__, bp)); 2266 2267 if ((flags & BX_VNDIRTY) != 0) 2268 bv = &bp->b_bufobj->bo_dirty; 2269 else 2270 bv = &bp->b_bufobj->bo_clean; 2271 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno); 2272 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 2273 bv->bv_cnt--; 2274 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 2275 } 2276 2277 /* 2278 * Add the buffer to the sorted clean or dirty block list. 2279 * 2280 * NOTE: xflags is passed as a constant, optimizing this inline function! 2281 */ 2282 static void 2283 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 2284 { 2285 struct bufv *bv; 2286 struct buf *n; 2287 int error; 2288 2289 ASSERT_BO_WLOCKED(bo); 2290 KASSERT((bo->bo_flag & BO_NOBUFS) == 0, 2291 ("buf_vlist_add: bo %p does not allow bufs", bo)); 2292 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0, 2293 ("dead bo %p", bo)); 2294 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 2295 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 2296 bp->b_xflags |= xflags; 2297 if (xflags & BX_VNDIRTY) 2298 bv = &bo->bo_dirty; 2299 else 2300 bv = &bo->bo_clean; 2301 2302 /* 2303 * Keep the list ordered. Optimize empty list insertion. Assume 2304 * we tend to grow at the tail so lookup_le should usually be cheaper 2305 * than _ge. 2306 */ 2307 if (bv->bv_cnt == 0 || 2308 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno) 2309 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 2310 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL) 2311 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs); 2312 else 2313 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs); 2314 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp); 2315 if (error) 2316 panic("buf_vlist_add: Preallocated nodes insufficient."); 2317 bv->bv_cnt++; 2318 } 2319 2320 /* 2321 * Look up a buffer using the buffer tries. 2322 */ 2323 struct buf * 2324 gbincore(struct bufobj *bo, daddr_t lblkno) 2325 { 2326 struct buf *bp; 2327 2328 ASSERT_BO_LOCKED(bo); 2329 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno); 2330 if (bp != NULL) 2331 return (bp); 2332 return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno)); 2333 } 2334 2335 /* 2336 * Look up a buf using the buffer tries, without the bufobj lock. This relies 2337 * on SMR for safe lookup, and bufs being in a no-free zone to provide type 2338 * stability of the result. Like other lockless lookups, the found buf may 2339 * already be invalid by the time this function returns. 2340 */ 2341 struct buf * 2342 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno) 2343 { 2344 struct buf *bp; 2345 2346 ASSERT_BO_UNLOCKED(bo); 2347 bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno); 2348 if (bp != NULL) 2349 return (bp); 2350 return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno)); 2351 } 2352 2353 /* 2354 * Associate a buffer with a vnode. 2355 */ 2356 void 2357 bgetvp(struct vnode *vp, struct buf *bp) 2358 { 2359 struct bufobj *bo; 2360 2361 bo = &vp->v_bufobj; 2362 ASSERT_BO_WLOCKED(bo); 2363 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 2364 2365 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 2366 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 2367 ("bgetvp: bp already attached! %p", bp)); 2368 2369 vhold(vp); 2370 bp->b_vp = vp; 2371 bp->b_bufobj = bo; 2372 /* 2373 * Insert onto list for new vnode. 2374 */ 2375 buf_vlist_add(bp, bo, BX_VNCLEAN); 2376 } 2377 2378 /* 2379 * Disassociate a buffer from a vnode. 2380 */ 2381 void 2382 brelvp(struct buf *bp) 2383 { 2384 struct bufobj *bo; 2385 struct vnode *vp; 2386 2387 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2388 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 2389 2390 /* 2391 * Delete from old vnode list, if on one. 2392 */ 2393 vp = bp->b_vp; /* XXX */ 2394 bo = bp->b_bufobj; 2395 BO_LOCK(bo); 2396 buf_vlist_remove(bp); 2397 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2398 bo->bo_flag &= ~BO_ONWORKLST; 2399 mtx_lock(&sync_mtx); 2400 LIST_REMOVE(bo, bo_synclist); 2401 syncer_worklist_len--; 2402 mtx_unlock(&sync_mtx); 2403 } 2404 bp->b_vp = NULL; 2405 bp->b_bufobj = NULL; 2406 BO_UNLOCK(bo); 2407 vdrop(vp); 2408 } 2409 2410 /* 2411 * Add an item to the syncer work queue. 2412 */ 2413 static void 2414 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 2415 { 2416 int slot; 2417 2418 ASSERT_BO_WLOCKED(bo); 2419 2420 mtx_lock(&sync_mtx); 2421 if (bo->bo_flag & BO_ONWORKLST) 2422 LIST_REMOVE(bo, bo_synclist); 2423 else { 2424 bo->bo_flag |= BO_ONWORKLST; 2425 syncer_worklist_len++; 2426 } 2427 2428 if (delay > syncer_maxdelay - 2) 2429 delay = syncer_maxdelay - 2; 2430 slot = (syncer_delayno + delay) & syncer_mask; 2431 2432 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 2433 mtx_unlock(&sync_mtx); 2434 } 2435 2436 static int 2437 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 2438 { 2439 int error, len; 2440 2441 mtx_lock(&sync_mtx); 2442 len = syncer_worklist_len - sync_vnode_count; 2443 mtx_unlock(&sync_mtx); 2444 error = SYSCTL_OUT(req, &len, sizeof(len)); 2445 return (error); 2446 } 2447 2448 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, 2449 CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0, 2450 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 2451 2452 static struct proc *updateproc; 2453 static void sched_sync(void); 2454 static struct kproc_desc up_kp = { 2455 "syncer", 2456 sched_sync, 2457 &updateproc 2458 }; 2459 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); 2460 2461 static int 2462 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) 2463 { 2464 struct vnode *vp; 2465 struct mount *mp; 2466 2467 *bo = LIST_FIRST(slp); 2468 if (*bo == NULL) 2469 return (0); 2470 vp = bo2vnode(*bo); 2471 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) 2472 return (1); 2473 /* 2474 * We use vhold in case the vnode does not 2475 * successfully sync. vhold prevents the vnode from 2476 * going away when we unlock the sync_mtx so that 2477 * we can acquire the vnode interlock. 2478 */ 2479 vholdl(vp); 2480 mtx_unlock(&sync_mtx); 2481 VI_UNLOCK(vp); 2482 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2483 vdrop(vp); 2484 mtx_lock(&sync_mtx); 2485 return (*bo == LIST_FIRST(slp)); 2486 } 2487 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2488 (void) VOP_FSYNC(vp, MNT_LAZY, td); 2489 VOP_UNLOCK(vp); 2490 vn_finished_write(mp); 2491 BO_LOCK(*bo); 2492 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { 2493 /* 2494 * Put us back on the worklist. The worklist 2495 * routine will remove us from our current 2496 * position and then add us back in at a later 2497 * position. 2498 */ 2499 vn_syncer_add_to_worklist(*bo, syncdelay); 2500 } 2501 BO_UNLOCK(*bo); 2502 vdrop(vp); 2503 mtx_lock(&sync_mtx); 2504 return (0); 2505 } 2506 2507 static int first_printf = 1; 2508 2509 /* 2510 * System filesystem synchronizer daemon. 2511 */ 2512 static void 2513 sched_sync(void) 2514 { 2515 struct synclist *next, *slp; 2516 struct bufobj *bo; 2517 long starttime; 2518 struct thread *td = curthread; 2519 int last_work_seen; 2520 int net_worklist_len; 2521 int syncer_final_iter; 2522 int error; 2523 2524 last_work_seen = 0; 2525 syncer_final_iter = 0; 2526 syncer_state = SYNCER_RUNNING; 2527 starttime = time_uptime; 2528 td->td_pflags |= TDP_NORUNNINGBUF; 2529 2530 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 2531 SHUTDOWN_PRI_LAST); 2532 2533 mtx_lock(&sync_mtx); 2534 for (;;) { 2535 if (syncer_state == SYNCER_FINAL_DELAY && 2536 syncer_final_iter == 0) { 2537 mtx_unlock(&sync_mtx); 2538 kproc_suspend_check(td->td_proc); 2539 mtx_lock(&sync_mtx); 2540 } 2541 net_worklist_len = syncer_worklist_len - sync_vnode_count; 2542 if (syncer_state != SYNCER_RUNNING && 2543 starttime != time_uptime) { 2544 if (first_printf) { 2545 printf("\nSyncing disks, vnodes remaining... "); 2546 first_printf = 0; 2547 } 2548 printf("%d ", net_worklist_len); 2549 } 2550 starttime = time_uptime; 2551 2552 /* 2553 * Push files whose dirty time has expired. Be careful 2554 * of interrupt race on slp queue. 2555 * 2556 * Skip over empty worklist slots when shutting down. 2557 */ 2558 do { 2559 slp = &syncer_workitem_pending[syncer_delayno]; 2560 syncer_delayno += 1; 2561 if (syncer_delayno == syncer_maxdelay) 2562 syncer_delayno = 0; 2563 next = &syncer_workitem_pending[syncer_delayno]; 2564 /* 2565 * If the worklist has wrapped since the 2566 * it was emptied of all but syncer vnodes, 2567 * switch to the FINAL_DELAY state and run 2568 * for one more second. 2569 */ 2570 if (syncer_state == SYNCER_SHUTTING_DOWN && 2571 net_worklist_len == 0 && 2572 last_work_seen == syncer_delayno) { 2573 syncer_state = SYNCER_FINAL_DELAY; 2574 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 2575 } 2576 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 2577 syncer_worklist_len > 0); 2578 2579 /* 2580 * Keep track of the last time there was anything 2581 * on the worklist other than syncer vnodes. 2582 * Return to the SHUTTING_DOWN state if any 2583 * new work appears. 2584 */ 2585 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 2586 last_work_seen = syncer_delayno; 2587 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 2588 syncer_state = SYNCER_SHUTTING_DOWN; 2589 while (!LIST_EMPTY(slp)) { 2590 error = sync_vnode(slp, &bo, td); 2591 if (error == 1) { 2592 LIST_REMOVE(bo, bo_synclist); 2593 LIST_INSERT_HEAD(next, bo, bo_synclist); 2594 continue; 2595 } 2596 2597 if (first_printf == 0) { 2598 /* 2599 * Drop the sync mutex, because some watchdog 2600 * drivers need to sleep while patting 2601 */ 2602 mtx_unlock(&sync_mtx); 2603 wdog_kern_pat(WD_LASTVAL); 2604 mtx_lock(&sync_mtx); 2605 } 2606 2607 } 2608 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 2609 syncer_final_iter--; 2610 /* 2611 * The variable rushjob allows the kernel to speed up the 2612 * processing of the filesystem syncer process. A rushjob 2613 * value of N tells the filesystem syncer to process the next 2614 * N seconds worth of work on its queue ASAP. Currently rushjob 2615 * is used by the soft update code to speed up the filesystem 2616 * syncer process when the incore state is getting so far 2617 * ahead of the disk that the kernel memory pool is being 2618 * threatened with exhaustion. 2619 */ 2620 if (rushjob > 0) { 2621 rushjob -= 1; 2622 continue; 2623 } 2624 /* 2625 * Just sleep for a short period of time between 2626 * iterations when shutting down to allow some I/O 2627 * to happen. 2628 * 2629 * If it has taken us less than a second to process the 2630 * current work, then wait. Otherwise start right over 2631 * again. We can still lose time if any single round 2632 * takes more than two seconds, but it does not really 2633 * matter as we are just trying to generally pace the 2634 * filesystem activity. 2635 */ 2636 if (syncer_state != SYNCER_RUNNING || 2637 time_uptime == starttime) { 2638 thread_lock(td); 2639 sched_prio(td, PPAUSE); 2640 thread_unlock(td); 2641 } 2642 if (syncer_state != SYNCER_RUNNING) 2643 cv_timedwait(&sync_wakeup, &sync_mtx, 2644 hz / SYNCER_SHUTDOWN_SPEEDUP); 2645 else if (time_uptime == starttime) 2646 cv_timedwait(&sync_wakeup, &sync_mtx, hz); 2647 } 2648 } 2649 2650 /* 2651 * Request the syncer daemon to speed up its work. 2652 * We never push it to speed up more than half of its 2653 * normal turn time, otherwise it could take over the cpu. 2654 */ 2655 int 2656 speedup_syncer(void) 2657 { 2658 int ret = 0; 2659 2660 mtx_lock(&sync_mtx); 2661 if (rushjob < syncdelay / 2) { 2662 rushjob += 1; 2663 stat_rush_requests += 1; 2664 ret = 1; 2665 } 2666 mtx_unlock(&sync_mtx); 2667 cv_broadcast(&sync_wakeup); 2668 return (ret); 2669 } 2670 2671 /* 2672 * Tell the syncer to speed up its work and run though its work 2673 * list several times, then tell it to shut down. 2674 */ 2675 static void 2676 syncer_shutdown(void *arg, int howto) 2677 { 2678 2679 if (howto & RB_NOSYNC) 2680 return; 2681 mtx_lock(&sync_mtx); 2682 syncer_state = SYNCER_SHUTTING_DOWN; 2683 rushjob = 0; 2684 mtx_unlock(&sync_mtx); 2685 cv_broadcast(&sync_wakeup); 2686 kproc_shutdown(arg, howto); 2687 } 2688 2689 void 2690 syncer_suspend(void) 2691 { 2692 2693 syncer_shutdown(updateproc, 0); 2694 } 2695 2696 void 2697 syncer_resume(void) 2698 { 2699 2700 mtx_lock(&sync_mtx); 2701 first_printf = 1; 2702 syncer_state = SYNCER_RUNNING; 2703 mtx_unlock(&sync_mtx); 2704 cv_broadcast(&sync_wakeup); 2705 kproc_resume(updateproc); 2706 } 2707 2708 /* 2709 * Move the buffer between the clean and dirty lists of its vnode. 2710 */ 2711 void 2712 reassignbuf(struct buf *bp) 2713 { 2714 struct vnode *vp; 2715 struct bufobj *bo; 2716 int delay; 2717 #ifdef INVARIANTS 2718 struct bufv *bv; 2719 #endif 2720 2721 vp = bp->b_vp; 2722 bo = bp->b_bufobj; 2723 2724 KASSERT((bp->b_flags & B_PAGING) == 0, 2725 ("%s: cannot reassign paging buffer %p", __func__, bp)); 2726 2727 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 2728 bp, bp->b_vp, bp->b_flags); 2729 2730 BO_LOCK(bo); 2731 buf_vlist_remove(bp); 2732 2733 /* 2734 * If dirty, put on list of dirty buffers; otherwise insert onto list 2735 * of clean buffers. 2736 */ 2737 if (bp->b_flags & B_DELWRI) { 2738 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 2739 switch (vp->v_type) { 2740 case VDIR: 2741 delay = dirdelay; 2742 break; 2743 case VCHR: 2744 delay = metadelay; 2745 break; 2746 default: 2747 delay = filedelay; 2748 } 2749 vn_syncer_add_to_worklist(bo, delay); 2750 } 2751 buf_vlist_add(bp, bo, BX_VNDIRTY); 2752 } else { 2753 buf_vlist_add(bp, bo, BX_VNCLEAN); 2754 2755 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2756 mtx_lock(&sync_mtx); 2757 LIST_REMOVE(bo, bo_synclist); 2758 syncer_worklist_len--; 2759 mtx_unlock(&sync_mtx); 2760 bo->bo_flag &= ~BO_ONWORKLST; 2761 } 2762 } 2763 #ifdef INVARIANTS 2764 bv = &bo->bo_clean; 2765 bp = TAILQ_FIRST(&bv->bv_hd); 2766 KASSERT(bp == NULL || bp->b_bufobj == bo, 2767 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2768 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2769 KASSERT(bp == NULL || bp->b_bufobj == bo, 2770 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2771 bv = &bo->bo_dirty; 2772 bp = TAILQ_FIRST(&bv->bv_hd); 2773 KASSERT(bp == NULL || bp->b_bufobj == bo, 2774 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2775 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2776 KASSERT(bp == NULL || bp->b_bufobj == bo, 2777 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2778 #endif 2779 BO_UNLOCK(bo); 2780 } 2781 2782 static void 2783 v_init_counters(struct vnode *vp) 2784 { 2785 2786 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0, 2787 vp, ("%s called for an initialized vnode", __FUNCTION__)); 2788 ASSERT_VI_UNLOCKED(vp, __FUNCTION__); 2789 2790 refcount_init(&vp->v_holdcnt, 1); 2791 refcount_init(&vp->v_usecount, 1); 2792 } 2793 2794 /* 2795 * Grab a particular vnode from the free list, increment its 2796 * reference count and lock it. VIRF_DOOMED is set if the vnode 2797 * is being destroyed. Only callers who specify LK_RETRY will 2798 * see doomed vnodes. If inactive processing was delayed in 2799 * vput try to do it here. 2800 * 2801 * usecount is manipulated using atomics without holding any locks. 2802 * 2803 * holdcnt can be manipulated using atomics without holding any locks, 2804 * except when transitioning 1<->0, in which case the interlock is held. 2805 * 2806 * Consumers which don't guarantee liveness of the vnode can use SMR to 2807 * try to get a reference. Note this operation can fail since the vnode 2808 * may be awaiting getting freed by the time they get to it. 2809 */ 2810 enum vgetstate 2811 vget_prep_smr(struct vnode *vp) 2812 { 2813 enum vgetstate vs; 2814 2815 VFS_SMR_ASSERT_ENTERED(); 2816 2817 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2818 vs = VGET_USECOUNT; 2819 } else { 2820 if (vhold_smr(vp)) 2821 vs = VGET_HOLDCNT; 2822 else 2823 vs = VGET_NONE; 2824 } 2825 return (vs); 2826 } 2827 2828 enum vgetstate 2829 vget_prep(struct vnode *vp) 2830 { 2831 enum vgetstate vs; 2832 2833 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2834 vs = VGET_USECOUNT; 2835 } else { 2836 vhold(vp); 2837 vs = VGET_HOLDCNT; 2838 } 2839 return (vs); 2840 } 2841 2842 void 2843 vget_abort(struct vnode *vp, enum vgetstate vs) 2844 { 2845 2846 switch (vs) { 2847 case VGET_USECOUNT: 2848 vrele(vp); 2849 break; 2850 case VGET_HOLDCNT: 2851 vdrop(vp); 2852 break; 2853 default: 2854 __assert_unreachable(); 2855 } 2856 } 2857 2858 int 2859 vget(struct vnode *vp, int flags, struct thread *td) 2860 { 2861 enum vgetstate vs; 2862 2863 MPASS(td == curthread); 2864 2865 vs = vget_prep(vp); 2866 return (vget_finish(vp, flags, vs)); 2867 } 2868 2869 int 2870 vget_finish(struct vnode *vp, int flags, enum vgetstate vs) 2871 { 2872 int error; 2873 2874 if ((flags & LK_INTERLOCK) != 0) 2875 ASSERT_VI_LOCKED(vp, __func__); 2876 else 2877 ASSERT_VI_UNLOCKED(vp, __func__); 2878 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); 2879 VNPASS(vp->v_holdcnt > 0, vp); 2880 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); 2881 2882 error = vn_lock(vp, flags); 2883 if (__predict_false(error != 0)) { 2884 vget_abort(vp, vs); 2885 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, 2886 vp); 2887 return (error); 2888 } 2889 2890 vget_finish_ref(vp, vs); 2891 return (0); 2892 } 2893 2894 void 2895 vget_finish_ref(struct vnode *vp, enum vgetstate vs) 2896 { 2897 int old; 2898 2899 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); 2900 VNPASS(vp->v_holdcnt > 0, vp); 2901 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); 2902 2903 if (vs == VGET_USECOUNT) 2904 return; 2905 2906 /* 2907 * We hold the vnode. If the usecount is 0 it will be utilized to keep 2908 * the vnode around. Otherwise someone else lended their hold count and 2909 * we have to drop ours. 2910 */ 2911 old = atomic_fetchadd_int(&vp->v_usecount, 1); 2912 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old)); 2913 if (old != 0) { 2914 #ifdef INVARIANTS 2915 old = atomic_fetchadd_int(&vp->v_holdcnt, -1); 2916 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); 2917 #else 2918 refcount_release(&vp->v_holdcnt); 2919 #endif 2920 } 2921 } 2922 2923 void 2924 vref(struct vnode *vp) 2925 { 2926 enum vgetstate vs; 2927 2928 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2929 vs = vget_prep(vp); 2930 vget_finish_ref(vp, vs); 2931 } 2932 2933 void 2934 vrefl(struct vnode *vp) 2935 { 2936 2937 ASSERT_VI_LOCKED(vp, __func__); 2938 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2939 vref(vp); 2940 } 2941 2942 void 2943 vrefact(struct vnode *vp) 2944 { 2945 2946 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2947 #ifdef INVARIANTS 2948 int old = atomic_fetchadd_int(&vp->v_usecount, 1); 2949 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); 2950 #else 2951 refcount_acquire(&vp->v_usecount); 2952 #endif 2953 } 2954 2955 void 2956 vlazy(struct vnode *vp) 2957 { 2958 struct mount *mp; 2959 2960 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__)); 2961 2962 if ((vp->v_mflag & VMP_LAZYLIST) != 0) 2963 return; 2964 /* 2965 * We may get here for inactive routines after the vnode got doomed. 2966 */ 2967 if (VN_IS_DOOMED(vp)) 2968 return; 2969 mp = vp->v_mount; 2970 mtx_lock(&mp->mnt_listmtx); 2971 if ((vp->v_mflag & VMP_LAZYLIST) == 0) { 2972 vp->v_mflag |= VMP_LAZYLIST; 2973 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist); 2974 mp->mnt_lazyvnodelistsize++; 2975 } 2976 mtx_unlock(&mp->mnt_listmtx); 2977 } 2978 2979 /* 2980 * This routine is only meant to be called from vgonel prior to dooming 2981 * the vnode. 2982 */ 2983 static void 2984 vunlazy_gone(struct vnode *vp) 2985 { 2986 struct mount *mp; 2987 2988 ASSERT_VOP_ELOCKED(vp, __func__); 2989 ASSERT_VI_LOCKED(vp, __func__); 2990 VNPASS(!VN_IS_DOOMED(vp), vp); 2991 2992 if (vp->v_mflag & VMP_LAZYLIST) { 2993 mp = vp->v_mount; 2994 mtx_lock(&mp->mnt_listmtx); 2995 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 2996 vp->v_mflag &= ~VMP_LAZYLIST; 2997 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 2998 mp->mnt_lazyvnodelistsize--; 2999 mtx_unlock(&mp->mnt_listmtx); 3000 } 3001 } 3002 3003 static void 3004 vdefer_inactive(struct vnode *vp) 3005 { 3006 3007 ASSERT_VI_LOCKED(vp, __func__); 3008 VNASSERT(vp->v_holdcnt > 0, vp, 3009 ("%s: vnode without hold count", __func__)); 3010 if (VN_IS_DOOMED(vp)) { 3011 vdropl(vp); 3012 return; 3013 } 3014 if (vp->v_iflag & VI_DEFINACT) { 3015 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 3016 vdropl(vp); 3017 return; 3018 } 3019 if (vp->v_usecount > 0) { 3020 vp->v_iflag &= ~VI_OWEINACT; 3021 vdropl(vp); 3022 return; 3023 } 3024 vlazy(vp); 3025 vp->v_iflag |= VI_DEFINACT; 3026 VI_UNLOCK(vp); 3027 counter_u64_add(deferred_inact, 1); 3028 } 3029 3030 static void 3031 vdefer_inactive_unlocked(struct vnode *vp) 3032 { 3033 3034 VI_LOCK(vp); 3035 if ((vp->v_iflag & VI_OWEINACT) == 0) { 3036 vdropl(vp); 3037 return; 3038 } 3039 vdefer_inactive(vp); 3040 } 3041 3042 enum vput_op { VRELE, VPUT, VUNREF }; 3043 3044 /* 3045 * Handle ->v_usecount transitioning to 0. 3046 * 3047 * By releasing the last usecount we take ownership of the hold count which 3048 * provides liveness of the vnode, meaning we have to vdrop. 3049 * 3050 * For all vnodes we may need to perform inactive processing. It requires an 3051 * exclusive lock on the vnode, while it is legal to call here with only a 3052 * shared lock (or no locks). If locking the vnode in an expected manner fails, 3053 * inactive processing gets deferred to the syncer. 3054 * 3055 * XXX Some filesystems pass in an exclusively locked vnode and strongly depend 3056 * on the lock being held all the way until VOP_INACTIVE. This in particular 3057 * happens with UFS which adds half-constructed vnodes to the hash, where they 3058 * can be found by other code. 3059 */ 3060 static void 3061 vput_final(struct vnode *vp, enum vput_op func) 3062 { 3063 int error; 3064 bool want_unlock; 3065 3066 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3067 VNPASS(vp->v_holdcnt > 0, vp); 3068 3069 VI_LOCK(vp); 3070 3071 /* 3072 * By the time we got here someone else might have transitioned 3073 * the count back to > 0. 3074 */ 3075 if (vp->v_usecount > 0) 3076 goto out; 3077 3078 /* 3079 * If the vnode is doomed vgone already performed inactive processing 3080 * (if needed). 3081 */ 3082 if (VN_IS_DOOMED(vp)) 3083 goto out; 3084 3085 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0)) 3086 goto out; 3087 3088 if (vp->v_iflag & VI_DOINGINACT) 3089 goto out; 3090 3091 /* 3092 * Locking operations here will drop the interlock and possibly the 3093 * vnode lock, opening a window where the vnode can get doomed all the 3094 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to 3095 * perform inactive. 3096 */ 3097 vp->v_iflag |= VI_OWEINACT; 3098 want_unlock = false; 3099 error = 0; 3100 switch (func) { 3101 case VRELE: 3102 switch (VOP_ISLOCKED(vp)) { 3103 case LK_EXCLUSIVE: 3104 break; 3105 case LK_EXCLOTHER: 3106 case 0: 3107 want_unlock = true; 3108 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 3109 VI_LOCK(vp); 3110 break; 3111 default: 3112 /* 3113 * The lock has at least one sharer, but we have no way 3114 * to conclude whether this is us. Play it safe and 3115 * defer processing. 3116 */ 3117 error = EAGAIN; 3118 break; 3119 } 3120 break; 3121 case VPUT: 3122 want_unlock = true; 3123 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3124 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | 3125 LK_NOWAIT); 3126 VI_LOCK(vp); 3127 } 3128 break; 3129 case VUNREF: 3130 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3131 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); 3132 VI_LOCK(vp); 3133 } 3134 break; 3135 } 3136 if (error == 0) { 3137 vinactive(vp); 3138 if (want_unlock) 3139 VOP_UNLOCK(vp); 3140 vdropl(vp); 3141 } else { 3142 vdefer_inactive(vp); 3143 } 3144 return; 3145 out: 3146 if (func == VPUT) 3147 VOP_UNLOCK(vp); 3148 vdropl(vp); 3149 } 3150 3151 /* 3152 * Decrement ->v_usecount for a vnode. 3153 * 3154 * Releasing the last use count requires additional processing, see vput_final 3155 * above for details. 3156 * 3157 * Comment above each variant denotes lock state on entry and exit. 3158 */ 3159 3160 /* 3161 * in: any 3162 * out: same as passed in 3163 */ 3164 void 3165 vrele(struct vnode *vp) 3166 { 3167 3168 ASSERT_VI_UNLOCKED(vp, __func__); 3169 if (!refcount_release(&vp->v_usecount)) 3170 return; 3171 vput_final(vp, VRELE); 3172 } 3173 3174 /* 3175 * in: locked 3176 * out: unlocked 3177 */ 3178 void 3179 vput(struct vnode *vp) 3180 { 3181 3182 ASSERT_VOP_LOCKED(vp, __func__); 3183 ASSERT_VI_UNLOCKED(vp, __func__); 3184 if (!refcount_release(&vp->v_usecount)) { 3185 VOP_UNLOCK(vp); 3186 return; 3187 } 3188 vput_final(vp, VPUT); 3189 } 3190 3191 /* 3192 * in: locked 3193 * out: locked 3194 */ 3195 void 3196 vunref(struct vnode *vp) 3197 { 3198 3199 ASSERT_VOP_LOCKED(vp, __func__); 3200 ASSERT_VI_UNLOCKED(vp, __func__); 3201 if (!refcount_release(&vp->v_usecount)) 3202 return; 3203 vput_final(vp, VUNREF); 3204 } 3205 3206 void 3207 vhold(struct vnode *vp) 3208 { 3209 struct vdbatch *vd; 3210 int old; 3211 3212 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3213 old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3214 VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3215 ("%s: wrong hold count %d", __func__, old)); 3216 if (old != 0) 3217 return; 3218 critical_enter(); 3219 vd = DPCPU_PTR(vd); 3220 vd->freevnodes--; 3221 critical_exit(); 3222 } 3223 3224 void 3225 vholdl(struct vnode *vp) 3226 { 3227 3228 ASSERT_VI_LOCKED(vp, __func__); 3229 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3230 vhold(vp); 3231 } 3232 3233 void 3234 vholdnz(struct vnode *vp) 3235 { 3236 3237 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3238 #ifdef INVARIANTS 3239 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3240 VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3241 ("%s: wrong hold count %d", __func__, old)); 3242 #else 3243 atomic_add_int(&vp->v_holdcnt, 1); 3244 #endif 3245 } 3246 3247 /* 3248 * Grab a hold count unless the vnode is freed. 3249 * 3250 * Only use this routine if vfs smr is the only protection you have against 3251 * freeing the vnode. 3252 * 3253 * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag 3254 * is not set. After the flag is set the vnode becomes immutable to anyone but 3255 * the thread which managed to set the flag. 3256 * 3257 * It may be tempting to replace the loop with: 3258 * count = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3259 * if (count & VHOLD_NO_SMR) { 3260 * backpedal and error out; 3261 * } 3262 * 3263 * However, while this is more performant, it hinders debugging by eliminating 3264 * the previously mentioned invariant. 3265 */ 3266 bool 3267 vhold_smr(struct vnode *vp) 3268 { 3269 int count; 3270 3271 VFS_SMR_ASSERT_ENTERED(); 3272 3273 count = atomic_load_int(&vp->v_holdcnt); 3274 for (;;) { 3275 if (count & VHOLD_NO_SMR) { 3276 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, 3277 ("non-zero hold count with flags %d\n", count)); 3278 return (false); 3279 } 3280 3281 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); 3282 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) 3283 return (true); 3284 } 3285 } 3286 3287 static void __noinline 3288 vdbatch_process(struct vdbatch *vd) 3289 { 3290 struct vnode *vp; 3291 int i; 3292 3293 mtx_assert(&vd->lock, MA_OWNED); 3294 MPASS(curthread->td_pinned > 0); 3295 MPASS(vd->index == VDBATCH_SIZE); 3296 3297 mtx_lock(&vnode_list_mtx); 3298 critical_enter(); 3299 freevnodes += vd->freevnodes; 3300 for (i = 0; i < VDBATCH_SIZE; i++) { 3301 vp = vd->tab[i]; 3302 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 3303 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist); 3304 MPASS(vp->v_dbatchcpu != NOCPU); 3305 vp->v_dbatchcpu = NOCPU; 3306 } 3307 mtx_unlock(&vnode_list_mtx); 3308 vd->freevnodes = 0; 3309 bzero(vd->tab, sizeof(vd->tab)); 3310 vd->index = 0; 3311 critical_exit(); 3312 } 3313 3314 static void 3315 vdbatch_enqueue(struct vnode *vp) 3316 { 3317 struct vdbatch *vd; 3318 3319 ASSERT_VI_LOCKED(vp, __func__); 3320 VNASSERT(!VN_IS_DOOMED(vp), vp, 3321 ("%s: deferring requeue of a doomed vnode", __func__)); 3322 3323 critical_enter(); 3324 vd = DPCPU_PTR(vd); 3325 vd->freevnodes++; 3326 if (vp->v_dbatchcpu != NOCPU) { 3327 VI_UNLOCK(vp); 3328 critical_exit(); 3329 return; 3330 } 3331 3332 sched_pin(); 3333 critical_exit(); 3334 mtx_lock(&vd->lock); 3335 MPASS(vd->index < VDBATCH_SIZE); 3336 MPASS(vd->tab[vd->index] == NULL); 3337 /* 3338 * A hack: we depend on being pinned so that we know what to put in 3339 * ->v_dbatchcpu. 3340 */ 3341 vp->v_dbatchcpu = curcpu; 3342 vd->tab[vd->index] = vp; 3343 vd->index++; 3344 VI_UNLOCK(vp); 3345 if (vd->index == VDBATCH_SIZE) 3346 vdbatch_process(vd); 3347 mtx_unlock(&vd->lock); 3348 sched_unpin(); 3349 } 3350 3351 /* 3352 * This routine must only be called for vnodes which are about to be 3353 * deallocated. Supporting dequeue for arbitrary vndoes would require 3354 * validating that the locked batch matches. 3355 */ 3356 static void 3357 vdbatch_dequeue(struct vnode *vp) 3358 { 3359 struct vdbatch *vd; 3360 int i; 3361 short cpu; 3362 3363 VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp, 3364 ("%s: called for a used vnode\n", __func__)); 3365 3366 cpu = vp->v_dbatchcpu; 3367 if (cpu == NOCPU) 3368 return; 3369 3370 vd = DPCPU_ID_PTR(cpu, vd); 3371 mtx_lock(&vd->lock); 3372 for (i = 0; i < vd->index; i++) { 3373 if (vd->tab[i] != vp) 3374 continue; 3375 vp->v_dbatchcpu = NOCPU; 3376 vd->index--; 3377 vd->tab[i] = vd->tab[vd->index]; 3378 vd->tab[vd->index] = NULL; 3379 break; 3380 } 3381 mtx_unlock(&vd->lock); 3382 /* 3383 * Either we dequeued the vnode above or the target CPU beat us to it. 3384 */ 3385 MPASS(vp->v_dbatchcpu == NOCPU); 3386 } 3387 3388 /* 3389 * Drop the hold count of the vnode. If this is the last reference to 3390 * the vnode we place it on the free list unless it has been vgone'd 3391 * (marked VIRF_DOOMED) in which case we will free it. 3392 * 3393 * Because the vnode vm object keeps a hold reference on the vnode if 3394 * there is at least one resident non-cached page, the vnode cannot 3395 * leave the active list without the page cleanup done. 3396 */ 3397 static void 3398 vdrop_deactivate(struct vnode *vp) 3399 { 3400 struct mount *mp; 3401 3402 ASSERT_VI_LOCKED(vp, __func__); 3403 /* 3404 * Mark a vnode as free: remove it from its active list 3405 * and put it up for recycling on the freelist. 3406 */ 3407 VNASSERT(!VN_IS_DOOMED(vp), vp, 3408 ("vdrop: returning doomed vnode")); 3409 VNASSERT(vp->v_op != NULL, vp, 3410 ("vdrop: vnode already reclaimed.")); 3411 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 3412 ("vnode with VI_OWEINACT set")); 3413 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, 3414 ("vnode with VI_DEFINACT set")); 3415 if (vp->v_mflag & VMP_LAZYLIST) { 3416 mp = vp->v_mount; 3417 mtx_lock(&mp->mnt_listmtx); 3418 VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST")); 3419 /* 3420 * Don't remove the vnode from the lazy list if another thread 3421 * has increased the hold count. It may have re-enqueued the 3422 * vnode to the lazy list and is now responsible for its 3423 * removal. 3424 */ 3425 if (vp->v_holdcnt == 0) { 3426 vp->v_mflag &= ~VMP_LAZYLIST; 3427 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3428 mp->mnt_lazyvnodelistsize--; 3429 } 3430 mtx_unlock(&mp->mnt_listmtx); 3431 } 3432 vdbatch_enqueue(vp); 3433 } 3434 3435 void 3436 vdrop(struct vnode *vp) 3437 { 3438 3439 ASSERT_VI_UNLOCKED(vp, __func__); 3440 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3441 if (refcount_release_if_not_last(&vp->v_holdcnt)) 3442 return; 3443 VI_LOCK(vp); 3444 vdropl(vp); 3445 } 3446 3447 void 3448 vdropl(struct vnode *vp) 3449 { 3450 3451 ASSERT_VI_LOCKED(vp, __func__); 3452 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3453 if (!refcount_release(&vp->v_holdcnt)) { 3454 VI_UNLOCK(vp); 3455 return; 3456 } 3457 if (!VN_IS_DOOMED(vp)) { 3458 vdrop_deactivate(vp); 3459 /* 3460 * Also unlocks the interlock. We can't assert on it as we 3461 * released our hold and by now the vnode might have been 3462 * freed. 3463 */ 3464 return; 3465 } 3466 /* 3467 * Set the VHOLD_NO_SMR flag. 3468 * 3469 * We may be racing against vhold_smr. If they win we can just pretend 3470 * we never got this far, they will vdrop later. 3471 */ 3472 if (!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR)) { 3473 VI_UNLOCK(vp); 3474 /* 3475 * We lost the aforementioned race. Any subsequent access is 3476 * invalid as they might have managed to vdropl on their own. 3477 */ 3478 return; 3479 } 3480 freevnode(vp); 3481 } 3482 3483 /* 3484 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 3485 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 3486 */ 3487 static void 3488 vinactivef(struct vnode *vp) 3489 { 3490 struct vm_object *obj; 3491 3492 ASSERT_VOP_ELOCKED(vp, "vinactive"); 3493 ASSERT_VI_LOCKED(vp, "vinactive"); 3494 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, 3495 ("vinactive: recursed on VI_DOINGINACT")); 3496 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3497 vp->v_iflag |= VI_DOINGINACT; 3498 vp->v_iflag &= ~VI_OWEINACT; 3499 VI_UNLOCK(vp); 3500 /* 3501 * Before moving off the active list, we must be sure that any 3502 * modified pages are converted into the vnode's dirty 3503 * buffers, since these will no longer be checked once the 3504 * vnode is on the inactive list. 3505 * 3506 * The write-out of the dirty pages is asynchronous. At the 3507 * point that VOP_INACTIVE() is called, there could still be 3508 * pending I/O and dirty pages in the object. 3509 */ 3510 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 3511 vm_object_mightbedirty(obj)) { 3512 VM_OBJECT_WLOCK(obj); 3513 vm_object_page_clean(obj, 0, 0, 0); 3514 VM_OBJECT_WUNLOCK(obj); 3515 } 3516 VOP_INACTIVE(vp, curthread); 3517 VI_LOCK(vp); 3518 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, 3519 ("vinactive: lost VI_DOINGINACT")); 3520 vp->v_iflag &= ~VI_DOINGINACT; 3521 } 3522 3523 void 3524 vinactive(struct vnode *vp) 3525 { 3526 3527 ASSERT_VOP_ELOCKED(vp, "vinactive"); 3528 ASSERT_VI_LOCKED(vp, "vinactive"); 3529 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3530 3531 if ((vp->v_iflag & VI_OWEINACT) == 0) 3532 return; 3533 if (vp->v_iflag & VI_DOINGINACT) 3534 return; 3535 if (vp->v_usecount > 0) { 3536 vp->v_iflag &= ~VI_OWEINACT; 3537 return; 3538 } 3539 vinactivef(vp); 3540 } 3541 3542 /* 3543 * Remove any vnodes in the vnode table belonging to mount point mp. 3544 * 3545 * If FORCECLOSE is not specified, there should not be any active ones, 3546 * return error if any are found (nb: this is a user error, not a 3547 * system error). If FORCECLOSE is specified, detach any active vnodes 3548 * that are found. 3549 * 3550 * If WRITECLOSE is set, only flush out regular file vnodes open for 3551 * writing. 3552 * 3553 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 3554 * 3555 * `rootrefs' specifies the base reference count for the root vnode 3556 * of this filesystem. The root vnode is considered busy if its 3557 * v_usecount exceeds this value. On a successful return, vflush(, td) 3558 * will call vrele() on the root vnode exactly rootrefs times. 3559 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 3560 * be zero. 3561 */ 3562 #ifdef DIAGNOSTIC 3563 static int busyprt = 0; /* print out busy vnodes */ 3564 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 3565 #endif 3566 3567 int 3568 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 3569 { 3570 struct vnode *vp, *mvp, *rootvp = NULL; 3571 struct vattr vattr; 3572 int busy = 0, error; 3573 3574 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 3575 rootrefs, flags); 3576 if (rootrefs > 0) { 3577 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 3578 ("vflush: bad args")); 3579 /* 3580 * Get the filesystem root vnode. We can vput() it 3581 * immediately, since with rootrefs > 0, it won't go away. 3582 */ 3583 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 3584 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 3585 __func__, error); 3586 return (error); 3587 } 3588 vput(rootvp); 3589 } 3590 loop: 3591 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 3592 vholdl(vp); 3593 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 3594 if (error) { 3595 vdrop(vp); 3596 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3597 goto loop; 3598 } 3599 /* 3600 * Skip over a vnodes marked VV_SYSTEM. 3601 */ 3602 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 3603 VOP_UNLOCK(vp); 3604 vdrop(vp); 3605 continue; 3606 } 3607 /* 3608 * If WRITECLOSE is set, flush out unlinked but still open 3609 * files (even if open only for reading) and regular file 3610 * vnodes open for writing. 3611 */ 3612 if (flags & WRITECLOSE) { 3613 if (vp->v_object != NULL) { 3614 VM_OBJECT_WLOCK(vp->v_object); 3615 vm_object_page_clean(vp->v_object, 0, 0, 0); 3616 VM_OBJECT_WUNLOCK(vp->v_object); 3617 } 3618 error = VOP_FSYNC(vp, MNT_WAIT, td); 3619 if (error != 0) { 3620 VOP_UNLOCK(vp); 3621 vdrop(vp); 3622 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3623 return (error); 3624 } 3625 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 3626 VI_LOCK(vp); 3627 3628 if ((vp->v_type == VNON || 3629 (error == 0 && vattr.va_nlink > 0)) && 3630 (vp->v_writecount <= 0 || vp->v_type != VREG)) { 3631 VOP_UNLOCK(vp); 3632 vdropl(vp); 3633 continue; 3634 } 3635 } else 3636 VI_LOCK(vp); 3637 /* 3638 * With v_usecount == 0, all we need to do is clear out the 3639 * vnode data structures and we are done. 3640 * 3641 * If FORCECLOSE is set, forcibly close the vnode. 3642 */ 3643 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 3644 vgonel(vp); 3645 } else { 3646 busy++; 3647 #ifdef DIAGNOSTIC 3648 if (busyprt) 3649 vn_printf(vp, "vflush: busy vnode "); 3650 #endif 3651 } 3652 VOP_UNLOCK(vp); 3653 vdropl(vp); 3654 } 3655 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 3656 /* 3657 * If just the root vnode is busy, and if its refcount 3658 * is equal to `rootrefs', then go ahead and kill it. 3659 */ 3660 VI_LOCK(rootvp); 3661 KASSERT(busy > 0, ("vflush: not busy")); 3662 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 3663 ("vflush: usecount %d < rootrefs %d", 3664 rootvp->v_usecount, rootrefs)); 3665 if (busy == 1 && rootvp->v_usecount == rootrefs) { 3666 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 3667 vgone(rootvp); 3668 VOP_UNLOCK(rootvp); 3669 busy = 0; 3670 } else 3671 VI_UNLOCK(rootvp); 3672 } 3673 if (busy) { 3674 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 3675 busy); 3676 return (EBUSY); 3677 } 3678 for (; rootrefs > 0; rootrefs--) 3679 vrele(rootvp); 3680 return (0); 3681 } 3682 3683 /* 3684 * Recycle an unused vnode to the front of the free list. 3685 */ 3686 int 3687 vrecycle(struct vnode *vp) 3688 { 3689 int recycled; 3690 3691 VI_LOCK(vp); 3692 recycled = vrecyclel(vp); 3693 VI_UNLOCK(vp); 3694 return (recycled); 3695 } 3696 3697 /* 3698 * vrecycle, with the vp interlock held. 3699 */ 3700 int 3701 vrecyclel(struct vnode *vp) 3702 { 3703 int recycled; 3704 3705 ASSERT_VOP_ELOCKED(vp, __func__); 3706 ASSERT_VI_LOCKED(vp, __func__); 3707 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3708 recycled = 0; 3709 if (vp->v_usecount == 0) { 3710 recycled = 1; 3711 vgonel(vp); 3712 } 3713 return (recycled); 3714 } 3715 3716 /* 3717 * Eliminate all activity associated with a vnode 3718 * in preparation for reuse. 3719 */ 3720 void 3721 vgone(struct vnode *vp) 3722 { 3723 VI_LOCK(vp); 3724 vgonel(vp); 3725 VI_UNLOCK(vp); 3726 } 3727 3728 static void 3729 notify_lowervp_vfs_dummy(struct mount *mp __unused, 3730 struct vnode *lowervp __unused) 3731 { 3732 } 3733 3734 /* 3735 * Notify upper mounts about reclaimed or unlinked vnode. 3736 */ 3737 void 3738 vfs_notify_upper(struct vnode *vp, int event) 3739 { 3740 static struct vfsops vgonel_vfsops = { 3741 .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy, 3742 .vfs_unlink_lowervp = notify_lowervp_vfs_dummy, 3743 }; 3744 struct mount *mp, *ump, *mmp; 3745 3746 mp = vp->v_mount; 3747 if (mp == NULL) 3748 return; 3749 if (TAILQ_EMPTY(&mp->mnt_uppers)) 3750 return; 3751 3752 mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO); 3753 mmp->mnt_op = &vgonel_vfsops; 3754 mmp->mnt_kern_flag |= MNTK_MARKER; 3755 MNT_ILOCK(mp); 3756 mp->mnt_kern_flag |= MNTK_VGONE_UPPER; 3757 for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) { 3758 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) { 3759 ump = TAILQ_NEXT(ump, mnt_upper_link); 3760 continue; 3761 } 3762 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link); 3763 MNT_IUNLOCK(mp); 3764 switch (event) { 3765 case VFS_NOTIFY_UPPER_RECLAIM: 3766 VFS_RECLAIM_LOWERVP(ump, vp); 3767 break; 3768 case VFS_NOTIFY_UPPER_UNLINK: 3769 VFS_UNLINK_LOWERVP(ump, vp); 3770 break; 3771 default: 3772 KASSERT(0, ("invalid event %d", event)); 3773 break; 3774 } 3775 MNT_ILOCK(mp); 3776 ump = TAILQ_NEXT(mmp, mnt_upper_link); 3777 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link); 3778 } 3779 free(mmp, M_TEMP); 3780 mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER; 3781 if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) { 3782 mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER; 3783 wakeup(&mp->mnt_uppers); 3784 } 3785 MNT_IUNLOCK(mp); 3786 } 3787 3788 /* 3789 * vgone, with the vp interlock held. 3790 */ 3791 static void 3792 vgonel(struct vnode *vp) 3793 { 3794 struct thread *td; 3795 struct mount *mp; 3796 vm_object_t object; 3797 bool active, oweinact; 3798 3799 ASSERT_VOP_ELOCKED(vp, "vgonel"); 3800 ASSERT_VI_LOCKED(vp, "vgonel"); 3801 VNASSERT(vp->v_holdcnt, vp, 3802 ("vgonel: vp %p has no reference.", vp)); 3803 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3804 td = curthread; 3805 3806 /* 3807 * Don't vgonel if we're already doomed. 3808 */ 3809 if (vp->v_irflag & VIRF_DOOMED) 3810 return; 3811 /* 3812 * Paired with freevnode. 3813 */ 3814 vn_seqc_write_begin_locked(vp); 3815 vunlazy_gone(vp); 3816 vp->v_irflag |= VIRF_DOOMED; 3817 3818 /* 3819 * Check to see if the vnode is in use. If so, we have to call 3820 * VOP_CLOSE() and VOP_INACTIVE(). 3821 */ 3822 active = vp->v_usecount > 0; 3823 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 3824 /* 3825 * If we need to do inactive VI_OWEINACT will be set. 3826 */ 3827 if (vp->v_iflag & VI_DEFINACT) { 3828 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 3829 vp->v_iflag &= ~VI_DEFINACT; 3830 vdropl(vp); 3831 } else { 3832 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); 3833 VI_UNLOCK(vp); 3834 } 3835 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); 3836 3837 /* 3838 * If purging an active vnode, it must be closed and 3839 * deactivated before being reclaimed. 3840 */ 3841 if (active) 3842 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 3843 if (oweinact || active) { 3844 VI_LOCK(vp); 3845 vinactivef(vp); 3846 VI_UNLOCK(vp); 3847 } 3848 if (vp->v_type == VSOCK) 3849 vfs_unp_reclaim(vp); 3850 3851 /* 3852 * Clean out any buffers associated with the vnode. 3853 * If the flush fails, just toss the buffers. 3854 */ 3855 mp = NULL; 3856 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 3857 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 3858 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { 3859 while (vinvalbuf(vp, 0, 0, 0) != 0) 3860 ; 3861 } 3862 3863 BO_LOCK(&vp->v_bufobj); 3864 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && 3865 vp->v_bufobj.bo_dirty.bv_cnt == 0 && 3866 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && 3867 vp->v_bufobj.bo_clean.bv_cnt == 0, 3868 ("vp %p bufobj not invalidated", vp)); 3869 3870 /* 3871 * For VMIO bufobj, BO_DEAD is set later, or in 3872 * vm_object_terminate() after the object's page queue is 3873 * flushed. 3874 */ 3875 object = vp->v_bufobj.bo_object; 3876 if (object == NULL) 3877 vp->v_bufobj.bo_flag |= BO_DEAD; 3878 BO_UNLOCK(&vp->v_bufobj); 3879 3880 /* 3881 * Handle the VM part. Tmpfs handles v_object on its own (the 3882 * OBJT_VNODE check). Nullfs or other bypassing filesystems 3883 * should not touch the object borrowed from the lower vnode 3884 * (the handle check). 3885 */ 3886 if (object != NULL && object->type == OBJT_VNODE && 3887 object->handle == vp) 3888 vnode_destroy_vobject(vp); 3889 3890 /* 3891 * Reclaim the vnode. 3892 */ 3893 if (VOP_RECLAIM(vp, td)) 3894 panic("vgone: cannot reclaim"); 3895 if (mp != NULL) 3896 vn_finished_secondary_write(mp); 3897 VNASSERT(vp->v_object == NULL, vp, 3898 ("vop_reclaim left v_object vp=%p", vp)); 3899 /* 3900 * Clear the advisory locks and wake up waiting threads. 3901 */ 3902 (void)VOP_ADVLOCKPURGE(vp); 3903 vp->v_lockf = NULL; 3904 /* 3905 * Delete from old mount point vnode list. 3906 */ 3907 delmntque(vp); 3908 cache_purge_vgone(vp); 3909 /* 3910 * Done with purge, reset to the standard lock and invalidate 3911 * the vnode. 3912 */ 3913 VI_LOCK(vp); 3914 vp->v_vnlock = &vp->v_lock; 3915 vp->v_op = &dead_vnodeops; 3916 vp->v_type = VBAD; 3917 } 3918 3919 /* 3920 * Print out a description of a vnode. 3921 */ 3922 static const char * const typename[] = 3923 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", 3924 "VMARKER"}; 3925 3926 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0, 3927 "new hold count flag not added to vn_printf"); 3928 3929 void 3930 vn_printf(struct vnode *vp, const char *fmt, ...) 3931 { 3932 va_list ap; 3933 char buf[256], buf2[16]; 3934 u_long flags; 3935 u_int holdcnt; 3936 3937 va_start(ap, fmt); 3938 vprintf(fmt, ap); 3939 va_end(ap); 3940 printf("%p: ", (void *)vp); 3941 printf("type %s\n", typename[vp->v_type]); 3942 holdcnt = atomic_load_int(&vp->v_holdcnt); 3943 printf(" usecount %d, writecount %d, refcount %d seqc users %d", 3944 vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS, 3945 vp->v_seqc_users); 3946 switch (vp->v_type) { 3947 case VDIR: 3948 printf(" mountedhere %p\n", vp->v_mountedhere); 3949 break; 3950 case VCHR: 3951 printf(" rdev %p\n", vp->v_rdev); 3952 break; 3953 case VSOCK: 3954 printf(" socket %p\n", vp->v_unpcb); 3955 break; 3956 case VFIFO: 3957 printf(" fifoinfo %p\n", vp->v_fifoinfo); 3958 break; 3959 default: 3960 printf("\n"); 3961 break; 3962 } 3963 buf[0] = '\0'; 3964 buf[1] = '\0'; 3965 if (holdcnt & VHOLD_NO_SMR) 3966 strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf)); 3967 printf(" hold count flags (%s)\n", buf + 1); 3968 3969 buf[0] = '\0'; 3970 buf[1] = '\0'; 3971 if (vp->v_irflag & VIRF_DOOMED) 3972 strlcat(buf, "|VIRF_DOOMED", sizeof(buf)); 3973 flags = vp->v_irflag & ~(VIRF_DOOMED); 3974 if (flags != 0) { 3975 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags); 3976 strlcat(buf, buf2, sizeof(buf)); 3977 } 3978 if (vp->v_vflag & VV_ROOT) 3979 strlcat(buf, "|VV_ROOT", sizeof(buf)); 3980 if (vp->v_vflag & VV_ISTTY) 3981 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 3982 if (vp->v_vflag & VV_NOSYNC) 3983 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 3984 if (vp->v_vflag & VV_ETERNALDEV) 3985 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 3986 if (vp->v_vflag & VV_CACHEDLABEL) 3987 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 3988 if (vp->v_vflag & VV_VMSIZEVNLOCK) 3989 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf)); 3990 if (vp->v_vflag & VV_COPYONWRITE) 3991 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 3992 if (vp->v_vflag & VV_SYSTEM) 3993 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 3994 if (vp->v_vflag & VV_PROCDEP) 3995 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 3996 if (vp->v_vflag & VV_NOKNOTE) 3997 strlcat(buf, "|VV_NOKNOTE", sizeof(buf)); 3998 if (vp->v_vflag & VV_DELETED) 3999 strlcat(buf, "|VV_DELETED", sizeof(buf)); 4000 if (vp->v_vflag & VV_MD) 4001 strlcat(buf, "|VV_MD", sizeof(buf)); 4002 if (vp->v_vflag & VV_FORCEINSMQ) 4003 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 4004 if (vp->v_vflag & VV_READLINK) 4005 strlcat(buf, "|VV_READLINK", sizeof(buf)); 4006 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 4007 VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | 4008 VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ); 4009 if (flags != 0) { 4010 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 4011 strlcat(buf, buf2, sizeof(buf)); 4012 } 4013 if (vp->v_iflag & VI_TEXT_REF) 4014 strlcat(buf, "|VI_TEXT_REF", sizeof(buf)); 4015 if (vp->v_iflag & VI_MOUNT) 4016 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 4017 if (vp->v_iflag & VI_DOINGINACT) 4018 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 4019 if (vp->v_iflag & VI_OWEINACT) 4020 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 4021 if (vp->v_iflag & VI_DEFINACT) 4022 strlcat(buf, "|VI_DEFINACT", sizeof(buf)); 4023 flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT | 4024 VI_OWEINACT | VI_DEFINACT); 4025 if (flags != 0) { 4026 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 4027 strlcat(buf, buf2, sizeof(buf)); 4028 } 4029 if (vp->v_mflag & VMP_LAZYLIST) 4030 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf)); 4031 flags = vp->v_mflag & ~(VMP_LAZYLIST); 4032 if (flags != 0) { 4033 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags); 4034 strlcat(buf, buf2, sizeof(buf)); 4035 } 4036 printf(" flags (%s)\n", buf + 1); 4037 if (mtx_owned(VI_MTX(vp))) 4038 printf(" VI_LOCKed"); 4039 if (vp->v_object != NULL) 4040 printf(" v_object %p ref %d pages %d " 4041 "cleanbuf %d dirtybuf %d\n", 4042 vp->v_object, vp->v_object->ref_count, 4043 vp->v_object->resident_page_count, 4044 vp->v_bufobj.bo_clean.bv_cnt, 4045 vp->v_bufobj.bo_dirty.bv_cnt); 4046 printf(" "); 4047 lockmgr_printinfo(vp->v_vnlock); 4048 if (vp->v_data != NULL) 4049 VOP_PRINT(vp); 4050 } 4051 4052 #ifdef DDB 4053 /* 4054 * List all of the locked vnodes in the system. 4055 * Called when debugging the kernel. 4056 */ 4057 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 4058 { 4059 struct mount *mp; 4060 struct vnode *vp; 4061 4062 /* 4063 * Note: because this is DDB, we can't obey the locking semantics 4064 * for these structures, which means we could catch an inconsistent 4065 * state and dereference a nasty pointer. Not much to be done 4066 * about that. 4067 */ 4068 db_printf("Locked vnodes\n"); 4069 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4070 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4071 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) 4072 vn_printf(vp, "vnode "); 4073 } 4074 } 4075 } 4076 4077 /* 4078 * Show details about the given vnode. 4079 */ 4080 DB_SHOW_COMMAND(vnode, db_show_vnode) 4081 { 4082 struct vnode *vp; 4083 4084 if (!have_addr) 4085 return; 4086 vp = (struct vnode *)addr; 4087 vn_printf(vp, "vnode "); 4088 } 4089 4090 /* 4091 * Show details about the given mount point. 4092 */ 4093 DB_SHOW_COMMAND(mount, db_show_mount) 4094 { 4095 struct mount *mp; 4096 struct vfsopt *opt; 4097 struct statfs *sp; 4098 struct vnode *vp; 4099 char buf[512]; 4100 uint64_t mflags; 4101 u_int flags; 4102 4103 if (!have_addr) { 4104 /* No address given, print short info about all mount points. */ 4105 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4106 db_printf("%p %s on %s (%s)\n", mp, 4107 mp->mnt_stat.f_mntfromname, 4108 mp->mnt_stat.f_mntonname, 4109 mp->mnt_stat.f_fstypename); 4110 if (db_pager_quit) 4111 break; 4112 } 4113 db_printf("\nMore info: show mount <addr>\n"); 4114 return; 4115 } 4116 4117 mp = (struct mount *)addr; 4118 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 4119 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 4120 4121 buf[0] = '\0'; 4122 mflags = mp->mnt_flag; 4123 #define MNT_FLAG(flag) do { \ 4124 if (mflags & (flag)) { \ 4125 if (buf[0] != '\0') \ 4126 strlcat(buf, ", ", sizeof(buf)); \ 4127 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 4128 mflags &= ~(flag); \ 4129 } \ 4130 } while (0) 4131 MNT_FLAG(MNT_RDONLY); 4132 MNT_FLAG(MNT_SYNCHRONOUS); 4133 MNT_FLAG(MNT_NOEXEC); 4134 MNT_FLAG(MNT_NOSUID); 4135 MNT_FLAG(MNT_NFS4ACLS); 4136 MNT_FLAG(MNT_UNION); 4137 MNT_FLAG(MNT_ASYNC); 4138 MNT_FLAG(MNT_SUIDDIR); 4139 MNT_FLAG(MNT_SOFTDEP); 4140 MNT_FLAG(MNT_NOSYMFOLLOW); 4141 MNT_FLAG(MNT_GJOURNAL); 4142 MNT_FLAG(MNT_MULTILABEL); 4143 MNT_FLAG(MNT_ACLS); 4144 MNT_FLAG(MNT_NOATIME); 4145 MNT_FLAG(MNT_NOCLUSTERR); 4146 MNT_FLAG(MNT_NOCLUSTERW); 4147 MNT_FLAG(MNT_SUJ); 4148 MNT_FLAG(MNT_EXRDONLY); 4149 MNT_FLAG(MNT_EXPORTED); 4150 MNT_FLAG(MNT_DEFEXPORTED); 4151 MNT_FLAG(MNT_EXPORTANON); 4152 MNT_FLAG(MNT_EXKERB); 4153 MNT_FLAG(MNT_EXPUBLIC); 4154 MNT_FLAG(MNT_LOCAL); 4155 MNT_FLAG(MNT_QUOTA); 4156 MNT_FLAG(MNT_ROOTFS); 4157 MNT_FLAG(MNT_USER); 4158 MNT_FLAG(MNT_IGNORE); 4159 MNT_FLAG(MNT_UPDATE); 4160 MNT_FLAG(MNT_DELEXPORT); 4161 MNT_FLAG(MNT_RELOAD); 4162 MNT_FLAG(MNT_FORCE); 4163 MNT_FLAG(MNT_SNAPSHOT); 4164 MNT_FLAG(MNT_BYFSID); 4165 #undef MNT_FLAG 4166 if (mflags != 0) { 4167 if (buf[0] != '\0') 4168 strlcat(buf, ", ", sizeof(buf)); 4169 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4170 "0x%016jx", mflags); 4171 } 4172 db_printf(" mnt_flag = %s\n", buf); 4173 4174 buf[0] = '\0'; 4175 flags = mp->mnt_kern_flag; 4176 #define MNT_KERN_FLAG(flag) do { \ 4177 if (flags & (flag)) { \ 4178 if (buf[0] != '\0') \ 4179 strlcat(buf, ", ", sizeof(buf)); \ 4180 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 4181 flags &= ~(flag); \ 4182 } \ 4183 } while (0) 4184 MNT_KERN_FLAG(MNTK_UNMOUNTF); 4185 MNT_KERN_FLAG(MNTK_ASYNC); 4186 MNT_KERN_FLAG(MNTK_SOFTDEP); 4187 MNT_KERN_FLAG(MNTK_DRAINING); 4188 MNT_KERN_FLAG(MNTK_REFEXPIRE); 4189 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 4190 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 4191 MNT_KERN_FLAG(MNTK_NO_IOPF); 4192 MNT_KERN_FLAG(MNTK_VGONE_UPPER); 4193 MNT_KERN_FLAG(MNTK_VGONE_WAITER); 4194 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT); 4195 MNT_KERN_FLAG(MNTK_MARKER); 4196 MNT_KERN_FLAG(MNTK_USES_BCACHE); 4197 MNT_KERN_FLAG(MNTK_FPLOOKUP); 4198 MNT_KERN_FLAG(MNTK_NOASYNC); 4199 MNT_KERN_FLAG(MNTK_UNMOUNT); 4200 MNT_KERN_FLAG(MNTK_MWAIT); 4201 MNT_KERN_FLAG(MNTK_SUSPEND); 4202 MNT_KERN_FLAG(MNTK_SUSPEND2); 4203 MNT_KERN_FLAG(MNTK_SUSPENDED); 4204 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 4205 MNT_KERN_FLAG(MNTK_NOKNOTE); 4206 #undef MNT_KERN_FLAG 4207 if (flags != 0) { 4208 if (buf[0] != '\0') 4209 strlcat(buf, ", ", sizeof(buf)); 4210 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4211 "0x%08x", flags); 4212 } 4213 db_printf(" mnt_kern_flag = %s\n", buf); 4214 4215 db_printf(" mnt_opt = "); 4216 opt = TAILQ_FIRST(mp->mnt_opt); 4217 if (opt != NULL) { 4218 db_printf("%s", opt->name); 4219 opt = TAILQ_NEXT(opt, link); 4220 while (opt != NULL) { 4221 db_printf(", %s", opt->name); 4222 opt = TAILQ_NEXT(opt, link); 4223 } 4224 } 4225 db_printf("\n"); 4226 4227 sp = &mp->mnt_stat; 4228 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 4229 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 4230 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 4231 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 4232 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 4233 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 4234 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 4235 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 4236 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 4237 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 4238 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 4239 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 4240 4241 db_printf(" mnt_cred = { uid=%u ruid=%u", 4242 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 4243 if (jailed(mp->mnt_cred)) 4244 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 4245 db_printf(" }\n"); 4246 db_printf(" mnt_ref = %d (with %d in the struct)\n", 4247 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref); 4248 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 4249 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 4250 db_printf(" mnt_lazyvnodelistsize = %d\n", 4251 mp->mnt_lazyvnodelistsize); 4252 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n", 4253 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount); 4254 db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen); 4255 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 4256 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 4257 db_printf(" mnt_lockref = %d (with %d in the struct)\n", 4258 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref); 4259 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 4260 db_printf(" mnt_secondary_accwrites = %d\n", 4261 mp->mnt_secondary_accwrites); 4262 db_printf(" mnt_gjprovider = %s\n", 4263 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 4264 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops); 4265 4266 db_printf("\n\nList of active vnodes\n"); 4267 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4268 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) { 4269 vn_printf(vp, "vnode "); 4270 if (db_pager_quit) 4271 break; 4272 } 4273 } 4274 db_printf("\n\nList of inactive vnodes\n"); 4275 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4276 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) { 4277 vn_printf(vp, "vnode "); 4278 if (db_pager_quit) 4279 break; 4280 } 4281 } 4282 } 4283 #endif /* DDB */ 4284 4285 /* 4286 * Fill in a struct xvfsconf based on a struct vfsconf. 4287 */ 4288 static int 4289 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 4290 { 4291 struct xvfsconf xvfsp; 4292 4293 bzero(&xvfsp, sizeof(xvfsp)); 4294 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4295 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4296 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4297 xvfsp.vfc_flags = vfsp->vfc_flags; 4298 /* 4299 * These are unused in userland, we keep them 4300 * to not break binary compatibility. 4301 */ 4302 xvfsp.vfc_vfsops = NULL; 4303 xvfsp.vfc_next = NULL; 4304 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4305 } 4306 4307 #ifdef COMPAT_FREEBSD32 4308 struct xvfsconf32 { 4309 uint32_t vfc_vfsops; 4310 char vfc_name[MFSNAMELEN]; 4311 int32_t vfc_typenum; 4312 int32_t vfc_refcount; 4313 int32_t vfc_flags; 4314 uint32_t vfc_next; 4315 }; 4316 4317 static int 4318 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 4319 { 4320 struct xvfsconf32 xvfsp; 4321 4322 bzero(&xvfsp, sizeof(xvfsp)); 4323 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4324 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4325 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4326 xvfsp.vfc_flags = vfsp->vfc_flags; 4327 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4328 } 4329 #endif 4330 4331 /* 4332 * Top level filesystem related information gathering. 4333 */ 4334 static int 4335 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 4336 { 4337 struct vfsconf *vfsp; 4338 int error; 4339 4340 error = 0; 4341 vfsconf_slock(); 4342 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4343 #ifdef COMPAT_FREEBSD32 4344 if (req->flags & SCTL_MASK32) 4345 error = vfsconf2x32(req, vfsp); 4346 else 4347 #endif 4348 error = vfsconf2x(req, vfsp); 4349 if (error) 4350 break; 4351 } 4352 vfsconf_sunlock(); 4353 return (error); 4354 } 4355 4356 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | 4357 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, 4358 "S,xvfsconf", "List of all configured filesystems"); 4359 4360 #ifndef BURN_BRIDGES 4361 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 4362 4363 static int 4364 vfs_sysctl(SYSCTL_HANDLER_ARGS) 4365 { 4366 int *name = (int *)arg1 - 1; /* XXX */ 4367 u_int namelen = arg2 + 1; /* XXX */ 4368 struct vfsconf *vfsp; 4369 4370 log(LOG_WARNING, "userland calling deprecated sysctl, " 4371 "please rebuild world\n"); 4372 4373 #if 1 || defined(COMPAT_PRELITE2) 4374 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 4375 if (namelen == 1) 4376 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 4377 #endif 4378 4379 switch (name[1]) { 4380 case VFS_MAXTYPENUM: 4381 if (namelen != 2) 4382 return (ENOTDIR); 4383 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 4384 case VFS_CONF: 4385 if (namelen != 3) 4386 return (ENOTDIR); /* overloaded */ 4387 vfsconf_slock(); 4388 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4389 if (vfsp->vfc_typenum == name[2]) 4390 break; 4391 } 4392 vfsconf_sunlock(); 4393 if (vfsp == NULL) 4394 return (EOPNOTSUPP); 4395 #ifdef COMPAT_FREEBSD32 4396 if (req->flags & SCTL_MASK32) 4397 return (vfsconf2x32(req, vfsp)); 4398 else 4399 #endif 4400 return (vfsconf2x(req, vfsp)); 4401 } 4402 return (EOPNOTSUPP); 4403 } 4404 4405 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | 4406 CTLFLAG_MPSAFE, vfs_sysctl, 4407 "Generic filesystem"); 4408 4409 #if 1 || defined(COMPAT_PRELITE2) 4410 4411 static int 4412 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 4413 { 4414 int error; 4415 struct vfsconf *vfsp; 4416 struct ovfsconf ovfs; 4417 4418 vfsconf_slock(); 4419 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4420 bzero(&ovfs, sizeof(ovfs)); 4421 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 4422 strcpy(ovfs.vfc_name, vfsp->vfc_name); 4423 ovfs.vfc_index = vfsp->vfc_typenum; 4424 ovfs.vfc_refcount = vfsp->vfc_refcount; 4425 ovfs.vfc_flags = vfsp->vfc_flags; 4426 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 4427 if (error != 0) { 4428 vfsconf_sunlock(); 4429 return (error); 4430 } 4431 } 4432 vfsconf_sunlock(); 4433 return (0); 4434 } 4435 4436 #endif /* 1 || COMPAT_PRELITE2 */ 4437 #endif /* !BURN_BRIDGES */ 4438 4439 #define KINFO_VNODESLOP 10 4440 #ifdef notyet 4441 /* 4442 * Dump vnode list (via sysctl). 4443 */ 4444 /* ARGSUSED */ 4445 static int 4446 sysctl_vnode(SYSCTL_HANDLER_ARGS) 4447 { 4448 struct xvnode *xvn; 4449 struct mount *mp; 4450 struct vnode *vp; 4451 int error, len, n; 4452 4453 /* 4454 * Stale numvnodes access is not fatal here. 4455 */ 4456 req->lock = 0; 4457 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 4458 if (!req->oldptr) 4459 /* Make an estimate */ 4460 return (SYSCTL_OUT(req, 0, len)); 4461 4462 error = sysctl_wire_old_buffer(req, 0); 4463 if (error != 0) 4464 return (error); 4465 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 4466 n = 0; 4467 mtx_lock(&mountlist_mtx); 4468 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4469 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) 4470 continue; 4471 MNT_ILOCK(mp); 4472 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4473 if (n == len) 4474 break; 4475 vref(vp); 4476 xvn[n].xv_size = sizeof *xvn; 4477 xvn[n].xv_vnode = vp; 4478 xvn[n].xv_id = 0; /* XXX compat */ 4479 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 4480 XV_COPY(usecount); 4481 XV_COPY(writecount); 4482 XV_COPY(holdcnt); 4483 XV_COPY(mount); 4484 XV_COPY(numoutput); 4485 XV_COPY(type); 4486 #undef XV_COPY 4487 xvn[n].xv_flag = vp->v_vflag; 4488 4489 switch (vp->v_type) { 4490 case VREG: 4491 case VDIR: 4492 case VLNK: 4493 break; 4494 case VBLK: 4495 case VCHR: 4496 if (vp->v_rdev == NULL) { 4497 vrele(vp); 4498 continue; 4499 } 4500 xvn[n].xv_dev = dev2udev(vp->v_rdev); 4501 break; 4502 case VSOCK: 4503 xvn[n].xv_socket = vp->v_socket; 4504 break; 4505 case VFIFO: 4506 xvn[n].xv_fifo = vp->v_fifoinfo; 4507 break; 4508 case VNON: 4509 case VBAD: 4510 default: 4511 /* shouldn't happen? */ 4512 vrele(vp); 4513 continue; 4514 } 4515 vrele(vp); 4516 ++n; 4517 } 4518 MNT_IUNLOCK(mp); 4519 mtx_lock(&mountlist_mtx); 4520 vfs_unbusy(mp); 4521 if (n == len) 4522 break; 4523 } 4524 mtx_unlock(&mountlist_mtx); 4525 4526 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 4527 free(xvn, M_TEMP); 4528 return (error); 4529 } 4530 4531 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD | 4532 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode", 4533 ""); 4534 #endif 4535 4536 static void 4537 unmount_or_warn(struct mount *mp) 4538 { 4539 int error; 4540 4541 error = dounmount(mp, MNT_FORCE, curthread); 4542 if (error != 0) { 4543 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); 4544 if (error == EBUSY) 4545 printf("BUSY)\n"); 4546 else 4547 printf("%d)\n", error); 4548 } 4549 } 4550 4551 /* 4552 * Unmount all filesystems. The list is traversed in reverse order 4553 * of mounting to avoid dependencies. 4554 */ 4555 void 4556 vfs_unmountall(void) 4557 { 4558 struct mount *mp, *tmp; 4559 4560 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 4561 4562 /* 4563 * Since this only runs when rebooting, it is not interlocked. 4564 */ 4565 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { 4566 vfs_ref(mp); 4567 4568 /* 4569 * Forcibly unmounting "/dev" before "/" would prevent clean 4570 * unmount of the latter. 4571 */ 4572 if (mp == rootdevmp) 4573 continue; 4574 4575 unmount_or_warn(mp); 4576 } 4577 4578 if (rootdevmp != NULL) 4579 unmount_or_warn(rootdevmp); 4580 } 4581 4582 static void 4583 vfs_deferred_inactive(struct vnode *vp, int lkflags) 4584 { 4585 4586 ASSERT_VI_LOCKED(vp, __func__); 4587 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set")); 4588 if ((vp->v_iflag & VI_OWEINACT) == 0) { 4589 vdropl(vp); 4590 return; 4591 } 4592 if (vn_lock(vp, lkflags) == 0) { 4593 VI_LOCK(vp); 4594 vinactive(vp); 4595 VOP_UNLOCK(vp); 4596 vdropl(vp); 4597 return; 4598 } 4599 vdefer_inactive_unlocked(vp); 4600 } 4601 4602 static int 4603 vfs_periodic_inactive_filter(struct vnode *vp, void *arg) 4604 { 4605 4606 return (vp->v_iflag & VI_DEFINACT); 4607 } 4608 4609 static void __noinline 4610 vfs_periodic_inactive(struct mount *mp, int flags) 4611 { 4612 struct vnode *vp, *mvp; 4613 int lkflags; 4614 4615 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 4616 if (flags != MNT_WAIT) 4617 lkflags |= LK_NOWAIT; 4618 4619 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) { 4620 if ((vp->v_iflag & VI_DEFINACT) == 0) { 4621 VI_UNLOCK(vp); 4622 continue; 4623 } 4624 vp->v_iflag &= ~VI_DEFINACT; 4625 vfs_deferred_inactive(vp, lkflags); 4626 } 4627 } 4628 4629 static inline bool 4630 vfs_want_msync(struct vnode *vp) 4631 { 4632 struct vm_object *obj; 4633 4634 /* 4635 * This test may be performed without any locks held. 4636 * We rely on vm_object's type stability. 4637 */ 4638 if (vp->v_vflag & VV_NOSYNC) 4639 return (false); 4640 obj = vp->v_object; 4641 return (obj != NULL && vm_object_mightbedirty(obj)); 4642 } 4643 4644 static int 4645 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused) 4646 { 4647 4648 if (vp->v_vflag & VV_NOSYNC) 4649 return (false); 4650 if (vp->v_iflag & VI_DEFINACT) 4651 return (true); 4652 return (vfs_want_msync(vp)); 4653 } 4654 4655 static void __noinline 4656 vfs_periodic_msync_inactive(struct mount *mp, int flags) 4657 { 4658 struct vnode *vp, *mvp; 4659 struct vm_object *obj; 4660 struct thread *td; 4661 int lkflags, objflags; 4662 bool seen_defer; 4663 4664 td = curthread; 4665 4666 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 4667 if (flags != MNT_WAIT) { 4668 lkflags |= LK_NOWAIT; 4669 objflags = OBJPC_NOSYNC; 4670 } else { 4671 objflags = OBJPC_SYNC; 4672 } 4673 4674 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) { 4675 seen_defer = false; 4676 if (vp->v_iflag & VI_DEFINACT) { 4677 vp->v_iflag &= ~VI_DEFINACT; 4678 seen_defer = true; 4679 } 4680 if (!vfs_want_msync(vp)) { 4681 if (seen_defer) 4682 vfs_deferred_inactive(vp, lkflags); 4683 else 4684 VI_UNLOCK(vp); 4685 continue; 4686 } 4687 if (vget(vp, lkflags, td) == 0) { 4688 obj = vp->v_object; 4689 if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) { 4690 VM_OBJECT_WLOCK(obj); 4691 vm_object_page_clean(obj, 0, 0, objflags); 4692 VM_OBJECT_WUNLOCK(obj); 4693 } 4694 vput(vp); 4695 if (seen_defer) 4696 vdrop(vp); 4697 } else { 4698 if (seen_defer) 4699 vdefer_inactive_unlocked(vp); 4700 } 4701 } 4702 } 4703 4704 void 4705 vfs_periodic(struct mount *mp, int flags) 4706 { 4707 4708 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 4709 4710 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0) 4711 vfs_periodic_inactive(mp, flags); 4712 else 4713 vfs_periodic_msync_inactive(mp, flags); 4714 } 4715 4716 static void 4717 destroy_vpollinfo_free(struct vpollinfo *vi) 4718 { 4719 4720 knlist_destroy(&vi->vpi_selinfo.si_note); 4721 mtx_destroy(&vi->vpi_lock); 4722 uma_zfree(vnodepoll_zone, vi); 4723 } 4724 4725 static void 4726 destroy_vpollinfo(struct vpollinfo *vi) 4727 { 4728 4729 knlist_clear(&vi->vpi_selinfo.si_note, 1); 4730 seldrain(&vi->vpi_selinfo); 4731 destroy_vpollinfo_free(vi); 4732 } 4733 4734 /* 4735 * Initialize per-vnode helper structure to hold poll-related state. 4736 */ 4737 void 4738 v_addpollinfo(struct vnode *vp) 4739 { 4740 struct vpollinfo *vi; 4741 4742 if (vp->v_pollinfo != NULL) 4743 return; 4744 vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO); 4745 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 4746 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 4747 vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked); 4748 VI_LOCK(vp); 4749 if (vp->v_pollinfo != NULL) { 4750 VI_UNLOCK(vp); 4751 destroy_vpollinfo_free(vi); 4752 return; 4753 } 4754 vp->v_pollinfo = vi; 4755 VI_UNLOCK(vp); 4756 } 4757 4758 /* 4759 * Record a process's interest in events which might happen to 4760 * a vnode. Because poll uses the historic select-style interface 4761 * internally, this routine serves as both the ``check for any 4762 * pending events'' and the ``record my interest in future events'' 4763 * functions. (These are done together, while the lock is held, 4764 * to avoid race conditions.) 4765 */ 4766 int 4767 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 4768 { 4769 4770 v_addpollinfo(vp); 4771 mtx_lock(&vp->v_pollinfo->vpi_lock); 4772 if (vp->v_pollinfo->vpi_revents & events) { 4773 /* 4774 * This leaves events we are not interested 4775 * in available for the other process which 4776 * which presumably had requested them 4777 * (otherwise they would never have been 4778 * recorded). 4779 */ 4780 events &= vp->v_pollinfo->vpi_revents; 4781 vp->v_pollinfo->vpi_revents &= ~events; 4782 4783 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4784 return (events); 4785 } 4786 vp->v_pollinfo->vpi_events |= events; 4787 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 4788 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4789 return (0); 4790 } 4791 4792 /* 4793 * Routine to create and manage a filesystem syncer vnode. 4794 */ 4795 #define sync_close ((int (*)(struct vop_close_args *))nullop) 4796 static int sync_fsync(struct vop_fsync_args *); 4797 static int sync_inactive(struct vop_inactive_args *); 4798 static int sync_reclaim(struct vop_reclaim_args *); 4799 4800 static struct vop_vector sync_vnodeops = { 4801 .vop_bypass = VOP_EOPNOTSUPP, 4802 .vop_close = sync_close, /* close */ 4803 .vop_fsync = sync_fsync, /* fsync */ 4804 .vop_inactive = sync_inactive, /* inactive */ 4805 .vop_need_inactive = vop_stdneed_inactive, /* need_inactive */ 4806 .vop_reclaim = sync_reclaim, /* reclaim */ 4807 .vop_lock1 = vop_stdlock, /* lock */ 4808 .vop_unlock = vop_stdunlock, /* unlock */ 4809 .vop_islocked = vop_stdislocked, /* islocked */ 4810 }; 4811 VFS_VOP_VECTOR_REGISTER(sync_vnodeops); 4812 4813 /* 4814 * Create a new filesystem syncer vnode for the specified mount point. 4815 */ 4816 void 4817 vfs_allocate_syncvnode(struct mount *mp) 4818 { 4819 struct vnode *vp; 4820 struct bufobj *bo; 4821 static long start, incr, next; 4822 int error; 4823 4824 /* Allocate a new vnode */ 4825 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 4826 if (error != 0) 4827 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 4828 vp->v_type = VNON; 4829 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4830 vp->v_vflag |= VV_FORCEINSMQ; 4831 error = insmntque(vp, mp); 4832 if (error != 0) 4833 panic("vfs_allocate_syncvnode: insmntque() failed"); 4834 vp->v_vflag &= ~VV_FORCEINSMQ; 4835 VOP_UNLOCK(vp); 4836 /* 4837 * Place the vnode onto the syncer worklist. We attempt to 4838 * scatter them about on the list so that they will go off 4839 * at evenly distributed times even if all the filesystems 4840 * are mounted at once. 4841 */ 4842 next += incr; 4843 if (next == 0 || next > syncer_maxdelay) { 4844 start /= 2; 4845 incr /= 2; 4846 if (start == 0) { 4847 start = syncer_maxdelay / 2; 4848 incr = syncer_maxdelay; 4849 } 4850 next = start; 4851 } 4852 bo = &vp->v_bufobj; 4853 BO_LOCK(bo); 4854 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 4855 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 4856 mtx_lock(&sync_mtx); 4857 sync_vnode_count++; 4858 if (mp->mnt_syncer == NULL) { 4859 mp->mnt_syncer = vp; 4860 vp = NULL; 4861 } 4862 mtx_unlock(&sync_mtx); 4863 BO_UNLOCK(bo); 4864 if (vp != NULL) { 4865 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4866 vgone(vp); 4867 vput(vp); 4868 } 4869 } 4870 4871 void 4872 vfs_deallocate_syncvnode(struct mount *mp) 4873 { 4874 struct vnode *vp; 4875 4876 mtx_lock(&sync_mtx); 4877 vp = mp->mnt_syncer; 4878 if (vp != NULL) 4879 mp->mnt_syncer = NULL; 4880 mtx_unlock(&sync_mtx); 4881 if (vp != NULL) 4882 vrele(vp); 4883 } 4884 4885 /* 4886 * Do a lazy sync of the filesystem. 4887 */ 4888 static int 4889 sync_fsync(struct vop_fsync_args *ap) 4890 { 4891 struct vnode *syncvp = ap->a_vp; 4892 struct mount *mp = syncvp->v_mount; 4893 int error, save; 4894 struct bufobj *bo; 4895 4896 /* 4897 * We only need to do something if this is a lazy evaluation. 4898 */ 4899 if (ap->a_waitfor != MNT_LAZY) 4900 return (0); 4901 4902 /* 4903 * Move ourselves to the back of the sync list. 4904 */ 4905 bo = &syncvp->v_bufobj; 4906 BO_LOCK(bo); 4907 vn_syncer_add_to_worklist(bo, syncdelay); 4908 BO_UNLOCK(bo); 4909 4910 /* 4911 * Walk the list of vnodes pushing all that are dirty and 4912 * not already on the sync list. 4913 */ 4914 if (vfs_busy(mp, MBF_NOWAIT) != 0) 4915 return (0); 4916 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 4917 vfs_unbusy(mp); 4918 return (0); 4919 } 4920 save = curthread_pflags_set(TDP_SYNCIO); 4921 /* 4922 * The filesystem at hand may be idle with free vnodes stored in the 4923 * batch. Return them instead of letting them stay there indefinitely. 4924 */ 4925 vfs_periodic(mp, MNT_NOWAIT); 4926 error = VFS_SYNC(mp, MNT_LAZY); 4927 curthread_pflags_restore(save); 4928 vn_finished_write(mp); 4929 vfs_unbusy(mp); 4930 return (error); 4931 } 4932 4933 /* 4934 * The syncer vnode is no referenced. 4935 */ 4936 static int 4937 sync_inactive(struct vop_inactive_args *ap) 4938 { 4939 4940 vgone(ap->a_vp); 4941 return (0); 4942 } 4943 4944 /* 4945 * The syncer vnode is no longer needed and is being decommissioned. 4946 * 4947 * Modifications to the worklist must be protected by sync_mtx. 4948 */ 4949 static int 4950 sync_reclaim(struct vop_reclaim_args *ap) 4951 { 4952 struct vnode *vp = ap->a_vp; 4953 struct bufobj *bo; 4954 4955 bo = &vp->v_bufobj; 4956 BO_LOCK(bo); 4957 mtx_lock(&sync_mtx); 4958 if (vp->v_mount->mnt_syncer == vp) 4959 vp->v_mount->mnt_syncer = NULL; 4960 if (bo->bo_flag & BO_ONWORKLST) { 4961 LIST_REMOVE(bo, bo_synclist); 4962 syncer_worklist_len--; 4963 sync_vnode_count--; 4964 bo->bo_flag &= ~BO_ONWORKLST; 4965 } 4966 mtx_unlock(&sync_mtx); 4967 BO_UNLOCK(bo); 4968 4969 return (0); 4970 } 4971 4972 int 4973 vn_need_pageq_flush(struct vnode *vp) 4974 { 4975 struct vm_object *obj; 4976 int need; 4977 4978 MPASS(mtx_owned(VI_MTX(vp))); 4979 need = 0; 4980 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 4981 vm_object_mightbedirty(obj)) 4982 need = 1; 4983 return (need); 4984 } 4985 4986 /* 4987 * Check if vnode represents a disk device 4988 */ 4989 int 4990 vn_isdisk(struct vnode *vp, int *errp) 4991 { 4992 int error; 4993 4994 if (vp->v_type != VCHR) { 4995 error = ENOTBLK; 4996 goto out; 4997 } 4998 error = 0; 4999 dev_lock(); 5000 if (vp->v_rdev == NULL) 5001 error = ENXIO; 5002 else if (vp->v_rdev->si_devsw == NULL) 5003 error = ENXIO; 5004 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 5005 error = ENOTBLK; 5006 dev_unlock(); 5007 out: 5008 if (errp != NULL) 5009 *errp = error; 5010 return (error == 0); 5011 } 5012 5013 /* 5014 * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see 5015 * the comment above cache_fplookup for details. 5016 * 5017 * We never deny as priv_check_cred calls are not yet supported, see vaccess. 5018 */ 5019 int 5020 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred) 5021 { 5022 5023 VFS_SMR_ASSERT_ENTERED(); 5024 5025 /* Check the owner. */ 5026 if (cred->cr_uid == file_uid) { 5027 if (file_mode & S_IXUSR) 5028 return (0); 5029 return (EAGAIN); 5030 } 5031 5032 /* Otherwise, check the groups (first match) */ 5033 if (groupmember(file_gid, cred)) { 5034 if (file_mode & S_IXGRP) 5035 return (0); 5036 return (EAGAIN); 5037 } 5038 5039 /* Otherwise, check everyone else. */ 5040 if (file_mode & S_IXOTH) 5041 return (0); 5042 return (EAGAIN); 5043 } 5044 5045 /* 5046 * Common filesystem object access control check routine. Accepts a 5047 * vnode's type, "mode", uid and gid, requested access mode, and credentials. 5048 * Returns 0 on success, or an errno on failure. 5049 */ 5050 int 5051 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 5052 accmode_t accmode, struct ucred *cred) 5053 { 5054 accmode_t dac_granted; 5055 accmode_t priv_granted; 5056 5057 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 5058 ("invalid bit in accmode")); 5059 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 5060 ("VAPPEND without VWRITE")); 5061 5062 /* 5063 * Look for a normal, non-privileged way to access the file/directory 5064 * as requested. If it exists, go with that. 5065 */ 5066 5067 dac_granted = 0; 5068 5069 /* Check the owner. */ 5070 if (cred->cr_uid == file_uid) { 5071 dac_granted |= VADMIN; 5072 if (file_mode & S_IXUSR) 5073 dac_granted |= VEXEC; 5074 if (file_mode & S_IRUSR) 5075 dac_granted |= VREAD; 5076 if (file_mode & S_IWUSR) 5077 dac_granted |= (VWRITE | VAPPEND); 5078 5079 if ((accmode & dac_granted) == accmode) 5080 return (0); 5081 5082 goto privcheck; 5083 } 5084 5085 /* Otherwise, check the groups (first match) */ 5086 if (groupmember(file_gid, cred)) { 5087 if (file_mode & S_IXGRP) 5088 dac_granted |= VEXEC; 5089 if (file_mode & S_IRGRP) 5090 dac_granted |= VREAD; 5091 if (file_mode & S_IWGRP) 5092 dac_granted |= (VWRITE | VAPPEND); 5093 5094 if ((accmode & dac_granted) == accmode) 5095 return (0); 5096 5097 goto privcheck; 5098 } 5099 5100 /* Otherwise, check everyone else. */ 5101 if (file_mode & S_IXOTH) 5102 dac_granted |= VEXEC; 5103 if (file_mode & S_IROTH) 5104 dac_granted |= VREAD; 5105 if (file_mode & S_IWOTH) 5106 dac_granted |= (VWRITE | VAPPEND); 5107 if ((accmode & dac_granted) == accmode) 5108 return (0); 5109 5110 privcheck: 5111 /* 5112 * Build a privilege mask to determine if the set of privileges 5113 * satisfies the requirements when combined with the granted mask 5114 * from above. For each privilege, if the privilege is required, 5115 * bitwise or the request type onto the priv_granted mask. 5116 */ 5117 priv_granted = 0; 5118 5119 if (type == VDIR) { 5120 /* 5121 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 5122 * requests, instead of PRIV_VFS_EXEC. 5123 */ 5124 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5125 !priv_check_cred(cred, PRIV_VFS_LOOKUP)) 5126 priv_granted |= VEXEC; 5127 } else { 5128 /* 5129 * Ensure that at least one execute bit is on. Otherwise, 5130 * a privileged user will always succeed, and we don't want 5131 * this to happen unless the file really is executable. 5132 */ 5133 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5134 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 5135 !priv_check_cred(cred, PRIV_VFS_EXEC)) 5136 priv_granted |= VEXEC; 5137 } 5138 5139 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 5140 !priv_check_cred(cred, PRIV_VFS_READ)) 5141 priv_granted |= VREAD; 5142 5143 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 5144 !priv_check_cred(cred, PRIV_VFS_WRITE)) 5145 priv_granted |= (VWRITE | VAPPEND); 5146 5147 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 5148 !priv_check_cred(cred, PRIV_VFS_ADMIN)) 5149 priv_granted |= VADMIN; 5150 5151 if ((accmode & (priv_granted | dac_granted)) == accmode) { 5152 return (0); 5153 } 5154 5155 return ((accmode & VADMIN) ? EPERM : EACCES); 5156 } 5157 5158 /* 5159 * Credential check based on process requesting service, and per-attribute 5160 * permissions. 5161 */ 5162 int 5163 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 5164 struct thread *td, accmode_t accmode) 5165 { 5166 5167 /* 5168 * Kernel-invoked always succeeds. 5169 */ 5170 if (cred == NOCRED) 5171 return (0); 5172 5173 /* 5174 * Do not allow privileged processes in jail to directly manipulate 5175 * system attributes. 5176 */ 5177 switch (attrnamespace) { 5178 case EXTATTR_NAMESPACE_SYSTEM: 5179 /* Potentially should be: return (EPERM); */ 5180 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM)); 5181 case EXTATTR_NAMESPACE_USER: 5182 return (VOP_ACCESS(vp, accmode, cred, td)); 5183 default: 5184 return (EPERM); 5185 } 5186 } 5187 5188 #ifdef DEBUG_VFS_LOCKS 5189 /* 5190 * This only exists to suppress warnings from unlocked specfs accesses. It is 5191 * no longer ok to have an unlocked VFS. 5192 */ 5193 #define IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL || \ 5194 (vp)->v_type == VCHR || (vp)->v_type == VBAD) 5195 5196 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 5197 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, 5198 "Drop into debugger on lock violation"); 5199 5200 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 5201 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 5202 0, "Check for interlock across VOPs"); 5203 5204 int vfs_badlock_print = 1; /* Print lock violations. */ 5205 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 5206 0, "Print lock violations"); 5207 5208 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */ 5209 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode, 5210 0, "Print vnode details on lock violations"); 5211 5212 #ifdef KDB 5213 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 5214 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, 5215 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations"); 5216 #endif 5217 5218 static void 5219 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 5220 { 5221 5222 #ifdef KDB 5223 if (vfs_badlock_backtrace) 5224 kdb_backtrace(); 5225 #endif 5226 if (vfs_badlock_vnode) 5227 vn_printf(vp, "vnode "); 5228 if (vfs_badlock_print) 5229 printf("%s: %p %s\n", str, (void *)vp, msg); 5230 if (vfs_badlock_ddb) 5231 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 5232 } 5233 5234 void 5235 assert_vi_locked(struct vnode *vp, const char *str) 5236 { 5237 5238 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 5239 vfs_badlock("interlock is not locked but should be", str, vp); 5240 } 5241 5242 void 5243 assert_vi_unlocked(struct vnode *vp, const char *str) 5244 { 5245 5246 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 5247 vfs_badlock("interlock is locked but should not be", str, vp); 5248 } 5249 5250 void 5251 assert_vop_locked(struct vnode *vp, const char *str) 5252 { 5253 int locked; 5254 5255 if (!IGNORE_LOCK(vp)) { 5256 locked = VOP_ISLOCKED(vp); 5257 if (locked == 0 || locked == LK_EXCLOTHER) 5258 vfs_badlock("is not locked but should be", str, vp); 5259 } 5260 } 5261 5262 void 5263 assert_vop_unlocked(struct vnode *vp, const char *str) 5264 { 5265 5266 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE) 5267 vfs_badlock("is locked but should not be", str, vp); 5268 } 5269 5270 void 5271 assert_vop_elocked(struct vnode *vp, const char *str) 5272 { 5273 5274 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 5275 vfs_badlock("is not exclusive locked but should be", str, vp); 5276 } 5277 #endif /* DEBUG_VFS_LOCKS */ 5278 5279 void 5280 vop_rename_fail(struct vop_rename_args *ap) 5281 { 5282 5283 if (ap->a_tvp != NULL) 5284 vput(ap->a_tvp); 5285 if (ap->a_tdvp == ap->a_tvp) 5286 vrele(ap->a_tdvp); 5287 else 5288 vput(ap->a_tdvp); 5289 vrele(ap->a_fdvp); 5290 vrele(ap->a_fvp); 5291 } 5292 5293 void 5294 vop_rename_pre(void *ap) 5295 { 5296 struct vop_rename_args *a = ap; 5297 5298 #ifdef DEBUG_VFS_LOCKS 5299 if (a->a_tvp) 5300 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 5301 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 5302 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 5303 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 5304 5305 /* Check the source (from). */ 5306 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 5307 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 5308 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 5309 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 5310 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 5311 5312 /* Check the target. */ 5313 if (a->a_tvp) 5314 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 5315 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 5316 #endif 5317 /* 5318 * It may be tempting to add vn_seqc_write_begin/end calls here and 5319 * in vop_rename_post but that's not going to work out since some 5320 * filesystems relookup vnodes mid-rename. This is probably a bug. 5321 * 5322 * For now filesystems are expected to do the relevant calls after they 5323 * decide what vnodes to operate on. 5324 */ 5325 if (a->a_tdvp != a->a_fdvp) 5326 vhold(a->a_fdvp); 5327 if (a->a_tvp != a->a_fvp) 5328 vhold(a->a_fvp); 5329 vhold(a->a_tdvp); 5330 if (a->a_tvp) 5331 vhold(a->a_tvp); 5332 } 5333 5334 #ifdef DEBUG_VFS_LOCKS 5335 void 5336 vop_fplookup_vexec_debugpre(void *ap __unused) 5337 { 5338 5339 VFS_SMR_ASSERT_ENTERED(); 5340 } 5341 5342 void 5343 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused) 5344 { 5345 5346 VFS_SMR_ASSERT_ENTERED(); 5347 } 5348 5349 void 5350 vop_strategy_debugpre(void *ap) 5351 { 5352 struct vop_strategy_args *a; 5353 struct buf *bp; 5354 5355 a = ap; 5356 bp = a->a_bp; 5357 5358 /* 5359 * Cluster ops lock their component buffers but not the IO container. 5360 */ 5361 if ((bp->b_flags & B_CLUSTER) != 0) 5362 return; 5363 5364 if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) { 5365 if (vfs_badlock_print) 5366 printf( 5367 "VOP_STRATEGY: bp is not locked but should be\n"); 5368 if (vfs_badlock_ddb) 5369 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 5370 } 5371 } 5372 5373 void 5374 vop_lock_debugpre(void *ap) 5375 { 5376 struct vop_lock1_args *a = ap; 5377 5378 if ((a->a_flags & LK_INTERLOCK) == 0) 5379 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5380 else 5381 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 5382 } 5383 5384 void 5385 vop_lock_debugpost(void *ap, int rc) 5386 { 5387 struct vop_lock1_args *a = ap; 5388 5389 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5390 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) 5391 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 5392 } 5393 5394 void 5395 vop_unlock_debugpre(void *ap) 5396 { 5397 struct vop_unlock_args *a = ap; 5398 5399 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 5400 } 5401 5402 void 5403 vop_need_inactive_debugpre(void *ap) 5404 { 5405 struct vop_need_inactive_args *a = ap; 5406 5407 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 5408 } 5409 5410 void 5411 vop_need_inactive_debugpost(void *ap, int rc) 5412 { 5413 struct vop_need_inactive_args *a = ap; 5414 5415 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 5416 } 5417 #endif 5418 5419 void 5420 vop_create_pre(void *ap) 5421 { 5422 struct vop_create_args *a; 5423 struct vnode *dvp; 5424 5425 a = ap; 5426 dvp = a->a_dvp; 5427 vn_seqc_write_begin(dvp); 5428 } 5429 5430 void 5431 vop_create_post(void *ap, int rc) 5432 { 5433 struct vop_create_args *a; 5434 struct vnode *dvp; 5435 5436 a = ap; 5437 dvp = a->a_dvp; 5438 vn_seqc_write_end(dvp); 5439 if (!rc) 5440 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 5441 } 5442 5443 void 5444 vop_whiteout_pre(void *ap) 5445 { 5446 struct vop_whiteout_args *a; 5447 struct vnode *dvp; 5448 5449 a = ap; 5450 dvp = a->a_dvp; 5451 vn_seqc_write_begin(dvp); 5452 } 5453 5454 void 5455 vop_whiteout_post(void *ap, int rc) 5456 { 5457 struct vop_whiteout_args *a; 5458 struct vnode *dvp; 5459 5460 a = ap; 5461 dvp = a->a_dvp; 5462 vn_seqc_write_end(dvp); 5463 } 5464 5465 void 5466 vop_deleteextattr_pre(void *ap) 5467 { 5468 struct vop_deleteextattr_args *a; 5469 struct vnode *vp; 5470 5471 a = ap; 5472 vp = a->a_vp; 5473 vn_seqc_write_begin(vp); 5474 } 5475 5476 void 5477 vop_deleteextattr_post(void *ap, int rc) 5478 { 5479 struct vop_deleteextattr_args *a; 5480 struct vnode *vp; 5481 5482 a = ap; 5483 vp = a->a_vp; 5484 vn_seqc_write_end(vp); 5485 if (!rc) 5486 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 5487 } 5488 5489 void 5490 vop_link_pre(void *ap) 5491 { 5492 struct vop_link_args *a; 5493 struct vnode *vp, *tdvp; 5494 5495 a = ap; 5496 vp = a->a_vp; 5497 tdvp = a->a_tdvp; 5498 vn_seqc_write_begin(vp); 5499 vn_seqc_write_begin(tdvp); 5500 } 5501 5502 void 5503 vop_link_post(void *ap, int rc) 5504 { 5505 struct vop_link_args *a; 5506 struct vnode *vp, *tdvp; 5507 5508 a = ap; 5509 vp = a->a_vp; 5510 tdvp = a->a_tdvp; 5511 vn_seqc_write_end(vp); 5512 vn_seqc_write_end(tdvp); 5513 if (!rc) { 5514 VFS_KNOTE_LOCKED(vp, NOTE_LINK); 5515 VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE); 5516 } 5517 } 5518 5519 void 5520 vop_mkdir_pre(void *ap) 5521 { 5522 struct vop_mkdir_args *a; 5523 struct vnode *dvp; 5524 5525 a = ap; 5526 dvp = a->a_dvp; 5527 vn_seqc_write_begin(dvp); 5528 } 5529 5530 void 5531 vop_mkdir_post(void *ap, int rc) 5532 { 5533 struct vop_mkdir_args *a; 5534 struct vnode *dvp; 5535 5536 a = ap; 5537 dvp = a->a_dvp; 5538 vn_seqc_write_end(dvp); 5539 if (!rc) 5540 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 5541 } 5542 5543 void 5544 vop_mknod_pre(void *ap) 5545 { 5546 struct vop_mknod_args *a; 5547 struct vnode *dvp; 5548 5549 a = ap; 5550 dvp = a->a_dvp; 5551 vn_seqc_write_begin(dvp); 5552 } 5553 5554 void 5555 vop_mknod_post(void *ap, int rc) 5556 { 5557 struct vop_mknod_args *a; 5558 struct vnode *dvp; 5559 5560 a = ap; 5561 dvp = a->a_dvp; 5562 vn_seqc_write_end(dvp); 5563 if (!rc) 5564 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 5565 } 5566 5567 void 5568 vop_reclaim_post(void *ap, int rc) 5569 { 5570 struct vop_reclaim_args *a; 5571 struct vnode *vp; 5572 5573 a = ap; 5574 vp = a->a_vp; 5575 ASSERT_VOP_IN_SEQC(vp); 5576 if (!rc) 5577 VFS_KNOTE_LOCKED(vp, NOTE_REVOKE); 5578 } 5579 5580 void 5581 vop_remove_pre(void *ap) 5582 { 5583 struct vop_remove_args *a; 5584 struct vnode *dvp, *vp; 5585 5586 a = ap; 5587 dvp = a->a_dvp; 5588 vp = a->a_vp; 5589 vn_seqc_write_begin(dvp); 5590 vn_seqc_write_begin(vp); 5591 } 5592 5593 void 5594 vop_remove_post(void *ap, int rc) 5595 { 5596 struct vop_remove_args *a; 5597 struct vnode *dvp, *vp; 5598 5599 a = ap; 5600 dvp = a->a_dvp; 5601 vp = a->a_vp; 5602 vn_seqc_write_end(dvp); 5603 vn_seqc_write_end(vp); 5604 if (!rc) { 5605 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 5606 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 5607 } 5608 } 5609 5610 void 5611 vop_rename_post(void *ap, int rc) 5612 { 5613 struct vop_rename_args *a = ap; 5614 long hint; 5615 5616 if (!rc) { 5617 hint = NOTE_WRITE; 5618 if (a->a_fdvp == a->a_tdvp) { 5619 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) 5620 hint |= NOTE_LINK; 5621 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 5622 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 5623 } else { 5624 hint |= NOTE_EXTEND; 5625 if (a->a_fvp->v_type == VDIR) 5626 hint |= NOTE_LINK; 5627 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 5628 5629 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && 5630 a->a_tvp->v_type == VDIR) 5631 hint &= ~NOTE_LINK; 5632 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 5633 } 5634 5635 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 5636 if (a->a_tvp) 5637 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 5638 } 5639 if (a->a_tdvp != a->a_fdvp) 5640 vdrop(a->a_fdvp); 5641 if (a->a_tvp != a->a_fvp) 5642 vdrop(a->a_fvp); 5643 vdrop(a->a_tdvp); 5644 if (a->a_tvp) 5645 vdrop(a->a_tvp); 5646 } 5647 5648 void 5649 vop_rmdir_pre(void *ap) 5650 { 5651 struct vop_rmdir_args *a; 5652 struct vnode *dvp, *vp; 5653 5654 a = ap; 5655 dvp = a->a_dvp; 5656 vp = a->a_vp; 5657 vn_seqc_write_begin(dvp); 5658 vn_seqc_write_begin(vp); 5659 } 5660 5661 void 5662 vop_rmdir_post(void *ap, int rc) 5663 { 5664 struct vop_rmdir_args *a; 5665 struct vnode *dvp, *vp; 5666 5667 a = ap; 5668 dvp = a->a_dvp; 5669 vp = a->a_vp; 5670 vn_seqc_write_end(dvp); 5671 vn_seqc_write_end(vp); 5672 if (!rc) { 5673 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 5674 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 5675 } 5676 } 5677 5678 void 5679 vop_setattr_pre(void *ap) 5680 { 5681 struct vop_setattr_args *a; 5682 struct vnode *vp; 5683 5684 a = ap; 5685 vp = a->a_vp; 5686 vn_seqc_write_begin(vp); 5687 } 5688 5689 void 5690 vop_setattr_post(void *ap, int rc) 5691 { 5692 struct vop_setattr_args *a; 5693 struct vnode *vp; 5694 5695 a = ap; 5696 vp = a->a_vp; 5697 vn_seqc_write_end(vp); 5698 if (!rc) 5699 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 5700 } 5701 5702 void 5703 vop_setacl_pre(void *ap) 5704 { 5705 struct vop_setacl_args *a; 5706 struct vnode *vp; 5707 5708 a = ap; 5709 vp = a->a_vp; 5710 vn_seqc_write_begin(vp); 5711 } 5712 5713 void 5714 vop_setacl_post(void *ap, int rc __unused) 5715 { 5716 struct vop_setacl_args *a; 5717 struct vnode *vp; 5718 5719 a = ap; 5720 vp = a->a_vp; 5721 vn_seqc_write_end(vp); 5722 } 5723 5724 void 5725 vop_setextattr_pre(void *ap) 5726 { 5727 struct vop_setextattr_args *a; 5728 struct vnode *vp; 5729 5730 a = ap; 5731 vp = a->a_vp; 5732 vn_seqc_write_begin(vp); 5733 } 5734 5735 void 5736 vop_setextattr_post(void *ap, int rc) 5737 { 5738 struct vop_setextattr_args *a; 5739 struct vnode *vp; 5740 5741 a = ap; 5742 vp = a->a_vp; 5743 vn_seqc_write_end(vp); 5744 if (!rc) 5745 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 5746 } 5747 5748 void 5749 vop_symlink_pre(void *ap) 5750 { 5751 struct vop_symlink_args *a; 5752 struct vnode *dvp; 5753 5754 a = ap; 5755 dvp = a->a_dvp; 5756 vn_seqc_write_begin(dvp); 5757 } 5758 5759 void 5760 vop_symlink_post(void *ap, int rc) 5761 { 5762 struct vop_symlink_args *a; 5763 struct vnode *dvp; 5764 5765 a = ap; 5766 dvp = a->a_dvp; 5767 vn_seqc_write_end(dvp); 5768 if (!rc) 5769 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 5770 } 5771 5772 void 5773 vop_open_post(void *ap, int rc) 5774 { 5775 struct vop_open_args *a = ap; 5776 5777 if (!rc) 5778 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); 5779 } 5780 5781 void 5782 vop_close_post(void *ap, int rc) 5783 { 5784 struct vop_close_args *a = ap; 5785 5786 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ 5787 !VN_IS_DOOMED(a->a_vp))) { 5788 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 5789 NOTE_CLOSE_WRITE : NOTE_CLOSE); 5790 } 5791 } 5792 5793 void 5794 vop_read_post(void *ap, int rc) 5795 { 5796 struct vop_read_args *a = ap; 5797 5798 if (!rc) 5799 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 5800 } 5801 5802 void 5803 vop_readdir_post(void *ap, int rc) 5804 { 5805 struct vop_readdir_args *a = ap; 5806 5807 if (!rc) 5808 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 5809 } 5810 5811 static struct knlist fs_knlist; 5812 5813 static void 5814 vfs_event_init(void *arg) 5815 { 5816 knlist_init_mtx(&fs_knlist, NULL); 5817 } 5818 /* XXX - correct order? */ 5819 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 5820 5821 void 5822 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 5823 { 5824 5825 KNOTE_UNLOCKED(&fs_knlist, event); 5826 } 5827 5828 static int filt_fsattach(struct knote *kn); 5829 static void filt_fsdetach(struct knote *kn); 5830 static int filt_fsevent(struct knote *kn, long hint); 5831 5832 struct filterops fs_filtops = { 5833 .f_isfd = 0, 5834 .f_attach = filt_fsattach, 5835 .f_detach = filt_fsdetach, 5836 .f_event = filt_fsevent 5837 }; 5838 5839 static int 5840 filt_fsattach(struct knote *kn) 5841 { 5842 5843 kn->kn_flags |= EV_CLEAR; 5844 knlist_add(&fs_knlist, kn, 0); 5845 return (0); 5846 } 5847 5848 static void 5849 filt_fsdetach(struct knote *kn) 5850 { 5851 5852 knlist_remove(&fs_knlist, kn, 0); 5853 } 5854 5855 static int 5856 filt_fsevent(struct knote *kn, long hint) 5857 { 5858 5859 kn->kn_fflags |= hint; 5860 return (kn->kn_fflags != 0); 5861 } 5862 5863 static int 5864 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 5865 { 5866 struct vfsidctl vc; 5867 int error; 5868 struct mount *mp; 5869 5870 error = SYSCTL_IN(req, &vc, sizeof(vc)); 5871 if (error) 5872 return (error); 5873 if (vc.vc_vers != VFS_CTL_VERS1) 5874 return (EINVAL); 5875 mp = vfs_getvfs(&vc.vc_fsid); 5876 if (mp == NULL) 5877 return (ENOENT); 5878 /* ensure that a specific sysctl goes to the right filesystem. */ 5879 if (strcmp(vc.vc_fstypename, "*") != 0 && 5880 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 5881 vfs_rel(mp); 5882 return (EINVAL); 5883 } 5884 VCTLTOREQ(&vc, req); 5885 error = VFS_SYSCTL(mp, vc.vc_op, req); 5886 vfs_rel(mp); 5887 return (error); 5888 } 5889 5890 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR, 5891 NULL, 0, sysctl_vfs_ctl, "", 5892 "Sysctl by fsid"); 5893 5894 /* 5895 * Function to initialize a va_filerev field sensibly. 5896 * XXX: Wouldn't a random number make a lot more sense ?? 5897 */ 5898 u_quad_t 5899 init_va_filerev(void) 5900 { 5901 struct bintime bt; 5902 5903 getbinuptime(&bt); 5904 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 5905 } 5906 5907 static int filt_vfsread(struct knote *kn, long hint); 5908 static int filt_vfswrite(struct knote *kn, long hint); 5909 static int filt_vfsvnode(struct knote *kn, long hint); 5910 static void filt_vfsdetach(struct knote *kn); 5911 static struct filterops vfsread_filtops = { 5912 .f_isfd = 1, 5913 .f_detach = filt_vfsdetach, 5914 .f_event = filt_vfsread 5915 }; 5916 static struct filterops vfswrite_filtops = { 5917 .f_isfd = 1, 5918 .f_detach = filt_vfsdetach, 5919 .f_event = filt_vfswrite 5920 }; 5921 static struct filterops vfsvnode_filtops = { 5922 .f_isfd = 1, 5923 .f_detach = filt_vfsdetach, 5924 .f_event = filt_vfsvnode 5925 }; 5926 5927 static void 5928 vfs_knllock(void *arg) 5929 { 5930 struct vnode *vp = arg; 5931 5932 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5933 } 5934 5935 static void 5936 vfs_knlunlock(void *arg) 5937 { 5938 struct vnode *vp = arg; 5939 5940 VOP_UNLOCK(vp); 5941 } 5942 5943 static void 5944 vfs_knl_assert_locked(void *arg) 5945 { 5946 #ifdef DEBUG_VFS_LOCKS 5947 struct vnode *vp = arg; 5948 5949 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 5950 #endif 5951 } 5952 5953 static void 5954 vfs_knl_assert_unlocked(void *arg) 5955 { 5956 #ifdef DEBUG_VFS_LOCKS 5957 struct vnode *vp = arg; 5958 5959 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 5960 #endif 5961 } 5962 5963 int 5964 vfs_kqfilter(struct vop_kqfilter_args *ap) 5965 { 5966 struct vnode *vp = ap->a_vp; 5967 struct knote *kn = ap->a_kn; 5968 struct knlist *knl; 5969 5970 switch (kn->kn_filter) { 5971 case EVFILT_READ: 5972 kn->kn_fop = &vfsread_filtops; 5973 break; 5974 case EVFILT_WRITE: 5975 kn->kn_fop = &vfswrite_filtops; 5976 break; 5977 case EVFILT_VNODE: 5978 kn->kn_fop = &vfsvnode_filtops; 5979 break; 5980 default: 5981 return (EINVAL); 5982 } 5983 5984 kn->kn_hook = (caddr_t)vp; 5985 5986 v_addpollinfo(vp); 5987 if (vp->v_pollinfo == NULL) 5988 return (ENOMEM); 5989 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 5990 vhold(vp); 5991 knlist_add(knl, kn, 0); 5992 5993 return (0); 5994 } 5995 5996 /* 5997 * Detach knote from vnode 5998 */ 5999 static void 6000 filt_vfsdetach(struct knote *kn) 6001 { 6002 struct vnode *vp = (struct vnode *)kn->kn_hook; 6003 6004 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 6005 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 6006 vdrop(vp); 6007 } 6008 6009 /*ARGSUSED*/ 6010 static int 6011 filt_vfsread(struct knote *kn, long hint) 6012 { 6013 struct vnode *vp = (struct vnode *)kn->kn_hook; 6014 struct vattr va; 6015 int res; 6016 6017 /* 6018 * filesystem is gone, so set the EOF flag and schedule 6019 * the knote for deletion. 6020 */ 6021 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6022 VI_LOCK(vp); 6023 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6024 VI_UNLOCK(vp); 6025 return (1); 6026 } 6027 6028 if (VOP_GETATTR(vp, &va, curthread->td_ucred)) 6029 return (0); 6030 6031 VI_LOCK(vp); 6032 kn->kn_data = va.va_size - kn->kn_fp->f_offset; 6033 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; 6034 VI_UNLOCK(vp); 6035 return (res); 6036 } 6037 6038 /*ARGSUSED*/ 6039 static int 6040 filt_vfswrite(struct knote *kn, long hint) 6041 { 6042 struct vnode *vp = (struct vnode *)kn->kn_hook; 6043 6044 VI_LOCK(vp); 6045 6046 /* 6047 * filesystem is gone, so set the EOF flag and schedule 6048 * the knote for deletion. 6049 */ 6050 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) 6051 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6052 6053 kn->kn_data = 0; 6054 VI_UNLOCK(vp); 6055 return (1); 6056 } 6057 6058 static int 6059 filt_vfsvnode(struct knote *kn, long hint) 6060 { 6061 struct vnode *vp = (struct vnode *)kn->kn_hook; 6062 int res; 6063 6064 VI_LOCK(vp); 6065 if (kn->kn_sfflags & hint) 6066 kn->kn_fflags |= hint; 6067 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6068 kn->kn_flags |= EV_EOF; 6069 VI_UNLOCK(vp); 6070 return (1); 6071 } 6072 res = (kn->kn_fflags != 0); 6073 VI_UNLOCK(vp); 6074 return (res); 6075 } 6076 6077 /* 6078 * Returns whether the directory is empty or not. 6079 * If it is empty, the return value is 0; otherwise 6080 * the return value is an error value (which may 6081 * be ENOTEMPTY). 6082 */ 6083 int 6084 vfs_emptydir(struct vnode *vp) 6085 { 6086 struct uio uio; 6087 struct iovec iov; 6088 struct dirent *dirent, *dp, *endp; 6089 int error, eof; 6090 6091 error = 0; 6092 eof = 0; 6093 6094 ASSERT_VOP_LOCKED(vp, "vfs_emptydir"); 6095 6096 dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK); 6097 iov.iov_base = dirent; 6098 iov.iov_len = sizeof(struct dirent); 6099 6100 uio.uio_iov = &iov; 6101 uio.uio_iovcnt = 1; 6102 uio.uio_offset = 0; 6103 uio.uio_resid = sizeof(struct dirent); 6104 uio.uio_segflg = UIO_SYSSPACE; 6105 uio.uio_rw = UIO_READ; 6106 uio.uio_td = curthread; 6107 6108 while (eof == 0 && error == 0) { 6109 error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof, 6110 NULL, NULL); 6111 if (error != 0) 6112 break; 6113 endp = (void *)((uint8_t *)dirent + 6114 sizeof(struct dirent) - uio.uio_resid); 6115 for (dp = dirent; dp < endp; 6116 dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) { 6117 if (dp->d_type == DT_WHT) 6118 continue; 6119 if (dp->d_namlen == 0) 6120 continue; 6121 if (dp->d_type != DT_DIR && 6122 dp->d_type != DT_UNKNOWN) { 6123 error = ENOTEMPTY; 6124 break; 6125 } 6126 if (dp->d_namlen > 2) { 6127 error = ENOTEMPTY; 6128 break; 6129 } 6130 if (dp->d_namlen == 1 && 6131 dp->d_name[0] != '.') { 6132 error = ENOTEMPTY; 6133 break; 6134 } 6135 if (dp->d_namlen == 2 && 6136 dp->d_name[1] != '.') { 6137 error = ENOTEMPTY; 6138 break; 6139 } 6140 uio.uio_resid = sizeof(struct dirent); 6141 } 6142 } 6143 free(dirent, M_TEMP); 6144 return (error); 6145 } 6146 6147 int 6148 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 6149 { 6150 int error; 6151 6152 if (dp->d_reclen > ap->a_uio->uio_resid) 6153 return (ENAMETOOLONG); 6154 error = uiomove(dp, dp->d_reclen, ap->a_uio); 6155 if (error) { 6156 if (ap->a_ncookies != NULL) { 6157 if (ap->a_cookies != NULL) 6158 free(ap->a_cookies, M_TEMP); 6159 ap->a_cookies = NULL; 6160 *ap->a_ncookies = 0; 6161 } 6162 return (error); 6163 } 6164 if (ap->a_ncookies == NULL) 6165 return (0); 6166 6167 KASSERT(ap->a_cookies, 6168 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 6169 6170 *ap->a_cookies = realloc(*ap->a_cookies, 6171 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); 6172 (*ap->a_cookies)[*ap->a_ncookies] = off; 6173 *ap->a_ncookies += 1; 6174 return (0); 6175 } 6176 6177 /* 6178 * The purpose of this routine is to remove granularity from accmode_t, 6179 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 6180 * VADMIN and VAPPEND. 6181 * 6182 * If it returns 0, the caller is supposed to continue with the usual 6183 * access checks using 'accmode' as modified by this routine. If it 6184 * returns nonzero value, the caller is supposed to return that value 6185 * as errno. 6186 * 6187 * Note that after this routine runs, accmode may be zero. 6188 */ 6189 int 6190 vfs_unixify_accmode(accmode_t *accmode) 6191 { 6192 /* 6193 * There is no way to specify explicit "deny" rule using 6194 * file mode or POSIX.1e ACLs. 6195 */ 6196 if (*accmode & VEXPLICIT_DENY) { 6197 *accmode = 0; 6198 return (0); 6199 } 6200 6201 /* 6202 * None of these can be translated into usual access bits. 6203 * Also, the common case for NFSv4 ACLs is to not contain 6204 * either of these bits. Caller should check for VWRITE 6205 * on the containing directory instead. 6206 */ 6207 if (*accmode & (VDELETE_CHILD | VDELETE)) 6208 return (EPERM); 6209 6210 if (*accmode & VADMIN_PERMS) { 6211 *accmode &= ~VADMIN_PERMS; 6212 *accmode |= VADMIN; 6213 } 6214 6215 /* 6216 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 6217 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 6218 */ 6219 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 6220 6221 return (0); 6222 } 6223 6224 /* 6225 * Clear out a doomed vnode (if any) and replace it with a new one as long 6226 * as the fs is not being unmounted. Return the root vnode to the caller. 6227 */ 6228 static int __noinline 6229 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp) 6230 { 6231 struct vnode *vp; 6232 int error; 6233 6234 restart: 6235 if (mp->mnt_rootvnode != NULL) { 6236 MNT_ILOCK(mp); 6237 vp = mp->mnt_rootvnode; 6238 if (vp != NULL) { 6239 if (!VN_IS_DOOMED(vp)) { 6240 vrefact(vp); 6241 MNT_IUNLOCK(mp); 6242 error = vn_lock(vp, flags); 6243 if (error == 0) { 6244 *vpp = vp; 6245 return (0); 6246 } 6247 vrele(vp); 6248 goto restart; 6249 } 6250 /* 6251 * Clear the old one. 6252 */ 6253 mp->mnt_rootvnode = NULL; 6254 } 6255 MNT_IUNLOCK(mp); 6256 if (vp != NULL) { 6257 vfs_op_barrier_wait(mp); 6258 vrele(vp); 6259 } 6260 } 6261 error = VFS_CACHEDROOT(mp, flags, vpp); 6262 if (error != 0) 6263 return (error); 6264 if (mp->mnt_vfs_ops == 0) { 6265 MNT_ILOCK(mp); 6266 if (mp->mnt_vfs_ops != 0) { 6267 MNT_IUNLOCK(mp); 6268 return (0); 6269 } 6270 if (mp->mnt_rootvnode == NULL) { 6271 vrefact(*vpp); 6272 mp->mnt_rootvnode = *vpp; 6273 } else { 6274 if (mp->mnt_rootvnode != *vpp) { 6275 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) { 6276 panic("%s: mismatch between vnode returned " 6277 " by VFS_CACHEDROOT and the one cached " 6278 " (%p != %p)", 6279 __func__, *vpp, mp->mnt_rootvnode); 6280 } 6281 } 6282 } 6283 MNT_IUNLOCK(mp); 6284 } 6285 return (0); 6286 } 6287 6288 int 6289 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp) 6290 { 6291 struct vnode *vp; 6292 int error; 6293 6294 if (!vfs_op_thread_enter(mp)) 6295 return (vfs_cache_root_fallback(mp, flags, vpp)); 6296 vp = atomic_load_ptr(&mp->mnt_rootvnode); 6297 if (vp == NULL || VN_IS_DOOMED(vp)) { 6298 vfs_op_thread_exit(mp); 6299 return (vfs_cache_root_fallback(mp, flags, vpp)); 6300 } 6301 vrefact(vp); 6302 vfs_op_thread_exit(mp); 6303 error = vn_lock(vp, flags); 6304 if (error != 0) { 6305 vrele(vp); 6306 return (vfs_cache_root_fallback(mp, flags, vpp)); 6307 } 6308 *vpp = vp; 6309 return (0); 6310 } 6311 6312 struct vnode * 6313 vfs_cache_root_clear(struct mount *mp) 6314 { 6315 struct vnode *vp; 6316 6317 /* 6318 * ops > 0 guarantees there is nobody who can see this vnode 6319 */ 6320 MPASS(mp->mnt_vfs_ops > 0); 6321 vp = mp->mnt_rootvnode; 6322 if (vp != NULL) 6323 vn_seqc_write_begin(vp); 6324 mp->mnt_rootvnode = NULL; 6325 return (vp); 6326 } 6327 6328 void 6329 vfs_cache_root_set(struct mount *mp, struct vnode *vp) 6330 { 6331 6332 MPASS(mp->mnt_vfs_ops > 0); 6333 vrefact(vp); 6334 mp->mnt_rootvnode = vp; 6335 } 6336 6337 /* 6338 * These are helper functions for filesystems to traverse all 6339 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 6340 * 6341 * This interface replaces MNT_VNODE_FOREACH. 6342 */ 6343 6344 struct vnode * 6345 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 6346 { 6347 struct vnode *vp; 6348 6349 if (should_yield()) 6350 kern_yield(PRI_USER); 6351 MNT_ILOCK(mp); 6352 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6353 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL; 6354 vp = TAILQ_NEXT(vp, v_nmntvnodes)) { 6355 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 6356 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 6357 continue; 6358 VI_LOCK(vp); 6359 if (VN_IS_DOOMED(vp)) { 6360 VI_UNLOCK(vp); 6361 continue; 6362 } 6363 break; 6364 } 6365 if (vp == NULL) { 6366 __mnt_vnode_markerfree_all(mvp, mp); 6367 /* MNT_IUNLOCK(mp); -- done in above function */ 6368 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 6369 return (NULL); 6370 } 6371 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 6372 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 6373 MNT_IUNLOCK(mp); 6374 return (vp); 6375 } 6376 6377 struct vnode * 6378 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 6379 { 6380 struct vnode *vp; 6381 6382 *mvp = vn_alloc_marker(mp); 6383 MNT_ILOCK(mp); 6384 MNT_REF(mp); 6385 6386 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 6387 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 6388 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 6389 continue; 6390 VI_LOCK(vp); 6391 if (VN_IS_DOOMED(vp)) { 6392 VI_UNLOCK(vp); 6393 continue; 6394 } 6395 break; 6396 } 6397 if (vp == NULL) { 6398 MNT_REL(mp); 6399 MNT_IUNLOCK(mp); 6400 vn_free_marker(*mvp); 6401 *mvp = NULL; 6402 return (NULL); 6403 } 6404 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 6405 MNT_IUNLOCK(mp); 6406 return (vp); 6407 } 6408 6409 void 6410 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 6411 { 6412 6413 if (*mvp == NULL) { 6414 MNT_IUNLOCK(mp); 6415 return; 6416 } 6417 6418 mtx_assert(MNT_MTX(mp), MA_OWNED); 6419 6420 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6421 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 6422 MNT_REL(mp); 6423 MNT_IUNLOCK(mp); 6424 vn_free_marker(*mvp); 6425 *mvp = NULL; 6426 } 6427 6428 /* 6429 * These are helper functions for filesystems to traverse their 6430 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h 6431 */ 6432 static void 6433 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 6434 { 6435 6436 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6437 6438 MNT_ILOCK(mp); 6439 MNT_REL(mp); 6440 MNT_IUNLOCK(mp); 6441 vn_free_marker(*mvp); 6442 *mvp = NULL; 6443 } 6444 6445 /* 6446 * Relock the mp mount vnode list lock with the vp vnode interlock in the 6447 * conventional lock order during mnt_vnode_next_lazy iteration. 6448 * 6449 * On entry, the mount vnode list lock is held and the vnode interlock is not. 6450 * The list lock is dropped and reacquired. On success, both locks are held. 6451 * On failure, the mount vnode list lock is held but the vnode interlock is 6452 * not, and the procedure may have yielded. 6453 */ 6454 static bool 6455 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp, 6456 struct vnode *vp) 6457 { 6458 6459 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER && 6460 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp, 6461 ("%s: bad marker", __func__)); 6462 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp, 6463 ("%s: inappropriate vnode", __func__)); 6464 ASSERT_VI_UNLOCKED(vp, __func__); 6465 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 6466 6467 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist); 6468 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist); 6469 6470 /* 6471 * Note we may be racing against vdrop which transitioned the hold 6472 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine, 6473 * if we are the only user after we get the interlock we will just 6474 * vdrop. 6475 */ 6476 vhold(vp); 6477 mtx_unlock(&mp->mnt_listmtx); 6478 VI_LOCK(vp); 6479 if (VN_IS_DOOMED(vp)) { 6480 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); 6481 goto out_lost; 6482 } 6483 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 6484 /* 6485 * There is nothing to do if we are the last user. 6486 */ 6487 if (!refcount_release_if_not_last(&vp->v_holdcnt)) 6488 goto out_lost; 6489 mtx_lock(&mp->mnt_listmtx); 6490 return (true); 6491 out_lost: 6492 vdropl(vp); 6493 maybe_yield(); 6494 mtx_lock(&mp->mnt_listmtx); 6495 return (false); 6496 } 6497 6498 static struct vnode * 6499 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 6500 void *cbarg) 6501 { 6502 struct vnode *vp; 6503 6504 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 6505 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6506 restart: 6507 vp = TAILQ_NEXT(*mvp, v_lazylist); 6508 while (vp != NULL) { 6509 if (vp->v_type == VMARKER) { 6510 vp = TAILQ_NEXT(vp, v_lazylist); 6511 continue; 6512 } 6513 /* 6514 * See if we want to process the vnode. Note we may encounter a 6515 * long string of vnodes we don't care about and hog the list 6516 * as a result. Check for it and requeue the marker. 6517 */ 6518 VNPASS(!VN_IS_DOOMED(vp), vp); 6519 if (!cb(vp, cbarg)) { 6520 if (!should_yield()) { 6521 vp = TAILQ_NEXT(vp, v_lazylist); 6522 continue; 6523 } 6524 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, 6525 v_lazylist); 6526 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, 6527 v_lazylist); 6528 mtx_unlock(&mp->mnt_listmtx); 6529 kern_yield(PRI_USER); 6530 mtx_lock(&mp->mnt_listmtx); 6531 goto restart; 6532 } 6533 /* 6534 * Try-lock because this is the wrong lock order. 6535 */ 6536 if (!VI_TRYLOCK(vp) && 6537 !mnt_vnode_next_lazy_relock(*mvp, mp, vp)) 6538 goto restart; 6539 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 6540 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 6541 ("alien vnode on the lazy list %p %p", vp, mp)); 6542 VNPASS(vp->v_mount == mp, vp); 6543 VNPASS(!VN_IS_DOOMED(vp), vp); 6544 break; 6545 } 6546 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 6547 6548 /* Check if we are done */ 6549 if (vp == NULL) { 6550 mtx_unlock(&mp->mnt_listmtx); 6551 mnt_vnode_markerfree_lazy(mvp, mp); 6552 return (NULL); 6553 } 6554 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist); 6555 mtx_unlock(&mp->mnt_listmtx); 6556 ASSERT_VI_LOCKED(vp, "lazy iter"); 6557 return (vp); 6558 } 6559 6560 struct vnode * 6561 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 6562 void *cbarg) 6563 { 6564 6565 if (should_yield()) 6566 kern_yield(PRI_USER); 6567 mtx_lock(&mp->mnt_listmtx); 6568 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 6569 } 6570 6571 struct vnode * 6572 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 6573 void *cbarg) 6574 { 6575 struct vnode *vp; 6576 6577 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist)) 6578 return (NULL); 6579 6580 *mvp = vn_alloc_marker(mp); 6581 MNT_ILOCK(mp); 6582 MNT_REF(mp); 6583 MNT_IUNLOCK(mp); 6584 6585 mtx_lock(&mp->mnt_listmtx); 6586 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist); 6587 if (vp == NULL) { 6588 mtx_unlock(&mp->mnt_listmtx); 6589 mnt_vnode_markerfree_lazy(mvp, mp); 6590 return (NULL); 6591 } 6592 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist); 6593 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 6594 } 6595 6596 void 6597 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 6598 { 6599 6600 if (*mvp == NULL) 6601 return; 6602 6603 mtx_lock(&mp->mnt_listmtx); 6604 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 6605 mtx_unlock(&mp->mnt_listmtx); 6606 mnt_vnode_markerfree_lazy(mvp, mp); 6607 } 6608 6609 int 6610 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp) 6611 { 6612 6613 if ((cnp->cn_flags & NOEXECCHECK) != 0) { 6614 cnp->cn_flags &= ~NOEXECCHECK; 6615 return (0); 6616 } 6617 6618 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread)); 6619 } 6620 6621 /* 6622 * Do not use this variant unless you have means other than the hold count 6623 * to prevent the vnode from getting freed. 6624 */ 6625 void 6626 vn_seqc_write_begin_unheld_locked(struct vnode *vp) 6627 { 6628 6629 ASSERT_VI_LOCKED(vp, __func__); 6630 VNPASS(vp->v_seqc_users >= 0, vp); 6631 vp->v_seqc_users++; 6632 if (vp->v_seqc_users == 1) 6633 seqc_sleepable_write_begin(&vp->v_seqc); 6634 } 6635 6636 void 6637 vn_seqc_write_begin_locked(struct vnode *vp) 6638 { 6639 6640 ASSERT_VI_LOCKED(vp, __func__); 6641 VNPASS(vp->v_holdcnt > 0, vp); 6642 vn_seqc_write_begin_unheld_locked(vp); 6643 } 6644 6645 void 6646 vn_seqc_write_begin(struct vnode *vp) 6647 { 6648 6649 VI_LOCK(vp); 6650 vn_seqc_write_begin_locked(vp); 6651 VI_UNLOCK(vp); 6652 } 6653 6654 void 6655 vn_seqc_write_begin_unheld(struct vnode *vp) 6656 { 6657 6658 VI_LOCK(vp); 6659 vn_seqc_write_begin_unheld_locked(vp); 6660 VI_UNLOCK(vp); 6661 } 6662 6663 void 6664 vn_seqc_write_end_locked(struct vnode *vp) 6665 { 6666 6667 ASSERT_VI_LOCKED(vp, __func__); 6668 VNPASS(vp->v_seqc_users > 0, vp); 6669 vp->v_seqc_users--; 6670 if (vp->v_seqc_users == 0) 6671 seqc_sleepable_write_end(&vp->v_seqc); 6672 } 6673 6674 void 6675 vn_seqc_write_end(struct vnode *vp) 6676 { 6677 6678 VI_LOCK(vp); 6679 vn_seqc_write_end_locked(vp); 6680 VI_UNLOCK(vp); 6681 } 6682