1 /*- 2 * Copyright (c) 1989, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 4. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 35 */ 36 37 /* 38 * External virtual filesystem routines 39 */ 40 41 #include <sys/cdefs.h> 42 __FBSDID("$FreeBSD$"); 43 44 #include "opt_compat.h" 45 #include "opt_ddb.h" 46 #include "opt_watchdog.h" 47 48 #include <sys/param.h> 49 #include <sys/systm.h> 50 #include <sys/bio.h> 51 #include <sys/buf.h> 52 #include <sys/condvar.h> 53 #include <sys/conf.h> 54 #include <sys/dirent.h> 55 #include <sys/event.h> 56 #include <sys/eventhandler.h> 57 #include <sys/extattr.h> 58 #include <sys/file.h> 59 #include <sys/fcntl.h> 60 #include <sys/jail.h> 61 #include <sys/kdb.h> 62 #include <sys/kernel.h> 63 #include <sys/kthread.h> 64 #include <sys/lockf.h> 65 #include <sys/malloc.h> 66 #include <sys/mount.h> 67 #include <sys/namei.h> 68 #include <sys/priv.h> 69 #include <sys/reboot.h> 70 #include <sys/rwlock.h> 71 #include <sys/sched.h> 72 #include <sys/sleepqueue.h> 73 #include <sys/smp.h> 74 #include <sys/stat.h> 75 #include <sys/sysctl.h> 76 #include <sys/syslog.h> 77 #include <sys/vmmeter.h> 78 #include <sys/vnode.h> 79 #include <sys/watchdog.h> 80 81 #include <machine/stdarg.h> 82 83 #include <security/mac/mac_framework.h> 84 85 #include <vm/vm.h> 86 #include <vm/vm_object.h> 87 #include <vm/vm_extern.h> 88 #include <vm/pmap.h> 89 #include <vm/vm_map.h> 90 #include <vm/vm_page.h> 91 #include <vm/vm_kern.h> 92 #include <vm/uma.h> 93 94 #ifdef DDB 95 #include <ddb/ddb.h> 96 #endif 97 98 static void delmntque(struct vnode *vp); 99 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, 100 int slpflag, int slptimeo); 101 static void syncer_shutdown(void *arg, int howto); 102 static int vtryrecycle(struct vnode *vp); 103 static void v_incr_usecount(struct vnode *); 104 static void v_decr_usecount(struct vnode *); 105 static void v_decr_useonly(struct vnode *); 106 static void v_upgrade_usecount(struct vnode *); 107 static void vnlru_free(int); 108 static void vgonel(struct vnode *); 109 static void vfs_knllock(void *arg); 110 static void vfs_knlunlock(void *arg); 111 static void vfs_knl_assert_locked(void *arg); 112 static void vfs_knl_assert_unlocked(void *arg); 113 static void destroy_vpollinfo(struct vpollinfo *vi); 114 115 /* 116 * Number of vnodes in existence. Increased whenever getnewvnode() 117 * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode. 118 */ 119 static unsigned long numvnodes; 120 121 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, 122 "Number of vnodes in existence"); 123 124 /* 125 * Conversion tables for conversion from vnode types to inode formats 126 * and back. 127 */ 128 enum vtype iftovt_tab[16] = { 129 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 130 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 131 }; 132 int vttoif_tab[10] = { 133 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 134 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT 135 }; 136 137 /* 138 * List of vnodes that are ready for recycling. 139 */ 140 static TAILQ_HEAD(freelst, vnode) vnode_free_list; 141 142 /* 143 * Free vnode target. Free vnodes may simply be files which have been stat'd 144 * but not read. This is somewhat common, and a small cache of such files 145 * should be kept to avoid recreation costs. 146 */ 147 static u_long wantfreevnodes; 148 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 149 /* Number of vnodes in the free list. */ 150 static u_long freevnodes; 151 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, 152 "Number of vnodes in the free list"); 153 154 static int vlru_allow_cache_src; 155 SYSCTL_INT(_vfs, OID_AUTO, vlru_allow_cache_src, CTLFLAG_RW, 156 &vlru_allow_cache_src, 0, "Allow vlru to reclaim source vnode"); 157 158 /* 159 * Various variables used for debugging the new implementation of 160 * reassignbuf(). 161 * XXX these are probably of (very) limited utility now. 162 */ 163 static int reassignbufcalls; 164 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, 165 "Number of calls to reassignbuf"); 166 167 /* 168 * Cache for the mount type id assigned to NFS. This is used for 169 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 170 */ 171 int nfs_mount_type = -1; 172 173 /* To keep more than one thread at a time from running vfs_getnewfsid */ 174 static struct mtx mntid_mtx; 175 176 /* 177 * Lock for any access to the following: 178 * vnode_free_list 179 * numvnodes 180 * freevnodes 181 */ 182 static struct mtx vnode_free_list_mtx; 183 184 /* Publicly exported FS */ 185 struct nfs_public nfs_pub; 186 187 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 188 static uma_zone_t vnode_zone; 189 static uma_zone_t vnodepoll_zone; 190 191 /* 192 * The workitem queue. 193 * 194 * It is useful to delay writes of file data and filesystem metadata 195 * for tens of seconds so that quickly created and deleted files need 196 * not waste disk bandwidth being created and removed. To realize this, 197 * we append vnodes to a "workitem" queue. When running with a soft 198 * updates implementation, most pending metadata dependencies should 199 * not wait for more than a few seconds. Thus, mounted on block devices 200 * are delayed only about a half the time that file data is delayed. 201 * Similarly, directory updates are more critical, so are only delayed 202 * about a third the time that file data is delayed. Thus, there are 203 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 204 * one each second (driven off the filesystem syncer process). The 205 * syncer_delayno variable indicates the next queue that is to be processed. 206 * Items that need to be processed soon are placed in this queue: 207 * 208 * syncer_workitem_pending[syncer_delayno] 209 * 210 * A delay of fifteen seconds is done by placing the request fifteen 211 * entries later in the queue: 212 * 213 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 214 * 215 */ 216 static int syncer_delayno; 217 static long syncer_mask; 218 LIST_HEAD(synclist, bufobj); 219 static struct synclist *syncer_workitem_pending; 220 /* 221 * The sync_mtx protects: 222 * bo->bo_synclist 223 * sync_vnode_count 224 * syncer_delayno 225 * syncer_state 226 * syncer_workitem_pending 227 * syncer_worklist_len 228 * rushjob 229 */ 230 static struct mtx sync_mtx; 231 static struct cv sync_wakeup; 232 233 #define SYNCER_MAXDELAY 32 234 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 235 static int syncdelay = 30; /* max time to delay syncing data */ 236 static int filedelay = 30; /* time to delay syncing files */ 237 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, 238 "Time to delay syncing files (in seconds)"); 239 static int dirdelay = 29; /* time to delay syncing directories */ 240 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, 241 "Time to delay syncing directories (in seconds)"); 242 static int metadelay = 28; /* time to delay syncing metadata */ 243 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, 244 "Time to delay syncing metadata (in seconds)"); 245 static int rushjob; /* number of slots to run ASAP */ 246 static int stat_rush_requests; /* number of times I/O speeded up */ 247 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, 248 "Number of times I/O speeded up (rush requests)"); 249 250 /* 251 * When shutting down the syncer, run it at four times normal speed. 252 */ 253 #define SYNCER_SHUTDOWN_SPEEDUP 4 254 static int sync_vnode_count; 255 static int syncer_worklist_len; 256 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 257 syncer_state; 258 259 /* 260 * Number of vnodes we want to exist at any one time. This is mostly used 261 * to size hash tables in vnode-related code. It is normally not used in 262 * getnewvnode(), as wantfreevnodes is normally nonzero.) 263 * 264 * XXX desiredvnodes is historical cruft and should not exist. 265 */ 266 int desiredvnodes; 267 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 268 &desiredvnodes, 0, "Maximum number of vnodes"); 269 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 270 &wantfreevnodes, 0, "Minimum number of vnodes (legacy)"); 271 static int vnlru_nowhere; 272 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 273 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 274 275 /* 276 * Macros to control when a vnode is freed and recycled. All require 277 * the vnode interlock. 278 */ 279 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt) 280 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt) 281 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt) 282 283 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */ 284 static int vnsz2log; 285 286 /* 287 * Initialize the vnode management data structures. 288 * 289 * Reevaluate the following cap on the number of vnodes after the physical 290 * memory size exceeds 512GB. In the limit, as the physical memory size 291 * grows, the ratio of physical pages to vnodes approaches sixteen to one. 292 */ 293 #ifndef MAXVNODES_MAX 294 #define MAXVNODES_MAX (512 * (1024 * 1024 * 1024 / (int)PAGE_SIZE / 16)) 295 #endif 296 static void 297 vntblinit(void *dummy __unused) 298 { 299 u_int i; 300 int physvnodes, virtvnodes; 301 302 /* 303 * Desiredvnodes is a function of the physical memory size and the 304 * kernel's heap size. Generally speaking, it scales with the 305 * physical memory size. The ratio of desiredvnodes to physical pages 306 * is one to four until desiredvnodes exceeds 98,304. Thereafter, the 307 * marginal ratio of desiredvnodes to physical pages is one to 308 * sixteen. However, desiredvnodes is limited by the kernel's heap 309 * size. The memory required by desiredvnodes vnodes and vm objects 310 * may not exceed one seventh of the kernel's heap size. 311 */ 312 physvnodes = maxproc + cnt.v_page_count / 16 + 3 * min(98304 * 4, 313 cnt.v_page_count) / 16; 314 virtvnodes = vm_kmem_size / (7 * (sizeof(struct vm_object) + 315 sizeof(struct vnode))); 316 desiredvnodes = min(physvnodes, virtvnodes); 317 if (desiredvnodes > MAXVNODES_MAX) { 318 if (bootverbose) 319 printf("Reducing kern.maxvnodes %d -> %d\n", 320 desiredvnodes, MAXVNODES_MAX); 321 desiredvnodes = MAXVNODES_MAX; 322 } 323 wantfreevnodes = desiredvnodes / 4; 324 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 325 TAILQ_INIT(&vnode_free_list); 326 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 327 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 328 NULL, NULL, UMA_ALIGN_PTR, 0); 329 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 330 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 331 /* 332 * Initialize the filesystem syncer. 333 */ 334 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 335 &syncer_mask); 336 syncer_maxdelay = syncer_mask + 1; 337 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 338 cv_init(&sync_wakeup, "syncer"); 339 for (i = 1; i <= sizeof(struct vnode); i <<= 1) 340 vnsz2log++; 341 vnsz2log--; 342 } 343 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL); 344 345 346 /* 347 * Mark a mount point as busy. Used to synchronize access and to delay 348 * unmounting. Eventually, mountlist_mtx is not released on failure. 349 * 350 * vfs_busy() is a custom lock, it can block the caller. 351 * vfs_busy() only sleeps if the unmount is active on the mount point. 352 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any 353 * vnode belonging to mp. 354 * 355 * Lookup uses vfs_busy() to traverse mount points. 356 * root fs var fs 357 * / vnode lock A / vnode lock (/var) D 358 * /var vnode lock B /log vnode lock(/var/log) E 359 * vfs_busy lock C vfs_busy lock F 360 * 361 * Within each file system, the lock order is C->A->B and F->D->E. 362 * 363 * When traversing across mounts, the system follows that lock order: 364 * 365 * C->A->B 366 * | 367 * +->F->D->E 368 * 369 * The lookup() process for namei("/var") illustrates the process: 370 * VOP_LOOKUP() obtains B while A is held 371 * vfs_busy() obtains a shared lock on F while A and B are held 372 * vput() releases lock on B 373 * vput() releases lock on A 374 * VFS_ROOT() obtains lock on D while shared lock on F is held 375 * vfs_unbusy() releases shared lock on F 376 * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A. 377 * Attempt to lock A (instead of vp_crossmp) while D is held would 378 * violate the global order, causing deadlocks. 379 * 380 * dounmount() locks B while F is drained. 381 */ 382 int 383 vfs_busy(struct mount *mp, int flags) 384 { 385 386 MPASS((flags & ~MBF_MASK) == 0); 387 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags); 388 389 MNT_ILOCK(mp); 390 MNT_REF(mp); 391 /* 392 * If mount point is currenly being unmounted, sleep until the 393 * mount point fate is decided. If thread doing the unmounting fails, 394 * it will clear MNTK_UNMOUNT flag before waking us up, indicating 395 * that this mount point has survived the unmount attempt and vfs_busy 396 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE 397 * flag in addition to MNTK_UNMOUNT, indicating that mount point is 398 * about to be really destroyed. vfs_busy needs to release its 399 * reference on the mount point in this case and return with ENOENT, 400 * telling the caller that mount mount it tried to busy is no longer 401 * valid. 402 */ 403 while (mp->mnt_kern_flag & MNTK_UNMOUNT) { 404 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) { 405 MNT_REL(mp); 406 MNT_IUNLOCK(mp); 407 CTR1(KTR_VFS, "%s: failed busying before sleeping", 408 __func__); 409 return (ENOENT); 410 } 411 if (flags & MBF_MNTLSTLOCK) 412 mtx_unlock(&mountlist_mtx); 413 mp->mnt_kern_flag |= MNTK_MWAIT; 414 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0); 415 if (flags & MBF_MNTLSTLOCK) 416 mtx_lock(&mountlist_mtx); 417 MNT_ILOCK(mp); 418 } 419 if (flags & MBF_MNTLSTLOCK) 420 mtx_unlock(&mountlist_mtx); 421 mp->mnt_lockref++; 422 MNT_IUNLOCK(mp); 423 return (0); 424 } 425 426 /* 427 * Free a busy filesystem. 428 */ 429 void 430 vfs_unbusy(struct mount *mp) 431 { 432 433 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 434 MNT_ILOCK(mp); 435 MNT_REL(mp); 436 KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref")); 437 mp->mnt_lockref--; 438 if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) { 439 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT); 440 CTR1(KTR_VFS, "%s: waking up waiters", __func__); 441 mp->mnt_kern_flag &= ~MNTK_DRAINING; 442 wakeup(&mp->mnt_lockref); 443 } 444 MNT_IUNLOCK(mp); 445 } 446 447 /* 448 * Lookup a mount point by filesystem identifier. 449 */ 450 struct mount * 451 vfs_getvfs(fsid_t *fsid) 452 { 453 struct mount *mp; 454 455 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 456 mtx_lock(&mountlist_mtx); 457 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 458 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 459 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 460 vfs_ref(mp); 461 mtx_unlock(&mountlist_mtx); 462 return (mp); 463 } 464 } 465 mtx_unlock(&mountlist_mtx); 466 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 467 return ((struct mount *) 0); 468 } 469 470 /* 471 * Lookup a mount point by filesystem identifier, busying it before 472 * returning. 473 */ 474 struct mount * 475 vfs_busyfs(fsid_t *fsid) 476 { 477 struct mount *mp; 478 int error; 479 480 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); 481 mtx_lock(&mountlist_mtx); 482 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 483 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 484 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 485 error = vfs_busy(mp, MBF_MNTLSTLOCK); 486 if (error) { 487 mtx_unlock(&mountlist_mtx); 488 return (NULL); 489 } 490 return (mp); 491 } 492 } 493 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); 494 mtx_unlock(&mountlist_mtx); 495 return ((struct mount *) 0); 496 } 497 498 /* 499 * Check if a user can access privileged mount options. 500 */ 501 int 502 vfs_suser(struct mount *mp, struct thread *td) 503 { 504 int error; 505 506 /* 507 * If the thread is jailed, but this is not a jail-friendly file 508 * system, deny immediately. 509 */ 510 if (!(mp->mnt_vfc->vfc_flags & VFCF_JAIL) && jailed(td->td_ucred)) 511 return (EPERM); 512 513 /* 514 * If the file system was mounted outside the jail of the calling 515 * thread, deny immediately. 516 */ 517 if (prison_check(td->td_ucred, mp->mnt_cred) != 0) 518 return (EPERM); 519 520 /* 521 * If file system supports delegated administration, we don't check 522 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified 523 * by the file system itself. 524 * If this is not the user that did original mount, we check for 525 * the PRIV_VFS_MOUNT_OWNER privilege. 526 */ 527 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) && 528 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 529 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) 530 return (error); 531 } 532 return (0); 533 } 534 535 /* 536 * Get a new unique fsid. Try to make its val[0] unique, since this value 537 * will be used to create fake device numbers for stat(). Also try (but 538 * not so hard) make its val[0] unique mod 2^16, since some emulators only 539 * support 16-bit device numbers. We end up with unique val[0]'s for the 540 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 541 * 542 * Keep in mind that several mounts may be running in parallel. Starting 543 * the search one past where the previous search terminated is both a 544 * micro-optimization and a defense against returning the same fsid to 545 * different mounts. 546 */ 547 void 548 vfs_getnewfsid(struct mount *mp) 549 { 550 static uint16_t mntid_base; 551 struct mount *nmp; 552 fsid_t tfsid; 553 int mtype; 554 555 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 556 mtx_lock(&mntid_mtx); 557 mtype = mp->mnt_vfc->vfc_typenum; 558 tfsid.val[1] = mtype; 559 mtype = (mtype & 0xFF) << 24; 560 for (;;) { 561 tfsid.val[0] = makedev(255, 562 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 563 mntid_base++; 564 if ((nmp = vfs_getvfs(&tfsid)) == NULL) 565 break; 566 vfs_rel(nmp); 567 } 568 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 569 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 570 mtx_unlock(&mntid_mtx); 571 } 572 573 /* 574 * Knob to control the precision of file timestamps: 575 * 576 * 0 = seconds only; nanoseconds zeroed. 577 * 1 = seconds and nanoseconds, accurate within 1/HZ. 578 * 2 = seconds and nanoseconds, truncated to microseconds. 579 * >=3 = seconds and nanoseconds, maximum precision. 580 */ 581 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 582 583 static int timestamp_precision = TSP_SEC; 584 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 585 ×tamp_precision, 0, "File timestamp precision (0: seconds, " 586 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to ms, " 587 "3+: sec + ns (max. precision))"); 588 589 /* 590 * Get a current timestamp. 591 */ 592 void 593 vfs_timestamp(struct timespec *tsp) 594 { 595 struct timeval tv; 596 597 switch (timestamp_precision) { 598 case TSP_SEC: 599 tsp->tv_sec = time_second; 600 tsp->tv_nsec = 0; 601 break; 602 case TSP_HZ: 603 getnanotime(tsp); 604 break; 605 case TSP_USEC: 606 microtime(&tv); 607 TIMEVAL_TO_TIMESPEC(&tv, tsp); 608 break; 609 case TSP_NSEC: 610 default: 611 nanotime(tsp); 612 break; 613 } 614 } 615 616 /* 617 * Set vnode attributes to VNOVAL 618 */ 619 void 620 vattr_null(struct vattr *vap) 621 { 622 623 vap->va_type = VNON; 624 vap->va_size = VNOVAL; 625 vap->va_bytes = VNOVAL; 626 vap->va_mode = VNOVAL; 627 vap->va_nlink = VNOVAL; 628 vap->va_uid = VNOVAL; 629 vap->va_gid = VNOVAL; 630 vap->va_fsid = VNOVAL; 631 vap->va_fileid = VNOVAL; 632 vap->va_blocksize = VNOVAL; 633 vap->va_rdev = VNOVAL; 634 vap->va_atime.tv_sec = VNOVAL; 635 vap->va_atime.tv_nsec = VNOVAL; 636 vap->va_mtime.tv_sec = VNOVAL; 637 vap->va_mtime.tv_nsec = VNOVAL; 638 vap->va_ctime.tv_sec = VNOVAL; 639 vap->va_ctime.tv_nsec = VNOVAL; 640 vap->va_birthtime.tv_sec = VNOVAL; 641 vap->va_birthtime.tv_nsec = VNOVAL; 642 vap->va_flags = VNOVAL; 643 vap->va_gen = VNOVAL; 644 vap->va_vaflags = 0; 645 } 646 647 /* 648 * This routine is called when we have too many vnodes. It attempts 649 * to free <count> vnodes and will potentially free vnodes that still 650 * have VM backing store (VM backing store is typically the cause 651 * of a vnode blowout so we want to do this). Therefore, this operation 652 * is not considered cheap. 653 * 654 * A number of conditions may prevent a vnode from being reclaimed. 655 * the buffer cache may have references on the vnode, a directory 656 * vnode may still have references due to the namei cache representing 657 * underlying files, or the vnode may be in active use. It is not 658 * desireable to reuse such vnodes. These conditions may cause the 659 * number of vnodes to reach some minimum value regardless of what 660 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 661 */ 662 static int 663 vlrureclaim(struct mount *mp) 664 { 665 struct vnode *vp; 666 int done; 667 int trigger; 668 int usevnodes; 669 int count; 670 671 /* 672 * Calculate the trigger point, don't allow user 673 * screwups to blow us up. This prevents us from 674 * recycling vnodes with lots of resident pages. We 675 * aren't trying to free memory, we are trying to 676 * free vnodes. 677 */ 678 usevnodes = desiredvnodes; 679 if (usevnodes <= 0) 680 usevnodes = 1; 681 trigger = cnt.v_page_count * 2 / usevnodes; 682 done = 0; 683 vn_start_write(NULL, &mp, V_WAIT); 684 MNT_ILOCK(mp); 685 count = mp->mnt_nvnodelistsize / 10 + 1; 686 while (count != 0) { 687 vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 688 while (vp != NULL && vp->v_type == VMARKER) 689 vp = TAILQ_NEXT(vp, v_nmntvnodes); 690 if (vp == NULL) 691 break; 692 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 693 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 694 --count; 695 if (!VI_TRYLOCK(vp)) 696 goto next_iter; 697 /* 698 * If it's been deconstructed already, it's still 699 * referenced, or it exceeds the trigger, skip it. 700 */ 701 if (vp->v_usecount || 702 (!vlru_allow_cache_src && 703 !LIST_EMPTY(&(vp)->v_cache_src)) || 704 (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL && 705 vp->v_object->resident_page_count > trigger)) { 706 VI_UNLOCK(vp); 707 goto next_iter; 708 } 709 MNT_IUNLOCK(mp); 710 vholdl(vp); 711 if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) { 712 vdrop(vp); 713 goto next_iter_mntunlocked; 714 } 715 VI_LOCK(vp); 716 /* 717 * v_usecount may have been bumped after VOP_LOCK() dropped 718 * the vnode interlock and before it was locked again. 719 * 720 * It is not necessary to recheck VI_DOOMED because it can 721 * only be set by another thread that holds both the vnode 722 * lock and vnode interlock. If another thread has the 723 * vnode lock before we get to VOP_LOCK() and obtains the 724 * vnode interlock after VOP_LOCK() drops the vnode 725 * interlock, the other thread will be unable to drop the 726 * vnode lock before our VOP_LOCK() call fails. 727 */ 728 if (vp->v_usecount || 729 (!vlru_allow_cache_src && 730 !LIST_EMPTY(&(vp)->v_cache_src)) || 731 (vp->v_object != NULL && 732 vp->v_object->resident_page_count > trigger)) { 733 VOP_UNLOCK(vp, LK_INTERLOCK); 734 goto next_iter_mntunlocked; 735 } 736 KASSERT((vp->v_iflag & VI_DOOMED) == 0, 737 ("VI_DOOMED unexpectedly detected in vlrureclaim()")); 738 vgonel(vp); 739 VOP_UNLOCK(vp, 0); 740 vdropl(vp); 741 done++; 742 next_iter_mntunlocked: 743 if (!should_yield()) 744 goto relock_mnt; 745 goto yield; 746 next_iter: 747 if (!should_yield()) 748 continue; 749 MNT_IUNLOCK(mp); 750 yield: 751 kern_yield(PRI_USER); 752 relock_mnt: 753 MNT_ILOCK(mp); 754 } 755 MNT_IUNLOCK(mp); 756 vn_finished_write(mp); 757 return done; 758 } 759 760 /* 761 * Attempt to keep the free list at wantfreevnodes length. 762 */ 763 static void 764 vnlru_free(int count) 765 { 766 struct vnode *vp; 767 768 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 769 for (; count > 0; count--) { 770 vp = TAILQ_FIRST(&vnode_free_list); 771 /* 772 * The list can be modified while the free_list_mtx 773 * has been dropped and vp could be NULL here. 774 */ 775 if (!vp) 776 break; 777 VNASSERT(vp->v_op != NULL, vp, 778 ("vnlru_free: vnode already reclaimed.")); 779 KASSERT((vp->v_iflag & VI_FREE) != 0, 780 ("Removing vnode not on freelist")); 781 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 782 ("Mangling active vnode")); 783 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist); 784 /* 785 * Don't recycle if we can't get the interlock. 786 */ 787 if (!VI_TRYLOCK(vp)) { 788 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist); 789 continue; 790 } 791 VNASSERT(VCANRECYCLE(vp), vp, 792 ("vp inconsistent on freelist")); 793 freevnodes--; 794 vp->v_iflag &= ~VI_FREE; 795 vholdl(vp); 796 mtx_unlock(&vnode_free_list_mtx); 797 VI_UNLOCK(vp); 798 vtryrecycle(vp); 799 /* 800 * If the recycled succeeded this vdrop will actually free 801 * the vnode. If not it will simply place it back on 802 * the free list. 803 */ 804 vdrop(vp); 805 mtx_lock(&vnode_free_list_mtx); 806 } 807 } 808 /* 809 * Attempt to recycle vnodes in a context that is always safe to block. 810 * Calling vlrurecycle() from the bowels of filesystem code has some 811 * interesting deadlock problems. 812 */ 813 static struct proc *vnlruproc; 814 static int vnlruproc_sig; 815 816 static void 817 vnlru_proc(void) 818 { 819 struct mount *mp, *nmp; 820 int done; 821 struct proc *p = vnlruproc; 822 823 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 824 SHUTDOWN_PRI_FIRST); 825 826 for (;;) { 827 kproc_suspend_check(p); 828 mtx_lock(&vnode_free_list_mtx); 829 if (freevnodes > wantfreevnodes) 830 vnlru_free(freevnodes - wantfreevnodes); 831 if (numvnodes <= desiredvnodes * 9 / 10) { 832 vnlruproc_sig = 0; 833 wakeup(&vnlruproc_sig); 834 msleep(vnlruproc, &vnode_free_list_mtx, 835 PVFS|PDROP, "vlruwt", hz); 836 continue; 837 } 838 mtx_unlock(&vnode_free_list_mtx); 839 done = 0; 840 mtx_lock(&mountlist_mtx); 841 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 842 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) { 843 nmp = TAILQ_NEXT(mp, mnt_list); 844 continue; 845 } 846 done += vlrureclaim(mp); 847 mtx_lock(&mountlist_mtx); 848 nmp = TAILQ_NEXT(mp, mnt_list); 849 vfs_unbusy(mp); 850 } 851 mtx_unlock(&mountlist_mtx); 852 if (done == 0) { 853 #if 0 854 /* These messages are temporary debugging aids */ 855 if (vnlru_nowhere < 5) 856 printf("vnlru process getting nowhere..\n"); 857 else if (vnlru_nowhere == 5) 858 printf("vnlru process messages stopped.\n"); 859 #endif 860 vnlru_nowhere++; 861 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 862 } else 863 kern_yield(PRI_USER); 864 } 865 } 866 867 static struct kproc_desc vnlru_kp = { 868 "vnlru", 869 vnlru_proc, 870 &vnlruproc 871 }; 872 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, 873 &vnlru_kp); 874 875 /* 876 * Routines having to do with the management of the vnode table. 877 */ 878 879 /* 880 * Try to recycle a freed vnode. We abort if anyone picks up a reference 881 * before we actually vgone(). This function must be called with the vnode 882 * held to prevent the vnode from being returned to the free list midway 883 * through vgone(). 884 */ 885 static int 886 vtryrecycle(struct vnode *vp) 887 { 888 struct mount *vnmp; 889 890 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 891 VNASSERT(vp->v_holdcnt, vp, 892 ("vtryrecycle: Recycling vp %p without a reference.", vp)); 893 /* 894 * This vnode may found and locked via some other list, if so we 895 * can't recycle it yet. 896 */ 897 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 898 CTR2(KTR_VFS, 899 "%s: impossible to recycle, vp %p lock is already held", 900 __func__, vp); 901 return (EWOULDBLOCK); 902 } 903 /* 904 * Don't recycle if its filesystem is being suspended. 905 */ 906 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 907 VOP_UNLOCK(vp, 0); 908 CTR2(KTR_VFS, 909 "%s: impossible to recycle, cannot start the write for %p", 910 __func__, vp); 911 return (EBUSY); 912 } 913 /* 914 * If we got this far, we need to acquire the interlock and see if 915 * anyone picked up this vnode from another list. If not, we will 916 * mark it with DOOMED via vgonel() so that anyone who does find it 917 * will skip over it. 918 */ 919 VI_LOCK(vp); 920 if (vp->v_usecount) { 921 VOP_UNLOCK(vp, LK_INTERLOCK); 922 vn_finished_write(vnmp); 923 CTR2(KTR_VFS, 924 "%s: impossible to recycle, %p is already referenced", 925 __func__, vp); 926 return (EBUSY); 927 } 928 if ((vp->v_iflag & VI_DOOMED) == 0) 929 vgonel(vp); 930 VOP_UNLOCK(vp, LK_INTERLOCK); 931 vn_finished_write(vnmp); 932 return (0); 933 } 934 935 /* 936 * Wait for available vnodes. 937 */ 938 static int 939 getnewvnode_wait(int suspended) 940 { 941 942 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 943 if (numvnodes > desiredvnodes) { 944 if (suspended) { 945 /* 946 * File system is beeing suspended, we cannot risk a 947 * deadlock here, so allocate new vnode anyway. 948 */ 949 if (freevnodes > wantfreevnodes) 950 vnlru_free(freevnodes - wantfreevnodes); 951 return (0); 952 } 953 if (vnlruproc_sig == 0) { 954 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 955 wakeup(vnlruproc); 956 } 957 msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS, 958 "vlruwk", hz); 959 } 960 return (numvnodes > desiredvnodes ? ENFILE : 0); 961 } 962 963 void 964 getnewvnode_reserve(u_int count) 965 { 966 struct thread *td; 967 968 td = curthread; 969 mtx_lock(&vnode_free_list_mtx); 970 while (count > 0) { 971 if (getnewvnode_wait(0) == 0) { 972 count--; 973 td->td_vp_reserv++; 974 numvnodes++; 975 } 976 } 977 mtx_unlock(&vnode_free_list_mtx); 978 } 979 980 void 981 getnewvnode_drop_reserve(void) 982 { 983 struct thread *td; 984 985 td = curthread; 986 mtx_lock(&vnode_free_list_mtx); 987 KASSERT(numvnodes >= td->td_vp_reserv, ("reserve too large")); 988 numvnodes -= td->td_vp_reserv; 989 mtx_unlock(&vnode_free_list_mtx); 990 td->td_vp_reserv = 0; 991 } 992 993 /* 994 * Return the next vnode from the free list. 995 */ 996 int 997 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, 998 struct vnode **vpp) 999 { 1000 struct vnode *vp; 1001 struct bufobj *bo; 1002 struct thread *td; 1003 int error; 1004 1005 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag); 1006 vp = NULL; 1007 td = curthread; 1008 if (td->td_vp_reserv > 0) { 1009 td->td_vp_reserv -= 1; 1010 goto alloc; 1011 } 1012 mtx_lock(&vnode_free_list_mtx); 1013 /* 1014 * Lend our context to reclaim vnodes if they've exceeded the max. 1015 */ 1016 if (freevnodes > wantfreevnodes) 1017 vnlru_free(1); 1018 error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag & 1019 MNTK_SUSPEND)); 1020 #if 0 /* XXX Not all VFS_VGET/ffs_vget callers check returns. */ 1021 if (error != 0) { 1022 mtx_unlock(&vnode_free_list_mtx); 1023 return (error); 1024 } 1025 #endif 1026 numvnodes++; 1027 mtx_unlock(&vnode_free_list_mtx); 1028 alloc: 1029 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); 1030 /* 1031 * Setup locks. 1032 */ 1033 vp->v_vnlock = &vp->v_lock; 1034 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 1035 /* 1036 * By default, don't allow shared locks unless filesystems 1037 * opt-in. 1038 */ 1039 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE); 1040 /* 1041 * Initialize bufobj. 1042 */ 1043 bo = &vp->v_bufobj; 1044 bo->__bo_vnode = vp; 1045 mtx_init(BO_MTX(bo), "bufobj interlock", NULL, MTX_DEF); 1046 bo->bo_ops = &buf_ops_bio; 1047 bo->bo_private = vp; 1048 TAILQ_INIT(&bo->bo_clean.bv_hd); 1049 TAILQ_INIT(&bo->bo_dirty.bv_hd); 1050 /* 1051 * Initialize namecache. 1052 */ 1053 LIST_INIT(&vp->v_cache_src); 1054 TAILQ_INIT(&vp->v_cache_dst); 1055 /* 1056 * Finalize various vnode identity bits. 1057 */ 1058 vp->v_type = VNON; 1059 vp->v_tag = tag; 1060 vp->v_op = vops; 1061 v_incr_usecount(vp); 1062 vp->v_data = NULL; 1063 #ifdef MAC 1064 mac_vnode_init(vp); 1065 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 1066 mac_vnode_associate_singlelabel(mp, vp); 1067 else if (mp == NULL && vops != &dead_vnodeops) 1068 printf("NULL mp in getnewvnode()\n"); 1069 #endif 1070 if (mp != NULL) { 1071 bo->bo_bsize = mp->mnt_stat.f_iosize; 1072 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) 1073 vp->v_vflag |= VV_NOKNOTE; 1074 } 1075 rangelock_init(&vp->v_rl); 1076 1077 /* 1078 * For the filesystems which do not use vfs_hash_insert(), 1079 * still initialize v_hash to have vfs_hash_index() useful. 1080 * E.g., nullfs uses vfs_hash_index() on the lower vnode for 1081 * its own hashing. 1082 */ 1083 vp->v_hash = (uintptr_t)vp >> vnsz2log; 1084 1085 *vpp = vp; 1086 return (0); 1087 } 1088 1089 /* 1090 * Delete from old mount point vnode list, if on one. 1091 */ 1092 static void 1093 delmntque(struct vnode *vp) 1094 { 1095 struct mount *mp; 1096 int active; 1097 1098 mp = vp->v_mount; 1099 if (mp == NULL) 1100 return; 1101 MNT_ILOCK(mp); 1102 VI_LOCK(vp); 1103 KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize, 1104 ("Active vnode list size %d > Vnode list size %d", 1105 mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize)); 1106 active = vp->v_iflag & VI_ACTIVE; 1107 vp->v_iflag &= ~VI_ACTIVE; 1108 if (active) { 1109 mtx_lock(&vnode_free_list_mtx); 1110 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist); 1111 mp->mnt_activevnodelistsize--; 1112 mtx_unlock(&vnode_free_list_mtx); 1113 } 1114 vp->v_mount = NULL; 1115 VI_UNLOCK(vp); 1116 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 1117 ("bad mount point vnode list size")); 1118 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1119 mp->mnt_nvnodelistsize--; 1120 MNT_REL(mp); 1121 MNT_IUNLOCK(mp); 1122 } 1123 1124 static void 1125 insmntque_stddtr(struct vnode *vp, void *dtr_arg) 1126 { 1127 1128 vp->v_data = NULL; 1129 vp->v_op = &dead_vnodeops; 1130 vgone(vp); 1131 vput(vp); 1132 } 1133 1134 /* 1135 * Insert into list of vnodes for the new mount point, if available. 1136 */ 1137 int 1138 insmntque1(struct vnode *vp, struct mount *mp, 1139 void (*dtr)(struct vnode *, void *), void *dtr_arg) 1140 { 1141 1142 KASSERT(vp->v_mount == NULL, 1143 ("insmntque: vnode already on per mount vnode list")); 1144 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 1145 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); 1146 1147 /* 1148 * We acquire the vnode interlock early to ensure that the 1149 * vnode cannot be recycled by another process releasing a 1150 * holdcnt on it before we get it on both the vnode list 1151 * and the active vnode list. The mount mutex protects only 1152 * manipulation of the vnode list and the vnode freelist 1153 * mutex protects only manipulation of the active vnode list. 1154 * Hence the need to hold the vnode interlock throughout. 1155 */ 1156 MNT_ILOCK(mp); 1157 VI_LOCK(vp); 1158 if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 && 1159 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || 1160 mp->mnt_nvnodelistsize == 0)) && 1161 (vp->v_vflag & VV_FORCEINSMQ) == 0) { 1162 VI_UNLOCK(vp); 1163 MNT_IUNLOCK(mp); 1164 if (dtr != NULL) 1165 dtr(vp, dtr_arg); 1166 return (EBUSY); 1167 } 1168 vp->v_mount = mp; 1169 MNT_REF(mp); 1170 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1171 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 1172 ("neg mount point vnode list size")); 1173 mp->mnt_nvnodelistsize++; 1174 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 1175 ("Activating already active vnode")); 1176 vp->v_iflag |= VI_ACTIVE; 1177 mtx_lock(&vnode_free_list_mtx); 1178 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist); 1179 mp->mnt_activevnodelistsize++; 1180 mtx_unlock(&vnode_free_list_mtx); 1181 VI_UNLOCK(vp); 1182 MNT_IUNLOCK(mp); 1183 return (0); 1184 } 1185 1186 int 1187 insmntque(struct vnode *vp, struct mount *mp) 1188 { 1189 1190 return (insmntque1(vp, mp, insmntque_stddtr, NULL)); 1191 } 1192 1193 /* 1194 * Flush out and invalidate all buffers associated with a bufobj 1195 * Called with the underlying object locked. 1196 */ 1197 int 1198 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) 1199 { 1200 int error; 1201 1202 BO_LOCK(bo); 1203 if (flags & V_SAVE) { 1204 error = bufobj_wwait(bo, slpflag, slptimeo); 1205 if (error) { 1206 BO_UNLOCK(bo); 1207 return (error); 1208 } 1209 if (bo->bo_dirty.bv_cnt > 0) { 1210 BO_UNLOCK(bo); 1211 if ((error = BO_SYNC(bo, MNT_WAIT)) != 0) 1212 return (error); 1213 /* 1214 * XXX We could save a lock/unlock if this was only 1215 * enabled under INVARIANTS 1216 */ 1217 BO_LOCK(bo); 1218 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 1219 panic("vinvalbuf: dirty bufs"); 1220 } 1221 } 1222 /* 1223 * If you alter this loop please notice that interlock is dropped and 1224 * reacquired in flushbuflist. Special care is needed to ensure that 1225 * no race conditions occur from this. 1226 */ 1227 do { 1228 error = flushbuflist(&bo->bo_clean, 1229 flags, bo, slpflag, slptimeo); 1230 if (error == 0 && !(flags & V_CLEANONLY)) 1231 error = flushbuflist(&bo->bo_dirty, 1232 flags, bo, slpflag, slptimeo); 1233 if (error != 0 && error != EAGAIN) { 1234 BO_UNLOCK(bo); 1235 return (error); 1236 } 1237 } while (error != 0); 1238 1239 /* 1240 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1241 * have write I/O in-progress but if there is a VM object then the 1242 * VM object can also have read-I/O in-progress. 1243 */ 1244 do { 1245 bufobj_wwait(bo, 0, 0); 1246 BO_UNLOCK(bo); 1247 if (bo->bo_object != NULL) { 1248 VM_OBJECT_WLOCK(bo->bo_object); 1249 vm_object_pip_wait(bo->bo_object, "bovlbx"); 1250 VM_OBJECT_WUNLOCK(bo->bo_object); 1251 } 1252 BO_LOCK(bo); 1253 } while (bo->bo_numoutput > 0); 1254 BO_UNLOCK(bo); 1255 1256 /* 1257 * Destroy the copy in the VM cache, too. 1258 */ 1259 if (bo->bo_object != NULL && 1260 (flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0) { 1261 VM_OBJECT_WLOCK(bo->bo_object); 1262 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? 1263 OBJPR_CLEANONLY : 0); 1264 VM_OBJECT_WUNLOCK(bo->bo_object); 1265 } 1266 1267 #ifdef INVARIANTS 1268 BO_LOCK(bo); 1269 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0 && 1270 (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0)) 1271 panic("vinvalbuf: flush failed"); 1272 BO_UNLOCK(bo); 1273 #endif 1274 return (0); 1275 } 1276 1277 /* 1278 * Flush out and invalidate all buffers associated with a vnode. 1279 * Called with the underlying object locked. 1280 */ 1281 int 1282 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 1283 { 1284 1285 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 1286 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1287 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); 1288 } 1289 1290 /* 1291 * Flush out buffers on the specified list. 1292 * 1293 */ 1294 static int 1295 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 1296 int slptimeo) 1297 { 1298 struct buf *bp, *nbp; 1299 int retval, error; 1300 daddr_t lblkno; 1301 b_xflags_t xflags; 1302 1303 ASSERT_BO_LOCKED(bo); 1304 1305 retval = 0; 1306 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 1307 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1308 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1309 continue; 1310 } 1311 lblkno = 0; 1312 xflags = 0; 1313 if (nbp != NULL) { 1314 lblkno = nbp->b_lblkno; 1315 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN); 1316 } 1317 retval = EAGAIN; 1318 error = BUF_TIMELOCK(bp, 1319 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo), 1320 "flushbuf", slpflag, slptimeo); 1321 if (error) { 1322 BO_LOCK(bo); 1323 return (error != ENOLCK ? error : EAGAIN); 1324 } 1325 KASSERT(bp->b_bufobj == bo, 1326 ("bp %p wrong b_bufobj %p should be %p", 1327 bp, bp->b_bufobj, bo)); 1328 if (bp->b_bufobj != bo) { /* XXX: necessary ? */ 1329 BUF_UNLOCK(bp); 1330 BO_LOCK(bo); 1331 return (EAGAIN); 1332 } 1333 /* 1334 * XXX Since there are no node locks for NFS, I 1335 * believe there is a slight chance that a delayed 1336 * write will occur while sleeping just above, so 1337 * check for it. 1338 */ 1339 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1340 (flags & V_SAVE)) { 1341 BO_LOCK(bo); 1342 bremfree(bp); 1343 BO_UNLOCK(bo); 1344 bp->b_flags |= B_ASYNC; 1345 bwrite(bp); 1346 BO_LOCK(bo); 1347 return (EAGAIN); /* XXX: why not loop ? */ 1348 } 1349 BO_LOCK(bo); 1350 bremfree(bp); 1351 BO_UNLOCK(bo); 1352 bp->b_flags |= (B_INVAL | B_RELBUF); 1353 bp->b_flags &= ~B_ASYNC; 1354 brelse(bp); 1355 BO_LOCK(bo); 1356 if (nbp != NULL && 1357 (nbp->b_bufobj != bo || 1358 nbp->b_lblkno != lblkno || 1359 (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) != xflags)) 1360 break; /* nbp invalid */ 1361 } 1362 return (retval); 1363 } 1364 1365 /* 1366 * Truncate a file's buffer and pages to a specified length. This 1367 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1368 * sync activity. 1369 */ 1370 int 1371 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize) 1372 { 1373 struct buf *bp, *nbp; 1374 int anyfreed; 1375 int trunclbn; 1376 struct bufobj *bo; 1377 1378 CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__, 1379 vp, cred, blksize, (uintmax_t)length); 1380 1381 /* 1382 * Round up to the *next* lbn. 1383 */ 1384 trunclbn = (length + blksize - 1) / blksize; 1385 1386 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1387 restart: 1388 bo = &vp->v_bufobj; 1389 BO_LOCK(bo); 1390 anyfreed = 1; 1391 for (;anyfreed;) { 1392 anyfreed = 0; 1393 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 1394 if (bp->b_lblkno < trunclbn) 1395 continue; 1396 if (BUF_LOCK(bp, 1397 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1398 BO_MTX(bo)) == ENOLCK) 1399 goto restart; 1400 1401 BO_LOCK(bo); 1402 bremfree(bp); 1403 BO_UNLOCK(bo); 1404 bp->b_flags |= (B_INVAL | B_RELBUF); 1405 bp->b_flags &= ~B_ASYNC; 1406 brelse(bp); 1407 anyfreed = 1; 1408 1409 BO_LOCK(bo); 1410 if (nbp != NULL && 1411 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1412 (nbp->b_vp != vp) || 1413 (nbp->b_flags & B_DELWRI))) { 1414 BO_UNLOCK(bo); 1415 goto restart; 1416 } 1417 } 1418 1419 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1420 if (bp->b_lblkno < trunclbn) 1421 continue; 1422 if (BUF_LOCK(bp, 1423 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1424 BO_MTX(bo)) == ENOLCK) 1425 goto restart; 1426 BO_LOCK(bo); 1427 bremfree(bp); 1428 BO_UNLOCK(bo); 1429 bp->b_flags |= (B_INVAL | B_RELBUF); 1430 bp->b_flags &= ~B_ASYNC; 1431 brelse(bp); 1432 anyfreed = 1; 1433 1434 BO_LOCK(bo); 1435 if (nbp != NULL && 1436 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1437 (nbp->b_vp != vp) || 1438 (nbp->b_flags & B_DELWRI) == 0)) { 1439 BO_UNLOCK(bo); 1440 goto restart; 1441 } 1442 } 1443 } 1444 1445 if (length > 0) { 1446 restartsync: 1447 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1448 if (bp->b_lblkno > 0) 1449 continue; 1450 /* 1451 * Since we hold the vnode lock this should only 1452 * fail if we're racing with the buf daemon. 1453 */ 1454 if (BUF_LOCK(bp, 1455 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1456 BO_MTX(bo)) == ENOLCK) { 1457 goto restart; 1458 } 1459 VNASSERT((bp->b_flags & B_DELWRI), vp, 1460 ("buf(%p) on dirty queue without DELWRI", bp)); 1461 1462 BO_LOCK(bo); 1463 bremfree(bp); 1464 BO_UNLOCK(bo); 1465 bawrite(bp); 1466 BO_LOCK(bo); 1467 goto restartsync; 1468 } 1469 } 1470 1471 bufobj_wwait(bo, 0, 0); 1472 BO_UNLOCK(bo); 1473 vnode_pager_setsize(vp, length); 1474 1475 return (0); 1476 } 1477 1478 /* 1479 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1480 * a vnode. 1481 * 1482 * NOTE: We have to deal with the special case of a background bitmap 1483 * buffer, a situation where two buffers will have the same logical 1484 * block offset. We want (1) only the foreground buffer to be accessed 1485 * in a lookup and (2) must differentiate between the foreground and 1486 * background buffer in the splay tree algorithm because the splay 1487 * tree cannot normally handle multiple entities with the same 'index'. 1488 * We accomplish this by adding differentiating flags to the splay tree's 1489 * numerical domain. 1490 */ 1491 static 1492 struct buf * 1493 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1494 { 1495 struct buf dummy; 1496 struct buf *lefttreemax, *righttreemin, *y; 1497 1498 if (root == NULL) 1499 return (NULL); 1500 lefttreemax = righttreemin = &dummy; 1501 for (;;) { 1502 if (lblkno < root->b_lblkno) { 1503 if ((y = root->b_left) == NULL) 1504 break; 1505 if (lblkno < y->b_lblkno) { 1506 /* Rotate right. */ 1507 root->b_left = y->b_right; 1508 y->b_right = root; 1509 root = y; 1510 if ((y = root->b_left) == NULL) 1511 break; 1512 } 1513 /* Link into the new root's right tree. */ 1514 righttreemin->b_left = root; 1515 righttreemin = root; 1516 } else if (lblkno > root->b_lblkno) { 1517 if ((y = root->b_right) == NULL) 1518 break; 1519 if (lblkno > y->b_lblkno) { 1520 /* Rotate left. */ 1521 root->b_right = y->b_left; 1522 y->b_left = root; 1523 root = y; 1524 if ((y = root->b_right) == NULL) 1525 break; 1526 } 1527 /* Link into the new root's left tree. */ 1528 lefttreemax->b_right = root; 1529 lefttreemax = root; 1530 } else { 1531 break; 1532 } 1533 root = y; 1534 } 1535 /* Assemble the new root. */ 1536 lefttreemax->b_right = root->b_left; 1537 righttreemin->b_left = root->b_right; 1538 root->b_left = dummy.b_right; 1539 root->b_right = dummy.b_left; 1540 return (root); 1541 } 1542 1543 static void 1544 buf_vlist_remove(struct buf *bp) 1545 { 1546 struct buf *root; 1547 struct bufv *bv; 1548 1549 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1550 ASSERT_BO_LOCKED(bp->b_bufobj); 1551 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) != 1552 (BX_VNDIRTY|BX_VNCLEAN), 1553 ("buf_vlist_remove: Buf %p is on two lists", bp)); 1554 if (bp->b_xflags & BX_VNDIRTY) 1555 bv = &bp->b_bufobj->bo_dirty; 1556 else 1557 bv = &bp->b_bufobj->bo_clean; 1558 if (bp != bv->bv_root) { 1559 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1560 KASSERT(root == bp, ("splay lookup failed in remove")); 1561 } 1562 if (bp->b_left == NULL) { 1563 root = bp->b_right; 1564 } else { 1565 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1566 root->b_right = bp->b_right; 1567 } 1568 bv->bv_root = root; 1569 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 1570 bv->bv_cnt--; 1571 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1572 } 1573 1574 /* 1575 * Add the buffer to the sorted clean or dirty block list using a 1576 * splay tree algorithm. 1577 * 1578 * NOTE: xflags is passed as a constant, optimizing this inline function! 1579 */ 1580 static void 1581 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 1582 { 1583 struct buf *root; 1584 struct bufv *bv; 1585 1586 ASSERT_BO_LOCKED(bo); 1587 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1588 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 1589 bp->b_xflags |= xflags; 1590 if (xflags & BX_VNDIRTY) 1591 bv = &bo->bo_dirty; 1592 else 1593 bv = &bo->bo_clean; 1594 1595 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1596 if (root == NULL) { 1597 bp->b_left = NULL; 1598 bp->b_right = NULL; 1599 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 1600 } else if (bp->b_lblkno < root->b_lblkno) { 1601 bp->b_left = root->b_left; 1602 bp->b_right = root; 1603 root->b_left = NULL; 1604 TAILQ_INSERT_BEFORE(root, bp, b_bobufs); 1605 } else { 1606 bp->b_right = root->b_right; 1607 bp->b_left = root; 1608 root->b_right = NULL; 1609 TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs); 1610 } 1611 bv->bv_cnt++; 1612 bv->bv_root = bp; 1613 } 1614 1615 /* 1616 * Lookup a buffer using the splay tree. Note that we specifically avoid 1617 * shadow buffers used in background bitmap writes. 1618 * 1619 * This code isn't quite efficient as it could be because we are maintaining 1620 * two sorted lists and do not know which list the block resides in. 1621 * 1622 * During a "make buildworld" the desired buffer is found at one of 1623 * the roots more than 60% of the time. Thus, checking both roots 1624 * before performing either splay eliminates unnecessary splays on the 1625 * first tree splayed. 1626 */ 1627 struct buf * 1628 gbincore(struct bufobj *bo, daddr_t lblkno) 1629 { 1630 struct buf *bp; 1631 1632 ASSERT_BO_LOCKED(bo); 1633 if ((bp = bo->bo_clean.bv_root) != NULL && bp->b_lblkno == lblkno) 1634 return (bp); 1635 if ((bp = bo->bo_dirty.bv_root) != NULL && bp->b_lblkno == lblkno) 1636 return (bp); 1637 if ((bp = bo->bo_clean.bv_root) != NULL) { 1638 bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp); 1639 if (bp->b_lblkno == lblkno) 1640 return (bp); 1641 } 1642 if ((bp = bo->bo_dirty.bv_root) != NULL) { 1643 bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp); 1644 if (bp->b_lblkno == lblkno) 1645 return (bp); 1646 } 1647 return (NULL); 1648 } 1649 1650 /* 1651 * Associate a buffer with a vnode. 1652 */ 1653 void 1654 bgetvp(struct vnode *vp, struct buf *bp) 1655 { 1656 struct bufobj *bo; 1657 1658 bo = &vp->v_bufobj; 1659 ASSERT_BO_LOCKED(bo); 1660 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 1661 1662 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 1663 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 1664 ("bgetvp: bp already attached! %p", bp)); 1665 1666 vhold(vp); 1667 bp->b_vp = vp; 1668 bp->b_bufobj = bo; 1669 /* 1670 * Insert onto list for new vnode. 1671 */ 1672 buf_vlist_add(bp, bo, BX_VNCLEAN); 1673 } 1674 1675 /* 1676 * Disassociate a buffer from a vnode. 1677 */ 1678 void 1679 brelvp(struct buf *bp) 1680 { 1681 struct bufobj *bo; 1682 struct vnode *vp; 1683 1684 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1685 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1686 1687 /* 1688 * Delete from old vnode list, if on one. 1689 */ 1690 vp = bp->b_vp; /* XXX */ 1691 bo = bp->b_bufobj; 1692 BO_LOCK(bo); 1693 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1694 buf_vlist_remove(bp); 1695 else 1696 panic("brelvp: Buffer %p not on queue.", bp); 1697 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1698 bo->bo_flag &= ~BO_ONWORKLST; 1699 mtx_lock(&sync_mtx); 1700 LIST_REMOVE(bo, bo_synclist); 1701 syncer_worklist_len--; 1702 mtx_unlock(&sync_mtx); 1703 } 1704 bp->b_vp = NULL; 1705 bp->b_bufobj = NULL; 1706 BO_UNLOCK(bo); 1707 vdrop(vp); 1708 } 1709 1710 /* 1711 * Add an item to the syncer work queue. 1712 */ 1713 static void 1714 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 1715 { 1716 int slot; 1717 1718 ASSERT_BO_LOCKED(bo); 1719 1720 mtx_lock(&sync_mtx); 1721 if (bo->bo_flag & BO_ONWORKLST) 1722 LIST_REMOVE(bo, bo_synclist); 1723 else { 1724 bo->bo_flag |= BO_ONWORKLST; 1725 syncer_worklist_len++; 1726 } 1727 1728 if (delay > syncer_maxdelay - 2) 1729 delay = syncer_maxdelay - 2; 1730 slot = (syncer_delayno + delay) & syncer_mask; 1731 1732 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); 1733 mtx_unlock(&sync_mtx); 1734 } 1735 1736 static int 1737 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 1738 { 1739 int error, len; 1740 1741 mtx_lock(&sync_mtx); 1742 len = syncer_worklist_len - sync_vnode_count; 1743 mtx_unlock(&sync_mtx); 1744 error = SYSCTL_OUT(req, &len, sizeof(len)); 1745 return (error); 1746 } 1747 1748 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0, 1749 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 1750 1751 static struct proc *updateproc; 1752 static void sched_sync(void); 1753 static struct kproc_desc up_kp = { 1754 "syncer", 1755 sched_sync, 1756 &updateproc 1757 }; 1758 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); 1759 1760 static int 1761 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) 1762 { 1763 struct vnode *vp; 1764 struct mount *mp; 1765 1766 *bo = LIST_FIRST(slp); 1767 if (*bo == NULL) 1768 return (0); 1769 vp = (*bo)->__bo_vnode; /* XXX */ 1770 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) 1771 return (1); 1772 /* 1773 * We use vhold in case the vnode does not 1774 * successfully sync. vhold prevents the vnode from 1775 * going away when we unlock the sync_mtx so that 1776 * we can acquire the vnode interlock. 1777 */ 1778 vholdl(vp); 1779 mtx_unlock(&sync_mtx); 1780 VI_UNLOCK(vp); 1781 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1782 vdrop(vp); 1783 mtx_lock(&sync_mtx); 1784 return (*bo == LIST_FIRST(slp)); 1785 } 1786 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1787 (void) VOP_FSYNC(vp, MNT_LAZY, td); 1788 VOP_UNLOCK(vp, 0); 1789 vn_finished_write(mp); 1790 BO_LOCK(*bo); 1791 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { 1792 /* 1793 * Put us back on the worklist. The worklist 1794 * routine will remove us from our current 1795 * position and then add us back in at a later 1796 * position. 1797 */ 1798 vn_syncer_add_to_worklist(*bo, syncdelay); 1799 } 1800 BO_UNLOCK(*bo); 1801 vdrop(vp); 1802 mtx_lock(&sync_mtx); 1803 return (0); 1804 } 1805 1806 /* 1807 * System filesystem synchronizer daemon. 1808 */ 1809 static void 1810 sched_sync(void) 1811 { 1812 struct synclist *next, *slp; 1813 struct bufobj *bo; 1814 long starttime; 1815 struct thread *td = curthread; 1816 int last_work_seen; 1817 int net_worklist_len; 1818 int syncer_final_iter; 1819 int first_printf; 1820 int error; 1821 1822 last_work_seen = 0; 1823 syncer_final_iter = 0; 1824 first_printf = 1; 1825 syncer_state = SYNCER_RUNNING; 1826 starttime = time_uptime; 1827 td->td_pflags |= TDP_NORUNNINGBUF; 1828 1829 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 1830 SHUTDOWN_PRI_LAST); 1831 1832 mtx_lock(&sync_mtx); 1833 for (;;) { 1834 if (syncer_state == SYNCER_FINAL_DELAY && 1835 syncer_final_iter == 0) { 1836 mtx_unlock(&sync_mtx); 1837 kproc_suspend_check(td->td_proc); 1838 mtx_lock(&sync_mtx); 1839 } 1840 net_worklist_len = syncer_worklist_len - sync_vnode_count; 1841 if (syncer_state != SYNCER_RUNNING && 1842 starttime != time_uptime) { 1843 if (first_printf) { 1844 printf("\nSyncing disks, vnodes remaining..."); 1845 first_printf = 0; 1846 } 1847 printf("%d ", net_worklist_len); 1848 } 1849 starttime = time_uptime; 1850 1851 /* 1852 * Push files whose dirty time has expired. Be careful 1853 * of interrupt race on slp queue. 1854 * 1855 * Skip over empty worklist slots when shutting down. 1856 */ 1857 do { 1858 slp = &syncer_workitem_pending[syncer_delayno]; 1859 syncer_delayno += 1; 1860 if (syncer_delayno == syncer_maxdelay) 1861 syncer_delayno = 0; 1862 next = &syncer_workitem_pending[syncer_delayno]; 1863 /* 1864 * If the worklist has wrapped since the 1865 * it was emptied of all but syncer vnodes, 1866 * switch to the FINAL_DELAY state and run 1867 * for one more second. 1868 */ 1869 if (syncer_state == SYNCER_SHUTTING_DOWN && 1870 net_worklist_len == 0 && 1871 last_work_seen == syncer_delayno) { 1872 syncer_state = SYNCER_FINAL_DELAY; 1873 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 1874 } 1875 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 1876 syncer_worklist_len > 0); 1877 1878 /* 1879 * Keep track of the last time there was anything 1880 * on the worklist other than syncer vnodes. 1881 * Return to the SHUTTING_DOWN state if any 1882 * new work appears. 1883 */ 1884 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 1885 last_work_seen = syncer_delayno; 1886 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 1887 syncer_state = SYNCER_SHUTTING_DOWN; 1888 while (!LIST_EMPTY(slp)) { 1889 error = sync_vnode(slp, &bo, td); 1890 if (error == 1) { 1891 LIST_REMOVE(bo, bo_synclist); 1892 LIST_INSERT_HEAD(next, bo, bo_synclist); 1893 continue; 1894 } 1895 1896 if (first_printf == 0) 1897 wdog_kern_pat(WD_LASTVAL); 1898 1899 } 1900 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 1901 syncer_final_iter--; 1902 /* 1903 * The variable rushjob allows the kernel to speed up the 1904 * processing of the filesystem syncer process. A rushjob 1905 * value of N tells the filesystem syncer to process the next 1906 * N seconds worth of work on its queue ASAP. Currently rushjob 1907 * is used by the soft update code to speed up the filesystem 1908 * syncer process when the incore state is getting so far 1909 * ahead of the disk that the kernel memory pool is being 1910 * threatened with exhaustion. 1911 */ 1912 if (rushjob > 0) { 1913 rushjob -= 1; 1914 continue; 1915 } 1916 /* 1917 * Just sleep for a short period of time between 1918 * iterations when shutting down to allow some I/O 1919 * to happen. 1920 * 1921 * If it has taken us less than a second to process the 1922 * current work, then wait. Otherwise start right over 1923 * again. We can still lose time if any single round 1924 * takes more than two seconds, but it does not really 1925 * matter as we are just trying to generally pace the 1926 * filesystem activity. 1927 */ 1928 if (syncer_state != SYNCER_RUNNING || 1929 time_uptime == starttime) { 1930 thread_lock(td); 1931 sched_prio(td, PPAUSE); 1932 thread_unlock(td); 1933 } 1934 if (syncer_state != SYNCER_RUNNING) 1935 cv_timedwait(&sync_wakeup, &sync_mtx, 1936 hz / SYNCER_SHUTDOWN_SPEEDUP); 1937 else if (time_uptime == starttime) 1938 cv_timedwait(&sync_wakeup, &sync_mtx, hz); 1939 } 1940 } 1941 1942 /* 1943 * Request the syncer daemon to speed up its work. 1944 * We never push it to speed up more than half of its 1945 * normal turn time, otherwise it could take over the cpu. 1946 */ 1947 int 1948 speedup_syncer(void) 1949 { 1950 int ret = 0; 1951 1952 mtx_lock(&sync_mtx); 1953 if (rushjob < syncdelay / 2) { 1954 rushjob += 1; 1955 stat_rush_requests += 1; 1956 ret = 1; 1957 } 1958 mtx_unlock(&sync_mtx); 1959 cv_broadcast(&sync_wakeup); 1960 return (ret); 1961 } 1962 1963 /* 1964 * Tell the syncer to speed up its work and run though its work 1965 * list several times, then tell it to shut down. 1966 */ 1967 static void 1968 syncer_shutdown(void *arg, int howto) 1969 { 1970 1971 if (howto & RB_NOSYNC) 1972 return; 1973 mtx_lock(&sync_mtx); 1974 syncer_state = SYNCER_SHUTTING_DOWN; 1975 rushjob = 0; 1976 mtx_unlock(&sync_mtx); 1977 cv_broadcast(&sync_wakeup); 1978 kproc_shutdown(arg, howto); 1979 } 1980 1981 /* 1982 * Reassign a buffer from one vnode to another. 1983 * Used to assign file specific control information 1984 * (indirect blocks) to the vnode to which they belong. 1985 */ 1986 void 1987 reassignbuf(struct buf *bp) 1988 { 1989 struct vnode *vp; 1990 struct bufobj *bo; 1991 int delay; 1992 #ifdef INVARIANTS 1993 struct bufv *bv; 1994 #endif 1995 1996 vp = bp->b_vp; 1997 bo = bp->b_bufobj; 1998 ++reassignbufcalls; 1999 2000 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 2001 bp, bp->b_vp, bp->b_flags); 2002 /* 2003 * B_PAGING flagged buffers cannot be reassigned because their vp 2004 * is not fully linked in. 2005 */ 2006 if (bp->b_flags & B_PAGING) 2007 panic("cannot reassign paging buffer"); 2008 2009 /* 2010 * Delete from old vnode list, if on one. 2011 */ 2012 BO_LOCK(bo); 2013 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 2014 buf_vlist_remove(bp); 2015 else 2016 panic("reassignbuf: Buffer %p not on queue.", bp); 2017 /* 2018 * If dirty, put on list of dirty buffers; otherwise insert onto list 2019 * of clean buffers. 2020 */ 2021 if (bp->b_flags & B_DELWRI) { 2022 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 2023 switch (vp->v_type) { 2024 case VDIR: 2025 delay = dirdelay; 2026 break; 2027 case VCHR: 2028 delay = metadelay; 2029 break; 2030 default: 2031 delay = filedelay; 2032 } 2033 vn_syncer_add_to_worklist(bo, delay); 2034 } 2035 buf_vlist_add(bp, bo, BX_VNDIRTY); 2036 } else { 2037 buf_vlist_add(bp, bo, BX_VNCLEAN); 2038 2039 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 2040 mtx_lock(&sync_mtx); 2041 LIST_REMOVE(bo, bo_synclist); 2042 syncer_worklist_len--; 2043 mtx_unlock(&sync_mtx); 2044 bo->bo_flag &= ~BO_ONWORKLST; 2045 } 2046 } 2047 #ifdef INVARIANTS 2048 bv = &bo->bo_clean; 2049 bp = TAILQ_FIRST(&bv->bv_hd); 2050 KASSERT(bp == NULL || bp->b_bufobj == bo, 2051 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2052 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2053 KASSERT(bp == NULL || bp->b_bufobj == bo, 2054 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2055 bv = &bo->bo_dirty; 2056 bp = TAILQ_FIRST(&bv->bv_hd); 2057 KASSERT(bp == NULL || bp->b_bufobj == bo, 2058 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2059 bp = TAILQ_LAST(&bv->bv_hd, buflists); 2060 KASSERT(bp == NULL || bp->b_bufobj == bo, 2061 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2062 #endif 2063 BO_UNLOCK(bo); 2064 } 2065 2066 /* 2067 * Increment the use and hold counts on the vnode, taking care to reference 2068 * the driver's usecount if this is a chardev. The vholdl() will remove 2069 * the vnode from the free list if it is presently free. Requires the 2070 * vnode interlock and returns with it held. 2071 */ 2072 static void 2073 v_incr_usecount(struct vnode *vp) 2074 { 2075 2076 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2077 vp->v_usecount++; 2078 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2079 dev_lock(); 2080 vp->v_rdev->si_usecount++; 2081 dev_unlock(); 2082 } 2083 vholdl(vp); 2084 } 2085 2086 /* 2087 * Turn a holdcnt into a use+holdcnt such that only one call to 2088 * v_decr_usecount is needed. 2089 */ 2090 static void 2091 v_upgrade_usecount(struct vnode *vp) 2092 { 2093 2094 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2095 vp->v_usecount++; 2096 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2097 dev_lock(); 2098 vp->v_rdev->si_usecount++; 2099 dev_unlock(); 2100 } 2101 } 2102 2103 /* 2104 * Decrement the vnode use and hold count along with the driver's usecount 2105 * if this is a chardev. The vdropl() below releases the vnode interlock 2106 * as it may free the vnode. 2107 */ 2108 static void 2109 v_decr_usecount(struct vnode *vp) 2110 { 2111 2112 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2113 VNASSERT(vp->v_usecount > 0, vp, 2114 ("v_decr_usecount: negative usecount")); 2115 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2116 vp->v_usecount--; 2117 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2118 dev_lock(); 2119 vp->v_rdev->si_usecount--; 2120 dev_unlock(); 2121 } 2122 vdropl(vp); 2123 } 2124 2125 /* 2126 * Decrement only the use count and driver use count. This is intended to 2127 * be paired with a follow on vdropl() to release the remaining hold count. 2128 * In this way we may vgone() a vnode with a 0 usecount without risk of 2129 * having it end up on a free list because the hold count is kept above 0. 2130 */ 2131 static void 2132 v_decr_useonly(struct vnode *vp) 2133 { 2134 2135 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2136 VNASSERT(vp->v_usecount > 0, vp, 2137 ("v_decr_useonly: negative usecount")); 2138 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2139 vp->v_usecount--; 2140 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2141 dev_lock(); 2142 vp->v_rdev->si_usecount--; 2143 dev_unlock(); 2144 } 2145 } 2146 2147 /* 2148 * Grab a particular vnode from the free list, increment its 2149 * reference count and lock it. VI_DOOMED is set if the vnode 2150 * is being destroyed. Only callers who specify LK_RETRY will 2151 * see doomed vnodes. If inactive processing was delayed in 2152 * vput try to do it here. 2153 */ 2154 int 2155 vget(struct vnode *vp, int flags, struct thread *td) 2156 { 2157 int error; 2158 2159 error = 0; 2160 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 2161 ("vget: invalid lock operation")); 2162 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); 2163 2164 if ((flags & LK_INTERLOCK) == 0) 2165 VI_LOCK(vp); 2166 vholdl(vp); 2167 if ((error = vn_lock(vp, flags | LK_INTERLOCK)) != 0) { 2168 vdrop(vp); 2169 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, 2170 vp); 2171 return (error); 2172 } 2173 if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0) 2174 panic("vget: vn_lock failed to return ENOENT\n"); 2175 VI_LOCK(vp); 2176 /* Upgrade our holdcnt to a usecount. */ 2177 v_upgrade_usecount(vp); 2178 /* 2179 * We don't guarantee that any particular close will 2180 * trigger inactive processing so just make a best effort 2181 * here at preventing a reference to a removed file. If 2182 * we don't succeed no harm is done. 2183 */ 2184 if (vp->v_iflag & VI_OWEINACT) { 2185 if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE && 2186 (flags & LK_NOWAIT) == 0) 2187 vinactive(vp, td); 2188 vp->v_iflag &= ~VI_OWEINACT; 2189 } 2190 VI_UNLOCK(vp); 2191 return (0); 2192 } 2193 2194 /* 2195 * Increase the reference count of a vnode. 2196 */ 2197 void 2198 vref(struct vnode *vp) 2199 { 2200 2201 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2202 VI_LOCK(vp); 2203 v_incr_usecount(vp); 2204 VI_UNLOCK(vp); 2205 } 2206 2207 /* 2208 * Return reference count of a vnode. 2209 * 2210 * The results of this call are only guaranteed when some mechanism other 2211 * than the VI lock is used to stop other processes from gaining references 2212 * to the vnode. This may be the case if the caller holds the only reference. 2213 * This is also useful when stale data is acceptable as race conditions may 2214 * be accounted for by some other means. 2215 */ 2216 int 2217 vrefcnt(struct vnode *vp) 2218 { 2219 int usecnt; 2220 2221 VI_LOCK(vp); 2222 usecnt = vp->v_usecount; 2223 VI_UNLOCK(vp); 2224 2225 return (usecnt); 2226 } 2227 2228 #define VPUTX_VRELE 1 2229 #define VPUTX_VPUT 2 2230 #define VPUTX_VUNREF 3 2231 2232 static void 2233 vputx(struct vnode *vp, int func) 2234 { 2235 int error; 2236 2237 KASSERT(vp != NULL, ("vputx: null vp")); 2238 if (func == VPUTX_VUNREF) 2239 ASSERT_VOP_LOCKED(vp, "vunref"); 2240 else if (func == VPUTX_VPUT) 2241 ASSERT_VOP_LOCKED(vp, "vput"); 2242 else 2243 KASSERT(func == VPUTX_VRELE, ("vputx: wrong func")); 2244 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2245 VI_LOCK(vp); 2246 2247 /* Skip this v_writecount check if we're going to panic below. */ 2248 VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp, 2249 ("vputx: missed vn_close")); 2250 error = 0; 2251 2252 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2253 vp->v_usecount == 1)) { 2254 if (func == VPUTX_VPUT) 2255 VOP_UNLOCK(vp, 0); 2256 v_decr_usecount(vp); 2257 return; 2258 } 2259 2260 if (vp->v_usecount != 1) { 2261 vprint("vputx: negative ref count", vp); 2262 panic("vputx: negative ref cnt"); 2263 } 2264 CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp); 2265 /* 2266 * We want to hold the vnode until the inactive finishes to 2267 * prevent vgone() races. We drop the use count here and the 2268 * hold count below when we're done. 2269 */ 2270 v_decr_useonly(vp); 2271 /* 2272 * We must call VOP_INACTIVE with the node locked. Mark 2273 * as VI_DOINGINACT to avoid recursion. 2274 */ 2275 vp->v_iflag |= VI_OWEINACT; 2276 switch (func) { 2277 case VPUTX_VRELE: 2278 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 2279 VI_LOCK(vp); 2280 break; 2281 case VPUTX_VPUT: 2282 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 2283 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | 2284 LK_NOWAIT); 2285 VI_LOCK(vp); 2286 } 2287 break; 2288 case VPUTX_VUNREF: 2289 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 2290 error = EBUSY; 2291 break; 2292 } 2293 if (vp->v_usecount > 0) 2294 vp->v_iflag &= ~VI_OWEINACT; 2295 if (error == 0) { 2296 if (vp->v_iflag & VI_OWEINACT) 2297 vinactive(vp, curthread); 2298 if (func != VPUTX_VUNREF) 2299 VOP_UNLOCK(vp, 0); 2300 } 2301 vdropl(vp); 2302 } 2303 2304 /* 2305 * Vnode put/release. 2306 * If count drops to zero, call inactive routine and return to freelist. 2307 */ 2308 void 2309 vrele(struct vnode *vp) 2310 { 2311 2312 vputx(vp, VPUTX_VRELE); 2313 } 2314 2315 /* 2316 * Release an already locked vnode. This give the same effects as 2317 * unlock+vrele(), but takes less time and avoids releasing and 2318 * re-aquiring the lock (as vrele() acquires the lock internally.) 2319 */ 2320 void 2321 vput(struct vnode *vp) 2322 { 2323 2324 vputx(vp, VPUTX_VPUT); 2325 } 2326 2327 /* 2328 * Release an exclusively locked vnode. Do not unlock the vnode lock. 2329 */ 2330 void 2331 vunref(struct vnode *vp) 2332 { 2333 2334 vputx(vp, VPUTX_VUNREF); 2335 } 2336 2337 /* 2338 * Somebody doesn't want the vnode recycled. 2339 */ 2340 void 2341 vhold(struct vnode *vp) 2342 { 2343 2344 VI_LOCK(vp); 2345 vholdl(vp); 2346 VI_UNLOCK(vp); 2347 } 2348 2349 /* 2350 * Increase the hold count and activate if this is the first reference. 2351 */ 2352 void 2353 vholdl(struct vnode *vp) 2354 { 2355 struct mount *mp; 2356 2357 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2358 vp->v_holdcnt++; 2359 if (!VSHOULDBUSY(vp)) 2360 return; 2361 ASSERT_VI_LOCKED(vp, "vholdl"); 2362 VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free")); 2363 VNASSERT(vp->v_op != NULL, vp, ("vholdl: vnode already reclaimed.")); 2364 /* 2365 * Remove a vnode from the free list, mark it as in use, 2366 * and put it on the active list. 2367 */ 2368 mtx_lock(&vnode_free_list_mtx); 2369 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist); 2370 freevnodes--; 2371 vp->v_iflag &= ~(VI_FREE|VI_AGE); 2372 KASSERT((vp->v_iflag & VI_ACTIVE) == 0, 2373 ("Activating already active vnode")); 2374 vp->v_iflag |= VI_ACTIVE; 2375 mp = vp->v_mount; 2376 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist); 2377 mp->mnt_activevnodelistsize++; 2378 mtx_unlock(&vnode_free_list_mtx); 2379 } 2380 2381 /* 2382 * Note that there is one less who cares about this vnode. 2383 * vdrop() is the opposite of vhold(). 2384 */ 2385 void 2386 vdrop(struct vnode *vp) 2387 { 2388 2389 VI_LOCK(vp); 2390 vdropl(vp); 2391 } 2392 2393 /* 2394 * Drop the hold count of the vnode. If this is the last reference to 2395 * the vnode we place it on the free list unless it has been vgone'd 2396 * (marked VI_DOOMED) in which case we will free it. 2397 */ 2398 void 2399 vdropl(struct vnode *vp) 2400 { 2401 struct bufobj *bo; 2402 struct mount *mp; 2403 int active; 2404 2405 ASSERT_VI_LOCKED(vp, "vdropl"); 2406 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2407 if (vp->v_holdcnt <= 0) 2408 panic("vdrop: holdcnt %d", vp->v_holdcnt); 2409 vp->v_holdcnt--; 2410 if (vp->v_holdcnt > 0) { 2411 VI_UNLOCK(vp); 2412 return; 2413 } 2414 if ((vp->v_iflag & VI_DOOMED) == 0) { 2415 /* 2416 * Mark a vnode as free: remove it from its active list 2417 * and put it up for recycling on the freelist. 2418 */ 2419 VNASSERT(vp->v_op != NULL, vp, 2420 ("vdropl: vnode already reclaimed.")); 2421 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 2422 ("vnode already free")); 2423 VNASSERT(VSHOULDFREE(vp), vp, 2424 ("vdropl: freeing when we shouldn't")); 2425 active = vp->v_iflag & VI_ACTIVE; 2426 vp->v_iflag &= ~VI_ACTIVE; 2427 mp = vp->v_mount; 2428 mtx_lock(&vnode_free_list_mtx); 2429 if (active) { 2430 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, 2431 v_actfreelist); 2432 mp->mnt_activevnodelistsize--; 2433 } 2434 if (vp->v_iflag & VI_AGE) { 2435 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_actfreelist); 2436 } else { 2437 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist); 2438 } 2439 freevnodes++; 2440 vp->v_iflag &= ~VI_AGE; 2441 vp->v_iflag |= VI_FREE; 2442 mtx_unlock(&vnode_free_list_mtx); 2443 VI_UNLOCK(vp); 2444 return; 2445 } 2446 /* 2447 * The vnode has been marked for destruction, so free it. 2448 */ 2449 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp); 2450 mtx_lock(&vnode_free_list_mtx); 2451 numvnodes--; 2452 mtx_unlock(&vnode_free_list_mtx); 2453 bo = &vp->v_bufobj; 2454 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 2455 ("cleaned vnode still on the free list.")); 2456 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 2457 VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count")); 2458 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 2459 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 2460 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 2461 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 2462 VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL")); 2463 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 2464 VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL")); 2465 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); 2466 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); 2467 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for ..")); 2468 VI_UNLOCK(vp); 2469 #ifdef MAC 2470 mac_vnode_destroy(vp); 2471 #endif 2472 if (vp->v_pollinfo != NULL) 2473 destroy_vpollinfo(vp->v_pollinfo); 2474 #ifdef INVARIANTS 2475 /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */ 2476 vp->v_op = NULL; 2477 #endif 2478 rangelock_destroy(&vp->v_rl); 2479 lockdestroy(vp->v_vnlock); 2480 mtx_destroy(&vp->v_interlock); 2481 mtx_destroy(BO_MTX(bo)); 2482 uma_zfree(vnode_zone, vp); 2483 } 2484 2485 /* 2486 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 2487 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 2488 * OWEINACT tracks whether a vnode missed a call to inactive due to a 2489 * failed lock upgrade. 2490 */ 2491 void 2492 vinactive(struct vnode *vp, struct thread *td) 2493 { 2494 struct vm_object *obj; 2495 2496 ASSERT_VOP_ELOCKED(vp, "vinactive"); 2497 ASSERT_VI_LOCKED(vp, "vinactive"); 2498 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, 2499 ("vinactive: recursed on VI_DOINGINACT")); 2500 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2501 vp->v_iflag |= VI_DOINGINACT; 2502 vp->v_iflag &= ~VI_OWEINACT; 2503 VI_UNLOCK(vp); 2504 /* 2505 * Before moving off the active list, we must be sure that any 2506 * modified pages are on the vnode's dirty list since these will 2507 * no longer be checked once the vnode is on the inactive list. 2508 * Because the vnode vm object keeps a hold reference on the vnode 2509 * if there is at least one resident non-cached page, the vnode 2510 * cannot leave the active list without the page cleanup done. 2511 */ 2512 obj = vp->v_object; 2513 if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0) { 2514 VM_OBJECT_WLOCK(obj); 2515 vm_object_page_clean(obj, 0, 0, OBJPC_NOSYNC); 2516 VM_OBJECT_WUNLOCK(obj); 2517 } 2518 VOP_INACTIVE(vp, td); 2519 VI_LOCK(vp); 2520 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, 2521 ("vinactive: lost VI_DOINGINACT")); 2522 vp->v_iflag &= ~VI_DOINGINACT; 2523 } 2524 2525 /* 2526 * Remove any vnodes in the vnode table belonging to mount point mp. 2527 * 2528 * If FORCECLOSE is not specified, there should not be any active ones, 2529 * return error if any are found (nb: this is a user error, not a 2530 * system error). If FORCECLOSE is specified, detach any active vnodes 2531 * that are found. 2532 * 2533 * If WRITECLOSE is set, only flush out regular file vnodes open for 2534 * writing. 2535 * 2536 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2537 * 2538 * `rootrefs' specifies the base reference count for the root vnode 2539 * of this filesystem. The root vnode is considered busy if its 2540 * v_usecount exceeds this value. On a successful return, vflush(, td) 2541 * will call vrele() on the root vnode exactly rootrefs times. 2542 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2543 * be zero. 2544 */ 2545 #ifdef DIAGNOSTIC 2546 static int busyprt = 0; /* print out busy vnodes */ 2547 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); 2548 #endif 2549 2550 int 2551 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) 2552 { 2553 struct vnode *vp, *mvp, *rootvp = NULL; 2554 struct vattr vattr; 2555 int busy = 0, error; 2556 2557 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, 2558 rootrefs, flags); 2559 if (rootrefs > 0) { 2560 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2561 ("vflush: bad args")); 2562 /* 2563 * Get the filesystem root vnode. We can vput() it 2564 * immediately, since with rootrefs > 0, it won't go away. 2565 */ 2566 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { 2567 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", 2568 __func__, error); 2569 return (error); 2570 } 2571 vput(rootvp); 2572 } 2573 loop: 2574 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 2575 vholdl(vp); 2576 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 2577 if (error) { 2578 vdrop(vp); 2579 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 2580 goto loop; 2581 } 2582 /* 2583 * Skip over a vnodes marked VV_SYSTEM. 2584 */ 2585 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2586 VOP_UNLOCK(vp, 0); 2587 vdrop(vp); 2588 continue; 2589 } 2590 /* 2591 * If WRITECLOSE is set, flush out unlinked but still open 2592 * files (even if open only for reading) and regular file 2593 * vnodes open for writing. 2594 */ 2595 if (flags & WRITECLOSE) { 2596 if (vp->v_object != NULL) { 2597 VM_OBJECT_WLOCK(vp->v_object); 2598 vm_object_page_clean(vp->v_object, 0, 0, 0); 2599 VM_OBJECT_WUNLOCK(vp->v_object); 2600 } 2601 error = VOP_FSYNC(vp, MNT_WAIT, td); 2602 if (error != 0) { 2603 VOP_UNLOCK(vp, 0); 2604 vdrop(vp); 2605 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 2606 return (error); 2607 } 2608 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 2609 VI_LOCK(vp); 2610 2611 if ((vp->v_type == VNON || 2612 (error == 0 && vattr.va_nlink > 0)) && 2613 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2614 VOP_UNLOCK(vp, 0); 2615 vdropl(vp); 2616 continue; 2617 } 2618 } else 2619 VI_LOCK(vp); 2620 /* 2621 * With v_usecount == 0, all we need to do is clear out the 2622 * vnode data structures and we are done. 2623 * 2624 * If FORCECLOSE is set, forcibly close the vnode. 2625 */ 2626 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 2627 VNASSERT(vp->v_usecount == 0 || 2628 (vp->v_type != VCHR && vp->v_type != VBLK), vp, 2629 ("device VNODE %p is FORCECLOSED", vp)); 2630 vgonel(vp); 2631 } else { 2632 busy++; 2633 #ifdef DIAGNOSTIC 2634 if (busyprt) 2635 vprint("vflush: busy vnode", vp); 2636 #endif 2637 } 2638 VOP_UNLOCK(vp, 0); 2639 vdropl(vp); 2640 } 2641 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2642 /* 2643 * If just the root vnode is busy, and if its refcount 2644 * is equal to `rootrefs', then go ahead and kill it. 2645 */ 2646 VI_LOCK(rootvp); 2647 KASSERT(busy > 0, ("vflush: not busy")); 2648 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 2649 ("vflush: usecount %d < rootrefs %d", 2650 rootvp->v_usecount, rootrefs)); 2651 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2652 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 2653 vgone(rootvp); 2654 VOP_UNLOCK(rootvp, 0); 2655 busy = 0; 2656 } else 2657 VI_UNLOCK(rootvp); 2658 } 2659 if (busy) { 2660 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, 2661 busy); 2662 return (EBUSY); 2663 } 2664 for (; rootrefs > 0; rootrefs--) 2665 vrele(rootvp); 2666 return (0); 2667 } 2668 2669 /* 2670 * Recycle an unused vnode to the front of the free list. 2671 */ 2672 int 2673 vrecycle(struct vnode *vp) 2674 { 2675 int recycled; 2676 2677 ASSERT_VOP_ELOCKED(vp, "vrecycle"); 2678 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2679 recycled = 0; 2680 VI_LOCK(vp); 2681 if (vp->v_usecount == 0) { 2682 recycled = 1; 2683 vgonel(vp); 2684 } 2685 VI_UNLOCK(vp); 2686 return (recycled); 2687 } 2688 2689 /* 2690 * Eliminate all activity associated with a vnode 2691 * in preparation for reuse. 2692 */ 2693 void 2694 vgone(struct vnode *vp) 2695 { 2696 VI_LOCK(vp); 2697 vgonel(vp); 2698 VI_UNLOCK(vp); 2699 } 2700 2701 static void 2702 vgonel_reclaim_lowervp_vfs(struct mount *mp __unused, 2703 struct vnode *lowervp __unused) 2704 { 2705 } 2706 2707 /* 2708 * Notify upper mounts about reclaimed vnode. 2709 */ 2710 static void 2711 vgonel_reclaim_lowervp(struct vnode *vp) 2712 { 2713 static struct vfsops vgonel_vfsops = { 2714 .vfs_reclaim_lowervp = vgonel_reclaim_lowervp_vfs 2715 }; 2716 struct mount *mp, *ump, *mmp; 2717 2718 mp = vp->v_mount; 2719 if (mp == NULL) 2720 return; 2721 2722 MNT_ILOCK(mp); 2723 if (TAILQ_EMPTY(&mp->mnt_uppers)) 2724 goto unlock; 2725 MNT_IUNLOCK(mp); 2726 mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO); 2727 mmp->mnt_op = &vgonel_vfsops; 2728 mmp->mnt_kern_flag |= MNTK_MARKER; 2729 MNT_ILOCK(mp); 2730 mp->mnt_kern_flag |= MNTK_VGONE_UPPER; 2731 for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) { 2732 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) { 2733 ump = TAILQ_NEXT(ump, mnt_upper_link); 2734 continue; 2735 } 2736 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link); 2737 MNT_IUNLOCK(mp); 2738 VFS_RECLAIM_LOWERVP(ump, vp); 2739 MNT_ILOCK(mp); 2740 ump = TAILQ_NEXT(mmp, mnt_upper_link); 2741 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link); 2742 } 2743 free(mmp, M_TEMP); 2744 mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER; 2745 if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) { 2746 mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER; 2747 wakeup(&mp->mnt_uppers); 2748 } 2749 unlock: 2750 MNT_IUNLOCK(mp); 2751 } 2752 2753 /* 2754 * vgone, with the vp interlock held. 2755 */ 2756 void 2757 vgonel(struct vnode *vp) 2758 { 2759 struct thread *td; 2760 int oweinact; 2761 int active; 2762 struct mount *mp; 2763 2764 ASSERT_VOP_ELOCKED(vp, "vgonel"); 2765 ASSERT_VI_LOCKED(vp, "vgonel"); 2766 VNASSERT(vp->v_holdcnt, vp, 2767 ("vgonel: vp %p has no reference.", vp)); 2768 CTR2(KTR_VFS, "%s: vp %p", __func__, vp); 2769 td = curthread; 2770 2771 /* 2772 * Don't vgonel if we're already doomed. 2773 */ 2774 if (vp->v_iflag & VI_DOOMED) 2775 return; 2776 vp->v_iflag |= VI_DOOMED; 2777 2778 /* 2779 * Check to see if the vnode is in use. If so, we have to call 2780 * VOP_CLOSE() and VOP_INACTIVE(). 2781 */ 2782 active = vp->v_usecount; 2783 oweinact = (vp->v_iflag & VI_OWEINACT); 2784 VI_UNLOCK(vp); 2785 vgonel_reclaim_lowervp(vp); 2786 2787 /* 2788 * Clean out any buffers associated with the vnode. 2789 * If the flush fails, just toss the buffers. 2790 */ 2791 mp = NULL; 2792 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 2793 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 2794 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) 2795 vinvalbuf(vp, 0, 0, 0); 2796 2797 /* 2798 * If purging an active vnode, it must be closed and 2799 * deactivated before being reclaimed. 2800 */ 2801 if (active) 2802 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2803 if (oweinact || active) { 2804 VI_LOCK(vp); 2805 if ((vp->v_iflag & VI_DOINGINACT) == 0) 2806 vinactive(vp, td); 2807 VI_UNLOCK(vp); 2808 } 2809 if (vp->v_type == VSOCK) 2810 vfs_unp_reclaim(vp); 2811 /* 2812 * Reclaim the vnode. 2813 */ 2814 if (VOP_RECLAIM(vp, td)) 2815 panic("vgone: cannot reclaim"); 2816 if (mp != NULL) 2817 vn_finished_secondary_write(mp); 2818 VNASSERT(vp->v_object == NULL, vp, 2819 ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag)); 2820 /* 2821 * Clear the advisory locks and wake up waiting threads. 2822 */ 2823 (void)VOP_ADVLOCKPURGE(vp); 2824 /* 2825 * Delete from old mount point vnode list. 2826 */ 2827 delmntque(vp); 2828 cache_purge(vp); 2829 /* 2830 * Done with purge, reset to the standard lock and invalidate 2831 * the vnode. 2832 */ 2833 VI_LOCK(vp); 2834 vp->v_vnlock = &vp->v_lock; 2835 vp->v_op = &dead_vnodeops; 2836 vp->v_tag = "none"; 2837 vp->v_type = VBAD; 2838 } 2839 2840 /* 2841 * Calculate the total number of references to a special device. 2842 */ 2843 int 2844 vcount(struct vnode *vp) 2845 { 2846 int count; 2847 2848 dev_lock(); 2849 count = vp->v_rdev->si_usecount; 2850 dev_unlock(); 2851 return (count); 2852 } 2853 2854 /* 2855 * Same as above, but using the struct cdev *as argument 2856 */ 2857 int 2858 count_dev(struct cdev *dev) 2859 { 2860 int count; 2861 2862 dev_lock(); 2863 count = dev->si_usecount; 2864 dev_unlock(); 2865 return(count); 2866 } 2867 2868 /* 2869 * Print out a description of a vnode. 2870 */ 2871 static char *typename[] = 2872 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", 2873 "VMARKER"}; 2874 2875 void 2876 vn_printf(struct vnode *vp, const char *fmt, ...) 2877 { 2878 va_list ap; 2879 char buf[256], buf2[16]; 2880 u_long flags; 2881 2882 va_start(ap, fmt); 2883 vprintf(fmt, ap); 2884 va_end(ap); 2885 printf("%p: ", (void *)vp); 2886 printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]); 2887 printf(" usecount %d, writecount %d, refcount %d mountedhere %p\n", 2888 vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere); 2889 buf[0] = '\0'; 2890 buf[1] = '\0'; 2891 if (vp->v_vflag & VV_ROOT) 2892 strlcat(buf, "|VV_ROOT", sizeof(buf)); 2893 if (vp->v_vflag & VV_ISTTY) 2894 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 2895 if (vp->v_vflag & VV_NOSYNC) 2896 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 2897 if (vp->v_vflag & VV_ETERNALDEV) 2898 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); 2899 if (vp->v_vflag & VV_CACHEDLABEL) 2900 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 2901 if (vp->v_vflag & VV_TEXT) 2902 strlcat(buf, "|VV_TEXT", sizeof(buf)); 2903 if (vp->v_vflag & VV_COPYONWRITE) 2904 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 2905 if (vp->v_vflag & VV_SYSTEM) 2906 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 2907 if (vp->v_vflag & VV_PROCDEP) 2908 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 2909 if (vp->v_vflag & VV_NOKNOTE) 2910 strlcat(buf, "|VV_NOKNOTE", sizeof(buf)); 2911 if (vp->v_vflag & VV_DELETED) 2912 strlcat(buf, "|VV_DELETED", sizeof(buf)); 2913 if (vp->v_vflag & VV_MD) 2914 strlcat(buf, "|VV_MD", sizeof(buf)); 2915 if (vp->v_vflag & VV_FORCEINSMQ) 2916 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); 2917 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | 2918 VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | 2919 VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ); 2920 if (flags != 0) { 2921 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 2922 strlcat(buf, buf2, sizeof(buf)); 2923 } 2924 if (vp->v_iflag & VI_MOUNT) 2925 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 2926 if (vp->v_iflag & VI_AGE) 2927 strlcat(buf, "|VI_AGE", sizeof(buf)); 2928 if (vp->v_iflag & VI_DOOMED) 2929 strlcat(buf, "|VI_DOOMED", sizeof(buf)); 2930 if (vp->v_iflag & VI_FREE) 2931 strlcat(buf, "|VI_FREE", sizeof(buf)); 2932 if (vp->v_iflag & VI_ACTIVE) 2933 strlcat(buf, "|VI_ACTIVE", sizeof(buf)); 2934 if (vp->v_iflag & VI_DOINGINACT) 2935 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 2936 if (vp->v_iflag & VI_OWEINACT) 2937 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 2938 flags = vp->v_iflag & ~(VI_MOUNT | VI_AGE | VI_DOOMED | VI_FREE | 2939 VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT); 2940 if (flags != 0) { 2941 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 2942 strlcat(buf, buf2, sizeof(buf)); 2943 } 2944 printf(" flags (%s)\n", buf + 1); 2945 if (mtx_owned(VI_MTX(vp))) 2946 printf(" VI_LOCKed"); 2947 if (vp->v_object != NULL) 2948 printf(" v_object %p ref %d pages %d\n", 2949 vp->v_object, vp->v_object->ref_count, 2950 vp->v_object->resident_page_count); 2951 printf(" "); 2952 lockmgr_printinfo(vp->v_vnlock); 2953 if (vp->v_data != NULL) 2954 VOP_PRINT(vp); 2955 } 2956 2957 #ifdef DDB 2958 /* 2959 * List all of the locked vnodes in the system. 2960 * Called when debugging the kernel. 2961 */ 2962 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 2963 { 2964 struct mount *mp, *nmp; 2965 struct vnode *vp; 2966 2967 /* 2968 * Note: because this is DDB, we can't obey the locking semantics 2969 * for these structures, which means we could catch an inconsistent 2970 * state and dereference a nasty pointer. Not much to be done 2971 * about that. 2972 */ 2973 db_printf("Locked vnodes\n"); 2974 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2975 nmp = TAILQ_NEXT(mp, mnt_list); 2976 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2977 if (vp->v_type != VMARKER && 2978 VOP_ISLOCKED(vp)) 2979 vprint("", vp); 2980 } 2981 nmp = TAILQ_NEXT(mp, mnt_list); 2982 } 2983 } 2984 2985 /* 2986 * Show details about the given vnode. 2987 */ 2988 DB_SHOW_COMMAND(vnode, db_show_vnode) 2989 { 2990 struct vnode *vp; 2991 2992 if (!have_addr) 2993 return; 2994 vp = (struct vnode *)addr; 2995 vn_printf(vp, "vnode "); 2996 } 2997 2998 /* 2999 * Show details about the given mount point. 3000 */ 3001 DB_SHOW_COMMAND(mount, db_show_mount) 3002 { 3003 struct mount *mp; 3004 struct vfsopt *opt; 3005 struct statfs *sp; 3006 struct vnode *vp; 3007 char buf[512]; 3008 uint64_t mflags; 3009 u_int flags; 3010 3011 if (!have_addr) { 3012 /* No address given, print short info about all mount points. */ 3013 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3014 db_printf("%p %s on %s (%s)\n", mp, 3015 mp->mnt_stat.f_mntfromname, 3016 mp->mnt_stat.f_mntonname, 3017 mp->mnt_stat.f_fstypename); 3018 if (db_pager_quit) 3019 break; 3020 } 3021 db_printf("\nMore info: show mount <addr>\n"); 3022 return; 3023 } 3024 3025 mp = (struct mount *)addr; 3026 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 3027 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 3028 3029 buf[0] = '\0'; 3030 mflags = mp->mnt_flag; 3031 #define MNT_FLAG(flag) do { \ 3032 if (mflags & (flag)) { \ 3033 if (buf[0] != '\0') \ 3034 strlcat(buf, ", ", sizeof(buf)); \ 3035 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 3036 mflags &= ~(flag); \ 3037 } \ 3038 } while (0) 3039 MNT_FLAG(MNT_RDONLY); 3040 MNT_FLAG(MNT_SYNCHRONOUS); 3041 MNT_FLAG(MNT_NOEXEC); 3042 MNT_FLAG(MNT_NOSUID); 3043 MNT_FLAG(MNT_NFS4ACLS); 3044 MNT_FLAG(MNT_UNION); 3045 MNT_FLAG(MNT_ASYNC); 3046 MNT_FLAG(MNT_SUIDDIR); 3047 MNT_FLAG(MNT_SOFTDEP); 3048 MNT_FLAG(MNT_NOSYMFOLLOW); 3049 MNT_FLAG(MNT_GJOURNAL); 3050 MNT_FLAG(MNT_MULTILABEL); 3051 MNT_FLAG(MNT_ACLS); 3052 MNT_FLAG(MNT_NOATIME); 3053 MNT_FLAG(MNT_NOCLUSTERR); 3054 MNT_FLAG(MNT_NOCLUSTERW); 3055 MNT_FLAG(MNT_SUJ); 3056 MNT_FLAG(MNT_EXRDONLY); 3057 MNT_FLAG(MNT_EXPORTED); 3058 MNT_FLAG(MNT_DEFEXPORTED); 3059 MNT_FLAG(MNT_EXPORTANON); 3060 MNT_FLAG(MNT_EXKERB); 3061 MNT_FLAG(MNT_EXPUBLIC); 3062 MNT_FLAG(MNT_LOCAL); 3063 MNT_FLAG(MNT_QUOTA); 3064 MNT_FLAG(MNT_ROOTFS); 3065 MNT_FLAG(MNT_USER); 3066 MNT_FLAG(MNT_IGNORE); 3067 MNT_FLAG(MNT_UPDATE); 3068 MNT_FLAG(MNT_DELEXPORT); 3069 MNT_FLAG(MNT_RELOAD); 3070 MNT_FLAG(MNT_FORCE); 3071 MNT_FLAG(MNT_SNAPSHOT); 3072 MNT_FLAG(MNT_BYFSID); 3073 #undef MNT_FLAG 3074 if (mflags != 0) { 3075 if (buf[0] != '\0') 3076 strlcat(buf, ", ", sizeof(buf)); 3077 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 3078 "0x%016jx", mflags); 3079 } 3080 db_printf(" mnt_flag = %s\n", buf); 3081 3082 buf[0] = '\0'; 3083 flags = mp->mnt_kern_flag; 3084 #define MNT_KERN_FLAG(flag) do { \ 3085 if (flags & (flag)) { \ 3086 if (buf[0] != '\0') \ 3087 strlcat(buf, ", ", sizeof(buf)); \ 3088 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 3089 flags &= ~(flag); \ 3090 } \ 3091 } while (0) 3092 MNT_KERN_FLAG(MNTK_UNMOUNTF); 3093 MNT_KERN_FLAG(MNTK_ASYNC); 3094 MNT_KERN_FLAG(MNTK_SOFTDEP); 3095 MNT_KERN_FLAG(MNTK_NOINSMNTQ); 3096 MNT_KERN_FLAG(MNTK_DRAINING); 3097 MNT_KERN_FLAG(MNTK_REFEXPIRE); 3098 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); 3099 MNT_KERN_FLAG(MNTK_SHARED_WRITES); 3100 MNT_KERN_FLAG(MNTK_NO_IOPF); 3101 MNT_KERN_FLAG(MNTK_VGONE_UPPER); 3102 MNT_KERN_FLAG(MNTK_VGONE_WAITER); 3103 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT); 3104 MNT_KERN_FLAG(MNTK_MARKER); 3105 MNT_KERN_FLAG(MNTK_NOASYNC); 3106 MNT_KERN_FLAG(MNTK_UNMOUNT); 3107 MNT_KERN_FLAG(MNTK_MWAIT); 3108 MNT_KERN_FLAG(MNTK_SUSPEND); 3109 MNT_KERN_FLAG(MNTK_SUSPEND2); 3110 MNT_KERN_FLAG(MNTK_SUSPENDED); 3111 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 3112 MNT_KERN_FLAG(MNTK_NOKNOTE); 3113 #undef MNT_KERN_FLAG 3114 if (flags != 0) { 3115 if (buf[0] != '\0') 3116 strlcat(buf, ", ", sizeof(buf)); 3117 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 3118 "0x%08x", flags); 3119 } 3120 db_printf(" mnt_kern_flag = %s\n", buf); 3121 3122 db_printf(" mnt_opt = "); 3123 opt = TAILQ_FIRST(mp->mnt_opt); 3124 if (opt != NULL) { 3125 db_printf("%s", opt->name); 3126 opt = TAILQ_NEXT(opt, link); 3127 while (opt != NULL) { 3128 db_printf(", %s", opt->name); 3129 opt = TAILQ_NEXT(opt, link); 3130 } 3131 } 3132 db_printf("\n"); 3133 3134 sp = &mp->mnt_stat; 3135 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 3136 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 3137 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 3138 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 3139 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 3140 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 3141 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 3142 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 3143 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 3144 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 3145 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 3146 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 3147 3148 db_printf(" mnt_cred = { uid=%u ruid=%u", 3149 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 3150 if (jailed(mp->mnt_cred)) 3151 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 3152 db_printf(" }\n"); 3153 db_printf(" mnt_ref = %d\n", mp->mnt_ref); 3154 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 3155 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 3156 db_printf(" mnt_activevnodelistsize = %d\n", 3157 mp->mnt_activevnodelistsize); 3158 db_printf(" mnt_writeopcount = %d\n", mp->mnt_writeopcount); 3159 db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen); 3160 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 3161 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 3162 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 3163 db_printf(" mnt_secondary_accwrites = %d\n", 3164 mp->mnt_secondary_accwrites); 3165 db_printf(" mnt_gjprovider = %s\n", 3166 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 3167 3168 db_printf("\n\nList of active vnodes\n"); 3169 TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) { 3170 if (vp->v_type != VMARKER) { 3171 vn_printf(vp, "vnode "); 3172 if (db_pager_quit) 3173 break; 3174 } 3175 } 3176 db_printf("\n\nList of inactive vnodes\n"); 3177 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3178 if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) { 3179 vn_printf(vp, "vnode "); 3180 if (db_pager_quit) 3181 break; 3182 } 3183 } 3184 } 3185 #endif /* DDB */ 3186 3187 /* 3188 * Fill in a struct xvfsconf based on a struct vfsconf. 3189 */ 3190 static int 3191 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) 3192 { 3193 struct xvfsconf xvfsp; 3194 3195 bzero(&xvfsp, sizeof(xvfsp)); 3196 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 3197 xvfsp.vfc_typenum = vfsp->vfc_typenum; 3198 xvfsp.vfc_refcount = vfsp->vfc_refcount; 3199 xvfsp.vfc_flags = vfsp->vfc_flags; 3200 /* 3201 * These are unused in userland, we keep them 3202 * to not break binary compatibility. 3203 */ 3204 xvfsp.vfc_vfsops = NULL; 3205 xvfsp.vfc_next = NULL; 3206 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 3207 } 3208 3209 #ifdef COMPAT_FREEBSD32 3210 struct xvfsconf32 { 3211 uint32_t vfc_vfsops; 3212 char vfc_name[MFSNAMELEN]; 3213 int32_t vfc_typenum; 3214 int32_t vfc_refcount; 3215 int32_t vfc_flags; 3216 uint32_t vfc_next; 3217 }; 3218 3219 static int 3220 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) 3221 { 3222 struct xvfsconf32 xvfsp; 3223 3224 strcpy(xvfsp.vfc_name, vfsp->vfc_name); 3225 xvfsp.vfc_typenum = vfsp->vfc_typenum; 3226 xvfsp.vfc_refcount = vfsp->vfc_refcount; 3227 xvfsp.vfc_flags = vfsp->vfc_flags; 3228 xvfsp.vfc_vfsops = 0; 3229 xvfsp.vfc_next = 0; 3230 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 3231 } 3232 #endif 3233 3234 /* 3235 * Top level filesystem related information gathering. 3236 */ 3237 static int 3238 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 3239 { 3240 struct vfsconf *vfsp; 3241 int error; 3242 3243 error = 0; 3244 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 3245 #ifdef COMPAT_FREEBSD32 3246 if (req->flags & SCTL_MASK32) 3247 error = vfsconf2x32(req, vfsp); 3248 else 3249 #endif 3250 error = vfsconf2x(req, vfsp); 3251 if (error) 3252 break; 3253 } 3254 return (error); 3255 } 3256 3257 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD, 3258 NULL, 0, sysctl_vfs_conflist, 3259 "S,xvfsconf", "List of all configured filesystems"); 3260 3261 #ifndef BURN_BRIDGES 3262 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 3263 3264 static int 3265 vfs_sysctl(SYSCTL_HANDLER_ARGS) 3266 { 3267 int *name = (int *)arg1 - 1; /* XXX */ 3268 u_int namelen = arg2 + 1; /* XXX */ 3269 struct vfsconf *vfsp; 3270 3271 log(LOG_WARNING, "userland calling deprecated sysctl, " 3272 "please rebuild world\n"); 3273 3274 #if 1 || defined(COMPAT_PRELITE2) 3275 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 3276 if (namelen == 1) 3277 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 3278 #endif 3279 3280 switch (name[1]) { 3281 case VFS_MAXTYPENUM: 3282 if (namelen != 2) 3283 return (ENOTDIR); 3284 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 3285 case VFS_CONF: 3286 if (namelen != 3) 3287 return (ENOTDIR); /* overloaded */ 3288 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) 3289 if (vfsp->vfc_typenum == name[2]) 3290 break; 3291 if (vfsp == NULL) 3292 return (EOPNOTSUPP); 3293 #ifdef COMPAT_FREEBSD32 3294 if (req->flags & SCTL_MASK32) 3295 return (vfsconf2x32(req, vfsp)); 3296 else 3297 #endif 3298 return (vfsconf2x(req, vfsp)); 3299 } 3300 return (EOPNOTSUPP); 3301 } 3302 3303 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, 3304 vfs_sysctl, "Generic filesystem"); 3305 3306 #if 1 || defined(COMPAT_PRELITE2) 3307 3308 static int 3309 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 3310 { 3311 int error; 3312 struct vfsconf *vfsp; 3313 struct ovfsconf ovfs; 3314 3315 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 3316 bzero(&ovfs, sizeof(ovfs)); 3317 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 3318 strcpy(ovfs.vfc_name, vfsp->vfc_name); 3319 ovfs.vfc_index = vfsp->vfc_typenum; 3320 ovfs.vfc_refcount = vfsp->vfc_refcount; 3321 ovfs.vfc_flags = vfsp->vfc_flags; 3322 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 3323 if (error) 3324 return error; 3325 } 3326 return 0; 3327 } 3328 3329 #endif /* 1 || COMPAT_PRELITE2 */ 3330 #endif /* !BURN_BRIDGES */ 3331 3332 #define KINFO_VNODESLOP 10 3333 #ifdef notyet 3334 /* 3335 * Dump vnode list (via sysctl). 3336 */ 3337 /* ARGSUSED */ 3338 static int 3339 sysctl_vnode(SYSCTL_HANDLER_ARGS) 3340 { 3341 struct xvnode *xvn; 3342 struct mount *mp; 3343 struct vnode *vp; 3344 int error, len, n; 3345 3346 /* 3347 * Stale numvnodes access is not fatal here. 3348 */ 3349 req->lock = 0; 3350 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 3351 if (!req->oldptr) 3352 /* Make an estimate */ 3353 return (SYSCTL_OUT(req, 0, len)); 3354 3355 error = sysctl_wire_old_buffer(req, 0); 3356 if (error != 0) 3357 return (error); 3358 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 3359 n = 0; 3360 mtx_lock(&mountlist_mtx); 3361 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3362 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) 3363 continue; 3364 MNT_ILOCK(mp); 3365 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3366 if (n == len) 3367 break; 3368 vref(vp); 3369 xvn[n].xv_size = sizeof *xvn; 3370 xvn[n].xv_vnode = vp; 3371 xvn[n].xv_id = 0; /* XXX compat */ 3372 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 3373 XV_COPY(usecount); 3374 XV_COPY(writecount); 3375 XV_COPY(holdcnt); 3376 XV_COPY(mount); 3377 XV_COPY(numoutput); 3378 XV_COPY(type); 3379 #undef XV_COPY 3380 xvn[n].xv_flag = vp->v_vflag; 3381 3382 switch (vp->v_type) { 3383 case VREG: 3384 case VDIR: 3385 case VLNK: 3386 break; 3387 case VBLK: 3388 case VCHR: 3389 if (vp->v_rdev == NULL) { 3390 vrele(vp); 3391 continue; 3392 } 3393 xvn[n].xv_dev = dev2udev(vp->v_rdev); 3394 break; 3395 case VSOCK: 3396 xvn[n].xv_socket = vp->v_socket; 3397 break; 3398 case VFIFO: 3399 xvn[n].xv_fifo = vp->v_fifoinfo; 3400 break; 3401 case VNON: 3402 case VBAD: 3403 default: 3404 /* shouldn't happen? */ 3405 vrele(vp); 3406 continue; 3407 } 3408 vrele(vp); 3409 ++n; 3410 } 3411 MNT_IUNLOCK(mp); 3412 mtx_lock(&mountlist_mtx); 3413 vfs_unbusy(mp); 3414 if (n == len) 3415 break; 3416 } 3417 mtx_unlock(&mountlist_mtx); 3418 3419 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 3420 free(xvn, M_TEMP); 3421 return (error); 3422 } 3423 3424 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 3425 0, 0, sysctl_vnode, "S,xvnode", ""); 3426 #endif 3427 3428 /* 3429 * Unmount all filesystems. The list is traversed in reverse order 3430 * of mounting to avoid dependencies. 3431 */ 3432 void 3433 vfs_unmountall(void) 3434 { 3435 struct mount *mp; 3436 struct thread *td; 3437 int error; 3438 3439 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); 3440 td = curthread; 3441 3442 /* 3443 * Since this only runs when rebooting, it is not interlocked. 3444 */ 3445 while(!TAILQ_EMPTY(&mountlist)) { 3446 mp = TAILQ_LAST(&mountlist, mntlist); 3447 error = dounmount(mp, MNT_FORCE, td); 3448 if (error) { 3449 TAILQ_REMOVE(&mountlist, mp, mnt_list); 3450 /* 3451 * XXX: Due to the way in which we mount the root 3452 * file system off of devfs, devfs will generate a 3453 * "busy" warning when we try to unmount it before 3454 * the root. Don't print a warning as a result in 3455 * order to avoid false positive errors that may 3456 * cause needless upset. 3457 */ 3458 if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) { 3459 printf("unmount of %s failed (", 3460 mp->mnt_stat.f_mntonname); 3461 if (error == EBUSY) 3462 printf("BUSY)\n"); 3463 else 3464 printf("%d)\n", error); 3465 } 3466 } else { 3467 /* The unmount has removed mp from the mountlist */ 3468 } 3469 } 3470 } 3471 3472 /* 3473 * perform msync on all vnodes under a mount point 3474 * the mount point must be locked. 3475 */ 3476 void 3477 vfs_msync(struct mount *mp, int flags) 3478 { 3479 struct vnode *vp, *mvp; 3480 struct vm_object *obj; 3481 3482 CTR2(KTR_VFS, "%s: mp %p", __func__, mp); 3483 MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) { 3484 obj = vp->v_object; 3485 if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 && 3486 (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) { 3487 if (!vget(vp, 3488 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 3489 curthread)) { 3490 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 3491 vput(vp); 3492 continue; 3493 } 3494 3495 obj = vp->v_object; 3496 if (obj != NULL) { 3497 VM_OBJECT_WLOCK(obj); 3498 vm_object_page_clean(obj, 0, 0, 3499 flags == MNT_WAIT ? 3500 OBJPC_SYNC : OBJPC_NOSYNC); 3501 VM_OBJECT_WUNLOCK(obj); 3502 } 3503 vput(vp); 3504 } 3505 } else 3506 VI_UNLOCK(vp); 3507 } 3508 } 3509 3510 static void 3511 destroy_vpollinfo(struct vpollinfo *vi) 3512 { 3513 seldrain(&vi->vpi_selinfo); 3514 knlist_destroy(&vi->vpi_selinfo.si_note); 3515 mtx_destroy(&vi->vpi_lock); 3516 uma_zfree(vnodepoll_zone, vi); 3517 } 3518 3519 /* 3520 * Initalize per-vnode helper structure to hold poll-related state. 3521 */ 3522 void 3523 v_addpollinfo(struct vnode *vp) 3524 { 3525 struct vpollinfo *vi; 3526 3527 if (vp->v_pollinfo != NULL) 3528 return; 3529 vi = uma_zalloc(vnodepoll_zone, M_WAITOK); 3530 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 3531 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 3532 vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked); 3533 VI_LOCK(vp); 3534 if (vp->v_pollinfo != NULL) { 3535 VI_UNLOCK(vp); 3536 destroy_vpollinfo(vi); 3537 return; 3538 } 3539 vp->v_pollinfo = vi; 3540 VI_UNLOCK(vp); 3541 } 3542 3543 /* 3544 * Record a process's interest in events which might happen to 3545 * a vnode. Because poll uses the historic select-style interface 3546 * internally, this routine serves as both the ``check for any 3547 * pending events'' and the ``record my interest in future events'' 3548 * functions. (These are done together, while the lock is held, 3549 * to avoid race conditions.) 3550 */ 3551 int 3552 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 3553 { 3554 3555 v_addpollinfo(vp); 3556 mtx_lock(&vp->v_pollinfo->vpi_lock); 3557 if (vp->v_pollinfo->vpi_revents & events) { 3558 /* 3559 * This leaves events we are not interested 3560 * in available for the other process which 3561 * which presumably had requested them 3562 * (otherwise they would never have been 3563 * recorded). 3564 */ 3565 events &= vp->v_pollinfo->vpi_revents; 3566 vp->v_pollinfo->vpi_revents &= ~events; 3567 3568 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3569 return (events); 3570 } 3571 vp->v_pollinfo->vpi_events |= events; 3572 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 3573 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3574 return (0); 3575 } 3576 3577 /* 3578 * Routine to create and manage a filesystem syncer vnode. 3579 */ 3580 #define sync_close ((int (*)(struct vop_close_args *))nullop) 3581 static int sync_fsync(struct vop_fsync_args *); 3582 static int sync_inactive(struct vop_inactive_args *); 3583 static int sync_reclaim(struct vop_reclaim_args *); 3584 3585 static struct vop_vector sync_vnodeops = { 3586 .vop_bypass = VOP_EOPNOTSUPP, 3587 .vop_close = sync_close, /* close */ 3588 .vop_fsync = sync_fsync, /* fsync */ 3589 .vop_inactive = sync_inactive, /* inactive */ 3590 .vop_reclaim = sync_reclaim, /* reclaim */ 3591 .vop_lock1 = vop_stdlock, /* lock */ 3592 .vop_unlock = vop_stdunlock, /* unlock */ 3593 .vop_islocked = vop_stdislocked, /* islocked */ 3594 }; 3595 3596 /* 3597 * Create a new filesystem syncer vnode for the specified mount point. 3598 */ 3599 void 3600 vfs_allocate_syncvnode(struct mount *mp) 3601 { 3602 struct vnode *vp; 3603 struct bufobj *bo; 3604 static long start, incr, next; 3605 int error; 3606 3607 /* Allocate a new vnode */ 3608 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); 3609 if (error != 0) 3610 panic("vfs_allocate_syncvnode: getnewvnode() failed"); 3611 vp->v_type = VNON; 3612 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3613 vp->v_vflag |= VV_FORCEINSMQ; 3614 error = insmntque(vp, mp); 3615 if (error != 0) 3616 panic("vfs_allocate_syncvnode: insmntque() failed"); 3617 vp->v_vflag &= ~VV_FORCEINSMQ; 3618 VOP_UNLOCK(vp, 0); 3619 /* 3620 * Place the vnode onto the syncer worklist. We attempt to 3621 * scatter them about on the list so that they will go off 3622 * at evenly distributed times even if all the filesystems 3623 * are mounted at once. 3624 */ 3625 next += incr; 3626 if (next == 0 || next > syncer_maxdelay) { 3627 start /= 2; 3628 incr /= 2; 3629 if (start == 0) { 3630 start = syncer_maxdelay / 2; 3631 incr = syncer_maxdelay; 3632 } 3633 next = start; 3634 } 3635 bo = &vp->v_bufobj; 3636 BO_LOCK(bo); 3637 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 3638 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 3639 mtx_lock(&sync_mtx); 3640 sync_vnode_count++; 3641 if (mp->mnt_syncer == NULL) { 3642 mp->mnt_syncer = vp; 3643 vp = NULL; 3644 } 3645 mtx_unlock(&sync_mtx); 3646 BO_UNLOCK(bo); 3647 if (vp != NULL) { 3648 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3649 vgone(vp); 3650 vput(vp); 3651 } 3652 } 3653 3654 void 3655 vfs_deallocate_syncvnode(struct mount *mp) 3656 { 3657 struct vnode *vp; 3658 3659 mtx_lock(&sync_mtx); 3660 vp = mp->mnt_syncer; 3661 if (vp != NULL) 3662 mp->mnt_syncer = NULL; 3663 mtx_unlock(&sync_mtx); 3664 if (vp != NULL) 3665 vrele(vp); 3666 } 3667 3668 /* 3669 * Do a lazy sync of the filesystem. 3670 */ 3671 static int 3672 sync_fsync(struct vop_fsync_args *ap) 3673 { 3674 struct vnode *syncvp = ap->a_vp; 3675 struct mount *mp = syncvp->v_mount; 3676 int error, save; 3677 struct bufobj *bo; 3678 3679 /* 3680 * We only need to do something if this is a lazy evaluation. 3681 */ 3682 if (ap->a_waitfor != MNT_LAZY) 3683 return (0); 3684 3685 /* 3686 * Move ourselves to the back of the sync list. 3687 */ 3688 bo = &syncvp->v_bufobj; 3689 BO_LOCK(bo); 3690 vn_syncer_add_to_worklist(bo, syncdelay); 3691 BO_UNLOCK(bo); 3692 3693 /* 3694 * Walk the list of vnodes pushing all that are dirty and 3695 * not already on the sync list. 3696 */ 3697 mtx_lock(&mountlist_mtx); 3698 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) != 0) { 3699 mtx_unlock(&mountlist_mtx); 3700 return (0); 3701 } 3702 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 3703 vfs_unbusy(mp); 3704 return (0); 3705 } 3706 save = curthread_pflags_set(TDP_SYNCIO); 3707 vfs_msync(mp, MNT_NOWAIT); 3708 error = VFS_SYNC(mp, MNT_LAZY); 3709 curthread_pflags_restore(save); 3710 vn_finished_write(mp); 3711 vfs_unbusy(mp); 3712 return (error); 3713 } 3714 3715 /* 3716 * The syncer vnode is no referenced. 3717 */ 3718 static int 3719 sync_inactive(struct vop_inactive_args *ap) 3720 { 3721 3722 vgone(ap->a_vp); 3723 return (0); 3724 } 3725 3726 /* 3727 * The syncer vnode is no longer needed and is being decommissioned. 3728 * 3729 * Modifications to the worklist must be protected by sync_mtx. 3730 */ 3731 static int 3732 sync_reclaim(struct vop_reclaim_args *ap) 3733 { 3734 struct vnode *vp = ap->a_vp; 3735 struct bufobj *bo; 3736 3737 bo = &vp->v_bufobj; 3738 BO_LOCK(bo); 3739 mtx_lock(&sync_mtx); 3740 if (vp->v_mount->mnt_syncer == vp) 3741 vp->v_mount->mnt_syncer = NULL; 3742 if (bo->bo_flag & BO_ONWORKLST) { 3743 LIST_REMOVE(bo, bo_synclist); 3744 syncer_worklist_len--; 3745 sync_vnode_count--; 3746 bo->bo_flag &= ~BO_ONWORKLST; 3747 } 3748 mtx_unlock(&sync_mtx); 3749 BO_UNLOCK(bo); 3750 3751 return (0); 3752 } 3753 3754 /* 3755 * Check if vnode represents a disk device 3756 */ 3757 int 3758 vn_isdisk(struct vnode *vp, int *errp) 3759 { 3760 int error; 3761 3762 error = 0; 3763 dev_lock(); 3764 if (vp->v_type != VCHR) 3765 error = ENOTBLK; 3766 else if (vp->v_rdev == NULL) 3767 error = ENXIO; 3768 else if (vp->v_rdev->si_devsw == NULL) 3769 error = ENXIO; 3770 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 3771 error = ENOTBLK; 3772 dev_unlock(); 3773 if (errp != NULL) 3774 *errp = error; 3775 return (error == 0); 3776 } 3777 3778 /* 3779 * Common filesystem object access control check routine. Accepts a 3780 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3781 * and optional call-by-reference privused argument allowing vaccess() 3782 * to indicate to the caller whether privilege was used to satisfy the 3783 * request (obsoleted). Returns 0 on success, or an errno on failure. 3784 */ 3785 int 3786 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 3787 accmode_t accmode, struct ucred *cred, int *privused) 3788 { 3789 accmode_t dac_granted; 3790 accmode_t priv_granted; 3791 3792 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, 3793 ("invalid bit in accmode")); 3794 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), 3795 ("VAPPEND without VWRITE")); 3796 3797 /* 3798 * Look for a normal, non-privileged way to access the file/directory 3799 * as requested. If it exists, go with that. 3800 */ 3801 3802 if (privused != NULL) 3803 *privused = 0; 3804 3805 dac_granted = 0; 3806 3807 /* Check the owner. */ 3808 if (cred->cr_uid == file_uid) { 3809 dac_granted |= VADMIN; 3810 if (file_mode & S_IXUSR) 3811 dac_granted |= VEXEC; 3812 if (file_mode & S_IRUSR) 3813 dac_granted |= VREAD; 3814 if (file_mode & S_IWUSR) 3815 dac_granted |= (VWRITE | VAPPEND); 3816 3817 if ((accmode & dac_granted) == accmode) 3818 return (0); 3819 3820 goto privcheck; 3821 } 3822 3823 /* Otherwise, check the groups (first match) */ 3824 if (groupmember(file_gid, cred)) { 3825 if (file_mode & S_IXGRP) 3826 dac_granted |= VEXEC; 3827 if (file_mode & S_IRGRP) 3828 dac_granted |= VREAD; 3829 if (file_mode & S_IWGRP) 3830 dac_granted |= (VWRITE | VAPPEND); 3831 3832 if ((accmode & dac_granted) == accmode) 3833 return (0); 3834 3835 goto privcheck; 3836 } 3837 3838 /* Otherwise, check everyone else. */ 3839 if (file_mode & S_IXOTH) 3840 dac_granted |= VEXEC; 3841 if (file_mode & S_IROTH) 3842 dac_granted |= VREAD; 3843 if (file_mode & S_IWOTH) 3844 dac_granted |= (VWRITE | VAPPEND); 3845 if ((accmode & dac_granted) == accmode) 3846 return (0); 3847 3848 privcheck: 3849 /* 3850 * Build a privilege mask to determine if the set of privileges 3851 * satisfies the requirements when combined with the granted mask 3852 * from above. For each privilege, if the privilege is required, 3853 * bitwise or the request type onto the priv_granted mask. 3854 */ 3855 priv_granted = 0; 3856 3857 if (type == VDIR) { 3858 /* 3859 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 3860 * requests, instead of PRIV_VFS_EXEC. 3861 */ 3862 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3863 !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0)) 3864 priv_granted |= VEXEC; 3865 } else { 3866 /* 3867 * Ensure that at least one execute bit is on. Otherwise, 3868 * a privileged user will always succeed, and we don't want 3869 * this to happen unless the file really is executable. 3870 */ 3871 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3872 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && 3873 !priv_check_cred(cred, PRIV_VFS_EXEC, 0)) 3874 priv_granted |= VEXEC; 3875 } 3876 3877 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 3878 !priv_check_cred(cred, PRIV_VFS_READ, 0)) 3879 priv_granted |= VREAD; 3880 3881 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3882 !priv_check_cred(cred, PRIV_VFS_WRITE, 0)) 3883 priv_granted |= (VWRITE | VAPPEND); 3884 3885 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3886 !priv_check_cred(cred, PRIV_VFS_ADMIN, 0)) 3887 priv_granted |= VADMIN; 3888 3889 if ((accmode & (priv_granted | dac_granted)) == accmode) { 3890 /* XXX audit: privilege used */ 3891 if (privused != NULL) 3892 *privused = 1; 3893 return (0); 3894 } 3895 3896 return ((accmode & VADMIN) ? EPERM : EACCES); 3897 } 3898 3899 /* 3900 * Credential check based on process requesting service, and per-attribute 3901 * permissions. 3902 */ 3903 int 3904 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 3905 struct thread *td, accmode_t accmode) 3906 { 3907 3908 /* 3909 * Kernel-invoked always succeeds. 3910 */ 3911 if (cred == NOCRED) 3912 return (0); 3913 3914 /* 3915 * Do not allow privileged processes in jail to directly manipulate 3916 * system attributes. 3917 */ 3918 switch (attrnamespace) { 3919 case EXTATTR_NAMESPACE_SYSTEM: 3920 /* Potentially should be: return (EPERM); */ 3921 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0)); 3922 case EXTATTR_NAMESPACE_USER: 3923 return (VOP_ACCESS(vp, accmode, cred, td)); 3924 default: 3925 return (EPERM); 3926 } 3927 } 3928 3929 #ifdef DEBUG_VFS_LOCKS 3930 /* 3931 * This only exists to supress warnings from unlocked specfs accesses. It is 3932 * no longer ok to have an unlocked VFS. 3933 */ 3934 #define IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL || \ 3935 (vp)->v_type == VCHR || (vp)->v_type == VBAD) 3936 3937 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 3938 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, 3939 "Drop into debugger on lock violation"); 3940 3941 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 3942 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 3943 0, "Check for interlock across VOPs"); 3944 3945 int vfs_badlock_print = 1; /* Print lock violations. */ 3946 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 3947 0, "Print lock violations"); 3948 3949 #ifdef KDB 3950 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 3951 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, 3952 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations"); 3953 #endif 3954 3955 static void 3956 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 3957 { 3958 3959 #ifdef KDB 3960 if (vfs_badlock_backtrace) 3961 kdb_backtrace(); 3962 #endif 3963 if (vfs_badlock_print) 3964 printf("%s: %p %s\n", str, (void *)vp, msg); 3965 if (vfs_badlock_ddb) 3966 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 3967 } 3968 3969 void 3970 assert_vi_locked(struct vnode *vp, const char *str) 3971 { 3972 3973 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 3974 vfs_badlock("interlock is not locked but should be", str, vp); 3975 } 3976 3977 void 3978 assert_vi_unlocked(struct vnode *vp, const char *str) 3979 { 3980 3981 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 3982 vfs_badlock("interlock is locked but should not be", str, vp); 3983 } 3984 3985 void 3986 assert_vop_locked(struct vnode *vp, const char *str) 3987 { 3988 int locked; 3989 3990 if (!IGNORE_LOCK(vp)) { 3991 locked = VOP_ISLOCKED(vp); 3992 if (locked == 0 || locked == LK_EXCLOTHER) 3993 vfs_badlock("is not locked but should be", str, vp); 3994 } 3995 } 3996 3997 void 3998 assert_vop_unlocked(struct vnode *vp, const char *str) 3999 { 4000 4001 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE) 4002 vfs_badlock("is locked but should not be", str, vp); 4003 } 4004 4005 void 4006 assert_vop_elocked(struct vnode *vp, const char *str) 4007 { 4008 4009 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 4010 vfs_badlock("is not exclusive locked but should be", str, vp); 4011 } 4012 4013 #if 0 4014 void 4015 assert_vop_elocked_other(struct vnode *vp, const char *str) 4016 { 4017 4018 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLOTHER) 4019 vfs_badlock("is not exclusive locked by another thread", 4020 str, vp); 4021 } 4022 4023 void 4024 assert_vop_slocked(struct vnode *vp, const char *str) 4025 { 4026 4027 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_SHARED) 4028 vfs_badlock("is not locked shared but should be", str, vp); 4029 } 4030 #endif /* 0 */ 4031 #endif /* DEBUG_VFS_LOCKS */ 4032 4033 void 4034 vop_rename_fail(struct vop_rename_args *ap) 4035 { 4036 4037 if (ap->a_tvp != NULL) 4038 vput(ap->a_tvp); 4039 if (ap->a_tdvp == ap->a_tvp) 4040 vrele(ap->a_tdvp); 4041 else 4042 vput(ap->a_tdvp); 4043 vrele(ap->a_fdvp); 4044 vrele(ap->a_fvp); 4045 } 4046 4047 void 4048 vop_rename_pre(void *ap) 4049 { 4050 struct vop_rename_args *a = ap; 4051 4052 #ifdef DEBUG_VFS_LOCKS 4053 if (a->a_tvp) 4054 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 4055 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 4056 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 4057 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 4058 4059 /* Check the source (from). */ 4060 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && 4061 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) 4062 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 4063 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) 4064 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 4065 4066 /* Check the target. */ 4067 if (a->a_tvp) 4068 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 4069 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 4070 #endif 4071 if (a->a_tdvp != a->a_fdvp) 4072 vhold(a->a_fdvp); 4073 if (a->a_tvp != a->a_fvp) 4074 vhold(a->a_fvp); 4075 vhold(a->a_tdvp); 4076 if (a->a_tvp) 4077 vhold(a->a_tvp); 4078 } 4079 4080 void 4081 vop_strategy_pre(void *ap) 4082 { 4083 #ifdef DEBUG_VFS_LOCKS 4084 struct vop_strategy_args *a; 4085 struct buf *bp; 4086 4087 a = ap; 4088 bp = a->a_bp; 4089 4090 /* 4091 * Cluster ops lock their component buffers but not the IO container. 4092 */ 4093 if ((bp->b_flags & B_CLUSTER) != 0) 4094 return; 4095 4096 if (panicstr == NULL && !BUF_ISLOCKED(bp)) { 4097 if (vfs_badlock_print) 4098 printf( 4099 "VOP_STRATEGY: bp is not locked but should be\n"); 4100 if (vfs_badlock_ddb) 4101 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 4102 } 4103 #endif 4104 } 4105 4106 void 4107 vop_lock_pre(void *ap) 4108 { 4109 #ifdef DEBUG_VFS_LOCKS 4110 struct vop_lock1_args *a = ap; 4111 4112 if ((a->a_flags & LK_INTERLOCK) == 0) 4113 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 4114 else 4115 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 4116 #endif 4117 } 4118 4119 void 4120 vop_lock_post(void *ap, int rc) 4121 { 4122 #ifdef DEBUG_VFS_LOCKS 4123 struct vop_lock1_args *a = ap; 4124 4125 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 4126 if (rc == 0) 4127 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 4128 #endif 4129 } 4130 4131 void 4132 vop_unlock_pre(void *ap) 4133 { 4134 #ifdef DEBUG_VFS_LOCKS 4135 struct vop_unlock_args *a = ap; 4136 4137 if (a->a_flags & LK_INTERLOCK) 4138 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 4139 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 4140 #endif 4141 } 4142 4143 void 4144 vop_unlock_post(void *ap, int rc) 4145 { 4146 #ifdef DEBUG_VFS_LOCKS 4147 struct vop_unlock_args *a = ap; 4148 4149 if (a->a_flags & LK_INTERLOCK) 4150 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 4151 #endif 4152 } 4153 4154 void 4155 vop_create_post(void *ap, int rc) 4156 { 4157 struct vop_create_args *a = ap; 4158 4159 if (!rc) 4160 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4161 } 4162 4163 void 4164 vop_deleteextattr_post(void *ap, int rc) 4165 { 4166 struct vop_deleteextattr_args *a = ap; 4167 4168 if (!rc) 4169 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 4170 } 4171 4172 void 4173 vop_link_post(void *ap, int rc) 4174 { 4175 struct vop_link_args *a = ap; 4176 4177 if (!rc) { 4178 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK); 4179 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE); 4180 } 4181 } 4182 4183 void 4184 vop_mkdir_post(void *ap, int rc) 4185 { 4186 struct vop_mkdir_args *a = ap; 4187 4188 if (!rc) 4189 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 4190 } 4191 4192 void 4193 vop_mknod_post(void *ap, int rc) 4194 { 4195 struct vop_mknod_args *a = ap; 4196 4197 if (!rc) 4198 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4199 } 4200 4201 void 4202 vop_remove_post(void *ap, int rc) 4203 { 4204 struct vop_remove_args *a = ap; 4205 4206 if (!rc) { 4207 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4208 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 4209 } 4210 } 4211 4212 void 4213 vop_rename_post(void *ap, int rc) 4214 { 4215 struct vop_rename_args *a = ap; 4216 4217 if (!rc) { 4218 VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE); 4219 VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE); 4220 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 4221 if (a->a_tvp) 4222 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 4223 } 4224 if (a->a_tdvp != a->a_fdvp) 4225 vdrop(a->a_fdvp); 4226 if (a->a_tvp != a->a_fvp) 4227 vdrop(a->a_fvp); 4228 vdrop(a->a_tdvp); 4229 if (a->a_tvp) 4230 vdrop(a->a_tvp); 4231 } 4232 4233 void 4234 vop_rmdir_post(void *ap, int rc) 4235 { 4236 struct vop_rmdir_args *a = ap; 4237 4238 if (!rc) { 4239 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 4240 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 4241 } 4242 } 4243 4244 void 4245 vop_setattr_post(void *ap, int rc) 4246 { 4247 struct vop_setattr_args *a = ap; 4248 4249 if (!rc) 4250 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 4251 } 4252 4253 void 4254 vop_setextattr_post(void *ap, int rc) 4255 { 4256 struct vop_setextattr_args *a = ap; 4257 4258 if (!rc) 4259 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 4260 } 4261 4262 void 4263 vop_symlink_post(void *ap, int rc) 4264 { 4265 struct vop_symlink_args *a = ap; 4266 4267 if (!rc) 4268 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 4269 } 4270 4271 static struct knlist fs_knlist; 4272 4273 static void 4274 vfs_event_init(void *arg) 4275 { 4276 knlist_init_mtx(&fs_knlist, NULL); 4277 } 4278 /* XXX - correct order? */ 4279 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 4280 4281 void 4282 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) 4283 { 4284 4285 KNOTE_UNLOCKED(&fs_knlist, event); 4286 } 4287 4288 static int filt_fsattach(struct knote *kn); 4289 static void filt_fsdetach(struct knote *kn); 4290 static int filt_fsevent(struct knote *kn, long hint); 4291 4292 struct filterops fs_filtops = { 4293 .f_isfd = 0, 4294 .f_attach = filt_fsattach, 4295 .f_detach = filt_fsdetach, 4296 .f_event = filt_fsevent 4297 }; 4298 4299 static int 4300 filt_fsattach(struct knote *kn) 4301 { 4302 4303 kn->kn_flags |= EV_CLEAR; 4304 knlist_add(&fs_knlist, kn, 0); 4305 return (0); 4306 } 4307 4308 static void 4309 filt_fsdetach(struct knote *kn) 4310 { 4311 4312 knlist_remove(&fs_knlist, kn, 0); 4313 } 4314 4315 static int 4316 filt_fsevent(struct knote *kn, long hint) 4317 { 4318 4319 kn->kn_fflags |= hint; 4320 return (kn->kn_fflags != 0); 4321 } 4322 4323 static int 4324 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 4325 { 4326 struct vfsidctl vc; 4327 int error; 4328 struct mount *mp; 4329 4330 error = SYSCTL_IN(req, &vc, sizeof(vc)); 4331 if (error) 4332 return (error); 4333 if (vc.vc_vers != VFS_CTL_VERS1) 4334 return (EINVAL); 4335 mp = vfs_getvfs(&vc.vc_fsid); 4336 if (mp == NULL) 4337 return (ENOENT); 4338 /* ensure that a specific sysctl goes to the right filesystem. */ 4339 if (strcmp(vc.vc_fstypename, "*") != 0 && 4340 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 4341 vfs_rel(mp); 4342 return (EINVAL); 4343 } 4344 VCTLTOREQ(&vc, req); 4345 error = VFS_SYSCTL(mp, vc.vc_op, req); 4346 vfs_rel(mp); 4347 return (error); 4348 } 4349 4350 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR, 4351 NULL, 0, sysctl_vfs_ctl, "", 4352 "Sysctl by fsid"); 4353 4354 /* 4355 * Function to initialize a va_filerev field sensibly. 4356 * XXX: Wouldn't a random number make a lot more sense ?? 4357 */ 4358 u_quad_t 4359 init_va_filerev(void) 4360 { 4361 struct bintime bt; 4362 4363 getbinuptime(&bt); 4364 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 4365 } 4366 4367 static int filt_vfsread(struct knote *kn, long hint); 4368 static int filt_vfswrite(struct knote *kn, long hint); 4369 static int filt_vfsvnode(struct knote *kn, long hint); 4370 static void filt_vfsdetach(struct knote *kn); 4371 static struct filterops vfsread_filtops = { 4372 .f_isfd = 1, 4373 .f_detach = filt_vfsdetach, 4374 .f_event = filt_vfsread 4375 }; 4376 static struct filterops vfswrite_filtops = { 4377 .f_isfd = 1, 4378 .f_detach = filt_vfsdetach, 4379 .f_event = filt_vfswrite 4380 }; 4381 static struct filterops vfsvnode_filtops = { 4382 .f_isfd = 1, 4383 .f_detach = filt_vfsdetach, 4384 .f_event = filt_vfsvnode 4385 }; 4386 4387 static void 4388 vfs_knllock(void *arg) 4389 { 4390 struct vnode *vp = arg; 4391 4392 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4393 } 4394 4395 static void 4396 vfs_knlunlock(void *arg) 4397 { 4398 struct vnode *vp = arg; 4399 4400 VOP_UNLOCK(vp, 0); 4401 } 4402 4403 static void 4404 vfs_knl_assert_locked(void *arg) 4405 { 4406 #ifdef DEBUG_VFS_LOCKS 4407 struct vnode *vp = arg; 4408 4409 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); 4410 #endif 4411 } 4412 4413 static void 4414 vfs_knl_assert_unlocked(void *arg) 4415 { 4416 #ifdef DEBUG_VFS_LOCKS 4417 struct vnode *vp = arg; 4418 4419 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); 4420 #endif 4421 } 4422 4423 int 4424 vfs_kqfilter(struct vop_kqfilter_args *ap) 4425 { 4426 struct vnode *vp = ap->a_vp; 4427 struct knote *kn = ap->a_kn; 4428 struct knlist *knl; 4429 4430 switch (kn->kn_filter) { 4431 case EVFILT_READ: 4432 kn->kn_fop = &vfsread_filtops; 4433 break; 4434 case EVFILT_WRITE: 4435 kn->kn_fop = &vfswrite_filtops; 4436 break; 4437 case EVFILT_VNODE: 4438 kn->kn_fop = &vfsvnode_filtops; 4439 break; 4440 default: 4441 return (EINVAL); 4442 } 4443 4444 kn->kn_hook = (caddr_t)vp; 4445 4446 v_addpollinfo(vp); 4447 if (vp->v_pollinfo == NULL) 4448 return (ENOMEM); 4449 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 4450 knlist_add(knl, kn, 0); 4451 4452 return (0); 4453 } 4454 4455 /* 4456 * Detach knote from vnode 4457 */ 4458 static void 4459 filt_vfsdetach(struct knote *kn) 4460 { 4461 struct vnode *vp = (struct vnode *)kn->kn_hook; 4462 4463 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 4464 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 4465 } 4466 4467 /*ARGSUSED*/ 4468 static int 4469 filt_vfsread(struct knote *kn, long hint) 4470 { 4471 struct vnode *vp = (struct vnode *)kn->kn_hook; 4472 struct vattr va; 4473 int res; 4474 4475 /* 4476 * filesystem is gone, so set the EOF flag and schedule 4477 * the knote for deletion. 4478 */ 4479 if (hint == NOTE_REVOKE) { 4480 VI_LOCK(vp); 4481 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 4482 VI_UNLOCK(vp); 4483 return (1); 4484 } 4485 4486 if (VOP_GETATTR(vp, &va, curthread->td_ucred)) 4487 return (0); 4488 4489 VI_LOCK(vp); 4490 kn->kn_data = va.va_size - kn->kn_fp->f_offset; 4491 res = (kn->kn_data != 0); 4492 VI_UNLOCK(vp); 4493 return (res); 4494 } 4495 4496 /*ARGSUSED*/ 4497 static int 4498 filt_vfswrite(struct knote *kn, long hint) 4499 { 4500 struct vnode *vp = (struct vnode *)kn->kn_hook; 4501 4502 VI_LOCK(vp); 4503 4504 /* 4505 * filesystem is gone, so set the EOF flag and schedule 4506 * the knote for deletion. 4507 */ 4508 if (hint == NOTE_REVOKE) 4509 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 4510 4511 kn->kn_data = 0; 4512 VI_UNLOCK(vp); 4513 return (1); 4514 } 4515 4516 static int 4517 filt_vfsvnode(struct knote *kn, long hint) 4518 { 4519 struct vnode *vp = (struct vnode *)kn->kn_hook; 4520 int res; 4521 4522 VI_LOCK(vp); 4523 if (kn->kn_sfflags & hint) 4524 kn->kn_fflags |= hint; 4525 if (hint == NOTE_REVOKE) { 4526 kn->kn_flags |= EV_EOF; 4527 VI_UNLOCK(vp); 4528 return (1); 4529 } 4530 res = (kn->kn_fflags != 0); 4531 VI_UNLOCK(vp); 4532 return (res); 4533 } 4534 4535 int 4536 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 4537 { 4538 int error; 4539 4540 if (dp->d_reclen > ap->a_uio->uio_resid) 4541 return (ENAMETOOLONG); 4542 error = uiomove(dp, dp->d_reclen, ap->a_uio); 4543 if (error) { 4544 if (ap->a_ncookies != NULL) { 4545 if (ap->a_cookies != NULL) 4546 free(ap->a_cookies, M_TEMP); 4547 ap->a_cookies = NULL; 4548 *ap->a_ncookies = 0; 4549 } 4550 return (error); 4551 } 4552 if (ap->a_ncookies == NULL) 4553 return (0); 4554 4555 KASSERT(ap->a_cookies, 4556 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 4557 4558 *ap->a_cookies = realloc(*ap->a_cookies, 4559 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); 4560 (*ap->a_cookies)[*ap->a_ncookies] = off; 4561 return (0); 4562 } 4563 4564 /* 4565 * Mark for update the access time of the file if the filesystem 4566 * supports VOP_MARKATIME. This functionality is used by execve and 4567 * mmap, so we want to avoid the I/O implied by directly setting 4568 * va_atime for the sake of efficiency. 4569 */ 4570 void 4571 vfs_mark_atime(struct vnode *vp, struct ucred *cred) 4572 { 4573 struct mount *mp; 4574 4575 mp = vp->v_mount; 4576 ASSERT_VOP_LOCKED(vp, "vfs_mark_atime"); 4577 if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) 4578 (void)VOP_MARKATIME(vp); 4579 } 4580 4581 /* 4582 * The purpose of this routine is to remove granularity from accmode_t, 4583 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, 4584 * VADMIN and VAPPEND. 4585 * 4586 * If it returns 0, the caller is supposed to continue with the usual 4587 * access checks using 'accmode' as modified by this routine. If it 4588 * returns nonzero value, the caller is supposed to return that value 4589 * as errno. 4590 * 4591 * Note that after this routine runs, accmode may be zero. 4592 */ 4593 int 4594 vfs_unixify_accmode(accmode_t *accmode) 4595 { 4596 /* 4597 * There is no way to specify explicit "deny" rule using 4598 * file mode or POSIX.1e ACLs. 4599 */ 4600 if (*accmode & VEXPLICIT_DENY) { 4601 *accmode = 0; 4602 return (0); 4603 } 4604 4605 /* 4606 * None of these can be translated into usual access bits. 4607 * Also, the common case for NFSv4 ACLs is to not contain 4608 * either of these bits. Caller should check for VWRITE 4609 * on the containing directory instead. 4610 */ 4611 if (*accmode & (VDELETE_CHILD | VDELETE)) 4612 return (EPERM); 4613 4614 if (*accmode & VADMIN_PERMS) { 4615 *accmode &= ~VADMIN_PERMS; 4616 *accmode |= VADMIN; 4617 } 4618 4619 /* 4620 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL 4621 * or VSYNCHRONIZE using file mode or POSIX.1e ACL. 4622 */ 4623 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); 4624 4625 return (0); 4626 } 4627 4628 /* 4629 * These are helper functions for filesystems to traverse all 4630 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. 4631 * 4632 * This interface replaces MNT_VNODE_FOREACH. 4633 */ 4634 4635 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker"); 4636 4637 struct vnode * 4638 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) 4639 { 4640 struct vnode *vp; 4641 4642 if (should_yield()) 4643 kern_yield(PRI_USER); 4644 MNT_ILOCK(mp); 4645 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 4646 vp = TAILQ_NEXT(*mvp, v_nmntvnodes); 4647 while (vp != NULL && (vp->v_type == VMARKER || 4648 (vp->v_iflag & VI_DOOMED) != 0)) 4649 vp = TAILQ_NEXT(vp, v_nmntvnodes); 4650 4651 /* Check if we are done */ 4652 if (vp == NULL) { 4653 __mnt_vnode_markerfree_all(mvp, mp); 4654 /* MNT_IUNLOCK(mp); -- done in above function */ 4655 mtx_assert(MNT_MTX(mp), MA_NOTOWNED); 4656 return (NULL); 4657 } 4658 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 4659 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 4660 VI_LOCK(vp); 4661 MNT_IUNLOCK(mp); 4662 return (vp); 4663 } 4664 4665 struct vnode * 4666 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) 4667 { 4668 struct vnode *vp; 4669 4670 *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 4671 MNT_ILOCK(mp); 4672 MNT_REF(mp); 4673 (*mvp)->v_type = VMARKER; 4674 4675 vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 4676 while (vp != NULL && (vp->v_type == VMARKER || 4677 (vp->v_iflag & VI_DOOMED) != 0)) 4678 vp = TAILQ_NEXT(vp, v_nmntvnodes); 4679 4680 /* Check if we are done */ 4681 if (vp == NULL) { 4682 MNT_REL(mp); 4683 MNT_IUNLOCK(mp); 4684 free(*mvp, M_VNODE_MARKER); 4685 *mvp = NULL; 4686 return (NULL); 4687 } 4688 (*mvp)->v_mount = mp; 4689 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); 4690 VI_LOCK(vp); 4691 MNT_IUNLOCK(mp); 4692 return (vp); 4693 } 4694 4695 4696 void 4697 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) 4698 { 4699 4700 if (*mvp == NULL) { 4701 MNT_IUNLOCK(mp); 4702 return; 4703 } 4704 4705 mtx_assert(MNT_MTX(mp), MA_OWNED); 4706 4707 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 4708 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); 4709 MNT_REL(mp); 4710 MNT_IUNLOCK(mp); 4711 free(*mvp, M_VNODE_MARKER); 4712 *mvp = NULL; 4713 } 4714 4715 /* 4716 * These are helper functions for filesystems to traverse their 4717 * active vnodes. See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h 4718 */ 4719 static void 4720 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp) 4721 { 4722 4723 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 4724 4725 MNT_ILOCK(mp); 4726 MNT_REL(mp); 4727 MNT_IUNLOCK(mp); 4728 free(*mvp, M_VNODE_MARKER); 4729 *mvp = NULL; 4730 } 4731 4732 static struct vnode * 4733 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp) 4734 { 4735 struct vnode *vp, *nvp; 4736 4737 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 4738 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); 4739 restart: 4740 vp = TAILQ_NEXT(*mvp, v_actfreelist); 4741 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist); 4742 while (vp != NULL) { 4743 if (vp->v_type == VMARKER) { 4744 vp = TAILQ_NEXT(vp, v_actfreelist); 4745 continue; 4746 } 4747 if (!VI_TRYLOCK(vp)) { 4748 if (mp_ncpus == 1 || should_yield()) { 4749 TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist); 4750 mtx_unlock(&vnode_free_list_mtx); 4751 kern_yield(PRI_USER); 4752 mtx_lock(&vnode_free_list_mtx); 4753 goto restart; 4754 } 4755 continue; 4756 } 4757 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); 4758 KASSERT(vp->v_mount == mp || vp->v_mount == NULL, 4759 ("alien vnode on the active list %p %p", vp, mp)); 4760 if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0) 4761 break; 4762 nvp = TAILQ_NEXT(vp, v_actfreelist); 4763 VI_UNLOCK(vp); 4764 vp = nvp; 4765 } 4766 4767 /* Check if we are done */ 4768 if (vp == NULL) { 4769 mtx_unlock(&vnode_free_list_mtx); 4770 mnt_vnode_markerfree_active(mvp, mp); 4771 return (NULL); 4772 } 4773 TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist); 4774 mtx_unlock(&vnode_free_list_mtx); 4775 ASSERT_VI_LOCKED(vp, "active iter"); 4776 KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp)); 4777 return (vp); 4778 } 4779 4780 struct vnode * 4781 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp) 4782 { 4783 4784 if (should_yield()) 4785 kern_yield(PRI_USER); 4786 mtx_lock(&vnode_free_list_mtx); 4787 return (mnt_vnode_next_active(mvp, mp)); 4788 } 4789 4790 struct vnode * 4791 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp) 4792 { 4793 struct vnode *vp; 4794 4795 *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); 4796 MNT_ILOCK(mp); 4797 MNT_REF(mp); 4798 MNT_IUNLOCK(mp); 4799 (*mvp)->v_type = VMARKER; 4800 (*mvp)->v_mount = mp; 4801 4802 mtx_lock(&vnode_free_list_mtx); 4803 vp = TAILQ_FIRST(&mp->mnt_activevnodelist); 4804 if (vp == NULL) { 4805 mtx_unlock(&vnode_free_list_mtx); 4806 mnt_vnode_markerfree_active(mvp, mp); 4807 return (NULL); 4808 } 4809 TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist); 4810 return (mnt_vnode_next_active(mvp, mp)); 4811 } 4812 4813 void 4814 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp) 4815 { 4816 4817 if (*mvp == NULL) 4818 return; 4819 4820 mtx_lock(&vnode_free_list_mtx); 4821 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist); 4822 mtx_unlock(&vnode_free_list_mtx); 4823 mnt_vnode_markerfree_active(mvp, mp); 4824 } 4825