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 /* 2924 * Increase the reference (use) and hold count of a vnode. 2925 * This will also remove the vnode from the free list if it is presently free. 2926 */ 2927 void 2928 vref(struct vnode *vp) 2929 { 2930 int old; 2931 2932 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2933 if (refcount_acquire_if_not_zero(&vp->v_usecount)) { 2934 VNODE_REFCOUNT_FENCE_ACQ(); 2935 VNASSERT(vp->v_holdcnt > 0, vp, 2936 ("%s: active vnode not held", __func__)); 2937 return; 2938 } 2939 vhold(vp); 2940 /* 2941 * See the comment in vget_finish. 2942 */ 2943 old = atomic_fetchadd_int(&vp->v_usecount, 1); 2944 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old)); 2945 if (old != 0) { 2946 #ifdef INVARIANTS 2947 old = atomic_fetchadd_int(&vp->v_holdcnt, -1); 2948 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); 2949 #else 2950 refcount_release(&vp->v_holdcnt); 2951 #endif 2952 } 2953 } 2954 2955 void 2956 vrefl(struct vnode *vp) 2957 { 2958 2959 ASSERT_VI_LOCKED(vp, __func__); 2960 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2961 vref(vp); 2962 } 2963 2964 void 2965 vrefact(struct vnode *vp) 2966 { 2967 2968 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2969 #ifdef INVARIANTS 2970 int old = atomic_fetchadd_int(&vp->v_usecount, 1); 2971 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); 2972 #else 2973 refcount_acquire(&vp->v_usecount); 2974 #endif 2975 } 2976 2977 void 2978 vrefactn(struct vnode *vp, u_int n) 2979 { 2980 2981 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2982 #ifdef INVARIANTS 2983 int old = atomic_fetchadd_int(&vp->v_usecount, n); 2984 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); 2985 #else 2986 atomic_add_int(&vp->v_usecount, n); 2987 #endif 2988 } 2989 2990 /* 2991 * Return reference count of a vnode. 2992 * 2993 * The results of this call are only guaranteed when some mechanism is used to 2994 * stop other processes from gaining references to the vnode. This may be the 2995 * case if the caller holds the only reference. This is also useful when stale 2996 * data is acceptable as race conditions may be accounted for by some other 2997 * means. 2998 */ 2999 int 3000 vrefcnt(struct vnode *vp) 3001 { 3002 3003 return (vp->v_usecount); 3004 } 3005 3006 void 3007 vlazy(struct vnode *vp) 3008 { 3009 struct mount *mp; 3010 3011 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__)); 3012 3013 if ((vp->v_mflag & VMP_LAZYLIST) != 0) 3014 return; 3015 /* 3016 * We may get here for inactive routines after the vnode got doomed. 3017 */ 3018 if (VN_IS_DOOMED(vp)) 3019 return; 3020 mp = vp->v_mount; 3021 mtx_lock(&mp->mnt_listmtx); 3022 if ((vp->v_mflag & VMP_LAZYLIST) == 0) { 3023 vp->v_mflag |= VMP_LAZYLIST; 3024 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3025 mp->mnt_lazyvnodelistsize++; 3026 } 3027 mtx_unlock(&mp->mnt_listmtx); 3028 } 3029 3030 /* 3031 * This routine is only meant to be called from vgonel prior to dooming 3032 * the vnode. 3033 */ 3034 static void 3035 vunlazy_gone(struct vnode *vp) 3036 { 3037 struct mount *mp; 3038 3039 ASSERT_VOP_ELOCKED(vp, __func__); 3040 ASSERT_VI_LOCKED(vp, __func__); 3041 VNPASS(!VN_IS_DOOMED(vp), vp); 3042 3043 if (vp->v_mflag & VMP_LAZYLIST) { 3044 mp = vp->v_mount; 3045 mtx_lock(&mp->mnt_listmtx); 3046 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 3047 vp->v_mflag &= ~VMP_LAZYLIST; 3048 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3049 mp->mnt_lazyvnodelistsize--; 3050 mtx_unlock(&mp->mnt_listmtx); 3051 } 3052 } 3053 3054 static void 3055 vdefer_inactive(struct vnode *vp) 3056 { 3057 3058 ASSERT_VI_LOCKED(vp, __func__); 3059 VNASSERT(vp->v_holdcnt > 0, vp, 3060 ("%s: vnode without hold count", __func__)); 3061 if (VN_IS_DOOMED(vp)) { 3062 vdropl(vp); 3063 return; 3064 } 3065 if (vp->v_iflag & VI_DEFINACT) { 3066 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 3067 vdropl(vp); 3068 return; 3069 } 3070 if (vp->v_usecount > 0) { 3071 vp->v_iflag &= ~VI_OWEINACT; 3072 vdropl(vp); 3073 return; 3074 } 3075 vlazy(vp); 3076 vp->v_iflag |= VI_DEFINACT; 3077 VI_UNLOCK(vp); 3078 counter_u64_add(deferred_inact, 1); 3079 } 3080 3081 static void 3082 vdefer_inactive_unlocked(struct vnode *vp) 3083 { 3084 3085 VI_LOCK(vp); 3086 if ((vp->v_iflag & VI_OWEINACT) == 0) { 3087 vdropl(vp); 3088 return; 3089 } 3090 vdefer_inactive(vp); 3091 } 3092 3093 enum vput_op { VRELE, VPUT, VUNREF }; 3094 3095 /* 3096 * Handle ->v_usecount transitioning to 0. 3097 * 3098 * By releasing the last usecount we take ownership of the hold count which 3099 * provides liveness of the vnode, meaning we have to vdrop. 3100 * 3101 * For all vnodes we may need to perform inactive processing. It requires an 3102 * exclusive lock on the vnode, while it is legal to call here with only a 3103 * shared lock (or no locks). If locking the vnode in an expected manner fails, 3104 * inactive processing gets deferred to the syncer. 3105 * 3106 * XXX Some filesystems pass in an exclusively locked vnode and strongly depend 3107 * on the lock being held all the way until VOP_INACTIVE. This in particular 3108 * happens with UFS which adds half-constructed vnodes to the hash, where they 3109 * can be found by other code. 3110 */ 3111 static void 3112 vput_final(struct vnode *vp, enum vput_op func) 3113 { 3114 int error; 3115 bool want_unlock; 3116 3117 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3118 VNPASS(vp->v_holdcnt > 0, vp); 3119 3120 VI_LOCK(vp); 3121 3122 /* 3123 * By the time we got here someone else might have transitioned 3124 * the count back to > 0. 3125 */ 3126 if (vp->v_usecount > 0) 3127 goto out; 3128 3129 /* 3130 * If the vnode is doomed vgone already performed inactive processing 3131 * (if needed). 3132 */ 3133 if (VN_IS_DOOMED(vp)) 3134 goto out; 3135 3136 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0)) 3137 goto out; 3138 3139 if (vp->v_iflag & VI_DOINGINACT) 3140 goto out; 3141 3142 /* 3143 * Locking operations here will drop the interlock and possibly the 3144 * vnode lock, opening a window where the vnode can get doomed all the 3145 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to 3146 * perform inactive. 3147 */ 3148 vp->v_iflag |= VI_OWEINACT; 3149 want_unlock = false; 3150 error = 0; 3151 switch (func) { 3152 case VRELE: 3153 switch (VOP_ISLOCKED(vp)) { 3154 case LK_EXCLUSIVE: 3155 break; 3156 case LK_EXCLOTHER: 3157 case 0: 3158 want_unlock = true; 3159 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 3160 VI_LOCK(vp); 3161 break; 3162 default: 3163 /* 3164 * The lock has at least one sharer, but we have no way 3165 * to conclude whether this is us. Play it safe and 3166 * defer processing. 3167 */ 3168 error = EAGAIN; 3169 break; 3170 } 3171 break; 3172 case VPUT: 3173 want_unlock = true; 3174 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3175 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | 3176 LK_NOWAIT); 3177 VI_LOCK(vp); 3178 } 3179 break; 3180 case VUNREF: 3181 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 3182 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); 3183 VI_LOCK(vp); 3184 } 3185 break; 3186 } 3187 if (error == 0) { 3188 vinactive(vp); 3189 if (want_unlock) 3190 VOP_UNLOCK(vp); 3191 vdropl(vp); 3192 } else { 3193 vdefer_inactive(vp); 3194 } 3195 return; 3196 out: 3197 if (func == VPUT) 3198 VOP_UNLOCK(vp); 3199 vdropl(vp); 3200 } 3201 3202 /* 3203 * Decrement ->v_usecount for a vnode. 3204 * 3205 * Releasing the last use count requires additional processing, see vput_final 3206 * above for details. 3207 * 3208 * Comment above each variant denotes lock state on entry and exit. 3209 */ 3210 3211 /* 3212 * in: any 3213 * out: same as passed in 3214 */ 3215 void 3216 vrele(struct vnode *vp) 3217 { 3218 3219 ASSERT_VI_UNLOCKED(vp, __func__); 3220 if (!refcount_release(&vp->v_usecount)) 3221 return; 3222 vput_final(vp, VRELE); 3223 } 3224 3225 /* 3226 * in: locked 3227 * out: unlocked 3228 */ 3229 void 3230 vput(struct vnode *vp) 3231 { 3232 3233 ASSERT_VOP_LOCKED(vp, __func__); 3234 ASSERT_VI_UNLOCKED(vp, __func__); 3235 if (!refcount_release(&vp->v_usecount)) { 3236 VOP_UNLOCK(vp); 3237 return; 3238 } 3239 vput_final(vp, VPUT); 3240 } 3241 3242 /* 3243 * in: locked 3244 * out: locked 3245 */ 3246 void 3247 vunref(struct vnode *vp) 3248 { 3249 3250 ASSERT_VOP_LOCKED(vp, __func__); 3251 ASSERT_VI_UNLOCKED(vp, __func__); 3252 if (!refcount_release(&vp->v_usecount)) 3253 return; 3254 vput_final(vp, VUNREF); 3255 } 3256 3257 void 3258 vhold(struct vnode *vp) 3259 { 3260 struct vdbatch *vd; 3261 int old; 3262 3263 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3264 old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3265 VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3266 ("%s: wrong hold count %d", __func__, old)); 3267 if (old != 0) 3268 return; 3269 critical_enter(); 3270 vd = DPCPU_PTR(vd); 3271 vd->freevnodes--; 3272 critical_exit(); 3273 } 3274 3275 void 3276 vholdl(struct vnode *vp) 3277 { 3278 3279 ASSERT_VI_LOCKED(vp, __func__); 3280 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3281 vhold(vp); 3282 } 3283 3284 void 3285 vholdnz(struct vnode *vp) 3286 { 3287 3288 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3289 #ifdef INVARIANTS 3290 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3291 VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, 3292 ("%s: wrong hold count %d", __func__, old)); 3293 #else 3294 atomic_add_int(&vp->v_holdcnt, 1); 3295 #endif 3296 } 3297 3298 /* 3299 * Grab a hold count unless the vnode is freed. 3300 * 3301 * Only use this routine if vfs smr is the only protection you have against 3302 * freeing the vnode. 3303 * 3304 * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag 3305 * is not set. After the flag is set the vnode becomes immutable to anyone but 3306 * the thread which managed to set the flag. 3307 * 3308 * It may be tempting to replace the loop with: 3309 * count = atomic_fetchadd_int(&vp->v_holdcnt, 1); 3310 * if (count & VHOLD_NO_SMR) { 3311 * backpedal and error out; 3312 * } 3313 * 3314 * However, while this is more performant, it hinders debugging by eliminating 3315 * the previously mentioned invariant. 3316 */ 3317 bool 3318 vhold_smr(struct vnode *vp) 3319 { 3320 int count; 3321 3322 VFS_SMR_ASSERT_ENTERED(); 3323 3324 count = atomic_load_int(&vp->v_holdcnt); 3325 for (;;) { 3326 if (count & VHOLD_NO_SMR) { 3327 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, 3328 ("non-zero hold count with flags %d\n", count)); 3329 return (false); 3330 } 3331 3332 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); 3333 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) 3334 return (true); 3335 } 3336 } 3337 3338 static void __noinline 3339 vdbatch_process(struct vdbatch *vd) 3340 { 3341 struct vnode *vp; 3342 int i; 3343 3344 mtx_assert(&vd->lock, MA_OWNED); 3345 MPASS(curthread->td_pinned > 0); 3346 MPASS(vd->index == VDBATCH_SIZE); 3347 3348 mtx_lock(&vnode_list_mtx); 3349 critical_enter(); 3350 freevnodes += vd->freevnodes; 3351 for (i = 0; i < VDBATCH_SIZE; i++) { 3352 vp = vd->tab[i]; 3353 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); 3354 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist); 3355 MPASS(vp->v_dbatchcpu != NOCPU); 3356 vp->v_dbatchcpu = NOCPU; 3357 } 3358 mtx_unlock(&vnode_list_mtx); 3359 vd->freevnodes = 0; 3360 bzero(vd->tab, sizeof(vd->tab)); 3361 vd->index = 0; 3362 critical_exit(); 3363 } 3364 3365 static void 3366 vdbatch_enqueue(struct vnode *vp) 3367 { 3368 struct vdbatch *vd; 3369 3370 ASSERT_VI_LOCKED(vp, __func__); 3371 VNASSERT(!VN_IS_DOOMED(vp), vp, 3372 ("%s: deferring requeue of a doomed vnode", __func__)); 3373 3374 critical_enter(); 3375 vd = DPCPU_PTR(vd); 3376 vd->freevnodes++; 3377 if (vp->v_dbatchcpu != NOCPU) { 3378 VI_UNLOCK(vp); 3379 critical_exit(); 3380 return; 3381 } 3382 3383 sched_pin(); 3384 critical_exit(); 3385 mtx_lock(&vd->lock); 3386 MPASS(vd->index < VDBATCH_SIZE); 3387 MPASS(vd->tab[vd->index] == NULL); 3388 /* 3389 * A hack: we depend on being pinned so that we know what to put in 3390 * ->v_dbatchcpu. 3391 */ 3392 vp->v_dbatchcpu = curcpu; 3393 vd->tab[vd->index] = vp; 3394 vd->index++; 3395 VI_UNLOCK(vp); 3396 if (vd->index == VDBATCH_SIZE) 3397 vdbatch_process(vd); 3398 mtx_unlock(&vd->lock); 3399 sched_unpin(); 3400 } 3401 3402 /* 3403 * This routine must only be called for vnodes which are about to be 3404 * deallocated. Supporting dequeue for arbitrary vndoes would require 3405 * validating that the locked batch matches. 3406 */ 3407 static void 3408 vdbatch_dequeue(struct vnode *vp) 3409 { 3410 struct vdbatch *vd; 3411 int i; 3412 short cpu; 3413 3414 VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp, 3415 ("%s: called for a used vnode\n", __func__)); 3416 3417 cpu = vp->v_dbatchcpu; 3418 if (cpu == NOCPU) 3419 return; 3420 3421 vd = DPCPU_ID_PTR(cpu, vd); 3422 mtx_lock(&vd->lock); 3423 for (i = 0; i < vd->index; i++) { 3424 if (vd->tab[i] != vp) 3425 continue; 3426 vp->v_dbatchcpu = NOCPU; 3427 vd->index--; 3428 vd->tab[i] = vd->tab[vd->index]; 3429 vd->tab[vd->index] = NULL; 3430 break; 3431 } 3432 mtx_unlock(&vd->lock); 3433 /* 3434 * Either we dequeued the vnode above or the target CPU beat us to it. 3435 */ 3436 MPASS(vp->v_dbatchcpu == NOCPU); 3437 } 3438 3439 /* 3440 * Drop the hold count of the vnode. If this is the last reference to 3441 * the vnode we place it on the free list unless it has been vgone'd 3442 * (marked VIRF_DOOMED) in which case we will free it. 3443 * 3444 * Because the vnode vm object keeps a hold reference on the vnode if 3445 * there is at least one resident non-cached page, the vnode cannot 3446 * leave the active list without the page cleanup done. 3447 */ 3448 static void 3449 vdrop_deactivate(struct vnode *vp) 3450 { 3451 struct mount *mp; 3452 3453 ASSERT_VI_LOCKED(vp, __func__); 3454 /* 3455 * Mark a vnode as free: remove it from its active list 3456 * and put it up for recycling on the freelist. 3457 */ 3458 VNASSERT(!VN_IS_DOOMED(vp), vp, 3459 ("vdrop: returning doomed vnode")); 3460 VNASSERT(vp->v_op != NULL, vp, 3461 ("vdrop: vnode already reclaimed.")); 3462 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 3463 ("vnode with VI_OWEINACT set")); 3464 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, 3465 ("vnode with VI_DEFINACT set")); 3466 if (vp->v_mflag & VMP_LAZYLIST) { 3467 mp = vp->v_mount; 3468 mtx_lock(&mp->mnt_listmtx); 3469 VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST")); 3470 /* 3471 * Don't remove the vnode from the lazy list if another thread 3472 * has increased the hold count. It may have re-enqueued the 3473 * vnode to the lazy list and is now responsible for its 3474 * removal. 3475 */ 3476 if (vp->v_holdcnt == 0) { 3477 vp->v_mflag &= ~VMP_LAZYLIST; 3478 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); 3479 mp->mnt_lazyvnodelistsize--; 3480 } 3481 mtx_unlock(&mp->mnt_listmtx); 3482 } 3483 vdbatch_enqueue(vp); 3484 } 3485 3486 void 3487 vdrop(struct vnode *vp) 3488 { 3489 3490 ASSERT_VI_UNLOCKED(vp, __func__); 3491 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3492 if (refcount_release_if_not_last(&vp->v_holdcnt)) 3493 return; 3494 VI_LOCK(vp); 3495 vdropl(vp); 3496 } 3497 3498 void 3499 vdropl(struct vnode *vp) 3500 { 3501 3502 ASSERT_VI_LOCKED(vp, __func__); 3503 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3504 if (!refcount_release(&vp->v_holdcnt)) { 3505 VI_UNLOCK(vp); 3506 return; 3507 } 3508 if (!VN_IS_DOOMED(vp)) { 3509 vdrop_deactivate(vp); 3510 /* 3511 * Also unlocks the interlock. We can't assert on it as we 3512 * released our hold and by now the vnode might have been 3513 * freed. 3514 */ 3515 return; 3516 } 3517 /* 3518 * Set the VHOLD_NO_SMR flag. 3519 * 3520 * We may be racing against vhold_smr. If they win we can just pretend 3521 * we never got this far, they will vdrop later. 3522 */ 3523 if (!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR)) { 3524 VI_UNLOCK(vp); 3525 /* 3526 * We lost the aforementioned race. Any subsequent access is 3527 * invalid as they might have managed to vdropl on their own. 3528 */ 3529 return; 3530 } 3531 freevnode(vp); 3532 } 3533 3534 /* 3535 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 3536 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 3537 */ 3538 static void 3539 vinactivef(struct vnode *vp) 3540 { 3541 struct vm_object *obj; 3542 3543 ASSERT_VOP_ELOCKED(vp, "vinactive"); 3544 ASSERT_VI_LOCKED(vp, "vinactive"); 3545 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, 3546 ("vinactive: recursed on VI_DOINGINACT")); 3547 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3548 vp->v_iflag |= VI_DOINGINACT; 3549 vp->v_iflag &= ~VI_OWEINACT; 3550 VI_UNLOCK(vp); 3551 /* 3552 * Before moving off the active list, we must be sure that any 3553 * modified pages are converted into the vnode's dirty 3554 * buffers, since these will no longer be checked once the 3555 * vnode is on the inactive list. 3556 * 3557 * The write-out of the dirty pages is asynchronous. At the 3558 * point that VOP_INACTIVE() is called, there could still be 3559 * pending I/O and dirty pages in the object. 3560 */ 3561 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 3562 vm_object_mightbedirty(obj)) { 3563 VM_OBJECT_WLOCK(obj); 3564 vm_object_page_clean(obj, 0, 0, 0); 3565 VM_OBJECT_WUNLOCK(obj); 3566 } 3567 VOP_INACTIVE(vp, curthread); 3568 VI_LOCK(vp); 3569 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, 3570 ("vinactive: lost VI_DOINGINACT")); 3571 vp->v_iflag &= ~VI_DOINGINACT; 3572 } 3573 3574 void 3575 vinactive(struct vnode *vp) 3576 { 3577 3578 ASSERT_VOP_ELOCKED(vp, "vinactive"); 3579 ASSERT_VI_LOCKED(vp, "vinactive"); 3580 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3581 3582 if ((vp->v_iflag & VI_OWEINACT) == 0) 3583 return; 3584 if (vp->v_iflag & VI_DOINGINACT) 3585 return; 3586 if (vp->v_usecount > 0) { 3587 vp->v_iflag &= ~VI_OWEINACT; 3588 return; 3589 } 3590 vinactivef(vp); 3591 } 3592 3593 /* 3594 * Remove any vnodes in the vnode table belonging to mount point mp. 3595 * 3596 * If FORCECLOSE is not specified, there should not be any active ones, 3597 * return error if any are found (nb: this is a user error, not a 3598 * system error). If FORCECLOSE is specified, detach any active vnodes 3599 * that are found. 3600 * 3601 * If WRITECLOSE is set, only flush out regular file vnodes open for 3602 * writing. 3603 * 3604 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 3605 * 3606 * `rootrefs' specifies the base reference count for the root vnode 3607 * of this filesystem. The root vnode is considered busy if its 3608 * v_usecount exceeds this value. On a successful return, vflush(, td) 3609 * will call vrele() on the root vnode exactly rootrefs times. 3610 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 3611 * be zero. 3612 */ 3613 #ifdef DIAGNOSTIC 3614 static int busyprt = 0; /* print out busy vnodes */ 3615 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 3616 #endif 3617 3618 int 3619 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 3620 { 3621 struct vnode *vp, *mvp, *rootvp = NULL; 3622 struct vattr vattr; 3623 int busy = 0, error; 3624 3625 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 3626 rootrefs, flags); 3627 if (rootrefs > 0) { 3628 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 3629 ("vflush: bad args")); 3630 /* 3631 * Get the filesystem root vnode. We can vput() it 3632 * immediately, since with rootrefs > 0, it won't go away. 3633 */ 3634 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 3635 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 3636 __func__, error); 3637 return (error); 3638 } 3639 vput(rootvp); 3640 } 3641 loop: 3642 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 3643 vholdl(vp); 3644 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 3645 if (error) { 3646 vdrop(vp); 3647 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3648 goto loop; 3649 } 3650 /* 3651 * Skip over a vnodes marked VV_SYSTEM. 3652 */ 3653 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 3654 VOP_UNLOCK(vp); 3655 vdrop(vp); 3656 continue; 3657 } 3658 /* 3659 * If WRITECLOSE is set, flush out unlinked but still open 3660 * files (even if open only for reading) and regular file 3661 * vnodes open for writing. 3662 */ 3663 if (flags & WRITECLOSE) { 3664 if (vp->v_object != NULL) { 3665 VM_OBJECT_WLOCK(vp->v_object); 3666 vm_object_page_clean(vp->v_object, 0, 0, 0); 3667 VM_OBJECT_WUNLOCK(vp->v_object); 3668 } 3669 error = VOP_FSYNC(vp, MNT_WAIT, td); 3670 if (error != 0) { 3671 VOP_UNLOCK(vp); 3672 vdrop(vp); 3673 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 3674 return (error); 3675 } 3676 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 3677 VI_LOCK(vp); 3678 3679 if ((vp->v_type == VNON || 3680 (error == 0 && vattr.va_nlink > 0)) && 3681 (vp->v_writecount <= 0 || vp->v_type != VREG)) { 3682 VOP_UNLOCK(vp); 3683 vdropl(vp); 3684 continue; 3685 } 3686 } else 3687 VI_LOCK(vp); 3688 /* 3689 * With v_usecount == 0, all we need to do is clear out the 3690 * vnode data structures and we are done. 3691 * 3692 * If FORCECLOSE is set, forcibly close the vnode. 3693 */ 3694 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 3695 vgonel(vp); 3696 } else { 3697 busy++; 3698 #ifdef DIAGNOSTIC 3699 if (busyprt) 3700 vn_printf(vp, "vflush: busy vnode "); 3701 #endif 3702 } 3703 VOP_UNLOCK(vp); 3704 vdropl(vp); 3705 } 3706 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 3707 /* 3708 * If just the root vnode is busy, and if its refcount 3709 * is equal to `rootrefs', then go ahead and kill it. 3710 */ 3711 VI_LOCK(rootvp); 3712 KASSERT(busy > 0, ("vflush: not busy")); 3713 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 3714 ("vflush: usecount %d < rootrefs %d", 3715 rootvp->v_usecount, rootrefs)); 3716 if (busy == 1 && rootvp->v_usecount == rootrefs) { 3717 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 3718 vgone(rootvp); 3719 VOP_UNLOCK(rootvp); 3720 busy = 0; 3721 } else 3722 VI_UNLOCK(rootvp); 3723 } 3724 if (busy) { 3725 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 3726 busy); 3727 return (EBUSY); 3728 } 3729 for (; rootrefs > 0; rootrefs--) 3730 vrele(rootvp); 3731 return (0); 3732 } 3733 3734 /* 3735 * Recycle an unused vnode to the front of the free list. 3736 */ 3737 int 3738 vrecycle(struct vnode *vp) 3739 { 3740 int recycled; 3741 3742 VI_LOCK(vp); 3743 recycled = vrecyclel(vp); 3744 VI_UNLOCK(vp); 3745 return (recycled); 3746 } 3747 3748 /* 3749 * vrecycle, with the vp interlock held. 3750 */ 3751 int 3752 vrecyclel(struct vnode *vp) 3753 { 3754 int recycled; 3755 3756 ASSERT_VOP_ELOCKED(vp, __func__); 3757 ASSERT_VI_LOCKED(vp, __func__); 3758 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3759 recycled = 0; 3760 if (vp->v_usecount == 0) { 3761 recycled = 1; 3762 vgonel(vp); 3763 } 3764 return (recycled); 3765 } 3766 3767 /* 3768 * Eliminate all activity associated with a vnode 3769 * in preparation for reuse. 3770 */ 3771 void 3772 vgone(struct vnode *vp) 3773 { 3774 VI_LOCK(vp); 3775 vgonel(vp); 3776 VI_UNLOCK(vp); 3777 } 3778 3779 static void 3780 notify_lowervp_vfs_dummy(struct mount *mp __unused, 3781 struct vnode *lowervp __unused) 3782 { 3783 } 3784 3785 /* 3786 * Notify upper mounts about reclaimed or unlinked vnode. 3787 */ 3788 void 3789 vfs_notify_upper(struct vnode *vp, int event) 3790 { 3791 static struct vfsops vgonel_vfsops = { 3792 .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy, 3793 .vfs_unlink_lowervp = notify_lowervp_vfs_dummy, 3794 }; 3795 struct mount *mp, *ump, *mmp; 3796 3797 mp = vp->v_mount; 3798 if (mp == NULL) 3799 return; 3800 if (TAILQ_EMPTY(&mp->mnt_uppers)) 3801 return; 3802 3803 mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO); 3804 mmp->mnt_op = &vgonel_vfsops; 3805 mmp->mnt_kern_flag |= MNTK_MARKER; 3806 MNT_ILOCK(mp); 3807 mp->mnt_kern_flag |= MNTK_VGONE_UPPER; 3808 for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) { 3809 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) { 3810 ump = TAILQ_NEXT(ump, mnt_upper_link); 3811 continue; 3812 } 3813 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link); 3814 MNT_IUNLOCK(mp); 3815 switch (event) { 3816 case VFS_NOTIFY_UPPER_RECLAIM: 3817 VFS_RECLAIM_LOWERVP(ump, vp); 3818 break; 3819 case VFS_NOTIFY_UPPER_UNLINK: 3820 VFS_UNLINK_LOWERVP(ump, vp); 3821 break; 3822 default: 3823 KASSERT(0, ("invalid event %d", event)); 3824 break; 3825 } 3826 MNT_ILOCK(mp); 3827 ump = TAILQ_NEXT(mmp, mnt_upper_link); 3828 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link); 3829 } 3830 free(mmp, M_TEMP); 3831 mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER; 3832 if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) { 3833 mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER; 3834 wakeup(&mp->mnt_uppers); 3835 } 3836 MNT_IUNLOCK(mp); 3837 } 3838 3839 /* 3840 * vgone, with the vp interlock held. 3841 */ 3842 static void 3843 vgonel(struct vnode *vp) 3844 { 3845 struct thread *td; 3846 struct mount *mp; 3847 vm_object_t object; 3848 bool active, oweinact; 3849 3850 ASSERT_VOP_ELOCKED(vp, "vgonel"); 3851 ASSERT_VI_LOCKED(vp, "vgonel"); 3852 VNASSERT(vp->v_holdcnt, vp, 3853 ("vgonel: vp %p has no reference.", vp)); 3854 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 3855 td = curthread; 3856 3857 /* 3858 * Don't vgonel if we're already doomed. 3859 */ 3860 if (vp->v_irflag & VIRF_DOOMED) 3861 return; 3862 /* 3863 * Paired with freevnode. 3864 */ 3865 vn_seqc_write_begin_locked(vp); 3866 vunlazy_gone(vp); 3867 vp->v_irflag |= VIRF_DOOMED; 3868 3869 /* 3870 * Check to see if the vnode is in use. If so, we have to call 3871 * VOP_CLOSE() and VOP_INACTIVE(). 3872 */ 3873 active = vp->v_usecount > 0; 3874 oweinact = (vp->v_iflag & VI_OWEINACT) != 0; 3875 /* 3876 * If we need to do inactive VI_OWEINACT will be set. 3877 */ 3878 if (vp->v_iflag & VI_DEFINACT) { 3879 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); 3880 vp->v_iflag &= ~VI_DEFINACT; 3881 vdropl(vp); 3882 } else { 3883 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); 3884 VI_UNLOCK(vp); 3885 } 3886 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); 3887 3888 /* 3889 * If purging an active vnode, it must be closed and 3890 * deactivated before being reclaimed. 3891 */ 3892 if (active) 3893 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 3894 if (oweinact || active) { 3895 VI_LOCK(vp); 3896 vinactivef(vp); 3897 VI_UNLOCK(vp); 3898 } 3899 if (vp->v_type == VSOCK) 3900 vfs_unp_reclaim(vp); 3901 3902 /* 3903 * Clean out any buffers associated with the vnode. 3904 * If the flush fails, just toss the buffers. 3905 */ 3906 mp = NULL; 3907 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 3908 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 3909 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { 3910 while (vinvalbuf(vp, 0, 0, 0) != 0) 3911 ; 3912 } 3913 3914 BO_LOCK(&vp->v_bufobj); 3915 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && 3916 vp->v_bufobj.bo_dirty.bv_cnt == 0 && 3917 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && 3918 vp->v_bufobj.bo_clean.bv_cnt == 0, 3919 ("vp %p bufobj not invalidated", vp)); 3920 3921 /* 3922 * For VMIO bufobj, BO_DEAD is set later, or in 3923 * vm_object_terminate() after the object's page queue is 3924 * flushed. 3925 */ 3926 object = vp->v_bufobj.bo_object; 3927 if (object == NULL) 3928 vp->v_bufobj.bo_flag |= BO_DEAD; 3929 BO_UNLOCK(&vp->v_bufobj); 3930 3931 /* 3932 * Handle the VM part. Tmpfs handles v_object on its own (the 3933 * OBJT_VNODE check). Nullfs or other bypassing filesystems 3934 * should not touch the object borrowed from the lower vnode 3935 * (the handle check). 3936 */ 3937 if (object != NULL && object->type == OBJT_VNODE && 3938 object->handle == vp) 3939 vnode_destroy_vobject(vp); 3940 3941 /* 3942 * Reclaim the vnode. 3943 */ 3944 if (VOP_RECLAIM(vp, td)) 3945 panic("vgone: cannot reclaim"); 3946 if (mp != NULL) 3947 vn_finished_secondary_write(mp); 3948 VNASSERT(vp->v_object == NULL, vp, 3949 ("vop_reclaim left v_object vp=%p", vp)); 3950 /* 3951 * Clear the advisory locks and wake up waiting threads. 3952 */ 3953 (void)VOP_ADVLOCKPURGE(vp); 3954 vp->v_lockf = NULL; 3955 /* 3956 * Delete from old mount point vnode list. 3957 */ 3958 delmntque(vp); 3959 cache_purge_vgone(vp); 3960 /* 3961 * Done with purge, reset to the standard lock and invalidate 3962 * the vnode. 3963 */ 3964 VI_LOCK(vp); 3965 vp->v_vnlock = &vp->v_lock; 3966 vp->v_op = &dead_vnodeops; 3967 vp->v_type = VBAD; 3968 } 3969 3970 /* 3971 * Print out a description of a vnode. 3972 */ 3973 static const char * const typename[] = 3974 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", 3975 "VMARKER"}; 3976 3977 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0, 3978 "new hold count flag not added to vn_printf"); 3979 3980 void 3981 vn_printf(struct vnode *vp, const char *fmt, ...) 3982 { 3983 va_list ap; 3984 char buf[256], buf2[16]; 3985 u_long flags; 3986 u_int holdcnt; 3987 3988 va_start(ap, fmt); 3989 vprintf(fmt, ap); 3990 va_end(ap); 3991 printf("%p: ", (void *)vp); 3992 printf("type %s\n", typename[vp->v_type]); 3993 holdcnt = atomic_load_int(&vp->v_holdcnt); 3994 printf(" usecount %d, writecount %d, refcount %d seqc users %d", 3995 vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS, 3996 vp->v_seqc_users); 3997 switch (vp->v_type) { 3998 case VDIR: 3999 printf(" mountedhere %p\n", vp->v_mountedhere); 4000 break; 4001 case VCHR: 4002 printf(" rdev %p\n", vp->v_rdev); 4003 break; 4004 case VSOCK: 4005 printf(" socket %p\n", vp->v_unpcb); 4006 break; 4007 case VFIFO: 4008 printf(" fifoinfo %p\n", vp->v_fifoinfo); 4009 break; 4010 default: 4011 printf("\n"); 4012 break; 4013 } 4014 buf[0] = '\0'; 4015 buf[1] = '\0'; 4016 if (holdcnt & VHOLD_NO_SMR) 4017 strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf)); 4018 printf(" hold count flags (%s)\n", buf + 1); 4019 4020 buf[0] = '\0'; 4021 buf[1] = '\0'; 4022 if (vp->v_irflag & VIRF_DOOMED) 4023 strlcat(buf, "|VIRF_DOOMED", sizeof(buf)); 4024 flags = vp->v_irflag & ~(VIRF_DOOMED); 4025 if (flags != 0) { 4026 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags); 4027 strlcat(buf, buf2, sizeof(buf)); 4028 } 4029 if (vp->v_vflag & VV_ROOT) 4030 strlcat(buf, "|VV_ROOT", sizeof(buf)); 4031 if (vp->v_vflag & VV_ISTTY) 4032 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 4033 if (vp->v_vflag & VV_NOSYNC) 4034 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 4035 if (vp->v_vflag & VV_ETERNALDEV) 4036 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 4037 if (vp->v_vflag & VV_CACHEDLABEL) 4038 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 4039 if (vp->v_vflag & VV_VMSIZEVNLOCK) 4040 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf)); 4041 if (vp->v_vflag & VV_COPYONWRITE) 4042 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 4043 if (vp->v_vflag & VV_SYSTEM) 4044 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 4045 if (vp->v_vflag & VV_PROCDEP) 4046 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 4047 if (vp->v_vflag & VV_NOKNOTE) 4048 strlcat(buf, "|VV_NOKNOTE", sizeof(buf)); 4049 if (vp->v_vflag & VV_DELETED) 4050 strlcat(buf, "|VV_DELETED", sizeof(buf)); 4051 if (vp->v_vflag & VV_MD) 4052 strlcat(buf, "|VV_MD", sizeof(buf)); 4053 if (vp->v_vflag & VV_FORCEINSMQ) 4054 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 4055 if (vp->v_vflag & VV_READLINK) 4056 strlcat(buf, "|VV_READLINK", sizeof(buf)); 4057 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 4058 VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | 4059 VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ); 4060 if (flags != 0) { 4061 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 4062 strlcat(buf, buf2, sizeof(buf)); 4063 } 4064 if (vp->v_iflag & VI_TEXT_REF) 4065 strlcat(buf, "|VI_TEXT_REF", sizeof(buf)); 4066 if (vp->v_iflag & VI_MOUNT) 4067 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 4068 if (vp->v_iflag & VI_DOINGINACT) 4069 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 4070 if (vp->v_iflag & VI_OWEINACT) 4071 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 4072 if (vp->v_iflag & VI_DEFINACT) 4073 strlcat(buf, "|VI_DEFINACT", sizeof(buf)); 4074 flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT | 4075 VI_OWEINACT | VI_DEFINACT); 4076 if (flags != 0) { 4077 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 4078 strlcat(buf, buf2, sizeof(buf)); 4079 } 4080 if (vp->v_mflag & VMP_LAZYLIST) 4081 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf)); 4082 flags = vp->v_mflag & ~(VMP_LAZYLIST); 4083 if (flags != 0) { 4084 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags); 4085 strlcat(buf, buf2, sizeof(buf)); 4086 } 4087 printf(" flags (%s)\n", buf + 1); 4088 if (mtx_owned(VI_MTX(vp))) 4089 printf(" VI_LOCKed"); 4090 if (vp->v_object != NULL) 4091 printf(" v_object %p ref %d pages %d " 4092 "cleanbuf %d dirtybuf %d\n", 4093 vp->v_object, vp->v_object->ref_count, 4094 vp->v_object->resident_page_count, 4095 vp->v_bufobj.bo_clean.bv_cnt, 4096 vp->v_bufobj.bo_dirty.bv_cnt); 4097 printf(" "); 4098 lockmgr_printinfo(vp->v_vnlock); 4099 if (vp->v_data != NULL) 4100 VOP_PRINT(vp); 4101 } 4102 4103 #ifdef DDB 4104 /* 4105 * List all of the locked vnodes in the system. 4106 * Called when debugging the kernel. 4107 */ 4108 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 4109 { 4110 struct mount *mp; 4111 struct vnode *vp; 4112 4113 /* 4114 * Note: because this is DDB, we can't obey the locking semantics 4115 * for these structures, which means we could catch an inconsistent 4116 * state and dereference a nasty pointer. Not much to be done 4117 * about that. 4118 */ 4119 db_printf("Locked vnodes\n"); 4120 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4121 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4122 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) 4123 vn_printf(vp, "vnode "); 4124 } 4125 } 4126 } 4127 4128 /* 4129 * Show details about the given vnode. 4130 */ 4131 DB_SHOW_COMMAND(vnode, db_show_vnode) 4132 { 4133 struct vnode *vp; 4134 4135 if (!have_addr) 4136 return; 4137 vp = (struct vnode *)addr; 4138 vn_printf(vp, "vnode "); 4139 } 4140 4141 /* 4142 * Show details about the given mount point. 4143 */ 4144 DB_SHOW_COMMAND(mount, db_show_mount) 4145 { 4146 struct mount *mp; 4147 struct vfsopt *opt; 4148 struct statfs *sp; 4149 struct vnode *vp; 4150 char buf[512]; 4151 uint64_t mflags; 4152 u_int flags; 4153 4154 if (!have_addr) { 4155 /* No address given, print short info about all mount points. */ 4156 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4157 db_printf("%p %s on %s (%s)\n", mp, 4158 mp->mnt_stat.f_mntfromname, 4159 mp->mnt_stat.f_mntonname, 4160 mp->mnt_stat.f_fstypename); 4161 if (db_pager_quit) 4162 break; 4163 } 4164 db_printf("\nMore info: show mount <addr>\n"); 4165 return; 4166 } 4167 4168 mp = (struct mount *)addr; 4169 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 4170 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 4171 4172 buf[0] = '\0'; 4173 mflags = mp->mnt_flag; 4174 #define MNT_FLAG(flag) do { \ 4175 if (mflags & (flag)) { \ 4176 if (buf[0] != '\0') \ 4177 strlcat(buf, ", ", sizeof(buf)); \ 4178 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 4179 mflags &= ~(flag); \ 4180 } \ 4181 } while (0) 4182 MNT_FLAG(MNT_RDONLY); 4183 MNT_FLAG(MNT_SYNCHRONOUS); 4184 MNT_FLAG(MNT_NOEXEC); 4185 MNT_FLAG(MNT_NOSUID); 4186 MNT_FLAG(MNT_NFS4ACLS); 4187 MNT_FLAG(MNT_UNION); 4188 MNT_FLAG(MNT_ASYNC); 4189 MNT_FLAG(MNT_SUIDDIR); 4190 MNT_FLAG(MNT_SOFTDEP); 4191 MNT_FLAG(MNT_NOSYMFOLLOW); 4192 MNT_FLAG(MNT_GJOURNAL); 4193 MNT_FLAG(MNT_MULTILABEL); 4194 MNT_FLAG(MNT_ACLS); 4195 MNT_FLAG(MNT_NOATIME); 4196 MNT_FLAG(MNT_NOCLUSTERR); 4197 MNT_FLAG(MNT_NOCLUSTERW); 4198 MNT_FLAG(MNT_SUJ); 4199 MNT_FLAG(MNT_EXRDONLY); 4200 MNT_FLAG(MNT_EXPORTED); 4201 MNT_FLAG(MNT_DEFEXPORTED); 4202 MNT_FLAG(MNT_EXPORTANON); 4203 MNT_FLAG(MNT_EXKERB); 4204 MNT_FLAG(MNT_EXPUBLIC); 4205 MNT_FLAG(MNT_LOCAL); 4206 MNT_FLAG(MNT_QUOTA); 4207 MNT_FLAG(MNT_ROOTFS); 4208 MNT_FLAG(MNT_USER); 4209 MNT_FLAG(MNT_IGNORE); 4210 MNT_FLAG(MNT_UPDATE); 4211 MNT_FLAG(MNT_DELEXPORT); 4212 MNT_FLAG(MNT_RELOAD); 4213 MNT_FLAG(MNT_FORCE); 4214 MNT_FLAG(MNT_SNAPSHOT); 4215 MNT_FLAG(MNT_BYFSID); 4216 #undef MNT_FLAG 4217 if (mflags != 0) { 4218 if (buf[0] != '\0') 4219 strlcat(buf, ", ", sizeof(buf)); 4220 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4221 "0x%016jx", mflags); 4222 } 4223 db_printf(" mnt_flag = %s\n", buf); 4224 4225 buf[0] = '\0'; 4226 flags = mp->mnt_kern_flag; 4227 #define MNT_KERN_FLAG(flag) do { \ 4228 if (flags & (flag)) { \ 4229 if (buf[0] != '\0') \ 4230 strlcat(buf, ", ", sizeof(buf)); \ 4231 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 4232 flags &= ~(flag); \ 4233 } \ 4234 } while (0) 4235 MNT_KERN_FLAG(MNTK_UNMOUNTF); 4236 MNT_KERN_FLAG(MNTK_ASYNC); 4237 MNT_KERN_FLAG(MNTK_SOFTDEP); 4238 MNT_KERN_FLAG(MNTK_DRAINING); 4239 MNT_KERN_FLAG(MNTK_REFEXPIRE); 4240 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 4241 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 4242 MNT_KERN_FLAG(MNTK_NO_IOPF); 4243 MNT_KERN_FLAG(MNTK_VGONE_UPPER); 4244 MNT_KERN_FLAG(MNTK_VGONE_WAITER); 4245 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT); 4246 MNT_KERN_FLAG(MNTK_MARKER); 4247 MNT_KERN_FLAG(MNTK_USES_BCACHE); 4248 MNT_KERN_FLAG(MNTK_FPLOOKUP); 4249 MNT_KERN_FLAG(MNTK_NOASYNC); 4250 MNT_KERN_FLAG(MNTK_UNMOUNT); 4251 MNT_KERN_FLAG(MNTK_MWAIT); 4252 MNT_KERN_FLAG(MNTK_SUSPEND); 4253 MNT_KERN_FLAG(MNTK_SUSPEND2); 4254 MNT_KERN_FLAG(MNTK_SUSPENDED); 4255 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 4256 MNT_KERN_FLAG(MNTK_NOKNOTE); 4257 #undef MNT_KERN_FLAG 4258 if (flags != 0) { 4259 if (buf[0] != '\0') 4260 strlcat(buf, ", ", sizeof(buf)); 4261 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 4262 "0x%08x", flags); 4263 } 4264 db_printf(" mnt_kern_flag = %s\n", buf); 4265 4266 db_printf(" mnt_opt = "); 4267 opt = TAILQ_FIRST(mp->mnt_opt); 4268 if (opt != NULL) { 4269 db_printf("%s", opt->name); 4270 opt = TAILQ_NEXT(opt, link); 4271 while (opt != NULL) { 4272 db_printf(", %s", opt->name); 4273 opt = TAILQ_NEXT(opt, link); 4274 } 4275 } 4276 db_printf("\n"); 4277 4278 sp = &mp->mnt_stat; 4279 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 4280 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 4281 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 4282 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 4283 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 4284 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 4285 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 4286 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 4287 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 4288 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 4289 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 4290 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 4291 4292 db_printf(" mnt_cred = { uid=%u ruid=%u", 4293 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 4294 if (jailed(mp->mnt_cred)) 4295 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 4296 db_printf(" }\n"); 4297 db_printf(" mnt_ref = %d (with %d in the struct)\n", 4298 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref); 4299 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 4300 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 4301 db_printf(" mnt_lazyvnodelistsize = %d\n", 4302 mp->mnt_lazyvnodelistsize); 4303 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n", 4304 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount); 4305 db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen); 4306 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 4307 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 4308 db_printf(" mnt_lockref = %d (with %d in the struct)\n", 4309 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref); 4310 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 4311 db_printf(" mnt_secondary_accwrites = %d\n", 4312 mp->mnt_secondary_accwrites); 4313 db_printf(" mnt_gjprovider = %s\n", 4314 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 4315 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops); 4316 4317 db_printf("\n\nList of active vnodes\n"); 4318 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4319 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) { 4320 vn_printf(vp, "vnode "); 4321 if (db_pager_quit) 4322 break; 4323 } 4324 } 4325 db_printf("\n\nList of inactive vnodes\n"); 4326 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4327 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) { 4328 vn_printf(vp, "vnode "); 4329 if (db_pager_quit) 4330 break; 4331 } 4332 } 4333 } 4334 #endif /* DDB */ 4335 4336 /* 4337 * Fill in a struct xvfsconf based on a struct vfsconf. 4338 */ 4339 static int 4340 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 4341 { 4342 struct xvfsconf xvfsp; 4343 4344 bzero(&xvfsp, sizeof(xvfsp)); 4345 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4346 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4347 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4348 xvfsp.vfc_flags = vfsp->vfc_flags; 4349 /* 4350 * These are unused in userland, we keep them 4351 * to not break binary compatibility. 4352 */ 4353 xvfsp.vfc_vfsops = NULL; 4354 xvfsp.vfc_next = NULL; 4355 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4356 } 4357 4358 #ifdef COMPAT_FREEBSD32 4359 struct xvfsconf32 { 4360 uint32_t vfc_vfsops; 4361 char vfc_name[MFSNAMELEN]; 4362 int32_t vfc_typenum; 4363 int32_t vfc_refcount; 4364 int32_t vfc_flags; 4365 uint32_t vfc_next; 4366 }; 4367 4368 static int 4369 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 4370 { 4371 struct xvfsconf32 xvfsp; 4372 4373 bzero(&xvfsp, sizeof(xvfsp)); 4374 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 4375 xvfsp.vfc_typenum = vfsp->vfc_typenum; 4376 xvfsp.vfc_refcount = vfsp->vfc_refcount; 4377 xvfsp.vfc_flags = vfsp->vfc_flags; 4378 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 4379 } 4380 #endif 4381 4382 /* 4383 * Top level filesystem related information gathering. 4384 */ 4385 static int 4386 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 4387 { 4388 struct vfsconf *vfsp; 4389 int error; 4390 4391 error = 0; 4392 vfsconf_slock(); 4393 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4394 #ifdef COMPAT_FREEBSD32 4395 if (req->flags & SCTL_MASK32) 4396 error = vfsconf2x32(req, vfsp); 4397 else 4398 #endif 4399 error = vfsconf2x(req, vfsp); 4400 if (error) 4401 break; 4402 } 4403 vfsconf_sunlock(); 4404 return (error); 4405 } 4406 4407 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | 4408 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, 4409 "S,xvfsconf", "List of all configured filesystems"); 4410 4411 #ifndef BURN_BRIDGES 4412 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 4413 4414 static int 4415 vfs_sysctl(SYSCTL_HANDLER_ARGS) 4416 { 4417 int *name = (int *)arg1 - 1; /* XXX */ 4418 u_int namelen = arg2 + 1; /* XXX */ 4419 struct vfsconf *vfsp; 4420 4421 log(LOG_WARNING, "userland calling deprecated sysctl, " 4422 "please rebuild world\n"); 4423 4424 #if 1 || defined(COMPAT_PRELITE2) 4425 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 4426 if (namelen == 1) 4427 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 4428 #endif 4429 4430 switch (name[1]) { 4431 case VFS_MAXTYPENUM: 4432 if (namelen != 2) 4433 return (ENOTDIR); 4434 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 4435 case VFS_CONF: 4436 if (namelen != 3) 4437 return (ENOTDIR); /* overloaded */ 4438 vfsconf_slock(); 4439 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4440 if (vfsp->vfc_typenum == name[2]) 4441 break; 4442 } 4443 vfsconf_sunlock(); 4444 if (vfsp == NULL) 4445 return (EOPNOTSUPP); 4446 #ifdef COMPAT_FREEBSD32 4447 if (req->flags & SCTL_MASK32) 4448 return (vfsconf2x32(req, vfsp)); 4449 else 4450 #endif 4451 return (vfsconf2x(req, vfsp)); 4452 } 4453 return (EOPNOTSUPP); 4454 } 4455 4456 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | 4457 CTLFLAG_MPSAFE, vfs_sysctl, 4458 "Generic filesystem"); 4459 4460 #if 1 || defined(COMPAT_PRELITE2) 4461 4462 static int 4463 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 4464 { 4465 int error; 4466 struct vfsconf *vfsp; 4467 struct ovfsconf ovfs; 4468 4469 vfsconf_slock(); 4470 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 4471 bzero(&ovfs, sizeof(ovfs)); 4472 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 4473 strcpy(ovfs.vfc_name, vfsp->vfc_name); 4474 ovfs.vfc_index = vfsp->vfc_typenum; 4475 ovfs.vfc_refcount = vfsp->vfc_refcount; 4476 ovfs.vfc_flags = vfsp->vfc_flags; 4477 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 4478 if (error != 0) { 4479 vfsconf_sunlock(); 4480 return (error); 4481 } 4482 } 4483 vfsconf_sunlock(); 4484 return (0); 4485 } 4486 4487 #endif /* 1 || COMPAT_PRELITE2 */ 4488 #endif /* !BURN_BRIDGES */ 4489 4490 #define KINFO_VNODESLOP 10 4491 #ifdef notyet 4492 /* 4493 * Dump vnode list (via sysctl). 4494 */ 4495 /* ARGSUSED */ 4496 static int 4497 sysctl_vnode(SYSCTL_HANDLER_ARGS) 4498 { 4499 struct xvnode *xvn; 4500 struct mount *mp; 4501 struct vnode *vp; 4502 int error, len, n; 4503 4504 /* 4505 * Stale numvnodes access is not fatal here. 4506 */ 4507 req->lock = 0; 4508 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 4509 if (!req->oldptr) 4510 /* Make an estimate */ 4511 return (SYSCTL_OUT(req, 0, len)); 4512 4513 error = sysctl_wire_old_buffer(req, 0); 4514 if (error != 0) 4515 return (error); 4516 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 4517 n = 0; 4518 mtx_lock(&mountlist_mtx); 4519 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 4520 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) 4521 continue; 4522 MNT_ILOCK(mp); 4523 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 4524 if (n == len) 4525 break; 4526 vref(vp); 4527 xvn[n].xv_size = sizeof *xvn; 4528 xvn[n].xv_vnode = vp; 4529 xvn[n].xv_id = 0; /* XXX compat */ 4530 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 4531 XV_COPY(usecount); 4532 XV_COPY(writecount); 4533 XV_COPY(holdcnt); 4534 XV_COPY(mount); 4535 XV_COPY(numoutput); 4536 XV_COPY(type); 4537 #undef XV_COPY 4538 xvn[n].xv_flag = vp->v_vflag; 4539 4540 switch (vp->v_type) { 4541 case VREG: 4542 case VDIR: 4543 case VLNK: 4544 break; 4545 case VBLK: 4546 case VCHR: 4547 if (vp->v_rdev == NULL) { 4548 vrele(vp); 4549 continue; 4550 } 4551 xvn[n].xv_dev = dev2udev(vp->v_rdev); 4552 break; 4553 case VSOCK: 4554 xvn[n].xv_socket = vp->v_socket; 4555 break; 4556 case VFIFO: 4557 xvn[n].xv_fifo = vp->v_fifoinfo; 4558 break; 4559 case VNON: 4560 case VBAD: 4561 default: 4562 /* shouldn't happen? */ 4563 vrele(vp); 4564 continue; 4565 } 4566 vrele(vp); 4567 ++n; 4568 } 4569 MNT_IUNLOCK(mp); 4570 mtx_lock(&mountlist_mtx); 4571 vfs_unbusy(mp); 4572 if (n == len) 4573 break; 4574 } 4575 mtx_unlock(&mountlist_mtx); 4576 4577 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 4578 free(xvn, M_TEMP); 4579 return (error); 4580 } 4581 4582 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD | 4583 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode", 4584 ""); 4585 #endif 4586 4587 static void 4588 unmount_or_warn(struct mount *mp) 4589 { 4590 int error; 4591 4592 error = dounmount(mp, MNT_FORCE, curthread); 4593 if (error != 0) { 4594 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); 4595 if (error == EBUSY) 4596 printf("BUSY)\n"); 4597 else 4598 printf("%d)\n", error); 4599 } 4600 } 4601 4602 /* 4603 * Unmount all filesystems. The list is traversed in reverse order 4604 * of mounting to avoid dependencies. 4605 */ 4606 void 4607 vfs_unmountall(void) 4608 { 4609 struct mount *mp, *tmp; 4610 4611 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 4612 4613 /* 4614 * Since this only runs when rebooting, it is not interlocked. 4615 */ 4616 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { 4617 vfs_ref(mp); 4618 4619 /* 4620 * Forcibly unmounting "/dev" before "/" would prevent clean 4621 * unmount of the latter. 4622 */ 4623 if (mp == rootdevmp) 4624 continue; 4625 4626 unmount_or_warn(mp); 4627 } 4628 4629 if (rootdevmp != NULL) 4630 unmount_or_warn(rootdevmp); 4631 } 4632 4633 static void 4634 vfs_deferred_inactive(struct vnode *vp, int lkflags) 4635 { 4636 4637 ASSERT_VI_LOCKED(vp, __func__); 4638 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set")); 4639 if ((vp->v_iflag & VI_OWEINACT) == 0) { 4640 vdropl(vp); 4641 return; 4642 } 4643 if (vn_lock(vp, lkflags) == 0) { 4644 VI_LOCK(vp); 4645 vinactive(vp); 4646 VOP_UNLOCK(vp); 4647 vdropl(vp); 4648 return; 4649 } 4650 vdefer_inactive_unlocked(vp); 4651 } 4652 4653 static int 4654 vfs_periodic_inactive_filter(struct vnode *vp, void *arg) 4655 { 4656 4657 return (vp->v_iflag & VI_DEFINACT); 4658 } 4659 4660 static void __noinline 4661 vfs_periodic_inactive(struct mount *mp, int flags) 4662 { 4663 struct vnode *vp, *mvp; 4664 int lkflags; 4665 4666 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 4667 if (flags != MNT_WAIT) 4668 lkflags |= LK_NOWAIT; 4669 4670 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) { 4671 if ((vp->v_iflag & VI_DEFINACT) == 0) { 4672 VI_UNLOCK(vp); 4673 continue; 4674 } 4675 vp->v_iflag &= ~VI_DEFINACT; 4676 vfs_deferred_inactive(vp, lkflags); 4677 } 4678 } 4679 4680 static inline bool 4681 vfs_want_msync(struct vnode *vp) 4682 { 4683 struct vm_object *obj; 4684 4685 /* 4686 * This test may be performed without any locks held. 4687 * We rely on vm_object's type stability. 4688 */ 4689 if (vp->v_vflag & VV_NOSYNC) 4690 return (false); 4691 obj = vp->v_object; 4692 return (obj != NULL && vm_object_mightbedirty(obj)); 4693 } 4694 4695 static int 4696 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused) 4697 { 4698 4699 if (vp->v_vflag & VV_NOSYNC) 4700 return (false); 4701 if (vp->v_iflag & VI_DEFINACT) 4702 return (true); 4703 return (vfs_want_msync(vp)); 4704 } 4705 4706 static void __noinline 4707 vfs_periodic_msync_inactive(struct mount *mp, int flags) 4708 { 4709 struct vnode *vp, *mvp; 4710 struct vm_object *obj; 4711 struct thread *td; 4712 int lkflags, objflags; 4713 bool seen_defer; 4714 4715 td = curthread; 4716 4717 lkflags = LK_EXCLUSIVE | LK_INTERLOCK; 4718 if (flags != MNT_WAIT) { 4719 lkflags |= LK_NOWAIT; 4720 objflags = OBJPC_NOSYNC; 4721 } else { 4722 objflags = OBJPC_SYNC; 4723 } 4724 4725 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) { 4726 seen_defer = false; 4727 if (vp->v_iflag & VI_DEFINACT) { 4728 vp->v_iflag &= ~VI_DEFINACT; 4729 seen_defer = true; 4730 } 4731 if (!vfs_want_msync(vp)) { 4732 if (seen_defer) 4733 vfs_deferred_inactive(vp, lkflags); 4734 else 4735 VI_UNLOCK(vp); 4736 continue; 4737 } 4738 if (vget(vp, lkflags, td) == 0) { 4739 obj = vp->v_object; 4740 if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) { 4741 VM_OBJECT_WLOCK(obj); 4742 vm_object_page_clean(obj, 0, 0, objflags); 4743 VM_OBJECT_WUNLOCK(obj); 4744 } 4745 vput(vp); 4746 if (seen_defer) 4747 vdrop(vp); 4748 } else { 4749 if (seen_defer) 4750 vdefer_inactive_unlocked(vp); 4751 } 4752 } 4753 } 4754 4755 void 4756 vfs_periodic(struct mount *mp, int flags) 4757 { 4758 4759 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 4760 4761 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0) 4762 vfs_periodic_inactive(mp, flags); 4763 else 4764 vfs_periodic_msync_inactive(mp, flags); 4765 } 4766 4767 static void 4768 destroy_vpollinfo_free(struct vpollinfo *vi) 4769 { 4770 4771 knlist_destroy(&vi->vpi_selinfo.si_note); 4772 mtx_destroy(&vi->vpi_lock); 4773 uma_zfree(vnodepoll_zone, vi); 4774 } 4775 4776 static void 4777 destroy_vpollinfo(struct vpollinfo *vi) 4778 { 4779 4780 knlist_clear(&vi->vpi_selinfo.si_note, 1); 4781 seldrain(&vi->vpi_selinfo); 4782 destroy_vpollinfo_free(vi); 4783 } 4784 4785 /* 4786 * Initialize per-vnode helper structure to hold poll-related state. 4787 */ 4788 void 4789 v_addpollinfo(struct vnode *vp) 4790 { 4791 struct vpollinfo *vi; 4792 4793 if (vp->v_pollinfo != NULL) 4794 return; 4795 vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO); 4796 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 4797 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 4798 vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked); 4799 VI_LOCK(vp); 4800 if (vp->v_pollinfo != NULL) { 4801 VI_UNLOCK(vp); 4802 destroy_vpollinfo_free(vi); 4803 return; 4804 } 4805 vp->v_pollinfo = vi; 4806 VI_UNLOCK(vp); 4807 } 4808 4809 /* 4810 * Record a process's interest in events which might happen to 4811 * a vnode. Because poll uses the historic select-style interface 4812 * internally, this routine serves as both the ``check for any 4813 * pending events'' and the ``record my interest in future events'' 4814 * functions. (These are done together, while the lock is held, 4815 * to avoid race conditions.) 4816 */ 4817 int 4818 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 4819 { 4820 4821 v_addpollinfo(vp); 4822 mtx_lock(&vp->v_pollinfo->vpi_lock); 4823 if (vp->v_pollinfo->vpi_revents & events) { 4824 /* 4825 * This leaves events we are not interested 4826 * in available for the other process which 4827 * which presumably had requested them 4828 * (otherwise they would never have been 4829 * recorded). 4830 */ 4831 events &= vp->v_pollinfo->vpi_revents; 4832 vp->v_pollinfo->vpi_revents &= ~events; 4833 4834 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4835 return (events); 4836 } 4837 vp->v_pollinfo->vpi_events |= events; 4838 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 4839 mtx_unlock(&vp->v_pollinfo->vpi_lock); 4840 return (0); 4841 } 4842 4843 /* 4844 * Routine to create and manage a filesystem syncer vnode. 4845 */ 4846 #define sync_close ((int (*)(struct vop_close_args *))nullop) 4847 static int sync_fsync(struct vop_fsync_args *); 4848 static int sync_inactive(struct vop_inactive_args *); 4849 static int sync_reclaim(struct vop_reclaim_args *); 4850 4851 static struct vop_vector sync_vnodeops = { 4852 .vop_bypass = VOP_EOPNOTSUPP, 4853 .vop_close = sync_close, /* close */ 4854 .vop_fsync = sync_fsync, /* fsync */ 4855 .vop_inactive = sync_inactive, /* inactive */ 4856 .vop_need_inactive = vop_stdneed_inactive, /* need_inactive */ 4857 .vop_reclaim = sync_reclaim, /* reclaim */ 4858 .vop_lock1 = vop_stdlock, /* lock */ 4859 .vop_unlock = vop_stdunlock, /* unlock */ 4860 .vop_islocked = vop_stdislocked, /* islocked */ 4861 }; 4862 VFS_VOP_VECTOR_REGISTER(sync_vnodeops); 4863 4864 /* 4865 * Create a new filesystem syncer vnode for the specified mount point. 4866 */ 4867 void 4868 vfs_allocate_syncvnode(struct mount *mp) 4869 { 4870 struct vnode *vp; 4871 struct bufobj *bo; 4872 static long start, incr, next; 4873 int error; 4874 4875 /* Allocate a new vnode */ 4876 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 4877 if (error != 0) 4878 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 4879 vp->v_type = VNON; 4880 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4881 vp->v_vflag |= VV_FORCEINSMQ; 4882 error = insmntque(vp, mp); 4883 if (error != 0) 4884 panic("vfs_allocate_syncvnode: insmntque() failed"); 4885 vp->v_vflag &= ~VV_FORCEINSMQ; 4886 VOP_UNLOCK(vp); 4887 /* 4888 * Place the vnode onto the syncer worklist. We attempt to 4889 * scatter them about on the list so that they will go off 4890 * at evenly distributed times even if all the filesystems 4891 * are mounted at once. 4892 */ 4893 next += incr; 4894 if (next == 0 || next > syncer_maxdelay) { 4895 start /= 2; 4896 incr /= 2; 4897 if (start == 0) { 4898 start = syncer_maxdelay / 2; 4899 incr = syncer_maxdelay; 4900 } 4901 next = start; 4902 } 4903 bo = &vp->v_bufobj; 4904 BO_LOCK(bo); 4905 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 4906 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 4907 mtx_lock(&sync_mtx); 4908 sync_vnode_count++; 4909 if (mp->mnt_syncer == NULL) { 4910 mp->mnt_syncer = vp; 4911 vp = NULL; 4912 } 4913 mtx_unlock(&sync_mtx); 4914 BO_UNLOCK(bo); 4915 if (vp != NULL) { 4916 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4917 vgone(vp); 4918 vput(vp); 4919 } 4920 } 4921 4922 void 4923 vfs_deallocate_syncvnode(struct mount *mp) 4924 { 4925 struct vnode *vp; 4926 4927 mtx_lock(&sync_mtx); 4928 vp = mp->mnt_syncer; 4929 if (vp != NULL) 4930 mp->mnt_syncer = NULL; 4931 mtx_unlock(&sync_mtx); 4932 if (vp != NULL) 4933 vrele(vp); 4934 } 4935 4936 /* 4937 * Do a lazy sync of the filesystem. 4938 */ 4939 static int 4940 sync_fsync(struct vop_fsync_args *ap) 4941 { 4942 struct vnode *syncvp = ap->a_vp; 4943 struct mount *mp = syncvp->v_mount; 4944 int error, save; 4945 struct bufobj *bo; 4946 4947 /* 4948 * We only need to do something if this is a lazy evaluation. 4949 */ 4950 if (ap->a_waitfor != MNT_LAZY) 4951 return (0); 4952 4953 /* 4954 * Move ourselves to the back of the sync list. 4955 */ 4956 bo = &syncvp->v_bufobj; 4957 BO_LOCK(bo); 4958 vn_syncer_add_to_worklist(bo, syncdelay); 4959 BO_UNLOCK(bo); 4960 4961 /* 4962 * Walk the list of vnodes pushing all that are dirty and 4963 * not already on the sync list. 4964 */ 4965 if (vfs_busy(mp, MBF_NOWAIT) != 0) 4966 return (0); 4967 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 4968 vfs_unbusy(mp); 4969 return (0); 4970 } 4971 save = curthread_pflags_set(TDP_SYNCIO); 4972 /* 4973 * The filesystem at hand may be idle with free vnodes stored in the 4974 * batch. Return them instead of letting them stay there indefinitely. 4975 */ 4976 vfs_periodic(mp, MNT_NOWAIT); 4977 error = VFS_SYNC(mp, MNT_LAZY); 4978 curthread_pflags_restore(save); 4979 vn_finished_write(mp); 4980 vfs_unbusy(mp); 4981 return (error); 4982 } 4983 4984 /* 4985 * The syncer vnode is no referenced. 4986 */ 4987 static int 4988 sync_inactive(struct vop_inactive_args *ap) 4989 { 4990 4991 vgone(ap->a_vp); 4992 return (0); 4993 } 4994 4995 /* 4996 * The syncer vnode is no longer needed and is being decommissioned. 4997 * 4998 * Modifications to the worklist must be protected by sync_mtx. 4999 */ 5000 static int 5001 sync_reclaim(struct vop_reclaim_args *ap) 5002 { 5003 struct vnode *vp = ap->a_vp; 5004 struct bufobj *bo; 5005 5006 bo = &vp->v_bufobj; 5007 BO_LOCK(bo); 5008 mtx_lock(&sync_mtx); 5009 if (vp->v_mount->mnt_syncer == vp) 5010 vp->v_mount->mnt_syncer = NULL; 5011 if (bo->bo_flag & BO_ONWORKLST) { 5012 LIST_REMOVE(bo, bo_synclist); 5013 syncer_worklist_len--; 5014 sync_vnode_count--; 5015 bo->bo_flag &= ~BO_ONWORKLST; 5016 } 5017 mtx_unlock(&sync_mtx); 5018 BO_UNLOCK(bo); 5019 5020 return (0); 5021 } 5022 5023 int 5024 vn_need_pageq_flush(struct vnode *vp) 5025 { 5026 struct vm_object *obj; 5027 int need; 5028 5029 MPASS(mtx_owned(VI_MTX(vp))); 5030 need = 0; 5031 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && 5032 vm_object_mightbedirty(obj)) 5033 need = 1; 5034 return (need); 5035 } 5036 5037 /* 5038 * Check if vnode represents a disk device 5039 */ 5040 int 5041 vn_isdisk(struct vnode *vp, int *errp) 5042 { 5043 int error; 5044 5045 if (vp->v_type != VCHR) { 5046 error = ENOTBLK; 5047 goto out; 5048 } 5049 error = 0; 5050 dev_lock(); 5051 if (vp->v_rdev == NULL) 5052 error = ENXIO; 5053 else if (vp->v_rdev->si_devsw == NULL) 5054 error = ENXIO; 5055 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 5056 error = ENOTBLK; 5057 dev_unlock(); 5058 out: 5059 if (errp != NULL) 5060 *errp = error; 5061 return (error == 0); 5062 } 5063 5064 /* 5065 * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see 5066 * the comment above cache_fplookup for details. 5067 * 5068 * We never deny as priv_check_cred calls are not yet supported, see vaccess. 5069 */ 5070 int 5071 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred) 5072 { 5073 5074 VFS_SMR_ASSERT_ENTERED(); 5075 5076 /* Check the owner. */ 5077 if (cred->cr_uid == file_uid) { 5078 if (file_mode & S_IXUSR) 5079 return (0); 5080 return (EAGAIN); 5081 } 5082 5083 /* Otherwise, check the groups (first match) */ 5084 if (groupmember(file_gid, cred)) { 5085 if (file_mode & S_IXGRP) 5086 return (0); 5087 return (EAGAIN); 5088 } 5089 5090 /* Otherwise, check everyone else. */ 5091 if (file_mode & S_IXOTH) 5092 return (0); 5093 return (EAGAIN); 5094 } 5095 5096 /* 5097 * Common filesystem object access control check routine. Accepts a 5098 * vnode's type, "mode", uid and gid, requested access mode, and credentials. 5099 * Returns 0 on success, or an errno on failure. 5100 */ 5101 int 5102 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 5103 accmode_t accmode, struct ucred *cred) 5104 { 5105 accmode_t dac_granted; 5106 accmode_t priv_granted; 5107 5108 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 5109 ("invalid bit in accmode")); 5110 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 5111 ("VAPPEND without VWRITE")); 5112 5113 /* 5114 * Look for a normal, non-privileged way to access the file/directory 5115 * as requested. If it exists, go with that. 5116 */ 5117 5118 dac_granted = 0; 5119 5120 /* Check the owner. */ 5121 if (cred->cr_uid == file_uid) { 5122 dac_granted |= VADMIN; 5123 if (file_mode & S_IXUSR) 5124 dac_granted |= VEXEC; 5125 if (file_mode & S_IRUSR) 5126 dac_granted |= VREAD; 5127 if (file_mode & S_IWUSR) 5128 dac_granted |= (VWRITE | VAPPEND); 5129 5130 if ((accmode & dac_granted) == accmode) 5131 return (0); 5132 5133 goto privcheck; 5134 } 5135 5136 /* Otherwise, check the groups (first match) */ 5137 if (groupmember(file_gid, cred)) { 5138 if (file_mode & S_IXGRP) 5139 dac_granted |= VEXEC; 5140 if (file_mode & S_IRGRP) 5141 dac_granted |= VREAD; 5142 if (file_mode & S_IWGRP) 5143 dac_granted |= (VWRITE | VAPPEND); 5144 5145 if ((accmode & dac_granted) == accmode) 5146 return (0); 5147 5148 goto privcheck; 5149 } 5150 5151 /* Otherwise, check everyone else. */ 5152 if (file_mode & S_IXOTH) 5153 dac_granted |= VEXEC; 5154 if (file_mode & S_IROTH) 5155 dac_granted |= VREAD; 5156 if (file_mode & S_IWOTH) 5157 dac_granted |= (VWRITE | VAPPEND); 5158 if ((accmode & dac_granted) == accmode) 5159 return (0); 5160 5161 privcheck: 5162 /* 5163 * Build a privilege mask to determine if the set of privileges 5164 * satisfies the requirements when combined with the granted mask 5165 * from above. For each privilege, if the privilege is required, 5166 * bitwise or the request type onto the priv_granted mask. 5167 */ 5168 priv_granted = 0; 5169 5170 if (type == VDIR) { 5171 /* 5172 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 5173 * requests, instead of PRIV_VFS_EXEC. 5174 */ 5175 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5176 !priv_check_cred(cred, PRIV_VFS_LOOKUP)) 5177 priv_granted |= VEXEC; 5178 } else { 5179 /* 5180 * Ensure that at least one execute bit is on. Otherwise, 5181 * a privileged user will always succeed, and we don't want 5182 * this to happen unless the file really is executable. 5183 */ 5184 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 5185 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 5186 !priv_check_cred(cred, PRIV_VFS_EXEC)) 5187 priv_granted |= VEXEC; 5188 } 5189 5190 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 5191 !priv_check_cred(cred, PRIV_VFS_READ)) 5192 priv_granted |= VREAD; 5193 5194 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 5195 !priv_check_cred(cred, PRIV_VFS_WRITE)) 5196 priv_granted |= (VWRITE | VAPPEND); 5197 5198 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 5199 !priv_check_cred(cred, PRIV_VFS_ADMIN)) 5200 priv_granted |= VADMIN; 5201 5202 if ((accmode & (priv_granted | dac_granted)) == accmode) { 5203 return (0); 5204 } 5205 5206 return ((accmode & VADMIN) ? EPERM : EACCES); 5207 } 5208 5209 /* 5210 * Credential check based on process requesting service, and per-attribute 5211 * permissions. 5212 */ 5213 int 5214 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 5215 struct thread *td, accmode_t accmode) 5216 { 5217 5218 /* 5219 * Kernel-invoked always succeeds. 5220 */ 5221 if (cred == NOCRED) 5222 return (0); 5223 5224 /* 5225 * Do not allow privileged processes in jail to directly manipulate 5226 * system attributes. 5227 */ 5228 switch (attrnamespace) { 5229 case EXTATTR_NAMESPACE_SYSTEM: 5230 /* Potentially should be: return (EPERM); */ 5231 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM)); 5232 case EXTATTR_NAMESPACE_USER: 5233 return (VOP_ACCESS(vp, accmode, cred, td)); 5234 default: 5235 return (EPERM); 5236 } 5237 } 5238 5239 #ifdef DEBUG_VFS_LOCKS 5240 /* 5241 * This only exists to suppress warnings from unlocked specfs accesses. It is 5242 * no longer ok to have an unlocked VFS. 5243 */ 5244 #define IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL || \ 5245 (vp)->v_type == VCHR || (vp)->v_type == VBAD) 5246 5247 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 5248 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, 5249 "Drop into debugger on lock violation"); 5250 5251 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 5252 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 5253 0, "Check for interlock across VOPs"); 5254 5255 int vfs_badlock_print = 1; /* Print lock violations. */ 5256 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 5257 0, "Print lock violations"); 5258 5259 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */ 5260 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode, 5261 0, "Print vnode details on lock violations"); 5262 5263 #ifdef KDB 5264 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 5265 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, 5266 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations"); 5267 #endif 5268 5269 static void 5270 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 5271 { 5272 5273 #ifdef KDB 5274 if (vfs_badlock_backtrace) 5275 kdb_backtrace(); 5276 #endif 5277 if (vfs_badlock_vnode) 5278 vn_printf(vp, "vnode "); 5279 if (vfs_badlock_print) 5280 printf("%s: %p %s\n", str, (void *)vp, msg); 5281 if (vfs_badlock_ddb) 5282 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 5283 } 5284 5285 void 5286 assert_vi_locked(struct vnode *vp, const char *str) 5287 { 5288 5289 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 5290 vfs_badlock("interlock is not locked but should be", str, vp); 5291 } 5292 5293 void 5294 assert_vi_unlocked(struct vnode *vp, const char *str) 5295 { 5296 5297 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 5298 vfs_badlock("interlock is locked but should not be", str, vp); 5299 } 5300 5301 void 5302 assert_vop_locked(struct vnode *vp, const char *str) 5303 { 5304 int locked; 5305 5306 if (!IGNORE_LOCK(vp)) { 5307 locked = VOP_ISLOCKED(vp); 5308 if (locked == 0 || locked == LK_EXCLOTHER) 5309 vfs_badlock("is not locked but should be", str, vp); 5310 } 5311 } 5312 5313 void 5314 assert_vop_unlocked(struct vnode *vp, const char *str) 5315 { 5316 5317 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE) 5318 vfs_badlock("is locked but should not be", str, vp); 5319 } 5320 5321 void 5322 assert_vop_elocked(struct vnode *vp, const char *str) 5323 { 5324 5325 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 5326 vfs_badlock("is not exclusive locked but should be", str, vp); 5327 } 5328 #endif /* DEBUG_VFS_LOCKS */ 5329 5330 void 5331 vop_rename_fail(struct vop_rename_args *ap) 5332 { 5333 5334 if (ap->a_tvp != NULL) 5335 vput(ap->a_tvp); 5336 if (ap->a_tdvp == ap->a_tvp) 5337 vrele(ap->a_tdvp); 5338 else 5339 vput(ap->a_tdvp); 5340 vrele(ap->a_fdvp); 5341 vrele(ap->a_fvp); 5342 } 5343 5344 void 5345 vop_rename_pre(void *ap) 5346 { 5347 struct vop_rename_args *a = ap; 5348 5349 #ifdef DEBUG_VFS_LOCKS 5350 if (a->a_tvp) 5351 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 5352 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 5353 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 5354 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 5355 5356 /* Check the source (from). */ 5357 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 5358 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 5359 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 5360 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 5361 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 5362 5363 /* Check the target. */ 5364 if (a->a_tvp) 5365 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 5366 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 5367 #endif 5368 /* 5369 * It may be tempting to add vn_seqc_write_begin/end calls here and 5370 * in vop_rename_post but that's not going to work out since some 5371 * filesystems relookup vnodes mid-rename. This is probably a bug. 5372 * 5373 * For now filesystems are expected to do the relevant calls after they 5374 * decide what vnodes to operate on. 5375 */ 5376 if (a->a_tdvp != a->a_fdvp) 5377 vhold(a->a_fdvp); 5378 if (a->a_tvp != a->a_fvp) 5379 vhold(a->a_fvp); 5380 vhold(a->a_tdvp); 5381 if (a->a_tvp) 5382 vhold(a->a_tvp); 5383 } 5384 5385 #ifdef DEBUG_VFS_LOCKS 5386 void 5387 vop_fplookup_vexec_debugpre(void *ap __unused) 5388 { 5389 5390 VFS_SMR_ASSERT_ENTERED(); 5391 } 5392 5393 void 5394 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused) 5395 { 5396 5397 VFS_SMR_ASSERT_ENTERED(); 5398 } 5399 5400 void 5401 vop_strategy_debugpre(void *ap) 5402 { 5403 struct vop_strategy_args *a; 5404 struct buf *bp; 5405 5406 a = ap; 5407 bp = a->a_bp; 5408 5409 /* 5410 * Cluster ops lock their component buffers but not the IO container. 5411 */ 5412 if ((bp->b_flags & B_CLUSTER) != 0) 5413 return; 5414 5415 if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) { 5416 if (vfs_badlock_print) 5417 printf( 5418 "VOP_STRATEGY: bp is not locked but should be\n"); 5419 if (vfs_badlock_ddb) 5420 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 5421 } 5422 } 5423 5424 void 5425 vop_lock_debugpre(void *ap) 5426 { 5427 struct vop_lock1_args *a = ap; 5428 5429 if ((a->a_flags & LK_INTERLOCK) == 0) 5430 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5431 else 5432 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 5433 } 5434 5435 void 5436 vop_lock_debugpost(void *ap, int rc) 5437 { 5438 struct vop_lock1_args *a = ap; 5439 5440 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 5441 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) 5442 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 5443 } 5444 5445 void 5446 vop_unlock_debugpre(void *ap) 5447 { 5448 struct vop_unlock_args *a = ap; 5449 5450 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 5451 } 5452 5453 void 5454 vop_need_inactive_debugpre(void *ap) 5455 { 5456 struct vop_need_inactive_args *a = ap; 5457 5458 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 5459 } 5460 5461 void 5462 vop_need_inactive_debugpost(void *ap, int rc) 5463 { 5464 struct vop_need_inactive_args *a = ap; 5465 5466 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); 5467 } 5468 #endif 5469 5470 void 5471 vop_create_pre(void *ap) 5472 { 5473 struct vop_create_args *a; 5474 struct vnode *dvp; 5475 5476 a = ap; 5477 dvp = a->a_dvp; 5478 vn_seqc_write_begin(dvp); 5479 } 5480 5481 void 5482 vop_create_post(void *ap, int rc) 5483 { 5484 struct vop_create_args *a; 5485 struct vnode *dvp; 5486 5487 a = ap; 5488 dvp = a->a_dvp; 5489 vn_seqc_write_end(dvp); 5490 if (!rc) 5491 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 5492 } 5493 5494 void 5495 vop_whiteout_pre(void *ap) 5496 { 5497 struct vop_whiteout_args *a; 5498 struct vnode *dvp; 5499 5500 a = ap; 5501 dvp = a->a_dvp; 5502 vn_seqc_write_begin(dvp); 5503 } 5504 5505 void 5506 vop_whiteout_post(void *ap, int rc) 5507 { 5508 struct vop_whiteout_args *a; 5509 struct vnode *dvp; 5510 5511 a = ap; 5512 dvp = a->a_dvp; 5513 vn_seqc_write_end(dvp); 5514 } 5515 5516 void 5517 vop_deleteextattr_pre(void *ap) 5518 { 5519 struct vop_deleteextattr_args *a; 5520 struct vnode *vp; 5521 5522 a = ap; 5523 vp = a->a_vp; 5524 vn_seqc_write_begin(vp); 5525 } 5526 5527 void 5528 vop_deleteextattr_post(void *ap, int rc) 5529 { 5530 struct vop_deleteextattr_args *a; 5531 struct vnode *vp; 5532 5533 a = ap; 5534 vp = a->a_vp; 5535 vn_seqc_write_end(vp); 5536 if (!rc) 5537 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 5538 } 5539 5540 void 5541 vop_link_pre(void *ap) 5542 { 5543 struct vop_link_args *a; 5544 struct vnode *vp, *tdvp; 5545 5546 a = ap; 5547 vp = a->a_vp; 5548 tdvp = a->a_tdvp; 5549 vn_seqc_write_begin(vp); 5550 vn_seqc_write_begin(tdvp); 5551 } 5552 5553 void 5554 vop_link_post(void *ap, int rc) 5555 { 5556 struct vop_link_args *a; 5557 struct vnode *vp, *tdvp; 5558 5559 a = ap; 5560 vp = a->a_vp; 5561 tdvp = a->a_tdvp; 5562 vn_seqc_write_end(vp); 5563 vn_seqc_write_end(tdvp); 5564 if (!rc) { 5565 VFS_KNOTE_LOCKED(vp, NOTE_LINK); 5566 VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE); 5567 } 5568 } 5569 5570 void 5571 vop_mkdir_pre(void *ap) 5572 { 5573 struct vop_mkdir_args *a; 5574 struct vnode *dvp; 5575 5576 a = ap; 5577 dvp = a->a_dvp; 5578 vn_seqc_write_begin(dvp); 5579 } 5580 5581 void 5582 vop_mkdir_post(void *ap, int rc) 5583 { 5584 struct vop_mkdir_args *a; 5585 struct vnode *dvp; 5586 5587 a = ap; 5588 dvp = a->a_dvp; 5589 vn_seqc_write_end(dvp); 5590 if (!rc) 5591 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 5592 } 5593 5594 void 5595 vop_mknod_pre(void *ap) 5596 { 5597 struct vop_mknod_args *a; 5598 struct vnode *dvp; 5599 5600 a = ap; 5601 dvp = a->a_dvp; 5602 vn_seqc_write_begin(dvp); 5603 } 5604 5605 void 5606 vop_mknod_post(void *ap, int rc) 5607 { 5608 struct vop_mknod_args *a; 5609 struct vnode *dvp; 5610 5611 a = ap; 5612 dvp = a->a_dvp; 5613 vn_seqc_write_end(dvp); 5614 if (!rc) 5615 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 5616 } 5617 5618 void 5619 vop_reclaim_post(void *ap, int rc) 5620 { 5621 struct vop_reclaim_args *a; 5622 struct vnode *vp; 5623 5624 a = ap; 5625 vp = a->a_vp; 5626 ASSERT_VOP_IN_SEQC(vp); 5627 if (!rc) 5628 VFS_KNOTE_LOCKED(vp, NOTE_REVOKE); 5629 } 5630 5631 void 5632 vop_remove_pre(void *ap) 5633 { 5634 struct vop_remove_args *a; 5635 struct vnode *dvp, *vp; 5636 5637 a = ap; 5638 dvp = a->a_dvp; 5639 vp = a->a_vp; 5640 vn_seqc_write_begin(dvp); 5641 vn_seqc_write_begin(vp); 5642 } 5643 5644 void 5645 vop_remove_post(void *ap, int rc) 5646 { 5647 struct vop_remove_args *a; 5648 struct vnode *dvp, *vp; 5649 5650 a = ap; 5651 dvp = a->a_dvp; 5652 vp = a->a_vp; 5653 vn_seqc_write_end(dvp); 5654 vn_seqc_write_end(vp); 5655 if (!rc) { 5656 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 5657 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 5658 } 5659 } 5660 5661 void 5662 vop_rename_post(void *ap, int rc) 5663 { 5664 struct vop_rename_args *a = ap; 5665 long hint; 5666 5667 if (!rc) { 5668 hint = NOTE_WRITE; 5669 if (a->a_fdvp == a->a_tdvp) { 5670 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) 5671 hint |= NOTE_LINK; 5672 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 5673 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 5674 } else { 5675 hint |= NOTE_EXTEND; 5676 if (a->a_fvp->v_type == VDIR) 5677 hint |= NOTE_LINK; 5678 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); 5679 5680 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && 5681 a->a_tvp->v_type == VDIR) 5682 hint &= ~NOTE_LINK; 5683 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); 5684 } 5685 5686 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 5687 if (a->a_tvp) 5688 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 5689 } 5690 if (a->a_tdvp != a->a_fdvp) 5691 vdrop(a->a_fdvp); 5692 if (a->a_tvp != a->a_fvp) 5693 vdrop(a->a_fvp); 5694 vdrop(a->a_tdvp); 5695 if (a->a_tvp) 5696 vdrop(a->a_tvp); 5697 } 5698 5699 void 5700 vop_rmdir_pre(void *ap) 5701 { 5702 struct vop_rmdir_args *a; 5703 struct vnode *dvp, *vp; 5704 5705 a = ap; 5706 dvp = a->a_dvp; 5707 vp = a->a_vp; 5708 vn_seqc_write_begin(dvp); 5709 vn_seqc_write_begin(vp); 5710 } 5711 5712 void 5713 vop_rmdir_post(void *ap, int rc) 5714 { 5715 struct vop_rmdir_args *a; 5716 struct vnode *dvp, *vp; 5717 5718 a = ap; 5719 dvp = a->a_dvp; 5720 vp = a->a_vp; 5721 vn_seqc_write_end(dvp); 5722 vn_seqc_write_end(vp); 5723 if (!rc) { 5724 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); 5725 VFS_KNOTE_LOCKED(vp, NOTE_DELETE); 5726 } 5727 } 5728 5729 void 5730 vop_setattr_pre(void *ap) 5731 { 5732 struct vop_setattr_args *a; 5733 struct vnode *vp; 5734 5735 a = ap; 5736 vp = a->a_vp; 5737 vn_seqc_write_begin(vp); 5738 } 5739 5740 void 5741 vop_setattr_post(void *ap, int rc) 5742 { 5743 struct vop_setattr_args *a; 5744 struct vnode *vp; 5745 5746 a = ap; 5747 vp = a->a_vp; 5748 vn_seqc_write_end(vp); 5749 if (!rc) 5750 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 5751 } 5752 5753 void 5754 vop_setacl_pre(void *ap) 5755 { 5756 struct vop_setacl_args *a; 5757 struct vnode *vp; 5758 5759 a = ap; 5760 vp = a->a_vp; 5761 vn_seqc_write_begin(vp); 5762 } 5763 5764 void 5765 vop_setacl_post(void *ap, int rc __unused) 5766 { 5767 struct vop_setacl_args *a; 5768 struct vnode *vp; 5769 5770 a = ap; 5771 vp = a->a_vp; 5772 vn_seqc_write_end(vp); 5773 } 5774 5775 void 5776 vop_setextattr_pre(void *ap) 5777 { 5778 struct vop_setextattr_args *a; 5779 struct vnode *vp; 5780 5781 a = ap; 5782 vp = a->a_vp; 5783 vn_seqc_write_begin(vp); 5784 } 5785 5786 void 5787 vop_setextattr_post(void *ap, int rc) 5788 { 5789 struct vop_setextattr_args *a; 5790 struct vnode *vp; 5791 5792 a = ap; 5793 vp = a->a_vp; 5794 vn_seqc_write_end(vp); 5795 if (!rc) 5796 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); 5797 } 5798 5799 void 5800 vop_symlink_pre(void *ap) 5801 { 5802 struct vop_symlink_args *a; 5803 struct vnode *dvp; 5804 5805 a = ap; 5806 dvp = a->a_dvp; 5807 vn_seqc_write_begin(dvp); 5808 } 5809 5810 void 5811 vop_symlink_post(void *ap, int rc) 5812 { 5813 struct vop_symlink_args *a; 5814 struct vnode *dvp; 5815 5816 a = ap; 5817 dvp = a->a_dvp; 5818 vn_seqc_write_end(dvp); 5819 if (!rc) 5820 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); 5821 } 5822 5823 void 5824 vop_open_post(void *ap, int rc) 5825 { 5826 struct vop_open_args *a = ap; 5827 5828 if (!rc) 5829 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); 5830 } 5831 5832 void 5833 vop_close_post(void *ap, int rc) 5834 { 5835 struct vop_close_args *a = ap; 5836 5837 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ 5838 !VN_IS_DOOMED(a->a_vp))) { 5839 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? 5840 NOTE_CLOSE_WRITE : NOTE_CLOSE); 5841 } 5842 } 5843 5844 void 5845 vop_read_post(void *ap, int rc) 5846 { 5847 struct vop_read_args *a = ap; 5848 5849 if (!rc) 5850 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 5851 } 5852 5853 void 5854 vop_readdir_post(void *ap, int rc) 5855 { 5856 struct vop_readdir_args *a = ap; 5857 5858 if (!rc) 5859 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); 5860 } 5861 5862 static struct knlist fs_knlist; 5863 5864 static void 5865 vfs_event_init(void *arg) 5866 { 5867 knlist_init_mtx(&fs_knlist, NULL); 5868 } 5869 /* XXX - correct order? */ 5870 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 5871 5872 void 5873 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 5874 { 5875 5876 KNOTE_UNLOCKED(&fs_knlist, event); 5877 } 5878 5879 static int filt_fsattach(struct knote *kn); 5880 static void filt_fsdetach(struct knote *kn); 5881 static int filt_fsevent(struct knote *kn, long hint); 5882 5883 struct filterops fs_filtops = { 5884 .f_isfd = 0, 5885 .f_attach = filt_fsattach, 5886 .f_detach = filt_fsdetach, 5887 .f_event = filt_fsevent 5888 }; 5889 5890 static int 5891 filt_fsattach(struct knote *kn) 5892 { 5893 5894 kn->kn_flags |= EV_CLEAR; 5895 knlist_add(&fs_knlist, kn, 0); 5896 return (0); 5897 } 5898 5899 static void 5900 filt_fsdetach(struct knote *kn) 5901 { 5902 5903 knlist_remove(&fs_knlist, kn, 0); 5904 } 5905 5906 static int 5907 filt_fsevent(struct knote *kn, long hint) 5908 { 5909 5910 kn->kn_fflags |= hint; 5911 return (kn->kn_fflags != 0); 5912 } 5913 5914 static int 5915 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 5916 { 5917 struct vfsidctl vc; 5918 int error; 5919 struct mount *mp; 5920 5921 error = SYSCTL_IN(req, &vc, sizeof(vc)); 5922 if (error) 5923 return (error); 5924 if (vc.vc_vers != VFS_CTL_VERS1) 5925 return (EINVAL); 5926 mp = vfs_getvfs(&vc.vc_fsid); 5927 if (mp == NULL) 5928 return (ENOENT); 5929 /* ensure that a specific sysctl goes to the right filesystem. */ 5930 if (strcmp(vc.vc_fstypename, "*") != 0 && 5931 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 5932 vfs_rel(mp); 5933 return (EINVAL); 5934 } 5935 VCTLTOREQ(&vc, req); 5936 error = VFS_SYSCTL(mp, vc.vc_op, req); 5937 vfs_rel(mp); 5938 return (error); 5939 } 5940 5941 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR, 5942 NULL, 0, sysctl_vfs_ctl, "", 5943 "Sysctl by fsid"); 5944 5945 /* 5946 * Function to initialize a va_filerev field sensibly. 5947 * XXX: Wouldn't a random number make a lot more sense ?? 5948 */ 5949 u_quad_t 5950 init_va_filerev(void) 5951 { 5952 struct bintime bt; 5953 5954 getbinuptime(&bt); 5955 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 5956 } 5957 5958 static int filt_vfsread(struct knote *kn, long hint); 5959 static int filt_vfswrite(struct knote *kn, long hint); 5960 static int filt_vfsvnode(struct knote *kn, long hint); 5961 static void filt_vfsdetach(struct knote *kn); 5962 static struct filterops vfsread_filtops = { 5963 .f_isfd = 1, 5964 .f_detach = filt_vfsdetach, 5965 .f_event = filt_vfsread 5966 }; 5967 static struct filterops vfswrite_filtops = { 5968 .f_isfd = 1, 5969 .f_detach = filt_vfsdetach, 5970 .f_event = filt_vfswrite 5971 }; 5972 static struct filterops vfsvnode_filtops = { 5973 .f_isfd = 1, 5974 .f_detach = filt_vfsdetach, 5975 .f_event = filt_vfsvnode 5976 }; 5977 5978 static void 5979 vfs_knllock(void *arg) 5980 { 5981 struct vnode *vp = arg; 5982 5983 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 5984 } 5985 5986 static void 5987 vfs_knlunlock(void *arg) 5988 { 5989 struct vnode *vp = arg; 5990 5991 VOP_UNLOCK(vp); 5992 } 5993 5994 static void 5995 vfs_knl_assert_locked(void *arg) 5996 { 5997 #ifdef DEBUG_VFS_LOCKS 5998 struct vnode *vp = arg; 5999 6000 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 6001 #endif 6002 } 6003 6004 static void 6005 vfs_knl_assert_unlocked(void *arg) 6006 { 6007 #ifdef DEBUG_VFS_LOCKS 6008 struct vnode *vp = arg; 6009 6010 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 6011 #endif 6012 } 6013 6014 int 6015 vfs_kqfilter(struct vop_kqfilter_args *ap) 6016 { 6017 struct vnode *vp = ap->a_vp; 6018 struct knote *kn = ap->a_kn; 6019 struct knlist *knl; 6020 6021 switch (kn->kn_filter) { 6022 case EVFILT_READ: 6023 kn->kn_fop = &vfsread_filtops; 6024 break; 6025 case EVFILT_WRITE: 6026 kn->kn_fop = &vfswrite_filtops; 6027 break; 6028 case EVFILT_VNODE: 6029 kn->kn_fop = &vfsvnode_filtops; 6030 break; 6031 default: 6032 return (EINVAL); 6033 } 6034 6035 kn->kn_hook = (caddr_t)vp; 6036 6037 v_addpollinfo(vp); 6038 if (vp->v_pollinfo == NULL) 6039 return (ENOMEM); 6040 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 6041 vhold(vp); 6042 knlist_add(knl, kn, 0); 6043 6044 return (0); 6045 } 6046 6047 /* 6048 * Detach knote from vnode 6049 */ 6050 static void 6051 filt_vfsdetach(struct knote *kn) 6052 { 6053 struct vnode *vp = (struct vnode *)kn->kn_hook; 6054 6055 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 6056 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 6057 vdrop(vp); 6058 } 6059 6060 /*ARGSUSED*/ 6061 static int 6062 filt_vfsread(struct knote *kn, long hint) 6063 { 6064 struct vnode *vp = (struct vnode *)kn->kn_hook; 6065 struct vattr va; 6066 int res; 6067 6068 /* 6069 * filesystem is gone, so set the EOF flag and schedule 6070 * the knote for deletion. 6071 */ 6072 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6073 VI_LOCK(vp); 6074 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6075 VI_UNLOCK(vp); 6076 return (1); 6077 } 6078 6079 if (VOP_GETATTR(vp, &va, curthread->td_ucred)) 6080 return (0); 6081 6082 VI_LOCK(vp); 6083 kn->kn_data = va.va_size - kn->kn_fp->f_offset; 6084 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; 6085 VI_UNLOCK(vp); 6086 return (res); 6087 } 6088 6089 /*ARGSUSED*/ 6090 static int 6091 filt_vfswrite(struct knote *kn, long hint) 6092 { 6093 struct vnode *vp = (struct vnode *)kn->kn_hook; 6094 6095 VI_LOCK(vp); 6096 6097 /* 6098 * filesystem is gone, so set the EOF flag and schedule 6099 * the knote for deletion. 6100 */ 6101 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) 6102 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 6103 6104 kn->kn_data = 0; 6105 VI_UNLOCK(vp); 6106 return (1); 6107 } 6108 6109 static int 6110 filt_vfsvnode(struct knote *kn, long hint) 6111 { 6112 struct vnode *vp = (struct vnode *)kn->kn_hook; 6113 int res; 6114 6115 VI_LOCK(vp); 6116 if (kn->kn_sfflags & hint) 6117 kn->kn_fflags |= hint; 6118 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { 6119 kn->kn_flags |= EV_EOF; 6120 VI_UNLOCK(vp); 6121 return (1); 6122 } 6123 res = (kn->kn_fflags != 0); 6124 VI_UNLOCK(vp); 6125 return (res); 6126 } 6127 6128 /* 6129 * Returns whether the directory is empty or not. 6130 * If it is empty, the return value is 0; otherwise 6131 * the return value is an error value (which may 6132 * be ENOTEMPTY). 6133 */ 6134 int 6135 vfs_emptydir(struct vnode *vp) 6136 { 6137 struct uio uio; 6138 struct iovec iov; 6139 struct dirent *dirent, *dp, *endp; 6140 int error, eof; 6141 6142 error = 0; 6143 eof = 0; 6144 6145 ASSERT_VOP_LOCKED(vp, "vfs_emptydir"); 6146 6147 dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK); 6148 iov.iov_base = dirent; 6149 iov.iov_len = sizeof(struct dirent); 6150 6151 uio.uio_iov = &iov; 6152 uio.uio_iovcnt = 1; 6153 uio.uio_offset = 0; 6154 uio.uio_resid = sizeof(struct dirent); 6155 uio.uio_segflg = UIO_SYSSPACE; 6156 uio.uio_rw = UIO_READ; 6157 uio.uio_td = curthread; 6158 6159 while (eof == 0 && error == 0) { 6160 error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof, 6161 NULL, NULL); 6162 if (error != 0) 6163 break; 6164 endp = (void *)((uint8_t *)dirent + 6165 sizeof(struct dirent) - uio.uio_resid); 6166 for (dp = dirent; dp < endp; 6167 dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) { 6168 if (dp->d_type == DT_WHT) 6169 continue; 6170 if (dp->d_namlen == 0) 6171 continue; 6172 if (dp->d_type != DT_DIR && 6173 dp->d_type != DT_UNKNOWN) { 6174 error = ENOTEMPTY; 6175 break; 6176 } 6177 if (dp->d_namlen > 2) { 6178 error = ENOTEMPTY; 6179 break; 6180 } 6181 if (dp->d_namlen == 1 && 6182 dp->d_name[0] != '.') { 6183 error = ENOTEMPTY; 6184 break; 6185 } 6186 if (dp->d_namlen == 2 && 6187 dp->d_name[1] != '.') { 6188 error = ENOTEMPTY; 6189 break; 6190 } 6191 uio.uio_resid = sizeof(struct dirent); 6192 } 6193 } 6194 free(dirent, M_TEMP); 6195 return (error); 6196 } 6197 6198 int 6199 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 6200 { 6201 int error; 6202 6203 if (dp->d_reclen > ap->a_uio->uio_resid) 6204 return (ENAMETOOLONG); 6205 error = uiomove(dp, dp->d_reclen, ap->a_uio); 6206 if (error) { 6207 if (ap->a_ncookies != NULL) { 6208 if (ap->a_cookies != NULL) 6209 free(ap->a_cookies, M_TEMP); 6210 ap->a_cookies = NULL; 6211 *ap->a_ncookies = 0; 6212 } 6213 return (error); 6214 } 6215 if (ap->a_ncookies == NULL) 6216 return (0); 6217 6218 KASSERT(ap->a_cookies, 6219 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 6220 6221 *ap->a_cookies = realloc(*ap->a_cookies, 6222 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); 6223 (*ap->a_cookies)[*ap->a_ncookies] = off; 6224 *ap->a_ncookies += 1; 6225 return (0); 6226 } 6227 6228 /* 6229 * The purpose of this routine is to remove granularity from accmode_t, 6230 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 6231 * VADMIN and VAPPEND. 6232 * 6233 * If it returns 0, the caller is supposed to continue with the usual 6234 * access checks using 'accmode' as modified by this routine. If it 6235 * returns nonzero value, the caller is supposed to return that value 6236 * as errno. 6237 * 6238 * Note that after this routine runs, accmode may be zero. 6239 */ 6240 int 6241 vfs_unixify_accmode(accmode_t *accmode) 6242 { 6243 /* 6244 * There is no way to specify explicit "deny" rule using 6245 * file mode or POSIX.1e ACLs. 6246 */ 6247 if (*accmode & VEXPLICIT_DENY) { 6248 *accmode = 0; 6249 return (0); 6250 } 6251 6252 /* 6253 * None of these can be translated into usual access bits. 6254 * Also, the common case for NFSv4 ACLs is to not contain 6255 * either of these bits. Caller should check for VWRITE 6256 * on the containing directory instead. 6257 */ 6258 if (*accmode & (VDELETE_CHILD | VDELETE)) 6259 return (EPERM); 6260 6261 if (*accmode & VADMIN_PERMS) { 6262 *accmode &= ~VADMIN_PERMS; 6263 *accmode |= VADMIN; 6264 } 6265 6266 /* 6267 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 6268 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 6269 */ 6270 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 6271 6272 return (0); 6273 } 6274 6275 /* 6276 * Clear out a doomed vnode (if any) and replace it with a new one as long 6277 * as the fs is not being unmounted. Return the root vnode to the caller. 6278 */ 6279 static int __noinline 6280 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp) 6281 { 6282 struct vnode *vp; 6283 int error; 6284 6285 restart: 6286 if (mp->mnt_rootvnode != NULL) { 6287 MNT_ILOCK(mp); 6288 vp = mp->mnt_rootvnode; 6289 if (vp != NULL) { 6290 if (!VN_IS_DOOMED(vp)) { 6291 vrefact(vp); 6292 MNT_IUNLOCK(mp); 6293 error = vn_lock(vp, flags); 6294 if (error == 0) { 6295 *vpp = vp; 6296 return (0); 6297 } 6298 vrele(vp); 6299 goto restart; 6300 } 6301 /* 6302 * Clear the old one. 6303 */ 6304 mp->mnt_rootvnode = NULL; 6305 } 6306 MNT_IUNLOCK(mp); 6307 if (vp != NULL) { 6308 vfs_op_barrier_wait(mp); 6309 vrele(vp); 6310 } 6311 } 6312 error = VFS_CACHEDROOT(mp, flags, vpp); 6313 if (error != 0) 6314 return (error); 6315 if (mp->mnt_vfs_ops == 0) { 6316 MNT_ILOCK(mp); 6317 if (mp->mnt_vfs_ops != 0) { 6318 MNT_IUNLOCK(mp); 6319 return (0); 6320 } 6321 if (mp->mnt_rootvnode == NULL) { 6322 vrefact(*vpp); 6323 mp->mnt_rootvnode = *vpp; 6324 } else { 6325 if (mp->mnt_rootvnode != *vpp) { 6326 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) { 6327 panic("%s: mismatch between vnode returned " 6328 " by VFS_CACHEDROOT and the one cached " 6329 " (%p != %p)", 6330 __func__, *vpp, mp->mnt_rootvnode); 6331 } 6332 } 6333 } 6334 MNT_IUNLOCK(mp); 6335 } 6336 return (0); 6337 } 6338 6339 int 6340 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp) 6341 { 6342 struct vnode *vp; 6343 int error; 6344 6345 if (!vfs_op_thread_enter(mp)) 6346 return (vfs_cache_root_fallback(mp, flags, vpp)); 6347 vp = atomic_load_ptr(&mp->mnt_rootvnode); 6348 if (vp == NULL || VN_IS_DOOMED(vp)) { 6349 vfs_op_thread_exit(mp); 6350 return (vfs_cache_root_fallback(mp, flags, vpp)); 6351 } 6352 vrefact(vp); 6353 vfs_op_thread_exit(mp); 6354 error = vn_lock(vp, flags); 6355 if (error != 0) { 6356 vrele(vp); 6357 return (vfs_cache_root_fallback(mp, flags, vpp)); 6358 } 6359 *vpp = vp; 6360 return (0); 6361 } 6362 6363 struct vnode * 6364 vfs_cache_root_clear(struct mount *mp) 6365 { 6366 struct vnode *vp; 6367 6368 /* 6369 * ops > 0 guarantees there is nobody who can see this vnode 6370 */ 6371 MPASS(mp->mnt_vfs_ops > 0); 6372 vp = mp->mnt_rootvnode; 6373 if (vp != NULL) 6374 vn_seqc_write_begin(vp); 6375 mp->mnt_rootvnode = NULL; 6376 return (vp); 6377 } 6378 6379 void 6380 vfs_cache_root_set(struct mount *mp, struct vnode *vp) 6381 { 6382 6383 MPASS(mp->mnt_vfs_ops > 0); 6384 vrefact(vp); 6385 mp->mnt_rootvnode = vp; 6386 } 6387 6388 /* 6389 * These are helper functions for filesystems to traverse all 6390 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 6391 * 6392 * This interface replaces MNT_VNODE_FOREACH. 6393 */ 6394 6395 struct vnode * 6396 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 6397 { 6398 struct vnode *vp; 6399 6400 if (should_yield()) 6401 kern_yield(PRI_USER); 6402 MNT_ILOCK(mp); 6403 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6404 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL; 6405 vp = TAILQ_NEXT(vp, v_nmntvnodes)) { 6406 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 6407 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 6408 continue; 6409 VI_LOCK(vp); 6410 if (VN_IS_DOOMED(vp)) { 6411 VI_UNLOCK(vp); 6412 continue; 6413 } 6414 break; 6415 } 6416 if (vp == NULL) { 6417 __mnt_vnode_markerfree_all(mvp, mp); 6418 /* MNT_IUNLOCK(mp); -- done in above function */ 6419 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 6420 return (NULL); 6421 } 6422 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 6423 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 6424 MNT_IUNLOCK(mp); 6425 return (vp); 6426 } 6427 6428 struct vnode * 6429 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 6430 { 6431 struct vnode *vp; 6432 6433 *mvp = vn_alloc_marker(mp); 6434 MNT_ILOCK(mp); 6435 MNT_REF(mp); 6436 6437 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 6438 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ 6439 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) 6440 continue; 6441 VI_LOCK(vp); 6442 if (VN_IS_DOOMED(vp)) { 6443 VI_UNLOCK(vp); 6444 continue; 6445 } 6446 break; 6447 } 6448 if (vp == NULL) { 6449 MNT_REL(mp); 6450 MNT_IUNLOCK(mp); 6451 vn_free_marker(*mvp); 6452 *mvp = NULL; 6453 return (NULL); 6454 } 6455 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 6456 MNT_IUNLOCK(mp); 6457 return (vp); 6458 } 6459 6460 void 6461 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 6462 { 6463 6464 if (*mvp == NULL) { 6465 MNT_IUNLOCK(mp); 6466 return; 6467 } 6468 6469 mtx_assert(MNT_MTX(mp), MA_OWNED); 6470 6471 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6472 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 6473 MNT_REL(mp); 6474 MNT_IUNLOCK(mp); 6475 vn_free_marker(*mvp); 6476 *mvp = NULL; 6477 } 6478 6479 /* 6480 * These are helper functions for filesystems to traverse their 6481 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h 6482 */ 6483 static void 6484 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 6485 { 6486 6487 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6488 6489 MNT_ILOCK(mp); 6490 MNT_REL(mp); 6491 MNT_IUNLOCK(mp); 6492 vn_free_marker(*mvp); 6493 *mvp = NULL; 6494 } 6495 6496 /* 6497 * Relock the mp mount vnode list lock with the vp vnode interlock in the 6498 * conventional lock order during mnt_vnode_next_lazy iteration. 6499 * 6500 * On entry, the mount vnode list lock is held and the vnode interlock is not. 6501 * The list lock is dropped and reacquired. On success, both locks are held. 6502 * On failure, the mount vnode list lock is held but the vnode interlock is 6503 * not, and the procedure may have yielded. 6504 */ 6505 static bool 6506 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp, 6507 struct vnode *vp) 6508 { 6509 6510 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER && 6511 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp, 6512 ("%s: bad marker", __func__)); 6513 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp, 6514 ("%s: inappropriate vnode", __func__)); 6515 ASSERT_VI_UNLOCKED(vp, __func__); 6516 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 6517 6518 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist); 6519 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist); 6520 6521 /* 6522 * Note we may be racing against vdrop which transitioned the hold 6523 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine, 6524 * if we are the only user after we get the interlock we will just 6525 * vdrop. 6526 */ 6527 vhold(vp); 6528 mtx_unlock(&mp->mnt_listmtx); 6529 VI_LOCK(vp); 6530 if (VN_IS_DOOMED(vp)) { 6531 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); 6532 goto out_lost; 6533 } 6534 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); 6535 /* 6536 * There is nothing to do if we are the last user. 6537 */ 6538 if (!refcount_release_if_not_last(&vp->v_holdcnt)) 6539 goto out_lost; 6540 mtx_lock(&mp->mnt_listmtx); 6541 return (true); 6542 out_lost: 6543 vdropl(vp); 6544 maybe_yield(); 6545 mtx_lock(&mp->mnt_listmtx); 6546 return (false); 6547 } 6548 6549 static struct vnode * 6550 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 6551 void *cbarg) 6552 { 6553 struct vnode *vp; 6554 6555 mtx_assert(&mp->mnt_listmtx, MA_OWNED); 6556 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 6557 restart: 6558 vp = TAILQ_NEXT(*mvp, v_lazylist); 6559 while (vp != NULL) { 6560 if (vp->v_type == VMARKER) { 6561 vp = TAILQ_NEXT(vp, v_lazylist); 6562 continue; 6563 } 6564 /* 6565 * See if we want to process the vnode. Note we may encounter a 6566 * long string of vnodes we don't care about and hog the list 6567 * as a result. Check for it and requeue the marker. 6568 */ 6569 VNPASS(!VN_IS_DOOMED(vp), vp); 6570 if (!cb(vp, cbarg)) { 6571 if (!should_yield()) { 6572 vp = TAILQ_NEXT(vp, v_lazylist); 6573 continue; 6574 } 6575 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, 6576 v_lazylist); 6577 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, 6578 v_lazylist); 6579 mtx_unlock(&mp->mnt_listmtx); 6580 kern_yield(PRI_USER); 6581 mtx_lock(&mp->mnt_listmtx); 6582 goto restart; 6583 } 6584 /* 6585 * Try-lock because this is the wrong lock order. 6586 */ 6587 if (!VI_TRYLOCK(vp) && 6588 !mnt_vnode_next_lazy_relock(*mvp, mp, vp)) 6589 goto restart; 6590 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 6591 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 6592 ("alien vnode on the lazy list %p %p", vp, mp)); 6593 VNPASS(vp->v_mount == mp, vp); 6594 VNPASS(!VN_IS_DOOMED(vp), vp); 6595 break; 6596 } 6597 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 6598 6599 /* Check if we are done */ 6600 if (vp == NULL) { 6601 mtx_unlock(&mp->mnt_listmtx); 6602 mnt_vnode_markerfree_lazy(mvp, mp); 6603 return (NULL); 6604 } 6605 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist); 6606 mtx_unlock(&mp->mnt_listmtx); 6607 ASSERT_VI_LOCKED(vp, "lazy iter"); 6608 return (vp); 6609 } 6610 6611 struct vnode * 6612 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 6613 void *cbarg) 6614 { 6615 6616 if (should_yield()) 6617 kern_yield(PRI_USER); 6618 mtx_lock(&mp->mnt_listmtx); 6619 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 6620 } 6621 6622 struct vnode * 6623 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, 6624 void *cbarg) 6625 { 6626 struct vnode *vp; 6627 6628 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist)) 6629 return (NULL); 6630 6631 *mvp = vn_alloc_marker(mp); 6632 MNT_ILOCK(mp); 6633 MNT_REF(mp); 6634 MNT_IUNLOCK(mp); 6635 6636 mtx_lock(&mp->mnt_listmtx); 6637 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist); 6638 if (vp == NULL) { 6639 mtx_unlock(&mp->mnt_listmtx); 6640 mnt_vnode_markerfree_lazy(mvp, mp); 6641 return (NULL); 6642 } 6643 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist); 6644 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); 6645 } 6646 6647 void 6648 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) 6649 { 6650 6651 if (*mvp == NULL) 6652 return; 6653 6654 mtx_lock(&mp->mnt_listmtx); 6655 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); 6656 mtx_unlock(&mp->mnt_listmtx); 6657 mnt_vnode_markerfree_lazy(mvp, mp); 6658 } 6659 6660 int 6661 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp) 6662 { 6663 6664 if ((cnp->cn_flags & NOEXECCHECK) != 0) { 6665 cnp->cn_flags &= ~NOEXECCHECK; 6666 return (0); 6667 } 6668 6669 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread)); 6670 } 6671 6672 /* 6673 * Do not use this variant unless you have means other than the hold count 6674 * to prevent the vnode from getting freed. 6675 */ 6676 void 6677 vn_seqc_write_begin_unheld_locked(struct vnode *vp) 6678 { 6679 6680 ASSERT_VI_LOCKED(vp, __func__); 6681 VNPASS(vp->v_seqc_users >= 0, vp); 6682 vp->v_seqc_users++; 6683 if (vp->v_seqc_users == 1) 6684 seqc_sleepable_write_begin(&vp->v_seqc); 6685 } 6686 6687 void 6688 vn_seqc_write_begin_locked(struct vnode *vp) 6689 { 6690 6691 ASSERT_VI_LOCKED(vp, __func__); 6692 VNPASS(vp->v_holdcnt > 0, vp); 6693 vn_seqc_write_begin_unheld_locked(vp); 6694 } 6695 6696 void 6697 vn_seqc_write_begin(struct vnode *vp) 6698 { 6699 6700 VI_LOCK(vp); 6701 vn_seqc_write_begin_locked(vp); 6702 VI_UNLOCK(vp); 6703 } 6704 6705 void 6706 vn_seqc_write_begin_unheld(struct vnode *vp) 6707 { 6708 6709 VI_LOCK(vp); 6710 vn_seqc_write_begin_unheld_locked(vp); 6711 VI_UNLOCK(vp); 6712 } 6713 6714 void 6715 vn_seqc_write_end_locked(struct vnode *vp) 6716 { 6717 6718 ASSERT_VI_LOCKED(vp, __func__); 6719 VNPASS(vp->v_seqc_users > 0, vp); 6720 vp->v_seqc_users--; 6721 if (vp->v_seqc_users == 0) 6722 seqc_sleepable_write_end(&vp->v_seqc); 6723 } 6724 6725 void 6726 vn_seqc_write_end(struct vnode *vp) 6727 { 6728 6729 VI_LOCK(vp); 6730 vn_seqc_write_end_locked(vp); 6731 VI_UNLOCK(vp); 6732 } 6733