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_ddb.h" 45 #include "opt_mac.h" 46 47 #include <sys/param.h> 48 #include <sys/systm.h> 49 #include <sys/bio.h> 50 #include <sys/buf.h> 51 #include <sys/condvar.h> 52 #include <sys/conf.h> 53 #include <sys/dirent.h> 54 #include <sys/event.h> 55 #include <sys/eventhandler.h> 56 #include <sys/extattr.h> 57 #include <sys/file.h> 58 #include <sys/fcntl.h> 59 #include <sys/jail.h> 60 #include <sys/kdb.h> 61 #include <sys/kernel.h> 62 #include <sys/kthread.h> 63 #include <sys/lockf.h> 64 #include <sys/malloc.h> 65 #include <sys/mount.h> 66 #include <sys/namei.h> 67 #include <sys/priv.h> 68 #include <sys/reboot.h> 69 #include <sys/sleepqueue.h> 70 #include <sys/stat.h> 71 #include <sys/sysctl.h> 72 #include <sys/syslog.h> 73 #include <sys/vmmeter.h> 74 #include <sys/vnode.h> 75 76 #include <machine/stdarg.h> 77 78 #include <security/mac/mac_framework.h> 79 80 #include <vm/vm.h> 81 #include <vm/vm_object.h> 82 #include <vm/vm_extern.h> 83 #include <vm/pmap.h> 84 #include <vm/vm_map.h> 85 #include <vm/vm_page.h> 86 #include <vm/vm_kern.h> 87 #include <vm/uma.h> 88 89 #ifdef DDB 90 #include <ddb/ddb.h> 91 #endif 92 93 #define WI_MPSAFEQ 0 94 #define WI_GIANTQ 1 95 96 static MALLOC_DEFINE(M_NETADDR, "subr_export_host", "Export host address structure"); 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 vbusy(struct vnode *vp); 104 static void vinactive(struct vnode *, struct thread *); 105 static void v_incr_usecount(struct vnode *); 106 static void v_decr_usecount(struct vnode *); 107 static void v_decr_useonly(struct vnode *); 108 static void v_upgrade_usecount(struct vnode *); 109 static void vfree(struct vnode *); 110 static void vnlru_free(int); 111 static void vgonel(struct vnode *); 112 static void vfs_knllock(void *arg); 113 static void vfs_knlunlock(void *arg); 114 static int vfs_knllocked(void *arg); 115 static void destroy_vpollinfo(struct vpollinfo *vi); 116 117 /* 118 * Enable Giant pushdown based on whether or not the vm is mpsafe in this 119 * build. Without mpsafevm the buffer cache can not run Giant free. 120 */ 121 int mpsafe_vfs = 1; 122 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs); 123 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0, 124 "MPSAFE VFS"); 125 126 /* 127 * Number of vnodes in existence. Increased whenever getnewvnode() 128 * allocates a new vnode, decreased on vdestroy() called on VI_DOOMed 129 * vnode. 130 */ 131 static unsigned long numvnodes; 132 133 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); 134 135 /* 136 * Conversion tables for conversion from vnode types to inode formats 137 * and back. 138 */ 139 enum vtype iftovt_tab[16] = { 140 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 141 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 142 }; 143 int vttoif_tab[10] = { 144 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 145 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT 146 }; 147 148 /* 149 * List of vnodes that are ready for recycling. 150 */ 151 static TAILQ_HEAD(freelst, vnode) vnode_free_list; 152 153 /* 154 * Free vnode target. Free vnodes may simply be files which have been stat'd 155 * but not read. This is somewhat common, and a small cache of such files 156 * should be kept to avoid recreation costs. 157 */ 158 static u_long wantfreevnodes; 159 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 160 /* Number of vnodes in the free list. */ 161 static u_long freevnodes; 162 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); 163 164 /* 165 * Various variables used for debugging the new implementation of 166 * reassignbuf(). 167 * XXX these are probably of (very) limited utility now. 168 */ 169 static int reassignbufcalls; 170 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); 171 172 /* 173 * Cache for the mount type id assigned to NFS. This is used for 174 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 175 */ 176 int nfs_mount_type = -1; 177 178 /* To keep more than one thread at a time from running vfs_getnewfsid */ 179 static struct mtx mntid_mtx; 180 181 /* 182 * Lock for any access to the following: 183 * vnode_free_list 184 * numvnodes 185 * freevnodes 186 */ 187 static struct mtx vnode_free_list_mtx; 188 189 /* Publicly exported FS */ 190 struct nfs_public nfs_pub; 191 192 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 193 static uma_zone_t vnode_zone; 194 static uma_zone_t vnodepoll_zone; 195 196 /* Set to 1 to print out reclaim of active vnodes */ 197 int prtactive; 198 199 /* 200 * The workitem queue. 201 * 202 * It is useful to delay writes of file data and filesystem metadata 203 * for tens of seconds so that quickly created and deleted files need 204 * not waste disk bandwidth being created and removed. To realize this, 205 * we append vnodes to a "workitem" queue. When running with a soft 206 * updates implementation, most pending metadata dependencies should 207 * not wait for more than a few seconds. Thus, mounted on block devices 208 * are delayed only about a half the time that file data is delayed. 209 * Similarly, directory updates are more critical, so are only delayed 210 * about a third the time that file data is delayed. Thus, there are 211 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 212 * one each second (driven off the filesystem syncer process). The 213 * syncer_delayno variable indicates the next queue that is to be processed. 214 * Items that need to be processed soon are placed in this queue: 215 * 216 * syncer_workitem_pending[syncer_delayno] 217 * 218 * A delay of fifteen seconds is done by placing the request fifteen 219 * entries later in the queue: 220 * 221 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 222 * 223 */ 224 static int syncer_delayno; 225 static long syncer_mask; 226 LIST_HEAD(synclist, bufobj); 227 static struct synclist *syncer_workitem_pending[2]; 228 /* 229 * The sync_mtx protects: 230 * bo->bo_synclist 231 * sync_vnode_count 232 * syncer_delayno 233 * syncer_state 234 * syncer_workitem_pending 235 * syncer_worklist_len 236 * rushjob 237 */ 238 static struct mtx sync_mtx; 239 static struct cv sync_wakeup; 240 241 #define SYNCER_MAXDELAY 32 242 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 243 static int syncdelay = 30; /* max time to delay syncing data */ 244 static int filedelay = 30; /* time to delay syncing files */ 245 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); 246 static int dirdelay = 29; /* time to delay syncing directories */ 247 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); 248 static int metadelay = 28; /* time to delay syncing metadata */ 249 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); 250 static int rushjob; /* number of slots to run ASAP */ 251 static int stat_rush_requests; /* number of times I/O speeded up */ 252 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); 253 254 /* 255 * When shutting down the syncer, run it at four times normal speed. 256 */ 257 #define SYNCER_SHUTDOWN_SPEEDUP 4 258 static int sync_vnode_count; 259 static int syncer_worklist_len; 260 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } 261 syncer_state; 262 263 /* 264 * Number of vnodes we want to exist at any one time. This is mostly used 265 * to size hash tables in vnode-related code. It is normally not used in 266 * getnewvnode(), as wantfreevnodes is normally nonzero.) 267 * 268 * XXX desiredvnodes is historical cruft and should not exist. 269 */ 270 int desiredvnodes; 271 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 272 &desiredvnodes, 0, "Maximum number of vnodes"); 273 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 274 &wantfreevnodes, 0, "Minimum number of vnodes (legacy)"); 275 static int vnlru_nowhere; 276 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 277 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 278 279 /* 280 * Macros to control when a vnode is freed and recycled. All require 281 * the vnode interlock. 282 */ 283 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt) 284 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt) 285 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt) 286 287 288 /* 289 * Initialize the vnode management data structures. 290 */ 291 #ifndef MAXVNODES_MAX 292 #define MAXVNODES_MAX 100000 293 #endif 294 static void 295 vntblinit(void *dummy __unused) 296 { 297 298 /* 299 * Desiredvnodes is a function of the physical memory size and 300 * the kernel's heap size. Specifically, desiredvnodes scales 301 * in proportion to the physical memory size until two fifths 302 * of the kernel's heap size is consumed by vnodes and vm 303 * objects. 304 */ 305 desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size / 306 (5 * (sizeof(struct vm_object) + sizeof(struct vnode)))); 307 if (desiredvnodes > MAXVNODES_MAX) { 308 if (bootverbose) 309 printf("Reducing kern.maxvnodes %d -> %d\n", 310 desiredvnodes, MAXVNODES_MAX); 311 desiredvnodes = MAXVNODES_MAX; 312 } 313 wantfreevnodes = desiredvnodes / 4; 314 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 315 TAILQ_INIT(&vnode_free_list); 316 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 317 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 318 NULL, NULL, UMA_ALIGN_PTR, 0); 319 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 320 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 321 /* 322 * Initialize the filesystem syncer. 323 */ 324 syncer_workitem_pending[WI_MPSAFEQ] = hashinit(syncer_maxdelay, M_VNODE, 325 &syncer_mask); 326 syncer_workitem_pending[WI_GIANTQ] = hashinit(syncer_maxdelay, M_VNODE, 327 &syncer_mask); 328 syncer_maxdelay = syncer_mask + 1; 329 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 330 cv_init(&sync_wakeup, "syncer"); 331 } 332 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL); 333 334 335 /* 336 * Mark a mount point as busy. Used to synchronize access and to delay 337 * unmounting. Eventually, mountlist_mtx is not released on failure. 338 */ 339 int 340 vfs_busy(struct mount *mp, int flags) 341 { 342 343 MPASS((flags & ~MBF_MASK) == 0); 344 345 MNT_ILOCK(mp); 346 MNT_REF(mp); 347 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 348 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) { 349 MNT_REL(mp); 350 MNT_IUNLOCK(mp); 351 return (ENOENT); 352 } 353 if (flags & MBF_MNTLSTLOCK) 354 mtx_unlock(&mountlist_mtx); 355 mp->mnt_kern_flag |= MNTK_MWAIT; 356 msleep(mp, MNT_MTX(mp), PVFS, "vfs_busy", 0); 357 MNT_REL(mp); 358 MNT_IUNLOCK(mp); 359 if (flags & MBF_MNTLSTLOCK) 360 mtx_lock(&mountlist_mtx); 361 return (ENOENT); 362 } 363 if (flags & MBF_MNTLSTLOCK) 364 mtx_unlock(&mountlist_mtx); 365 mp->mnt_lockref++; 366 MNT_IUNLOCK(mp); 367 return (0); 368 } 369 370 /* 371 * Free a busy filesystem. 372 */ 373 void 374 vfs_unbusy(struct mount *mp) 375 { 376 377 MNT_ILOCK(mp); 378 MNT_REL(mp); 379 mp->mnt_lockref--; 380 if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) { 381 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT); 382 mp->mnt_kern_flag &= ~MNTK_DRAINING; 383 wakeup(&mp->mnt_lockref); 384 } 385 MNT_IUNLOCK(mp); 386 } 387 388 /* 389 * Lookup a mount point by filesystem identifier. 390 */ 391 struct mount * 392 vfs_getvfs(fsid_t *fsid) 393 { 394 struct mount *mp; 395 396 mtx_lock(&mountlist_mtx); 397 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 398 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 399 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 400 vfs_ref(mp); 401 mtx_unlock(&mountlist_mtx); 402 return (mp); 403 } 404 } 405 mtx_unlock(&mountlist_mtx); 406 return ((struct mount *) 0); 407 } 408 409 /* 410 * Lookup a mount point by filesystem identifier, busying it before 411 * returning. 412 */ 413 struct mount * 414 vfs_busyfs(fsid_t *fsid) 415 { 416 struct mount *mp; 417 int error; 418 419 mtx_lock(&mountlist_mtx); 420 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 421 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 422 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 423 error = vfs_busy(mp, MBF_MNTLSTLOCK); 424 if (error) { 425 mtx_unlock(&mountlist_mtx); 426 return (NULL); 427 } 428 return (mp); 429 } 430 } 431 mtx_unlock(&mountlist_mtx); 432 return ((struct mount *) 0); 433 } 434 435 /* 436 * Check if a user can access privileged mount options. 437 */ 438 int 439 vfs_suser(struct mount *mp, struct thread *td) 440 { 441 int error; 442 443 /* 444 * If the thread is jailed, but this is not a jail-friendly file 445 * system, deny immediately. 446 */ 447 if (!(mp->mnt_vfc->vfc_flags & VFCF_JAIL) && jailed(td->td_ucred)) 448 return (EPERM); 449 450 /* 451 * If the file system was mounted outside a jail and a jailed thread 452 * tries to access it, deny immediately. 453 */ 454 if (!jailed(mp->mnt_cred) && jailed(td->td_ucred)) 455 return (EPERM); 456 457 /* 458 * If the file system was mounted inside different jail that the jail of 459 * the calling thread, deny immediately. 460 */ 461 if (jailed(mp->mnt_cred) && jailed(td->td_ucred) && 462 mp->mnt_cred->cr_prison != td->td_ucred->cr_prison) { 463 return (EPERM); 464 } 465 466 /* 467 * If file system supports delegated administration, we don't check 468 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified 469 * by the file system itself. 470 * If this is not the user that did original mount, we check for 471 * the PRIV_VFS_MOUNT_OWNER privilege. 472 */ 473 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) && 474 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { 475 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) 476 return (error); 477 } 478 return (0); 479 } 480 481 /* 482 * Get a new unique fsid. Try to make its val[0] unique, since this value 483 * will be used to create fake device numbers for stat(). Also try (but 484 * not so hard) make its val[0] unique mod 2^16, since some emulators only 485 * support 16-bit device numbers. We end up with unique val[0]'s for the 486 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 487 * 488 * Keep in mind that several mounts may be running in parallel. Starting 489 * the search one past where the previous search terminated is both a 490 * micro-optimization and a defense against returning the same fsid to 491 * different mounts. 492 */ 493 void 494 vfs_getnewfsid(struct mount *mp) 495 { 496 static u_int16_t mntid_base; 497 struct mount *nmp; 498 fsid_t tfsid; 499 int mtype; 500 501 mtx_lock(&mntid_mtx); 502 mtype = mp->mnt_vfc->vfc_typenum; 503 tfsid.val[1] = mtype; 504 mtype = (mtype & 0xFF) << 24; 505 for (;;) { 506 tfsid.val[0] = makedev(255, 507 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 508 mntid_base++; 509 if ((nmp = vfs_getvfs(&tfsid)) == NULL) 510 break; 511 vfs_rel(nmp); 512 } 513 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 514 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 515 mtx_unlock(&mntid_mtx); 516 } 517 518 /* 519 * Knob to control the precision of file timestamps: 520 * 521 * 0 = seconds only; nanoseconds zeroed. 522 * 1 = seconds and nanoseconds, accurate within 1/HZ. 523 * 2 = seconds and nanoseconds, truncated to microseconds. 524 * >=3 = seconds and nanoseconds, maximum precision. 525 */ 526 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 527 528 static int timestamp_precision = TSP_SEC; 529 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 530 ×tamp_precision, 0, ""); 531 532 /* 533 * Get a current timestamp. 534 */ 535 void 536 vfs_timestamp(struct timespec *tsp) 537 { 538 struct timeval tv; 539 540 switch (timestamp_precision) { 541 case TSP_SEC: 542 tsp->tv_sec = time_second; 543 tsp->tv_nsec = 0; 544 break; 545 case TSP_HZ: 546 getnanotime(tsp); 547 break; 548 case TSP_USEC: 549 microtime(&tv); 550 TIMEVAL_TO_TIMESPEC(&tv, tsp); 551 break; 552 case TSP_NSEC: 553 default: 554 nanotime(tsp); 555 break; 556 } 557 } 558 559 /* 560 * Set vnode attributes to VNOVAL 561 */ 562 void 563 vattr_null(struct vattr *vap) 564 { 565 566 vap->va_type = VNON; 567 vap->va_size = VNOVAL; 568 vap->va_bytes = VNOVAL; 569 vap->va_mode = VNOVAL; 570 vap->va_nlink = VNOVAL; 571 vap->va_uid = VNOVAL; 572 vap->va_gid = VNOVAL; 573 vap->va_fsid = VNOVAL; 574 vap->va_fileid = VNOVAL; 575 vap->va_blocksize = VNOVAL; 576 vap->va_rdev = VNOVAL; 577 vap->va_atime.tv_sec = VNOVAL; 578 vap->va_atime.tv_nsec = VNOVAL; 579 vap->va_mtime.tv_sec = VNOVAL; 580 vap->va_mtime.tv_nsec = VNOVAL; 581 vap->va_ctime.tv_sec = VNOVAL; 582 vap->va_ctime.tv_nsec = VNOVAL; 583 vap->va_birthtime.tv_sec = VNOVAL; 584 vap->va_birthtime.tv_nsec = VNOVAL; 585 vap->va_flags = VNOVAL; 586 vap->va_gen = VNOVAL; 587 vap->va_vaflags = 0; 588 } 589 590 /* 591 * This routine is called when we have too many vnodes. It attempts 592 * to free <count> vnodes and will potentially free vnodes that still 593 * have VM backing store (VM backing store is typically the cause 594 * of a vnode blowout so we want to do this). Therefore, this operation 595 * is not considered cheap. 596 * 597 * A number of conditions may prevent a vnode from being reclaimed. 598 * the buffer cache may have references on the vnode, a directory 599 * vnode may still have references due to the namei cache representing 600 * underlying files, or the vnode may be in active use. It is not 601 * desireable to reuse such vnodes. These conditions may cause the 602 * number of vnodes to reach some minimum value regardless of what 603 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 604 */ 605 static int 606 vlrureclaim(struct mount *mp) 607 { 608 struct vnode *vp; 609 int done; 610 int trigger; 611 int usevnodes; 612 int count; 613 614 /* 615 * Calculate the trigger point, don't allow user 616 * screwups to blow us up. This prevents us from 617 * recycling vnodes with lots of resident pages. We 618 * aren't trying to free memory, we are trying to 619 * free vnodes. 620 */ 621 usevnodes = desiredvnodes; 622 if (usevnodes <= 0) 623 usevnodes = 1; 624 trigger = cnt.v_page_count * 2 / usevnodes; 625 done = 0; 626 vn_start_write(NULL, &mp, V_WAIT); 627 MNT_ILOCK(mp); 628 count = mp->mnt_nvnodelistsize / 10 + 1; 629 while (count != 0) { 630 vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 631 while (vp != NULL && vp->v_type == VMARKER) 632 vp = TAILQ_NEXT(vp, v_nmntvnodes); 633 if (vp == NULL) 634 break; 635 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 636 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 637 --count; 638 if (!VI_TRYLOCK(vp)) 639 goto next_iter; 640 /* 641 * If it's been deconstructed already, it's still 642 * referenced, or it exceeds the trigger, skip it. 643 */ 644 if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) || 645 (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL && 646 vp->v_object->resident_page_count > trigger)) { 647 VI_UNLOCK(vp); 648 goto next_iter; 649 } 650 MNT_IUNLOCK(mp); 651 vholdl(vp); 652 if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) { 653 vdrop(vp); 654 goto next_iter_mntunlocked; 655 } 656 VI_LOCK(vp); 657 /* 658 * v_usecount may have been bumped after VOP_LOCK() dropped 659 * the vnode interlock and before it was locked again. 660 * 661 * It is not necessary to recheck VI_DOOMED because it can 662 * only be set by another thread that holds both the vnode 663 * lock and vnode interlock. If another thread has the 664 * vnode lock before we get to VOP_LOCK() and obtains the 665 * vnode interlock after VOP_LOCK() drops the vnode 666 * interlock, the other thread will be unable to drop the 667 * vnode lock before our VOP_LOCK() call fails. 668 */ 669 if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) || 670 (vp->v_object != NULL && 671 vp->v_object->resident_page_count > trigger)) { 672 VOP_UNLOCK(vp, LK_INTERLOCK); 673 goto next_iter_mntunlocked; 674 } 675 KASSERT((vp->v_iflag & VI_DOOMED) == 0, 676 ("VI_DOOMED unexpectedly detected in vlrureclaim()")); 677 vgonel(vp); 678 VOP_UNLOCK(vp, 0); 679 vdropl(vp); 680 done++; 681 next_iter_mntunlocked: 682 if ((count % 256) != 0) 683 goto relock_mnt; 684 goto yield; 685 next_iter: 686 if ((count % 256) != 0) 687 continue; 688 MNT_IUNLOCK(mp); 689 yield: 690 uio_yield(); 691 relock_mnt: 692 MNT_ILOCK(mp); 693 } 694 MNT_IUNLOCK(mp); 695 vn_finished_write(mp); 696 return done; 697 } 698 699 /* 700 * Attempt to keep the free list at wantfreevnodes length. 701 */ 702 static void 703 vnlru_free(int count) 704 { 705 struct vnode *vp; 706 int vfslocked; 707 708 mtx_assert(&vnode_free_list_mtx, MA_OWNED); 709 for (; count > 0; count--) { 710 vp = TAILQ_FIRST(&vnode_free_list); 711 /* 712 * The list can be modified while the free_list_mtx 713 * has been dropped and vp could be NULL here. 714 */ 715 if (!vp) 716 break; 717 VNASSERT(vp->v_op != NULL, vp, 718 ("vnlru_free: vnode already reclaimed.")); 719 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 720 /* 721 * Don't recycle if we can't get the interlock. 722 */ 723 if (!VI_TRYLOCK(vp)) { 724 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 725 continue; 726 } 727 VNASSERT(VCANRECYCLE(vp), vp, 728 ("vp inconsistent on freelist")); 729 freevnodes--; 730 vp->v_iflag &= ~VI_FREE; 731 vholdl(vp); 732 mtx_unlock(&vnode_free_list_mtx); 733 VI_UNLOCK(vp); 734 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 735 vtryrecycle(vp); 736 VFS_UNLOCK_GIANT(vfslocked); 737 /* 738 * If the recycled succeeded this vdrop will actually free 739 * the vnode. If not it will simply place it back on 740 * the free list. 741 */ 742 vdrop(vp); 743 mtx_lock(&vnode_free_list_mtx); 744 } 745 } 746 /* 747 * Attempt to recycle vnodes in a context that is always safe to block. 748 * Calling vlrurecycle() from the bowels of filesystem code has some 749 * interesting deadlock problems. 750 */ 751 static struct proc *vnlruproc; 752 static int vnlruproc_sig; 753 754 static void 755 vnlru_proc(void) 756 { 757 struct mount *mp, *nmp; 758 int done, vfslocked; 759 struct proc *p = vnlruproc; 760 761 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 762 SHUTDOWN_PRI_FIRST); 763 764 for (;;) { 765 kproc_suspend_check(p); 766 mtx_lock(&vnode_free_list_mtx); 767 if (freevnodes > wantfreevnodes) 768 vnlru_free(freevnodes - wantfreevnodes); 769 if (numvnodes <= desiredvnodes * 9 / 10) { 770 vnlruproc_sig = 0; 771 wakeup(&vnlruproc_sig); 772 msleep(vnlruproc, &vnode_free_list_mtx, 773 PVFS|PDROP, "vlruwt", hz); 774 continue; 775 } 776 mtx_unlock(&vnode_free_list_mtx); 777 done = 0; 778 mtx_lock(&mountlist_mtx); 779 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 780 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) { 781 nmp = TAILQ_NEXT(mp, mnt_list); 782 continue; 783 } 784 vfslocked = VFS_LOCK_GIANT(mp); 785 done += vlrureclaim(mp); 786 VFS_UNLOCK_GIANT(vfslocked); 787 mtx_lock(&mountlist_mtx); 788 nmp = TAILQ_NEXT(mp, mnt_list); 789 vfs_unbusy(mp); 790 } 791 mtx_unlock(&mountlist_mtx); 792 if (done == 0) { 793 EVENTHANDLER_INVOKE(vfs_lowvnodes, desiredvnodes / 10); 794 #if 0 795 /* These messages are temporary debugging aids */ 796 if (vnlru_nowhere < 5) 797 printf("vnlru process getting nowhere..\n"); 798 else if (vnlru_nowhere == 5) 799 printf("vnlru process messages stopped.\n"); 800 #endif 801 vnlru_nowhere++; 802 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 803 } else 804 uio_yield(); 805 } 806 } 807 808 static struct kproc_desc vnlru_kp = { 809 "vnlru", 810 vnlru_proc, 811 &vnlruproc 812 }; 813 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, 814 &vnlru_kp); 815 816 /* 817 * Routines having to do with the management of the vnode table. 818 */ 819 820 void 821 vdestroy(struct vnode *vp) 822 { 823 struct bufobj *bo; 824 825 CTR1(KTR_VFS, "vdestroy vp %p", vp); 826 mtx_lock(&vnode_free_list_mtx); 827 numvnodes--; 828 mtx_unlock(&vnode_free_list_mtx); 829 bo = &vp->v_bufobj; 830 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, 831 ("cleaned vnode still on the free list.")); 832 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); 833 VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count")); 834 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); 835 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); 836 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); 837 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); 838 VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL")); 839 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); 840 VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL")); 841 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); 842 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); 843 VI_UNLOCK(vp); 844 #ifdef MAC 845 mac_vnode_destroy(vp); 846 #endif 847 if (vp->v_pollinfo != NULL) 848 destroy_vpollinfo(vp->v_pollinfo); 849 #ifdef INVARIANTS 850 /* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */ 851 vp->v_op = NULL; 852 #endif 853 lockdestroy(vp->v_vnlock); 854 mtx_destroy(&vp->v_interlock); 855 mtx_destroy(BO_MTX(bo)); 856 uma_zfree(vnode_zone, vp); 857 } 858 859 /* 860 * Try to recycle a freed vnode. We abort if anyone picks up a reference 861 * before we actually vgone(). This function must be called with the vnode 862 * held to prevent the vnode from being returned to the free list midway 863 * through vgone(). 864 */ 865 static int 866 vtryrecycle(struct vnode *vp) 867 { 868 struct mount *vnmp; 869 870 CTR1(KTR_VFS, "vtryrecycle: trying vp %p", vp); 871 VNASSERT(vp->v_holdcnt, vp, 872 ("vtryrecycle: Recycling vp %p without a reference.", vp)); 873 /* 874 * This vnode may found and locked via some other list, if so we 875 * can't recycle it yet. 876 */ 877 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 878 return (EWOULDBLOCK); 879 /* 880 * Don't recycle if its filesystem is being suspended. 881 */ 882 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 883 VOP_UNLOCK(vp, 0); 884 return (EBUSY); 885 } 886 /* 887 * If we got this far, we need to acquire the interlock and see if 888 * anyone picked up this vnode from another list. If not, we will 889 * mark it with DOOMED via vgonel() so that anyone who does find it 890 * will skip over it. 891 */ 892 VI_LOCK(vp); 893 if (vp->v_usecount) { 894 VOP_UNLOCK(vp, LK_INTERLOCK); 895 vn_finished_write(vnmp); 896 return (EBUSY); 897 } 898 if ((vp->v_iflag & VI_DOOMED) == 0) 899 vgonel(vp); 900 VOP_UNLOCK(vp, LK_INTERLOCK); 901 vn_finished_write(vnmp); 902 CTR1(KTR_VFS, "vtryrecycle: recycled vp %p", vp); 903 return (0); 904 } 905 906 /* 907 * Return the next vnode from the free list. 908 */ 909 int 910 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, 911 struct vnode **vpp) 912 { 913 struct vnode *vp = NULL; 914 struct bufobj *bo; 915 916 mtx_lock(&vnode_free_list_mtx); 917 /* 918 * Lend our context to reclaim vnodes if they've exceeded the max. 919 */ 920 if (freevnodes > wantfreevnodes) 921 vnlru_free(1); 922 /* 923 * Wait for available vnodes. 924 */ 925 if (numvnodes > desiredvnodes) { 926 if (mp != NULL && (mp->mnt_kern_flag & MNTK_SUSPEND)) { 927 /* 928 * File system is beeing suspended, we cannot risk a 929 * deadlock here, so allocate new vnode anyway. 930 */ 931 if (freevnodes > wantfreevnodes) 932 vnlru_free(freevnodes - wantfreevnodes); 933 goto alloc; 934 } 935 if (vnlruproc_sig == 0) { 936 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 937 wakeup(vnlruproc); 938 } 939 msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS, 940 "vlruwk", hz); 941 #if 0 /* XXX Not all VFS_VGET/ffs_vget callers check returns. */ 942 if (numvnodes > desiredvnodes) { 943 mtx_unlock(&vnode_free_list_mtx); 944 return (ENFILE); 945 } 946 #endif 947 } 948 alloc: 949 numvnodes++; 950 mtx_unlock(&vnode_free_list_mtx); 951 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); 952 /* 953 * Setup locks. 954 */ 955 vp->v_vnlock = &vp->v_lock; 956 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 957 /* 958 * By default, don't allow shared locks unless filesystems 959 * opt-in. 960 */ 961 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE); 962 /* 963 * Initialize bufobj. 964 */ 965 bo = &vp->v_bufobj; 966 bo->__bo_vnode = vp; 967 mtx_init(BO_MTX(bo), "bufobj interlock", NULL, MTX_DEF); 968 bo->bo_ops = &buf_ops_bio; 969 bo->bo_private = vp; 970 TAILQ_INIT(&bo->bo_clean.bv_hd); 971 TAILQ_INIT(&bo->bo_dirty.bv_hd); 972 /* 973 * Initialize namecache. 974 */ 975 LIST_INIT(&vp->v_cache_src); 976 TAILQ_INIT(&vp->v_cache_dst); 977 /* 978 * Finalize various vnode identity bits. 979 */ 980 vp->v_type = VNON; 981 vp->v_tag = tag; 982 vp->v_op = vops; 983 v_incr_usecount(vp); 984 vp->v_data = 0; 985 #ifdef MAC 986 mac_vnode_init(vp); 987 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 988 mac_vnode_associate_singlelabel(mp, vp); 989 else if (mp == NULL && vops != &dead_vnodeops) 990 printf("NULL mp in getnewvnode()\n"); 991 #endif 992 if (mp != NULL) { 993 bo->bo_bsize = mp->mnt_stat.f_iosize; 994 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) 995 vp->v_vflag |= VV_NOKNOTE; 996 } 997 998 CTR2(KTR_VFS, "getnewvnode: mp %p vp %p", mp, vp); 999 *vpp = vp; 1000 return (0); 1001 } 1002 1003 /* 1004 * Delete from old mount point vnode list, if on one. 1005 */ 1006 static void 1007 delmntque(struct vnode *vp) 1008 { 1009 struct mount *mp; 1010 1011 mp = vp->v_mount; 1012 if (mp == NULL) 1013 return; 1014 MNT_ILOCK(mp); 1015 vp->v_mount = NULL; 1016 VNASSERT(mp->mnt_nvnodelistsize > 0, vp, 1017 ("bad mount point vnode list size")); 1018 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1019 mp->mnt_nvnodelistsize--; 1020 MNT_REL(mp); 1021 MNT_IUNLOCK(mp); 1022 } 1023 1024 static void 1025 insmntque_stddtr(struct vnode *vp, void *dtr_arg) 1026 { 1027 1028 vp->v_data = NULL; 1029 vp->v_op = &dead_vnodeops; 1030 /* XXX non mp-safe fs may still call insmntque with vnode 1031 unlocked */ 1032 if (!VOP_ISLOCKED(vp)) 1033 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1034 vgone(vp); 1035 vput(vp); 1036 } 1037 1038 /* 1039 * Insert into list of vnodes for the new mount point, if available. 1040 */ 1041 int 1042 insmntque1(struct vnode *vp, struct mount *mp, 1043 void (*dtr)(struct vnode *, void *), void *dtr_arg) 1044 { 1045 int locked; 1046 1047 KASSERT(vp->v_mount == NULL, 1048 ("insmntque: vnode already on per mount vnode list")); 1049 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); 1050 #ifdef DEBUG_VFS_LOCKS 1051 if (!VFS_NEEDSGIANT(mp)) 1052 ASSERT_VOP_ELOCKED(vp, 1053 "insmntque: mp-safe fs and non-locked vp"); 1054 #endif 1055 MNT_ILOCK(mp); 1056 if ((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 && 1057 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || 1058 mp->mnt_nvnodelistsize == 0)) { 1059 locked = VOP_ISLOCKED(vp); 1060 if (!locked || (locked == LK_EXCLUSIVE && 1061 (vp->v_vflag & VV_FORCEINSMQ) == 0)) { 1062 MNT_IUNLOCK(mp); 1063 if (dtr != NULL) 1064 dtr(vp, dtr_arg); 1065 return (EBUSY); 1066 } 1067 } 1068 vp->v_mount = mp; 1069 MNT_REF(mp); 1070 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 1071 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, 1072 ("neg mount point vnode list size")); 1073 mp->mnt_nvnodelistsize++; 1074 MNT_IUNLOCK(mp); 1075 return (0); 1076 } 1077 1078 int 1079 insmntque(struct vnode *vp, struct mount *mp) 1080 { 1081 1082 return (insmntque1(vp, mp, insmntque_stddtr, NULL)); 1083 } 1084 1085 /* 1086 * Flush out and invalidate all buffers associated with a bufobj 1087 * Called with the underlying object locked. 1088 */ 1089 int 1090 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) 1091 { 1092 int error; 1093 1094 BO_LOCK(bo); 1095 if (flags & V_SAVE) { 1096 error = bufobj_wwait(bo, slpflag, slptimeo); 1097 if (error) { 1098 BO_UNLOCK(bo); 1099 return (error); 1100 } 1101 if (bo->bo_dirty.bv_cnt > 0) { 1102 BO_UNLOCK(bo); 1103 if ((error = BO_SYNC(bo, MNT_WAIT)) != 0) 1104 return (error); 1105 /* 1106 * XXX We could save a lock/unlock if this was only 1107 * enabled under INVARIANTS 1108 */ 1109 BO_LOCK(bo); 1110 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) 1111 panic("vinvalbuf: dirty bufs"); 1112 } 1113 } 1114 /* 1115 * If you alter this loop please notice that interlock is dropped and 1116 * reacquired in flushbuflist. Special care is needed to ensure that 1117 * no race conditions occur from this. 1118 */ 1119 do { 1120 error = flushbuflist(&bo->bo_clean, 1121 flags, bo, slpflag, slptimeo); 1122 if (error == 0) 1123 error = flushbuflist(&bo->bo_dirty, 1124 flags, bo, slpflag, slptimeo); 1125 if (error != 0 && error != EAGAIN) { 1126 BO_UNLOCK(bo); 1127 return (error); 1128 } 1129 } while (error != 0); 1130 1131 /* 1132 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 1133 * have write I/O in-progress but if there is a VM object then the 1134 * VM object can also have read-I/O in-progress. 1135 */ 1136 do { 1137 bufobj_wwait(bo, 0, 0); 1138 BO_UNLOCK(bo); 1139 if (bo->bo_object != NULL) { 1140 VM_OBJECT_LOCK(bo->bo_object); 1141 vm_object_pip_wait(bo->bo_object, "bovlbx"); 1142 VM_OBJECT_UNLOCK(bo->bo_object); 1143 } 1144 BO_LOCK(bo); 1145 } while (bo->bo_numoutput > 0); 1146 BO_UNLOCK(bo); 1147 1148 /* 1149 * Destroy the copy in the VM cache, too. 1150 */ 1151 if (bo->bo_object != NULL && (flags & (V_ALT | V_NORMAL)) == 0) { 1152 VM_OBJECT_LOCK(bo->bo_object); 1153 vm_object_page_remove(bo->bo_object, 0, 0, 1154 (flags & V_SAVE) ? TRUE : FALSE); 1155 VM_OBJECT_UNLOCK(bo->bo_object); 1156 } 1157 1158 #ifdef INVARIANTS 1159 BO_LOCK(bo); 1160 if ((flags & (V_ALT | V_NORMAL)) == 0 && 1161 (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0)) 1162 panic("vinvalbuf: flush failed"); 1163 BO_UNLOCK(bo); 1164 #endif 1165 return (0); 1166 } 1167 1168 /* 1169 * Flush out and invalidate all buffers associated with a vnode. 1170 * Called with the underlying object locked. 1171 */ 1172 int 1173 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) 1174 { 1175 1176 CTR2(KTR_VFS, "vinvalbuf vp %p flags %d", vp, flags); 1177 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 1178 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); 1179 } 1180 1181 /* 1182 * Flush out buffers on the specified list. 1183 * 1184 */ 1185 static int 1186 flushbuflist( struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, 1187 int slptimeo) 1188 { 1189 struct buf *bp, *nbp; 1190 int retval, error; 1191 daddr_t lblkno; 1192 b_xflags_t xflags; 1193 1194 ASSERT_BO_LOCKED(bo); 1195 1196 retval = 0; 1197 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { 1198 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1199 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1200 continue; 1201 } 1202 lblkno = 0; 1203 xflags = 0; 1204 if (nbp != NULL) { 1205 lblkno = nbp->b_lblkno; 1206 xflags = nbp->b_xflags & 1207 (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN); 1208 } 1209 retval = EAGAIN; 1210 error = BUF_TIMELOCK(bp, 1211 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo), 1212 "flushbuf", slpflag, slptimeo); 1213 if (error) { 1214 BO_LOCK(bo); 1215 return (error != ENOLCK ? error : EAGAIN); 1216 } 1217 KASSERT(bp->b_bufobj == bo, 1218 ("bp %p wrong b_bufobj %p should be %p", 1219 bp, bp->b_bufobj, bo)); 1220 if (bp->b_bufobj != bo) { /* XXX: necessary ? */ 1221 BUF_UNLOCK(bp); 1222 BO_LOCK(bo); 1223 return (EAGAIN); 1224 } 1225 /* 1226 * XXX Since there are no node locks for NFS, I 1227 * believe there is a slight chance that a delayed 1228 * write will occur while sleeping just above, so 1229 * check for it. 1230 */ 1231 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1232 (flags & V_SAVE)) { 1233 bremfree(bp); 1234 bp->b_flags |= B_ASYNC; 1235 bwrite(bp); 1236 BO_LOCK(bo); 1237 return (EAGAIN); /* XXX: why not loop ? */ 1238 } 1239 bremfree(bp); 1240 bp->b_flags |= (B_INVAL | B_RELBUF); 1241 bp->b_flags &= ~B_ASYNC; 1242 brelse(bp); 1243 BO_LOCK(bo); 1244 if (nbp != NULL && 1245 (nbp->b_bufobj != bo || 1246 nbp->b_lblkno != lblkno || 1247 (nbp->b_xflags & 1248 (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN)) != xflags)) 1249 break; /* nbp invalid */ 1250 } 1251 return (retval); 1252 } 1253 1254 /* 1255 * Truncate a file's buffer and pages to a specified length. This 1256 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1257 * sync activity. 1258 */ 1259 int 1260 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, 1261 off_t length, int blksize) 1262 { 1263 struct buf *bp, *nbp; 1264 int anyfreed; 1265 int trunclbn; 1266 struct bufobj *bo; 1267 1268 CTR2(KTR_VFS, "vtruncbuf vp %p length %jd", vp, length); 1269 /* 1270 * Round up to the *next* lbn. 1271 */ 1272 trunclbn = (length + blksize - 1) / blksize; 1273 1274 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1275 restart: 1276 bo = &vp->v_bufobj; 1277 BO_LOCK(bo); 1278 anyfreed = 1; 1279 for (;anyfreed;) { 1280 anyfreed = 0; 1281 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { 1282 if (bp->b_lblkno < trunclbn) 1283 continue; 1284 if (BUF_LOCK(bp, 1285 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1286 BO_MTX(bo)) == ENOLCK) 1287 goto restart; 1288 1289 bremfree(bp); 1290 bp->b_flags |= (B_INVAL | B_RELBUF); 1291 bp->b_flags &= ~B_ASYNC; 1292 brelse(bp); 1293 anyfreed = 1; 1294 1295 if (nbp != NULL && 1296 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1297 (nbp->b_vp != vp) || 1298 (nbp->b_flags & B_DELWRI))) { 1299 goto restart; 1300 } 1301 BO_LOCK(bo); 1302 } 1303 1304 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1305 if (bp->b_lblkno < trunclbn) 1306 continue; 1307 if (BUF_LOCK(bp, 1308 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1309 BO_MTX(bo)) == ENOLCK) 1310 goto restart; 1311 bremfree(bp); 1312 bp->b_flags |= (B_INVAL | B_RELBUF); 1313 bp->b_flags &= ~B_ASYNC; 1314 brelse(bp); 1315 anyfreed = 1; 1316 if (nbp != NULL && 1317 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1318 (nbp->b_vp != vp) || 1319 (nbp->b_flags & B_DELWRI) == 0)) { 1320 goto restart; 1321 } 1322 BO_LOCK(bo); 1323 } 1324 } 1325 1326 if (length > 0) { 1327 restartsync: 1328 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 1329 if (bp->b_lblkno > 0) 1330 continue; 1331 /* 1332 * Since we hold the vnode lock this should only 1333 * fail if we're racing with the buf daemon. 1334 */ 1335 if (BUF_LOCK(bp, 1336 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1337 BO_MTX(bo)) == ENOLCK) { 1338 goto restart; 1339 } 1340 VNASSERT((bp->b_flags & B_DELWRI), vp, 1341 ("buf(%p) on dirty queue without DELWRI", bp)); 1342 1343 bremfree(bp); 1344 bawrite(bp); 1345 BO_LOCK(bo); 1346 goto restartsync; 1347 } 1348 } 1349 1350 bufobj_wwait(bo, 0, 0); 1351 BO_UNLOCK(bo); 1352 vnode_pager_setsize(vp, length); 1353 1354 return (0); 1355 } 1356 1357 /* 1358 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1359 * a vnode. 1360 * 1361 * NOTE: We have to deal with the special case of a background bitmap 1362 * buffer, a situation where two buffers will have the same logical 1363 * block offset. We want (1) only the foreground buffer to be accessed 1364 * in a lookup and (2) must differentiate between the foreground and 1365 * background buffer in the splay tree algorithm because the splay 1366 * tree cannot normally handle multiple entities with the same 'index'. 1367 * We accomplish this by adding differentiating flags to the splay tree's 1368 * numerical domain. 1369 */ 1370 static 1371 struct buf * 1372 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1373 { 1374 struct buf dummy; 1375 struct buf *lefttreemax, *righttreemin, *y; 1376 1377 if (root == NULL) 1378 return (NULL); 1379 lefttreemax = righttreemin = &dummy; 1380 for (;;) { 1381 if (lblkno < root->b_lblkno || 1382 (lblkno == root->b_lblkno && 1383 (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1384 if ((y = root->b_left) == NULL) 1385 break; 1386 if (lblkno < y->b_lblkno) { 1387 /* Rotate right. */ 1388 root->b_left = y->b_right; 1389 y->b_right = root; 1390 root = y; 1391 if ((y = root->b_left) == NULL) 1392 break; 1393 } 1394 /* Link into the new root's right tree. */ 1395 righttreemin->b_left = root; 1396 righttreemin = root; 1397 } else if (lblkno > root->b_lblkno || 1398 (lblkno == root->b_lblkno && 1399 (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { 1400 if ((y = root->b_right) == NULL) 1401 break; 1402 if (lblkno > y->b_lblkno) { 1403 /* Rotate left. */ 1404 root->b_right = y->b_left; 1405 y->b_left = root; 1406 root = y; 1407 if ((y = root->b_right) == NULL) 1408 break; 1409 } 1410 /* Link into the new root's left tree. */ 1411 lefttreemax->b_right = root; 1412 lefttreemax = root; 1413 } else { 1414 break; 1415 } 1416 root = y; 1417 } 1418 /* Assemble the new root. */ 1419 lefttreemax->b_right = root->b_left; 1420 righttreemin->b_left = root->b_right; 1421 root->b_left = dummy.b_right; 1422 root->b_right = dummy.b_left; 1423 return (root); 1424 } 1425 1426 static void 1427 buf_vlist_remove(struct buf *bp) 1428 { 1429 struct buf *root; 1430 struct bufv *bv; 1431 1432 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1433 ASSERT_BO_LOCKED(bp->b_bufobj); 1434 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) != 1435 (BX_VNDIRTY|BX_VNCLEAN), 1436 ("buf_vlist_remove: Buf %p is on two lists", bp)); 1437 if (bp->b_xflags & BX_VNDIRTY) 1438 bv = &bp->b_bufobj->bo_dirty; 1439 else 1440 bv = &bp->b_bufobj->bo_clean; 1441 if (bp != bv->bv_root) { 1442 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1443 KASSERT(root == bp, ("splay lookup failed in remove")); 1444 } 1445 if (bp->b_left == NULL) { 1446 root = bp->b_right; 1447 } else { 1448 root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left); 1449 root->b_right = bp->b_right; 1450 } 1451 bv->bv_root = root; 1452 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); 1453 bv->bv_cnt--; 1454 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1455 } 1456 1457 /* 1458 * Add the buffer to the sorted clean or dirty block list using a 1459 * splay tree algorithm. 1460 * 1461 * NOTE: xflags is passed as a constant, optimizing this inline function! 1462 */ 1463 static void 1464 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) 1465 { 1466 struct buf *root; 1467 struct bufv *bv; 1468 1469 ASSERT_BO_LOCKED(bo); 1470 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1471 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); 1472 bp->b_xflags |= xflags; 1473 if (xflags & BX_VNDIRTY) 1474 bv = &bo->bo_dirty; 1475 else 1476 bv = &bo->bo_clean; 1477 1478 root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root); 1479 if (root == NULL) { 1480 bp->b_left = NULL; 1481 bp->b_right = NULL; 1482 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); 1483 } else if (bp->b_lblkno < root->b_lblkno || 1484 (bp->b_lblkno == root->b_lblkno && 1485 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1486 bp->b_left = root->b_left; 1487 bp->b_right = root; 1488 root->b_left = NULL; 1489 TAILQ_INSERT_BEFORE(root, bp, b_bobufs); 1490 } else { 1491 bp->b_right = root->b_right; 1492 bp->b_left = root; 1493 root->b_right = NULL; 1494 TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs); 1495 } 1496 bv->bv_cnt++; 1497 bv->bv_root = bp; 1498 } 1499 1500 /* 1501 * Lookup a buffer using the splay tree. Note that we specifically avoid 1502 * shadow buffers used in background bitmap writes. 1503 * 1504 * This code isn't quite efficient as it could be because we are maintaining 1505 * two sorted lists and do not know which list the block resides in. 1506 * 1507 * During a "make buildworld" the desired buffer is found at one of 1508 * the roots more than 60% of the time. Thus, checking both roots 1509 * before performing either splay eliminates unnecessary splays on the 1510 * first tree splayed. 1511 */ 1512 struct buf * 1513 gbincore(struct bufobj *bo, daddr_t lblkno) 1514 { 1515 struct buf *bp; 1516 1517 ASSERT_BO_LOCKED(bo); 1518 if ((bp = bo->bo_clean.bv_root) != NULL && 1519 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1520 return (bp); 1521 if ((bp = bo->bo_dirty.bv_root) != NULL && 1522 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1523 return (bp); 1524 if ((bp = bo->bo_clean.bv_root) != NULL) { 1525 bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp); 1526 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1527 return (bp); 1528 } 1529 if ((bp = bo->bo_dirty.bv_root) != NULL) { 1530 bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp); 1531 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1532 return (bp); 1533 } 1534 return (NULL); 1535 } 1536 1537 /* 1538 * Associate a buffer with a vnode. 1539 */ 1540 void 1541 bgetvp(struct vnode *vp, struct buf *bp) 1542 { 1543 struct bufobj *bo; 1544 1545 bo = &vp->v_bufobj; 1546 ASSERT_BO_LOCKED(bo); 1547 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); 1548 1549 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); 1550 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, 1551 ("bgetvp: bp already attached! %p", bp)); 1552 1553 vhold(vp); 1554 if (VFS_NEEDSGIANT(vp->v_mount) || bo->bo_flag & BO_NEEDSGIANT) 1555 bp->b_flags |= B_NEEDSGIANT; 1556 bp->b_vp = vp; 1557 bp->b_bufobj = bo; 1558 /* 1559 * Insert onto list for new vnode. 1560 */ 1561 buf_vlist_add(bp, bo, BX_VNCLEAN); 1562 } 1563 1564 /* 1565 * Disassociate a buffer from a vnode. 1566 */ 1567 void 1568 brelvp(struct buf *bp) 1569 { 1570 struct bufobj *bo; 1571 struct vnode *vp; 1572 1573 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1574 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1575 1576 /* 1577 * Delete from old vnode list, if on one. 1578 */ 1579 vp = bp->b_vp; /* XXX */ 1580 bo = bp->b_bufobj; 1581 BO_LOCK(bo); 1582 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1583 buf_vlist_remove(bp); 1584 else 1585 panic("brelvp: Buffer %p not on queue.", bp); 1586 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1587 bo->bo_flag &= ~BO_ONWORKLST; 1588 mtx_lock(&sync_mtx); 1589 LIST_REMOVE(bo, bo_synclist); 1590 syncer_worklist_len--; 1591 mtx_unlock(&sync_mtx); 1592 } 1593 bp->b_flags &= ~B_NEEDSGIANT; 1594 bp->b_vp = NULL; 1595 bp->b_bufobj = NULL; 1596 BO_UNLOCK(bo); 1597 vdrop(vp); 1598 } 1599 1600 /* 1601 * Add an item to the syncer work queue. 1602 */ 1603 static void 1604 vn_syncer_add_to_worklist(struct bufobj *bo, int delay) 1605 { 1606 int queue, slot; 1607 1608 ASSERT_BO_LOCKED(bo); 1609 1610 mtx_lock(&sync_mtx); 1611 if (bo->bo_flag & BO_ONWORKLST) 1612 LIST_REMOVE(bo, bo_synclist); 1613 else { 1614 bo->bo_flag |= BO_ONWORKLST; 1615 syncer_worklist_len++; 1616 } 1617 1618 if (delay > syncer_maxdelay - 2) 1619 delay = syncer_maxdelay - 2; 1620 slot = (syncer_delayno + delay) & syncer_mask; 1621 1622 queue = VFS_NEEDSGIANT(bo->__bo_vnode->v_mount) ? WI_GIANTQ : 1623 WI_MPSAFEQ; 1624 LIST_INSERT_HEAD(&syncer_workitem_pending[queue][slot], bo, 1625 bo_synclist); 1626 mtx_unlock(&sync_mtx); 1627 } 1628 1629 static int 1630 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) 1631 { 1632 int error, len; 1633 1634 mtx_lock(&sync_mtx); 1635 len = syncer_worklist_len - sync_vnode_count; 1636 mtx_unlock(&sync_mtx); 1637 error = SYSCTL_OUT(req, &len, sizeof(len)); 1638 return (error); 1639 } 1640 1641 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0, 1642 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); 1643 1644 static struct proc *updateproc; 1645 static void sched_sync(void); 1646 static struct kproc_desc up_kp = { 1647 "syncer", 1648 sched_sync, 1649 &updateproc 1650 }; 1651 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); 1652 1653 static int 1654 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) 1655 { 1656 struct vnode *vp; 1657 struct mount *mp; 1658 1659 *bo = LIST_FIRST(slp); 1660 if (*bo == NULL) 1661 return (0); 1662 vp = (*bo)->__bo_vnode; /* XXX */ 1663 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) 1664 return (1); 1665 /* 1666 * We use vhold in case the vnode does not 1667 * successfully sync. vhold prevents the vnode from 1668 * going away when we unlock the sync_mtx so that 1669 * we can acquire the vnode interlock. 1670 */ 1671 vholdl(vp); 1672 mtx_unlock(&sync_mtx); 1673 VI_UNLOCK(vp); 1674 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1675 vdrop(vp); 1676 mtx_lock(&sync_mtx); 1677 return (*bo == LIST_FIRST(slp)); 1678 } 1679 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1680 (void) VOP_FSYNC(vp, MNT_LAZY, td); 1681 VOP_UNLOCK(vp, 0); 1682 vn_finished_write(mp); 1683 BO_LOCK(*bo); 1684 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { 1685 /* 1686 * Put us back on the worklist. The worklist 1687 * routine will remove us from our current 1688 * position and then add us back in at a later 1689 * position. 1690 */ 1691 vn_syncer_add_to_worklist(*bo, syncdelay); 1692 } 1693 BO_UNLOCK(*bo); 1694 vdrop(vp); 1695 mtx_lock(&sync_mtx); 1696 return (0); 1697 } 1698 1699 /* 1700 * System filesystem synchronizer daemon. 1701 */ 1702 static void 1703 sched_sync(void) 1704 { 1705 struct synclist *gnext, *next; 1706 struct synclist *gslp, *slp; 1707 struct bufobj *bo; 1708 long starttime; 1709 struct thread *td = curthread; 1710 int last_work_seen; 1711 int net_worklist_len; 1712 int syncer_final_iter; 1713 int first_printf; 1714 int error; 1715 1716 last_work_seen = 0; 1717 syncer_final_iter = 0; 1718 first_printf = 1; 1719 syncer_state = SYNCER_RUNNING; 1720 starttime = time_uptime; 1721 td->td_pflags |= TDP_NORUNNINGBUF; 1722 1723 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 1724 SHUTDOWN_PRI_LAST); 1725 1726 mtx_lock(&sync_mtx); 1727 for (;;) { 1728 if (syncer_state == SYNCER_FINAL_DELAY && 1729 syncer_final_iter == 0) { 1730 mtx_unlock(&sync_mtx); 1731 kproc_suspend_check(td->td_proc); 1732 mtx_lock(&sync_mtx); 1733 } 1734 net_worklist_len = syncer_worklist_len - sync_vnode_count; 1735 if (syncer_state != SYNCER_RUNNING && 1736 starttime != time_uptime) { 1737 if (first_printf) { 1738 printf("\nSyncing disks, vnodes remaining..."); 1739 first_printf = 0; 1740 } 1741 printf("%d ", net_worklist_len); 1742 } 1743 starttime = time_uptime; 1744 1745 /* 1746 * Push files whose dirty time has expired. Be careful 1747 * of interrupt race on slp queue. 1748 * 1749 * Skip over empty worklist slots when shutting down. 1750 */ 1751 do { 1752 slp = &syncer_workitem_pending[WI_MPSAFEQ][syncer_delayno]; 1753 gslp = &syncer_workitem_pending[WI_GIANTQ][syncer_delayno]; 1754 syncer_delayno += 1; 1755 if (syncer_delayno == syncer_maxdelay) 1756 syncer_delayno = 0; 1757 next = &syncer_workitem_pending[WI_MPSAFEQ][syncer_delayno]; 1758 gnext = &syncer_workitem_pending[WI_GIANTQ][syncer_delayno]; 1759 /* 1760 * If the worklist has wrapped since the 1761 * it was emptied of all but syncer vnodes, 1762 * switch to the FINAL_DELAY state and run 1763 * for one more second. 1764 */ 1765 if (syncer_state == SYNCER_SHUTTING_DOWN && 1766 net_worklist_len == 0 && 1767 last_work_seen == syncer_delayno) { 1768 syncer_state = SYNCER_FINAL_DELAY; 1769 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; 1770 } 1771 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && 1772 LIST_EMPTY(gslp) && syncer_worklist_len > 0); 1773 1774 /* 1775 * Keep track of the last time there was anything 1776 * on the worklist other than syncer vnodes. 1777 * Return to the SHUTTING_DOWN state if any 1778 * new work appears. 1779 */ 1780 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) 1781 last_work_seen = syncer_delayno; 1782 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) 1783 syncer_state = SYNCER_SHUTTING_DOWN; 1784 while (!LIST_EMPTY(slp)) { 1785 error = sync_vnode(slp, &bo, td); 1786 if (error == 1) { 1787 LIST_REMOVE(bo, bo_synclist); 1788 LIST_INSERT_HEAD(next, bo, bo_synclist); 1789 continue; 1790 } 1791 } 1792 if (!LIST_EMPTY(gslp)) { 1793 mtx_unlock(&sync_mtx); 1794 mtx_lock(&Giant); 1795 mtx_lock(&sync_mtx); 1796 while (!LIST_EMPTY(gslp)) { 1797 error = sync_vnode(gslp, &bo, td); 1798 if (error == 1) { 1799 LIST_REMOVE(bo, bo_synclist); 1800 LIST_INSERT_HEAD(gnext, bo, 1801 bo_synclist); 1802 continue; 1803 } 1804 } 1805 mtx_unlock(&Giant); 1806 } 1807 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) 1808 syncer_final_iter--; 1809 /* 1810 * The variable rushjob allows the kernel to speed up the 1811 * processing of the filesystem syncer process. A rushjob 1812 * value of N tells the filesystem syncer to process the next 1813 * N seconds worth of work on its queue ASAP. Currently rushjob 1814 * is used by the soft update code to speed up the filesystem 1815 * syncer process when the incore state is getting so far 1816 * ahead of the disk that the kernel memory pool is being 1817 * threatened with exhaustion. 1818 */ 1819 if (rushjob > 0) { 1820 rushjob -= 1; 1821 continue; 1822 } 1823 /* 1824 * Just sleep for a short period of time between 1825 * iterations when shutting down to allow some I/O 1826 * to happen. 1827 * 1828 * If it has taken us less than a second to process the 1829 * current work, then wait. Otherwise start right over 1830 * again. We can still lose time if any single round 1831 * takes more than two seconds, but it does not really 1832 * matter as we are just trying to generally pace the 1833 * filesystem activity. 1834 */ 1835 if (syncer_state != SYNCER_RUNNING) 1836 cv_timedwait(&sync_wakeup, &sync_mtx, 1837 hz / SYNCER_SHUTDOWN_SPEEDUP); 1838 else if (time_uptime == starttime) 1839 cv_timedwait(&sync_wakeup, &sync_mtx, hz); 1840 } 1841 } 1842 1843 /* 1844 * Request the syncer daemon to speed up its work. 1845 * We never push it to speed up more than half of its 1846 * normal turn time, otherwise it could take over the cpu. 1847 */ 1848 int 1849 speedup_syncer(void) 1850 { 1851 int ret = 0; 1852 1853 mtx_lock(&sync_mtx); 1854 if (rushjob < syncdelay / 2) { 1855 rushjob += 1; 1856 stat_rush_requests += 1; 1857 ret = 1; 1858 } 1859 mtx_unlock(&sync_mtx); 1860 cv_broadcast(&sync_wakeup); 1861 return (ret); 1862 } 1863 1864 /* 1865 * Tell the syncer to speed up its work and run though its work 1866 * list several times, then tell it to shut down. 1867 */ 1868 static void 1869 syncer_shutdown(void *arg, int howto) 1870 { 1871 1872 if (howto & RB_NOSYNC) 1873 return; 1874 mtx_lock(&sync_mtx); 1875 syncer_state = SYNCER_SHUTTING_DOWN; 1876 rushjob = 0; 1877 mtx_unlock(&sync_mtx); 1878 cv_broadcast(&sync_wakeup); 1879 kproc_shutdown(arg, howto); 1880 } 1881 1882 /* 1883 * Reassign a buffer from one vnode to another. 1884 * Used to assign file specific control information 1885 * (indirect blocks) to the vnode to which they belong. 1886 */ 1887 void 1888 reassignbuf(struct buf *bp) 1889 { 1890 struct vnode *vp; 1891 struct bufobj *bo; 1892 int delay; 1893 #ifdef INVARIANTS 1894 struct bufv *bv; 1895 #endif 1896 1897 vp = bp->b_vp; 1898 bo = bp->b_bufobj; 1899 ++reassignbufcalls; 1900 1901 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", 1902 bp, bp->b_vp, bp->b_flags); 1903 /* 1904 * B_PAGING flagged buffers cannot be reassigned because their vp 1905 * is not fully linked in. 1906 */ 1907 if (bp->b_flags & B_PAGING) 1908 panic("cannot reassign paging buffer"); 1909 1910 /* 1911 * Delete from old vnode list, if on one. 1912 */ 1913 BO_LOCK(bo); 1914 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1915 buf_vlist_remove(bp); 1916 else 1917 panic("reassignbuf: Buffer %p not on queue.", bp); 1918 /* 1919 * If dirty, put on list of dirty buffers; otherwise insert onto list 1920 * of clean buffers. 1921 */ 1922 if (bp->b_flags & B_DELWRI) { 1923 if ((bo->bo_flag & BO_ONWORKLST) == 0) { 1924 switch (vp->v_type) { 1925 case VDIR: 1926 delay = dirdelay; 1927 break; 1928 case VCHR: 1929 delay = metadelay; 1930 break; 1931 default: 1932 delay = filedelay; 1933 } 1934 vn_syncer_add_to_worklist(bo, delay); 1935 } 1936 buf_vlist_add(bp, bo, BX_VNDIRTY); 1937 } else { 1938 buf_vlist_add(bp, bo, BX_VNCLEAN); 1939 1940 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { 1941 mtx_lock(&sync_mtx); 1942 LIST_REMOVE(bo, bo_synclist); 1943 syncer_worklist_len--; 1944 mtx_unlock(&sync_mtx); 1945 bo->bo_flag &= ~BO_ONWORKLST; 1946 } 1947 } 1948 #ifdef INVARIANTS 1949 bv = &bo->bo_clean; 1950 bp = TAILQ_FIRST(&bv->bv_hd); 1951 KASSERT(bp == NULL || bp->b_bufobj == bo, 1952 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 1953 bp = TAILQ_LAST(&bv->bv_hd, buflists); 1954 KASSERT(bp == NULL || bp->b_bufobj == bo, 1955 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 1956 bv = &bo->bo_dirty; 1957 bp = TAILQ_FIRST(&bv->bv_hd); 1958 KASSERT(bp == NULL || bp->b_bufobj == bo, 1959 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 1960 bp = TAILQ_LAST(&bv->bv_hd, buflists); 1961 KASSERT(bp == NULL || bp->b_bufobj == bo, 1962 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 1963 #endif 1964 BO_UNLOCK(bo); 1965 } 1966 1967 /* 1968 * Increment the use and hold counts on the vnode, taking care to reference 1969 * the driver's usecount if this is a chardev. The vholdl() will remove 1970 * the vnode from the free list if it is presently free. Requires the 1971 * vnode interlock and returns with it held. 1972 */ 1973 static void 1974 v_incr_usecount(struct vnode *vp) 1975 { 1976 1977 CTR3(KTR_VFS, "v_incr_usecount: vp %p holdcnt %d usecount %d\n", 1978 vp, vp->v_holdcnt, vp->v_usecount); 1979 vp->v_usecount++; 1980 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1981 dev_lock(); 1982 vp->v_rdev->si_usecount++; 1983 dev_unlock(); 1984 } 1985 vholdl(vp); 1986 } 1987 1988 /* 1989 * Turn a holdcnt into a use+holdcnt such that only one call to 1990 * v_decr_usecount is needed. 1991 */ 1992 static void 1993 v_upgrade_usecount(struct vnode *vp) 1994 { 1995 1996 CTR3(KTR_VFS, "v_upgrade_usecount: vp %p holdcnt %d usecount %d\n", 1997 vp, vp->v_holdcnt, vp->v_usecount); 1998 vp->v_usecount++; 1999 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2000 dev_lock(); 2001 vp->v_rdev->si_usecount++; 2002 dev_unlock(); 2003 } 2004 } 2005 2006 /* 2007 * Decrement the vnode use and hold count along with the driver's usecount 2008 * if this is a chardev. The vdropl() below releases the vnode interlock 2009 * as it may free the vnode. 2010 */ 2011 static void 2012 v_decr_usecount(struct vnode *vp) 2013 { 2014 2015 CTR3(KTR_VFS, "v_decr_usecount: vp %p holdcnt %d usecount %d\n", 2016 vp, vp->v_holdcnt, vp->v_usecount); 2017 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2018 VNASSERT(vp->v_usecount > 0, vp, 2019 ("v_decr_usecount: negative usecount")); 2020 vp->v_usecount--; 2021 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2022 dev_lock(); 2023 vp->v_rdev->si_usecount--; 2024 dev_unlock(); 2025 } 2026 vdropl(vp); 2027 } 2028 2029 /* 2030 * Decrement only the use count and driver use count. This is intended to 2031 * be paired with a follow on vdropl() to release the remaining hold count. 2032 * In this way we may vgone() a vnode with a 0 usecount without risk of 2033 * having it end up on a free list because the hold count is kept above 0. 2034 */ 2035 static void 2036 v_decr_useonly(struct vnode *vp) 2037 { 2038 2039 CTR3(KTR_VFS, "v_decr_useonly: vp %p holdcnt %d usecount %d\n", 2040 vp, vp->v_holdcnt, vp->v_usecount); 2041 ASSERT_VI_LOCKED(vp, __FUNCTION__); 2042 VNASSERT(vp->v_usecount > 0, vp, 2043 ("v_decr_useonly: negative usecount")); 2044 vp->v_usecount--; 2045 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2046 dev_lock(); 2047 vp->v_rdev->si_usecount--; 2048 dev_unlock(); 2049 } 2050 } 2051 2052 /* 2053 * Grab a particular vnode from the free list, increment its 2054 * reference count and lock it. VI_DOOMED is set if the vnode 2055 * is being destroyed. Only callers who specify LK_RETRY will 2056 * see doomed vnodes. If inactive processing was delayed in 2057 * vput try to do it here. 2058 */ 2059 int 2060 vget(struct vnode *vp, int flags, struct thread *td) 2061 { 2062 int error; 2063 2064 error = 0; 2065 VFS_ASSERT_GIANT(vp->v_mount); 2066 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 2067 ("vget: invalid lock operation")); 2068 if ((flags & LK_INTERLOCK) == 0) 2069 VI_LOCK(vp); 2070 vholdl(vp); 2071 if ((error = vn_lock(vp, flags | LK_INTERLOCK)) != 0) { 2072 vdrop(vp); 2073 return (error); 2074 } 2075 if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0) 2076 panic("vget: vn_lock failed to return ENOENT\n"); 2077 VI_LOCK(vp); 2078 /* Upgrade our holdcnt to a usecount. */ 2079 v_upgrade_usecount(vp); 2080 /* 2081 * We don't guarantee that any particular close will 2082 * trigger inactive processing so just make a best effort 2083 * here at preventing a reference to a removed file. If 2084 * we don't succeed no harm is done. 2085 */ 2086 if (vp->v_iflag & VI_OWEINACT) { 2087 if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE && 2088 (flags & LK_NOWAIT) == 0) 2089 vinactive(vp, td); 2090 vp->v_iflag &= ~VI_OWEINACT; 2091 } 2092 VI_UNLOCK(vp); 2093 return (0); 2094 } 2095 2096 /* 2097 * Increase the reference count of a vnode. 2098 */ 2099 void 2100 vref(struct vnode *vp) 2101 { 2102 2103 VI_LOCK(vp); 2104 v_incr_usecount(vp); 2105 VI_UNLOCK(vp); 2106 } 2107 2108 /* 2109 * Return reference count of a vnode. 2110 * 2111 * The results of this call are only guaranteed when some mechanism other 2112 * than the VI lock is used to stop other processes from gaining references 2113 * to the vnode. This may be the case if the caller holds the only reference. 2114 * This is also useful when stale data is acceptable as race conditions may 2115 * be accounted for by some other means. 2116 */ 2117 int 2118 vrefcnt(struct vnode *vp) 2119 { 2120 int usecnt; 2121 2122 VI_LOCK(vp); 2123 usecnt = vp->v_usecount; 2124 VI_UNLOCK(vp); 2125 2126 return (usecnt); 2127 } 2128 2129 2130 /* 2131 * Vnode put/release. 2132 * If count drops to zero, call inactive routine and return to freelist. 2133 */ 2134 void 2135 vrele(struct vnode *vp) 2136 { 2137 struct thread *td = curthread; /* XXX */ 2138 2139 KASSERT(vp != NULL, ("vrele: null vp")); 2140 VFS_ASSERT_GIANT(vp->v_mount); 2141 2142 VI_LOCK(vp); 2143 2144 /* Skip this v_writecount check if we're going to panic below. */ 2145 VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp, 2146 ("vrele: missed vn_close")); 2147 2148 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2149 vp->v_usecount == 1)) { 2150 v_decr_usecount(vp); 2151 return; 2152 } 2153 if (vp->v_usecount != 1) { 2154 #ifdef DIAGNOSTIC 2155 vprint("vrele: negative ref count", vp); 2156 #endif 2157 VI_UNLOCK(vp); 2158 panic("vrele: negative ref cnt"); 2159 } 2160 /* 2161 * We want to hold the vnode until the inactive finishes to 2162 * prevent vgone() races. We drop the use count here and the 2163 * hold count below when we're done. 2164 */ 2165 v_decr_useonly(vp); 2166 /* 2167 * We must call VOP_INACTIVE with the node locked. Mark 2168 * as VI_DOINGINACT to avoid recursion. 2169 */ 2170 vp->v_iflag |= VI_OWEINACT; 2171 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK) == 0) { 2172 VI_LOCK(vp); 2173 if (vp->v_usecount > 0) 2174 vp->v_iflag &= ~VI_OWEINACT; 2175 if (vp->v_iflag & VI_OWEINACT) 2176 vinactive(vp, td); 2177 VOP_UNLOCK(vp, 0); 2178 } else { 2179 VI_LOCK(vp); 2180 if (vp->v_usecount > 0) 2181 vp->v_iflag &= ~VI_OWEINACT; 2182 } 2183 vdropl(vp); 2184 } 2185 2186 /* 2187 * Release an already locked vnode. This give the same effects as 2188 * unlock+vrele(), but takes less time and avoids releasing and 2189 * re-aquiring the lock (as vrele() acquires the lock internally.) 2190 */ 2191 void 2192 vput(struct vnode *vp) 2193 { 2194 struct thread *td = curthread; /* XXX */ 2195 int error; 2196 2197 KASSERT(vp != NULL, ("vput: null vp")); 2198 ASSERT_VOP_LOCKED(vp, "vput"); 2199 VFS_ASSERT_GIANT(vp->v_mount); 2200 VI_LOCK(vp); 2201 /* Skip this v_writecount check if we're going to panic below. */ 2202 VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp, 2203 ("vput: missed vn_close")); 2204 error = 0; 2205 2206 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2207 vp->v_usecount == 1)) { 2208 VOP_UNLOCK(vp, 0); 2209 v_decr_usecount(vp); 2210 return; 2211 } 2212 2213 if (vp->v_usecount != 1) { 2214 #ifdef DIAGNOSTIC 2215 vprint("vput: negative ref count", vp); 2216 #endif 2217 panic("vput: negative ref cnt"); 2218 } 2219 /* 2220 * We want to hold the vnode until the inactive finishes to 2221 * prevent vgone() races. We drop the use count here and the 2222 * hold count below when we're done. 2223 */ 2224 v_decr_useonly(vp); 2225 vp->v_iflag |= VI_OWEINACT; 2226 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { 2227 error = VOP_LOCK(vp, LK_UPGRADE|LK_INTERLOCK|LK_NOWAIT); 2228 VI_LOCK(vp); 2229 if (error) { 2230 if (vp->v_usecount > 0) 2231 vp->v_iflag &= ~VI_OWEINACT; 2232 goto done; 2233 } 2234 } 2235 if (vp->v_usecount > 0) 2236 vp->v_iflag &= ~VI_OWEINACT; 2237 if (vp->v_iflag & VI_OWEINACT) 2238 vinactive(vp, td); 2239 VOP_UNLOCK(vp, 0); 2240 done: 2241 vdropl(vp); 2242 } 2243 2244 /* 2245 * Somebody doesn't want the vnode recycled. 2246 */ 2247 void 2248 vhold(struct vnode *vp) 2249 { 2250 2251 VI_LOCK(vp); 2252 vholdl(vp); 2253 VI_UNLOCK(vp); 2254 } 2255 2256 void 2257 vholdl(struct vnode *vp) 2258 { 2259 2260 vp->v_holdcnt++; 2261 if (VSHOULDBUSY(vp)) 2262 vbusy(vp); 2263 } 2264 2265 /* 2266 * Note that there is one less who cares about this vnode. vdrop() is the 2267 * opposite of vhold(). 2268 */ 2269 void 2270 vdrop(struct vnode *vp) 2271 { 2272 2273 VI_LOCK(vp); 2274 vdropl(vp); 2275 } 2276 2277 /* 2278 * Drop the hold count of the vnode. If this is the last reference to 2279 * the vnode we will free it if it has been vgone'd otherwise it is 2280 * placed on the free list. 2281 */ 2282 void 2283 vdropl(struct vnode *vp) 2284 { 2285 2286 ASSERT_VI_LOCKED(vp, "vdropl"); 2287 if (vp->v_holdcnt <= 0) 2288 panic("vdrop: holdcnt %d", vp->v_holdcnt); 2289 vp->v_holdcnt--; 2290 if (vp->v_holdcnt == 0) { 2291 if (vp->v_iflag & VI_DOOMED) { 2292 vdestroy(vp); 2293 return; 2294 } else 2295 vfree(vp); 2296 } 2297 VI_UNLOCK(vp); 2298 } 2299 2300 /* 2301 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT 2302 * flags. DOINGINACT prevents us from recursing in calls to vinactive. 2303 * OWEINACT tracks whether a vnode missed a call to inactive due to a 2304 * failed lock upgrade. 2305 */ 2306 static void 2307 vinactive(struct vnode *vp, struct thread *td) 2308 { 2309 2310 ASSERT_VOP_ELOCKED(vp, "vinactive"); 2311 ASSERT_VI_LOCKED(vp, "vinactive"); 2312 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, 2313 ("vinactive: recursed on VI_DOINGINACT")); 2314 vp->v_iflag |= VI_DOINGINACT; 2315 vp->v_iflag &= ~VI_OWEINACT; 2316 VI_UNLOCK(vp); 2317 VOP_INACTIVE(vp, td); 2318 VI_LOCK(vp); 2319 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, 2320 ("vinactive: lost VI_DOINGINACT")); 2321 vp->v_iflag &= ~VI_DOINGINACT; 2322 } 2323 2324 /* 2325 * Remove any vnodes in the vnode table belonging to mount point mp. 2326 * 2327 * If FORCECLOSE is not specified, there should not be any active ones, 2328 * return error if any are found (nb: this is a user error, not a 2329 * system error). If FORCECLOSE is specified, detach any active vnodes 2330 * that are found. 2331 * 2332 * If WRITECLOSE is set, only flush out regular file vnodes open for 2333 * writing. 2334 * 2335 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2336 * 2337 * `rootrefs' specifies the base reference count for the root vnode 2338 * of this filesystem. The root vnode is considered busy if its 2339 * v_usecount exceeds this value. On a successful return, vflush(, td) 2340 * will call vrele() on the root vnode exactly rootrefs times. 2341 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2342 * be zero. 2343 */ 2344 #ifdef DIAGNOSTIC 2345 static int busyprt = 0; /* print out busy vnodes */ 2346 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 2347 #endif 2348 2349 int 2350 vflush( struct mount *mp, int rootrefs, int flags, struct thread *td) 2351 { 2352 struct vnode *vp, *mvp, *rootvp = NULL; 2353 struct vattr vattr; 2354 int busy = 0, error; 2355 2356 CTR1(KTR_VFS, "vflush: mp %p", mp); 2357 if (rootrefs > 0) { 2358 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2359 ("vflush: bad args")); 2360 /* 2361 * Get the filesystem root vnode. We can vput() it 2362 * immediately, since with rootrefs > 0, it won't go away. 2363 */ 2364 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0) 2365 return (error); 2366 vput(rootvp); 2367 2368 } 2369 MNT_ILOCK(mp); 2370 loop: 2371 MNT_VNODE_FOREACH(vp, mp, mvp) { 2372 2373 VI_LOCK(vp); 2374 vholdl(vp); 2375 MNT_IUNLOCK(mp); 2376 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); 2377 if (error) { 2378 vdrop(vp); 2379 MNT_ILOCK(mp); 2380 MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp); 2381 goto loop; 2382 } 2383 /* 2384 * Skip over a vnodes marked VV_SYSTEM. 2385 */ 2386 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2387 VOP_UNLOCK(vp, 0); 2388 vdrop(vp); 2389 MNT_ILOCK(mp); 2390 continue; 2391 } 2392 /* 2393 * If WRITECLOSE is set, flush out unlinked but still open 2394 * files (even if open only for reading) and regular file 2395 * vnodes open for writing. 2396 */ 2397 if (flags & WRITECLOSE) { 2398 error = VOP_GETATTR(vp, &vattr, td->td_ucred); 2399 VI_LOCK(vp); 2400 2401 if ((vp->v_type == VNON || 2402 (error == 0 && vattr.va_nlink > 0)) && 2403 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2404 VOP_UNLOCK(vp, 0); 2405 vdropl(vp); 2406 MNT_ILOCK(mp); 2407 continue; 2408 } 2409 } else 2410 VI_LOCK(vp); 2411 /* 2412 * With v_usecount == 0, all we need to do is clear out the 2413 * vnode data structures and we are done. 2414 * 2415 * If FORCECLOSE is set, forcibly close the vnode. 2416 */ 2417 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { 2418 VNASSERT(vp->v_usecount == 0 || 2419 (vp->v_type != VCHR && vp->v_type != VBLK), vp, 2420 ("device VNODE %p is FORCECLOSED", vp)); 2421 vgonel(vp); 2422 } else { 2423 busy++; 2424 #ifdef DIAGNOSTIC 2425 if (busyprt) 2426 vprint("vflush: busy vnode", vp); 2427 #endif 2428 } 2429 VOP_UNLOCK(vp, 0); 2430 vdropl(vp); 2431 MNT_ILOCK(mp); 2432 } 2433 MNT_IUNLOCK(mp); 2434 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2435 /* 2436 * If just the root vnode is busy, and if its refcount 2437 * is equal to `rootrefs', then go ahead and kill it. 2438 */ 2439 VI_LOCK(rootvp); 2440 KASSERT(busy > 0, ("vflush: not busy")); 2441 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, 2442 ("vflush: usecount %d < rootrefs %d", 2443 rootvp->v_usecount, rootrefs)); 2444 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2445 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); 2446 vgone(rootvp); 2447 VOP_UNLOCK(rootvp, 0); 2448 busy = 0; 2449 } else 2450 VI_UNLOCK(rootvp); 2451 } 2452 if (busy) 2453 return (EBUSY); 2454 for (; rootrefs > 0; rootrefs--) 2455 vrele(rootvp); 2456 return (0); 2457 } 2458 2459 /* 2460 * Recycle an unused vnode to the front of the free list. 2461 */ 2462 int 2463 vrecycle(struct vnode *vp, struct thread *td) 2464 { 2465 int recycled; 2466 2467 ASSERT_VOP_ELOCKED(vp, "vrecycle"); 2468 recycled = 0; 2469 VI_LOCK(vp); 2470 if (vp->v_usecount == 0) { 2471 recycled = 1; 2472 vgonel(vp); 2473 } 2474 VI_UNLOCK(vp); 2475 return (recycled); 2476 } 2477 2478 /* 2479 * Eliminate all activity associated with a vnode 2480 * in preparation for reuse. 2481 */ 2482 void 2483 vgone(struct vnode *vp) 2484 { 2485 VI_LOCK(vp); 2486 vgonel(vp); 2487 VI_UNLOCK(vp); 2488 } 2489 2490 /* 2491 * vgone, with the vp interlock held. 2492 */ 2493 void 2494 vgonel(struct vnode *vp) 2495 { 2496 struct thread *td; 2497 int oweinact; 2498 int active; 2499 struct mount *mp; 2500 2501 CTR1(KTR_VFS, "vgonel: vp %p", vp); 2502 ASSERT_VOP_ELOCKED(vp, "vgonel"); 2503 ASSERT_VI_LOCKED(vp, "vgonel"); 2504 VNASSERT(vp->v_holdcnt, vp, 2505 ("vgonel: vp %p has no reference.", vp)); 2506 td = curthread; 2507 2508 /* 2509 * Don't vgonel if we're already doomed. 2510 */ 2511 if (vp->v_iflag & VI_DOOMED) 2512 return; 2513 vp->v_iflag |= VI_DOOMED; 2514 /* 2515 * Check to see if the vnode is in use. If so, we have to call 2516 * VOP_CLOSE() and VOP_INACTIVE(). 2517 */ 2518 active = vp->v_usecount; 2519 oweinact = (vp->v_iflag & VI_OWEINACT); 2520 VI_UNLOCK(vp); 2521 /* 2522 * Clean out any buffers associated with the vnode. 2523 * If the flush fails, just toss the buffers. 2524 */ 2525 mp = NULL; 2526 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) 2527 (void) vn_start_secondary_write(vp, &mp, V_WAIT); 2528 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) 2529 vinvalbuf(vp, 0, 0, 0); 2530 2531 /* 2532 * If purging an active vnode, it must be closed and 2533 * deactivated before being reclaimed. 2534 */ 2535 if (active) 2536 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2537 if (oweinact || active) { 2538 VI_LOCK(vp); 2539 if ((vp->v_iflag & VI_DOINGINACT) == 0) 2540 vinactive(vp, td); 2541 VI_UNLOCK(vp); 2542 } 2543 /* 2544 * Reclaim the vnode. 2545 */ 2546 if (VOP_RECLAIM(vp, td)) 2547 panic("vgone: cannot reclaim"); 2548 if (mp != NULL) 2549 vn_finished_secondary_write(mp); 2550 VNASSERT(vp->v_object == NULL, vp, 2551 ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag)); 2552 /* 2553 * Clear the advisory locks and wake up waiting threads. 2554 */ 2555 lf_purgelocks(vp, &(vp->v_lockf)); 2556 /* 2557 * Delete from old mount point vnode list. 2558 */ 2559 delmntque(vp); 2560 cache_purge(vp); 2561 /* 2562 * Done with purge, reset to the standard lock and invalidate 2563 * the vnode. 2564 */ 2565 VI_LOCK(vp); 2566 vp->v_vnlock = &vp->v_lock; 2567 vp->v_op = &dead_vnodeops; 2568 vp->v_tag = "none"; 2569 vp->v_type = VBAD; 2570 } 2571 2572 /* 2573 * Calculate the total number of references to a special device. 2574 */ 2575 int 2576 vcount(struct vnode *vp) 2577 { 2578 int count; 2579 2580 dev_lock(); 2581 count = vp->v_rdev->si_usecount; 2582 dev_unlock(); 2583 return (count); 2584 } 2585 2586 /* 2587 * Same as above, but using the struct cdev *as argument 2588 */ 2589 int 2590 count_dev(struct cdev *dev) 2591 { 2592 int count; 2593 2594 dev_lock(); 2595 count = dev->si_usecount; 2596 dev_unlock(); 2597 return(count); 2598 } 2599 2600 /* 2601 * Print out a description of a vnode. 2602 */ 2603 static char *typename[] = 2604 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", 2605 "VMARKER"}; 2606 2607 void 2608 vn_printf(struct vnode *vp, const char *fmt, ...) 2609 { 2610 va_list ap; 2611 char buf[256], buf2[16]; 2612 u_long flags; 2613 2614 va_start(ap, fmt); 2615 vprintf(fmt, ap); 2616 va_end(ap); 2617 printf("%p: ", (void *)vp); 2618 printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]); 2619 printf(" usecount %d, writecount %d, refcount %d mountedhere %p\n", 2620 vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere); 2621 buf[0] = '\0'; 2622 buf[1] = '\0'; 2623 if (vp->v_vflag & VV_ROOT) 2624 strlcat(buf, "|VV_ROOT", sizeof(buf)); 2625 if (vp->v_vflag & VV_ISTTY) 2626 strlcat(buf, "|VV_ISTTY", sizeof(buf)); 2627 if (vp->v_vflag & VV_NOSYNC) 2628 strlcat(buf, "|VV_NOSYNC", sizeof(buf)); 2629 if (vp->v_vflag & VV_CACHEDLABEL) 2630 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); 2631 if (vp->v_vflag & VV_TEXT) 2632 strlcat(buf, "|VV_TEXT", sizeof(buf)); 2633 if (vp->v_vflag & VV_COPYONWRITE) 2634 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); 2635 if (vp->v_vflag & VV_SYSTEM) 2636 strlcat(buf, "|VV_SYSTEM", sizeof(buf)); 2637 if (vp->v_vflag & VV_PROCDEP) 2638 strlcat(buf, "|VV_PROCDEP", sizeof(buf)); 2639 if (vp->v_vflag & VV_NOKNOTE) 2640 strlcat(buf, "|VV_NOKNOTE", sizeof(buf)); 2641 if (vp->v_vflag & VV_DELETED) 2642 strlcat(buf, "|VV_DELETED", sizeof(buf)); 2643 if (vp->v_vflag & VV_MD) 2644 strlcat(buf, "|VV_MD", sizeof(buf)); 2645 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | 2646 VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | 2647 VV_NOKNOTE | VV_DELETED | VV_MD); 2648 if (flags != 0) { 2649 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); 2650 strlcat(buf, buf2, sizeof(buf)); 2651 } 2652 if (vp->v_iflag & VI_MOUNT) 2653 strlcat(buf, "|VI_MOUNT", sizeof(buf)); 2654 if (vp->v_iflag & VI_AGE) 2655 strlcat(buf, "|VI_AGE", sizeof(buf)); 2656 if (vp->v_iflag & VI_DOOMED) 2657 strlcat(buf, "|VI_DOOMED", sizeof(buf)); 2658 if (vp->v_iflag & VI_FREE) 2659 strlcat(buf, "|VI_FREE", sizeof(buf)); 2660 if (vp->v_iflag & VI_OBJDIRTY) 2661 strlcat(buf, "|VI_OBJDIRTY", sizeof(buf)); 2662 if (vp->v_iflag & VI_DOINGINACT) 2663 strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); 2664 if (vp->v_iflag & VI_OWEINACT) 2665 strlcat(buf, "|VI_OWEINACT", sizeof(buf)); 2666 flags = vp->v_iflag & ~(VI_MOUNT | VI_AGE | VI_DOOMED | VI_FREE | 2667 VI_OBJDIRTY | VI_DOINGINACT | VI_OWEINACT); 2668 if (flags != 0) { 2669 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); 2670 strlcat(buf, buf2, sizeof(buf)); 2671 } 2672 printf(" flags (%s)\n", buf + 1); 2673 if (mtx_owned(VI_MTX(vp))) 2674 printf(" VI_LOCKed"); 2675 if (vp->v_object != NULL) 2676 printf(" v_object %p ref %d pages %d\n", 2677 vp->v_object, vp->v_object->ref_count, 2678 vp->v_object->resident_page_count); 2679 printf(" "); 2680 lockmgr_printinfo(vp->v_vnlock); 2681 printf("\n"); 2682 if (vp->v_data != NULL) 2683 VOP_PRINT(vp); 2684 } 2685 2686 #ifdef DDB 2687 /* 2688 * List all of the locked vnodes in the system. 2689 * Called when debugging the kernel. 2690 */ 2691 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 2692 { 2693 struct mount *mp, *nmp; 2694 struct vnode *vp; 2695 2696 /* 2697 * Note: because this is DDB, we can't obey the locking semantics 2698 * for these structures, which means we could catch an inconsistent 2699 * state and dereference a nasty pointer. Not much to be done 2700 * about that. 2701 */ 2702 db_printf("Locked vnodes\n"); 2703 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2704 nmp = TAILQ_NEXT(mp, mnt_list); 2705 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2706 if (vp->v_type != VMARKER && 2707 VOP_ISLOCKED(vp)) 2708 vprint("", vp); 2709 } 2710 nmp = TAILQ_NEXT(mp, mnt_list); 2711 } 2712 } 2713 2714 /* 2715 * Show details about the given vnode. 2716 */ 2717 DB_SHOW_COMMAND(vnode, db_show_vnode) 2718 { 2719 struct vnode *vp; 2720 2721 if (!have_addr) 2722 return; 2723 vp = (struct vnode *)addr; 2724 vn_printf(vp, "vnode "); 2725 } 2726 2727 /* 2728 * Show details about the given mount point. 2729 */ 2730 DB_SHOW_COMMAND(mount, db_show_mount) 2731 { 2732 struct mount *mp; 2733 struct statfs *sp; 2734 struct vnode *vp; 2735 char buf[512]; 2736 u_int flags; 2737 2738 if (!have_addr) { 2739 /* No address given, print short info about all mount points. */ 2740 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 2741 db_printf("%p %s on %s (%s)\n", mp, 2742 mp->mnt_stat.f_mntfromname, 2743 mp->mnt_stat.f_mntonname, 2744 mp->mnt_stat.f_fstypename); 2745 if (db_pager_quit) 2746 break; 2747 } 2748 db_printf("\nMore info: show mount <addr>\n"); 2749 return; 2750 } 2751 2752 mp = (struct mount *)addr; 2753 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, 2754 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); 2755 2756 buf[0] = '\0'; 2757 flags = mp->mnt_flag; 2758 #define MNT_FLAG(flag) do { \ 2759 if (flags & (flag)) { \ 2760 if (buf[0] != '\0') \ 2761 strlcat(buf, ", ", sizeof(buf)); \ 2762 strlcat(buf, (#flag) + 4, sizeof(buf)); \ 2763 flags &= ~(flag); \ 2764 } \ 2765 } while (0) 2766 MNT_FLAG(MNT_RDONLY); 2767 MNT_FLAG(MNT_SYNCHRONOUS); 2768 MNT_FLAG(MNT_NOEXEC); 2769 MNT_FLAG(MNT_NOSUID); 2770 MNT_FLAG(MNT_UNION); 2771 MNT_FLAG(MNT_ASYNC); 2772 MNT_FLAG(MNT_SUIDDIR); 2773 MNT_FLAG(MNT_SOFTDEP); 2774 MNT_FLAG(MNT_NOSYMFOLLOW); 2775 MNT_FLAG(MNT_GJOURNAL); 2776 MNT_FLAG(MNT_MULTILABEL); 2777 MNT_FLAG(MNT_ACLS); 2778 MNT_FLAG(MNT_NOATIME); 2779 MNT_FLAG(MNT_NOCLUSTERR); 2780 MNT_FLAG(MNT_NOCLUSTERW); 2781 MNT_FLAG(MNT_EXRDONLY); 2782 MNT_FLAG(MNT_EXPORTED); 2783 MNT_FLAG(MNT_DEFEXPORTED); 2784 MNT_FLAG(MNT_EXPORTANON); 2785 MNT_FLAG(MNT_EXKERB); 2786 MNT_FLAG(MNT_EXPUBLIC); 2787 MNT_FLAG(MNT_LOCAL); 2788 MNT_FLAG(MNT_QUOTA); 2789 MNT_FLAG(MNT_ROOTFS); 2790 MNT_FLAG(MNT_USER); 2791 MNT_FLAG(MNT_IGNORE); 2792 MNT_FLAG(MNT_UPDATE); 2793 MNT_FLAG(MNT_DELEXPORT); 2794 MNT_FLAG(MNT_RELOAD); 2795 MNT_FLAG(MNT_FORCE); 2796 MNT_FLAG(MNT_SNAPSHOT); 2797 MNT_FLAG(MNT_BYFSID); 2798 #undef MNT_FLAG 2799 if (flags != 0) { 2800 if (buf[0] != '\0') 2801 strlcat(buf, ", ", sizeof(buf)); 2802 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 2803 "0x%08x", flags); 2804 } 2805 db_printf(" mnt_flag = %s\n", buf); 2806 2807 buf[0] = '\0'; 2808 flags = mp->mnt_kern_flag; 2809 #define MNT_KERN_FLAG(flag) do { \ 2810 if (flags & (flag)) { \ 2811 if (buf[0] != '\0') \ 2812 strlcat(buf, ", ", sizeof(buf)); \ 2813 strlcat(buf, (#flag) + 5, sizeof(buf)); \ 2814 flags &= ~(flag); \ 2815 } \ 2816 } while (0) 2817 MNT_KERN_FLAG(MNTK_UNMOUNTF); 2818 MNT_KERN_FLAG(MNTK_ASYNC); 2819 MNT_KERN_FLAG(MNTK_SOFTDEP); 2820 MNT_KERN_FLAG(MNTK_NOINSMNTQ); 2821 MNT_KERN_FLAG(MNTK_UNMOUNT); 2822 MNT_KERN_FLAG(MNTK_MWAIT); 2823 MNT_KERN_FLAG(MNTK_SUSPEND); 2824 MNT_KERN_FLAG(MNTK_SUSPEND2); 2825 MNT_KERN_FLAG(MNTK_SUSPENDED); 2826 MNT_KERN_FLAG(MNTK_MPSAFE); 2827 MNT_KERN_FLAG(MNTK_NOKNOTE); 2828 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); 2829 #undef MNT_KERN_FLAG 2830 if (flags != 0) { 2831 if (buf[0] != '\0') 2832 strlcat(buf, ", ", sizeof(buf)); 2833 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), 2834 "0x%08x", flags); 2835 } 2836 db_printf(" mnt_kern_flag = %s\n", buf); 2837 2838 sp = &mp->mnt_stat; 2839 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " 2840 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " 2841 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " 2842 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", 2843 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, 2844 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, 2845 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, 2846 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, 2847 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, 2848 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, 2849 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, 2850 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); 2851 2852 db_printf(" mnt_cred = { uid=%u ruid=%u", 2853 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); 2854 if (mp->mnt_cred->cr_prison != NULL) 2855 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); 2856 db_printf(" }\n"); 2857 db_printf(" mnt_ref = %d\n", mp->mnt_ref); 2858 db_printf(" mnt_gen = %d\n", mp->mnt_gen); 2859 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); 2860 db_printf(" mnt_writeopcount = %d\n", mp->mnt_writeopcount); 2861 db_printf(" mnt_noasync = %u\n", mp->mnt_noasync); 2862 db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen); 2863 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); 2864 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); 2865 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); 2866 db_printf(" mnt_secondary_accwrites = %d\n", 2867 mp->mnt_secondary_accwrites); 2868 db_printf(" mnt_gjprovider = %s\n", 2869 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); 2870 db_printf("\n"); 2871 2872 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2873 if (vp->v_type != VMARKER) { 2874 vn_printf(vp, "vnode "); 2875 if (db_pager_quit) 2876 break; 2877 } 2878 } 2879 } 2880 #endif /* DDB */ 2881 2882 /* 2883 * Fill in a struct xvfsconf based on a struct vfsconf. 2884 */ 2885 static void 2886 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) 2887 { 2888 2889 strcpy(xvfsp->vfc_name, vfsp->vfc_name); 2890 xvfsp->vfc_typenum = vfsp->vfc_typenum; 2891 xvfsp->vfc_refcount = vfsp->vfc_refcount; 2892 xvfsp->vfc_flags = vfsp->vfc_flags; 2893 /* 2894 * These are unused in userland, we keep them 2895 * to not break binary compatibility. 2896 */ 2897 xvfsp->vfc_vfsops = NULL; 2898 xvfsp->vfc_next = NULL; 2899 } 2900 2901 /* 2902 * Top level filesystem related information gathering. 2903 */ 2904 static int 2905 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 2906 { 2907 struct vfsconf *vfsp; 2908 struct xvfsconf xvfsp; 2909 int error; 2910 2911 error = 0; 2912 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2913 bzero(&xvfsp, sizeof(xvfsp)); 2914 vfsconf2x(vfsp, &xvfsp); 2915 error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp); 2916 if (error) 2917 break; 2918 } 2919 return (error); 2920 } 2921 2922 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, 2923 "S,xvfsconf", "List of all configured filesystems"); 2924 2925 #ifndef BURN_BRIDGES 2926 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2927 2928 static int 2929 vfs_sysctl(SYSCTL_HANDLER_ARGS) 2930 { 2931 int *name = (int *)arg1 - 1; /* XXX */ 2932 u_int namelen = arg2 + 1; /* XXX */ 2933 struct vfsconf *vfsp; 2934 struct xvfsconf xvfsp; 2935 2936 printf("WARNING: userland calling deprecated sysctl, " 2937 "please rebuild world\n"); 2938 2939 #if 1 || defined(COMPAT_PRELITE2) 2940 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2941 if (namelen == 1) 2942 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2943 #endif 2944 2945 switch (name[1]) { 2946 case VFS_MAXTYPENUM: 2947 if (namelen != 2) 2948 return (ENOTDIR); 2949 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2950 case VFS_CONF: 2951 if (namelen != 3) 2952 return (ENOTDIR); /* overloaded */ 2953 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) 2954 if (vfsp->vfc_typenum == name[2]) 2955 break; 2956 if (vfsp == NULL) 2957 return (EOPNOTSUPP); 2958 bzero(&xvfsp, sizeof(xvfsp)); 2959 vfsconf2x(vfsp, &xvfsp); 2960 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 2961 } 2962 return (EOPNOTSUPP); 2963 } 2964 2965 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, 2966 vfs_sysctl, "Generic filesystem"); 2967 2968 #if 1 || defined(COMPAT_PRELITE2) 2969 2970 static int 2971 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 2972 { 2973 int error; 2974 struct vfsconf *vfsp; 2975 struct ovfsconf ovfs; 2976 2977 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { 2978 bzero(&ovfs, sizeof(ovfs)); 2979 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 2980 strcpy(ovfs.vfc_name, vfsp->vfc_name); 2981 ovfs.vfc_index = vfsp->vfc_typenum; 2982 ovfs.vfc_refcount = vfsp->vfc_refcount; 2983 ovfs.vfc_flags = vfsp->vfc_flags; 2984 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 2985 if (error) 2986 return error; 2987 } 2988 return 0; 2989 } 2990 2991 #endif /* 1 || COMPAT_PRELITE2 */ 2992 #endif /* !BURN_BRIDGES */ 2993 2994 #define KINFO_VNODESLOP 10 2995 #ifdef notyet 2996 /* 2997 * Dump vnode list (via sysctl). 2998 */ 2999 /* ARGSUSED */ 3000 static int 3001 sysctl_vnode(SYSCTL_HANDLER_ARGS) 3002 { 3003 struct xvnode *xvn; 3004 struct mount *mp; 3005 struct vnode *vp; 3006 int error, len, n; 3007 3008 /* 3009 * Stale numvnodes access is not fatal here. 3010 */ 3011 req->lock = 0; 3012 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 3013 if (!req->oldptr) 3014 /* Make an estimate */ 3015 return (SYSCTL_OUT(req, 0, len)); 3016 3017 error = sysctl_wire_old_buffer(req, 0); 3018 if (error != 0) 3019 return (error); 3020 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 3021 n = 0; 3022 mtx_lock(&mountlist_mtx); 3023 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 3024 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) 3025 continue; 3026 MNT_ILOCK(mp); 3027 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 3028 if (n == len) 3029 break; 3030 vref(vp); 3031 xvn[n].xv_size = sizeof *xvn; 3032 xvn[n].xv_vnode = vp; 3033 xvn[n].xv_id = 0; /* XXX compat */ 3034 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 3035 XV_COPY(usecount); 3036 XV_COPY(writecount); 3037 XV_COPY(holdcnt); 3038 XV_COPY(mount); 3039 XV_COPY(numoutput); 3040 XV_COPY(type); 3041 #undef XV_COPY 3042 xvn[n].xv_flag = vp->v_vflag; 3043 3044 switch (vp->v_type) { 3045 case VREG: 3046 case VDIR: 3047 case VLNK: 3048 break; 3049 case VBLK: 3050 case VCHR: 3051 if (vp->v_rdev == NULL) { 3052 vrele(vp); 3053 continue; 3054 } 3055 xvn[n].xv_dev = dev2udev(vp->v_rdev); 3056 break; 3057 case VSOCK: 3058 xvn[n].xv_socket = vp->v_socket; 3059 break; 3060 case VFIFO: 3061 xvn[n].xv_fifo = vp->v_fifoinfo; 3062 break; 3063 case VNON: 3064 case VBAD: 3065 default: 3066 /* shouldn't happen? */ 3067 vrele(vp); 3068 continue; 3069 } 3070 vrele(vp); 3071 ++n; 3072 } 3073 MNT_IUNLOCK(mp); 3074 mtx_lock(&mountlist_mtx); 3075 vfs_unbusy(mp); 3076 if (n == len) 3077 break; 3078 } 3079 mtx_unlock(&mountlist_mtx); 3080 3081 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 3082 free(xvn, M_TEMP); 3083 return (error); 3084 } 3085 3086 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 3087 0, 0, sysctl_vnode, "S,xvnode", ""); 3088 #endif 3089 3090 /* 3091 * Unmount all filesystems. The list is traversed in reverse order 3092 * of mounting to avoid dependencies. 3093 */ 3094 void 3095 vfs_unmountall(void) 3096 { 3097 struct mount *mp; 3098 struct thread *td; 3099 int error; 3100 3101 KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread")); 3102 td = curthread; 3103 /* 3104 * Since this only runs when rebooting, it is not interlocked. 3105 */ 3106 while(!TAILQ_EMPTY(&mountlist)) { 3107 mp = TAILQ_LAST(&mountlist, mntlist); 3108 error = dounmount(mp, MNT_FORCE, td); 3109 if (error) { 3110 TAILQ_REMOVE(&mountlist, mp, mnt_list); 3111 /* 3112 * XXX: Due to the way in which we mount the root 3113 * file system off of devfs, devfs will generate a 3114 * "busy" warning when we try to unmount it before 3115 * the root. Don't print a warning as a result in 3116 * order to avoid false positive errors that may 3117 * cause needless upset. 3118 */ 3119 if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) { 3120 printf("unmount of %s failed (", 3121 mp->mnt_stat.f_mntonname); 3122 if (error == EBUSY) 3123 printf("BUSY)\n"); 3124 else 3125 printf("%d)\n", error); 3126 } 3127 } else { 3128 /* The unmount has removed mp from the mountlist */ 3129 } 3130 } 3131 } 3132 3133 /* 3134 * perform msync on all vnodes under a mount point 3135 * the mount point must be locked. 3136 */ 3137 void 3138 vfs_msync(struct mount *mp, int flags) 3139 { 3140 struct vnode *vp, *mvp; 3141 struct vm_object *obj; 3142 3143 MNT_ILOCK(mp); 3144 MNT_VNODE_FOREACH(vp, mp, mvp) { 3145 VI_LOCK(vp); 3146 if ((vp->v_iflag & VI_OBJDIRTY) && 3147 (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) { 3148 MNT_IUNLOCK(mp); 3149 if (!vget(vp, 3150 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 3151 curthread)) { 3152 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 3153 vput(vp); 3154 MNT_ILOCK(mp); 3155 continue; 3156 } 3157 3158 obj = vp->v_object; 3159 if (obj != NULL) { 3160 VM_OBJECT_LOCK(obj); 3161 vm_object_page_clean(obj, 0, 0, 3162 flags == MNT_WAIT ? 3163 OBJPC_SYNC : OBJPC_NOSYNC); 3164 VM_OBJECT_UNLOCK(obj); 3165 } 3166 vput(vp); 3167 } 3168 MNT_ILOCK(mp); 3169 } else 3170 VI_UNLOCK(vp); 3171 } 3172 MNT_IUNLOCK(mp); 3173 } 3174 3175 /* 3176 * Mark a vnode as free, putting it up for recycling. 3177 */ 3178 static void 3179 vfree(struct vnode *vp) 3180 { 3181 3182 CTR1(KTR_VFS, "vfree vp %p", vp); 3183 ASSERT_VI_LOCKED(vp, "vfree"); 3184 mtx_lock(&vnode_free_list_mtx); 3185 VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed.")); 3186 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free")); 3187 VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't")); 3188 VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp, 3189 ("vfree: Freeing doomed vnode")); 3190 if (vp->v_iflag & VI_AGE) { 3191 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 3192 } else { 3193 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 3194 } 3195 freevnodes++; 3196 vp->v_iflag &= ~VI_AGE; 3197 vp->v_iflag |= VI_FREE; 3198 mtx_unlock(&vnode_free_list_mtx); 3199 } 3200 3201 /* 3202 * Opposite of vfree() - mark a vnode as in use. 3203 */ 3204 static void 3205 vbusy(struct vnode *vp) 3206 { 3207 CTR1(KTR_VFS, "vbusy vp %p", vp); 3208 ASSERT_VI_LOCKED(vp, "vbusy"); 3209 VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free")); 3210 VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed.")); 3211 3212 mtx_lock(&vnode_free_list_mtx); 3213 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 3214 freevnodes--; 3215 vp->v_iflag &= ~(VI_FREE|VI_AGE); 3216 mtx_unlock(&vnode_free_list_mtx); 3217 } 3218 3219 static void 3220 destroy_vpollinfo(struct vpollinfo *vi) 3221 { 3222 knlist_destroy(&vi->vpi_selinfo.si_note); 3223 mtx_destroy(&vi->vpi_lock); 3224 uma_zfree(vnodepoll_zone, vi); 3225 } 3226 3227 /* 3228 * Initalize per-vnode helper structure to hold poll-related state. 3229 */ 3230 void 3231 v_addpollinfo(struct vnode *vp) 3232 { 3233 struct vpollinfo *vi; 3234 3235 if (vp->v_pollinfo != NULL) 3236 return; 3237 vi = uma_zalloc(vnodepoll_zone, M_WAITOK); 3238 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 3239 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, 3240 vfs_knlunlock, vfs_knllocked); 3241 VI_LOCK(vp); 3242 if (vp->v_pollinfo != NULL) { 3243 VI_UNLOCK(vp); 3244 destroy_vpollinfo(vi); 3245 return; 3246 } 3247 vp->v_pollinfo = vi; 3248 VI_UNLOCK(vp); 3249 } 3250 3251 /* 3252 * Record a process's interest in events which might happen to 3253 * a vnode. Because poll uses the historic select-style interface 3254 * internally, this routine serves as both the ``check for any 3255 * pending events'' and the ``record my interest in future events'' 3256 * functions. (These are done together, while the lock is held, 3257 * to avoid race conditions.) 3258 */ 3259 int 3260 vn_pollrecord(struct vnode *vp, struct thread *td, int events) 3261 { 3262 3263 v_addpollinfo(vp); 3264 mtx_lock(&vp->v_pollinfo->vpi_lock); 3265 if (vp->v_pollinfo->vpi_revents & events) { 3266 /* 3267 * This leaves events we are not interested 3268 * in available for the other process which 3269 * which presumably had requested them 3270 * (otherwise they would never have been 3271 * recorded). 3272 */ 3273 events &= vp->v_pollinfo->vpi_revents; 3274 vp->v_pollinfo->vpi_revents &= ~events; 3275 3276 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3277 return (events); 3278 } 3279 vp->v_pollinfo->vpi_events |= events; 3280 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 3281 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3282 return (0); 3283 } 3284 3285 /* 3286 * Routine to create and manage a filesystem syncer vnode. 3287 */ 3288 #define sync_close ((int (*)(struct vop_close_args *))nullop) 3289 static int sync_fsync(struct vop_fsync_args *); 3290 static int sync_inactive(struct vop_inactive_args *); 3291 static int sync_reclaim(struct vop_reclaim_args *); 3292 3293 static struct vop_vector sync_vnodeops = { 3294 .vop_bypass = VOP_EOPNOTSUPP, 3295 .vop_close = sync_close, /* close */ 3296 .vop_fsync = sync_fsync, /* fsync */ 3297 .vop_inactive = sync_inactive, /* inactive */ 3298 .vop_reclaim = sync_reclaim, /* reclaim */ 3299 .vop_lock1 = vop_stdlock, /* lock */ 3300 .vop_unlock = vop_stdunlock, /* unlock */ 3301 .vop_islocked = vop_stdislocked, /* islocked */ 3302 }; 3303 3304 /* 3305 * Create a new filesystem syncer vnode for the specified mount point. 3306 */ 3307 int 3308 vfs_allocate_syncvnode(struct mount *mp) 3309 { 3310 struct vnode *vp; 3311 struct bufobj *bo; 3312 static long start, incr, next; 3313 int error; 3314 3315 /* Allocate a new vnode */ 3316 if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) { 3317 mp->mnt_syncer = NULL; 3318 return (error); 3319 } 3320 vp->v_type = VNON; 3321 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3322 vp->v_vflag |= VV_FORCEINSMQ; 3323 error = insmntque(vp, mp); 3324 if (error != 0) 3325 panic("vfs_allocate_syncvnode: insmntque failed"); 3326 vp->v_vflag &= ~VV_FORCEINSMQ; 3327 VOP_UNLOCK(vp, 0); 3328 /* 3329 * Place the vnode onto the syncer worklist. We attempt to 3330 * scatter them about on the list so that they will go off 3331 * at evenly distributed times even if all the filesystems 3332 * are mounted at once. 3333 */ 3334 next += incr; 3335 if (next == 0 || next > syncer_maxdelay) { 3336 start /= 2; 3337 incr /= 2; 3338 if (start == 0) { 3339 start = syncer_maxdelay / 2; 3340 incr = syncer_maxdelay; 3341 } 3342 next = start; 3343 } 3344 bo = &vp->v_bufobj; 3345 BO_LOCK(bo); 3346 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); 3347 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 3348 mtx_lock(&sync_mtx); 3349 sync_vnode_count++; 3350 mtx_unlock(&sync_mtx); 3351 BO_UNLOCK(bo); 3352 mp->mnt_syncer = vp; 3353 return (0); 3354 } 3355 3356 /* 3357 * Do a lazy sync of the filesystem. 3358 */ 3359 static int 3360 sync_fsync(struct vop_fsync_args *ap) 3361 { 3362 struct vnode *syncvp = ap->a_vp; 3363 struct mount *mp = syncvp->v_mount; 3364 int error; 3365 struct bufobj *bo; 3366 3367 /* 3368 * We only need to do something if this is a lazy evaluation. 3369 */ 3370 if (ap->a_waitfor != MNT_LAZY) 3371 return (0); 3372 3373 /* 3374 * Move ourselves to the back of the sync list. 3375 */ 3376 bo = &syncvp->v_bufobj; 3377 BO_LOCK(bo); 3378 vn_syncer_add_to_worklist(bo, syncdelay); 3379 BO_UNLOCK(bo); 3380 3381 /* 3382 * Walk the list of vnodes pushing all that are dirty and 3383 * not already on the sync list. 3384 */ 3385 mtx_lock(&mountlist_mtx); 3386 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) != 0) { 3387 mtx_unlock(&mountlist_mtx); 3388 return (0); 3389 } 3390 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 3391 vfs_unbusy(mp); 3392 return (0); 3393 } 3394 MNT_ILOCK(mp); 3395 mp->mnt_noasync++; 3396 mp->mnt_kern_flag &= ~MNTK_ASYNC; 3397 MNT_IUNLOCK(mp); 3398 vfs_msync(mp, MNT_NOWAIT); 3399 error = VFS_SYNC(mp, MNT_LAZY, ap->a_td); 3400 MNT_ILOCK(mp); 3401 mp->mnt_noasync--; 3402 if ((mp->mnt_flag & MNT_ASYNC) != 0 && mp->mnt_noasync == 0) 3403 mp->mnt_kern_flag |= MNTK_ASYNC; 3404 MNT_IUNLOCK(mp); 3405 vn_finished_write(mp); 3406 vfs_unbusy(mp); 3407 return (error); 3408 } 3409 3410 /* 3411 * The syncer vnode is no referenced. 3412 */ 3413 static int 3414 sync_inactive(struct vop_inactive_args *ap) 3415 { 3416 3417 vgone(ap->a_vp); 3418 return (0); 3419 } 3420 3421 /* 3422 * The syncer vnode is no longer needed and is being decommissioned. 3423 * 3424 * Modifications to the worklist must be protected by sync_mtx. 3425 */ 3426 static int 3427 sync_reclaim(struct vop_reclaim_args *ap) 3428 { 3429 struct vnode *vp = ap->a_vp; 3430 struct bufobj *bo; 3431 3432 bo = &vp->v_bufobj; 3433 BO_LOCK(bo); 3434 vp->v_mount->mnt_syncer = NULL; 3435 if (bo->bo_flag & BO_ONWORKLST) { 3436 mtx_lock(&sync_mtx); 3437 LIST_REMOVE(bo, bo_synclist); 3438 syncer_worklist_len--; 3439 sync_vnode_count--; 3440 mtx_unlock(&sync_mtx); 3441 bo->bo_flag &= ~BO_ONWORKLST; 3442 } 3443 BO_UNLOCK(bo); 3444 3445 return (0); 3446 } 3447 3448 /* 3449 * Check if vnode represents a disk device 3450 */ 3451 int 3452 vn_isdisk(struct vnode *vp, int *errp) 3453 { 3454 int error; 3455 3456 error = 0; 3457 dev_lock(); 3458 if (vp->v_type != VCHR) 3459 error = ENOTBLK; 3460 else if (vp->v_rdev == NULL) 3461 error = ENXIO; 3462 else if (vp->v_rdev->si_devsw == NULL) 3463 error = ENXIO; 3464 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) 3465 error = ENOTBLK; 3466 dev_unlock(); 3467 if (errp != NULL) 3468 *errp = error; 3469 return (error == 0); 3470 } 3471 3472 /* 3473 * Common filesystem object access control check routine. Accepts a 3474 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3475 * and optional call-by-reference privused argument allowing vaccess() 3476 * to indicate to the caller whether privilege was used to satisfy the 3477 * request (obsoleted). Returns 0 on success, or an errno on failure. 3478 * 3479 * The ifdef'd CAPABILITIES version is here for reference, but is not 3480 * actually used. 3481 */ 3482 int 3483 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, 3484 accmode_t accmode, struct ucred *cred, int *privused) 3485 { 3486 accmode_t dac_granted; 3487 accmode_t priv_granted; 3488 3489 /* 3490 * Look for a normal, non-privileged way to access the file/directory 3491 * as requested. If it exists, go with that. 3492 */ 3493 3494 if (privused != NULL) 3495 *privused = 0; 3496 3497 dac_granted = 0; 3498 3499 /* Check the owner. */ 3500 if (cred->cr_uid == file_uid) { 3501 dac_granted |= VADMIN; 3502 if (file_mode & S_IXUSR) 3503 dac_granted |= VEXEC; 3504 if (file_mode & S_IRUSR) 3505 dac_granted |= VREAD; 3506 if (file_mode & S_IWUSR) 3507 dac_granted |= (VWRITE | VAPPEND); 3508 3509 if ((accmode & dac_granted) == accmode) 3510 return (0); 3511 3512 goto privcheck; 3513 } 3514 3515 /* Otherwise, check the groups (first match) */ 3516 if (groupmember(file_gid, cred)) { 3517 if (file_mode & S_IXGRP) 3518 dac_granted |= VEXEC; 3519 if (file_mode & S_IRGRP) 3520 dac_granted |= VREAD; 3521 if (file_mode & S_IWGRP) 3522 dac_granted |= (VWRITE | VAPPEND); 3523 3524 if ((accmode & dac_granted) == accmode) 3525 return (0); 3526 3527 goto privcheck; 3528 } 3529 3530 /* Otherwise, check everyone else. */ 3531 if (file_mode & S_IXOTH) 3532 dac_granted |= VEXEC; 3533 if (file_mode & S_IROTH) 3534 dac_granted |= VREAD; 3535 if (file_mode & S_IWOTH) 3536 dac_granted |= (VWRITE | VAPPEND); 3537 if ((accmode & dac_granted) == accmode) 3538 return (0); 3539 3540 privcheck: 3541 /* 3542 * Build a privilege mask to determine if the set of privileges 3543 * satisfies the requirements when combined with the granted mask 3544 * from above. For each privilege, if the privilege is required, 3545 * bitwise or the request type onto the priv_granted mask. 3546 */ 3547 priv_granted = 0; 3548 3549 if (type == VDIR) { 3550 /* 3551 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC 3552 * requests, instead of PRIV_VFS_EXEC. 3553 */ 3554 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3555 !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0)) 3556 priv_granted |= VEXEC; 3557 } else { 3558 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3559 !priv_check_cred(cred, PRIV_VFS_EXEC, 0)) 3560 priv_granted |= VEXEC; 3561 } 3562 3563 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && 3564 !priv_check_cred(cred, PRIV_VFS_READ, 0)) 3565 priv_granted |= VREAD; 3566 3567 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3568 !priv_check_cred(cred, PRIV_VFS_WRITE, 0)) 3569 priv_granted |= (VWRITE | VAPPEND); 3570 3571 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3572 !priv_check_cred(cred, PRIV_VFS_ADMIN, 0)) 3573 priv_granted |= VADMIN; 3574 3575 if ((accmode & (priv_granted | dac_granted)) == accmode) { 3576 /* XXX audit: privilege used */ 3577 if (privused != NULL) 3578 *privused = 1; 3579 return (0); 3580 } 3581 3582 return ((accmode & VADMIN) ? EPERM : EACCES); 3583 } 3584 3585 /* 3586 * Credential check based on process requesting service, and per-attribute 3587 * permissions. 3588 */ 3589 int 3590 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, 3591 struct thread *td, accmode_t accmode) 3592 { 3593 3594 /* 3595 * Kernel-invoked always succeeds. 3596 */ 3597 if (cred == NOCRED) 3598 return (0); 3599 3600 /* 3601 * Do not allow privileged processes in jail to directly manipulate 3602 * system attributes. 3603 */ 3604 switch (attrnamespace) { 3605 case EXTATTR_NAMESPACE_SYSTEM: 3606 /* Potentially should be: return (EPERM); */ 3607 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0)); 3608 case EXTATTR_NAMESPACE_USER: 3609 return (VOP_ACCESS(vp, accmode, cred, td)); 3610 default: 3611 return (EPERM); 3612 } 3613 } 3614 3615 #ifdef DEBUG_VFS_LOCKS 3616 /* 3617 * This only exists to supress warnings from unlocked specfs accesses. It is 3618 * no longer ok to have an unlocked VFS. 3619 */ 3620 #define IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL || \ 3621 (vp)->v_type == VCHR || (vp)->v_type == VBAD) 3622 3623 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 3624 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, ""); 3625 3626 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 3627 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, ""); 3628 3629 int vfs_badlock_print = 1; /* Print lock violations. */ 3630 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, ""); 3631 3632 #ifdef KDB 3633 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ 3634 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, ""); 3635 #endif 3636 3637 static void 3638 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 3639 { 3640 3641 #ifdef KDB 3642 if (vfs_badlock_backtrace) 3643 kdb_backtrace(); 3644 #endif 3645 if (vfs_badlock_print) 3646 printf("%s: %p %s\n", str, (void *)vp, msg); 3647 if (vfs_badlock_ddb) 3648 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 3649 } 3650 3651 void 3652 assert_vi_locked(struct vnode *vp, const char *str) 3653 { 3654 3655 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 3656 vfs_badlock("interlock is not locked but should be", str, vp); 3657 } 3658 3659 void 3660 assert_vi_unlocked(struct vnode *vp, const char *str) 3661 { 3662 3663 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 3664 vfs_badlock("interlock is locked but should not be", str, vp); 3665 } 3666 3667 void 3668 assert_vop_locked(struct vnode *vp, const char *str) 3669 { 3670 3671 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == 0) 3672 vfs_badlock("is not locked but should be", str, vp); 3673 } 3674 3675 void 3676 assert_vop_unlocked(struct vnode *vp, const char *str) 3677 { 3678 3679 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE) 3680 vfs_badlock("is locked but should not be", str, vp); 3681 } 3682 3683 void 3684 assert_vop_elocked(struct vnode *vp, const char *str) 3685 { 3686 3687 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 3688 vfs_badlock("is not exclusive locked but should be", str, vp); 3689 } 3690 3691 #if 0 3692 void 3693 assert_vop_elocked_other(struct vnode *vp, const char *str) 3694 { 3695 3696 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLOTHER) 3697 vfs_badlock("is not exclusive locked by another thread", 3698 str, vp); 3699 } 3700 3701 void 3702 assert_vop_slocked(struct vnode *vp, const char *str) 3703 { 3704 3705 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_SHARED) 3706 vfs_badlock("is not locked shared but should be", str, vp); 3707 } 3708 #endif /* 0 */ 3709 #endif /* DEBUG_VFS_LOCKS */ 3710 3711 void 3712 vop_rename_pre(void *ap) 3713 { 3714 struct vop_rename_args *a = ap; 3715 3716 #ifdef DEBUG_VFS_LOCKS 3717 if (a->a_tvp) 3718 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 3719 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 3720 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 3721 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 3722 3723 /* Check the source (from). */ 3724 if (a->a_tdvp != a->a_fdvp && a->a_tvp != a->a_fdvp) 3725 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 3726 if (a->a_tvp != a->a_fvp) 3727 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); 3728 3729 /* Check the target. */ 3730 if (a->a_tvp) 3731 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 3732 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 3733 #endif 3734 if (a->a_tdvp != a->a_fdvp) 3735 vhold(a->a_fdvp); 3736 if (a->a_tvp != a->a_fvp) 3737 vhold(a->a_fvp); 3738 vhold(a->a_tdvp); 3739 if (a->a_tvp) 3740 vhold(a->a_tvp); 3741 } 3742 3743 void 3744 vop_strategy_pre(void *ap) 3745 { 3746 #ifdef DEBUG_VFS_LOCKS 3747 struct vop_strategy_args *a; 3748 struct buf *bp; 3749 3750 a = ap; 3751 bp = a->a_bp; 3752 3753 /* 3754 * Cluster ops lock their component buffers but not the IO container. 3755 */ 3756 if ((bp->b_flags & B_CLUSTER) != 0) 3757 return; 3758 3759 if (!BUF_ISLOCKED(bp)) { 3760 if (vfs_badlock_print) 3761 printf( 3762 "VOP_STRATEGY: bp is not locked but should be\n"); 3763 if (vfs_badlock_ddb) 3764 kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); 3765 } 3766 #endif 3767 } 3768 3769 void 3770 vop_lookup_pre(void *ap) 3771 { 3772 #ifdef DEBUG_VFS_LOCKS 3773 struct vop_lookup_args *a; 3774 struct vnode *dvp; 3775 3776 a = ap; 3777 dvp = a->a_dvp; 3778 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3779 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 3780 #endif 3781 } 3782 3783 void 3784 vop_lookup_post(void *ap, int rc) 3785 { 3786 #ifdef DEBUG_VFS_LOCKS 3787 struct vop_lookup_args *a; 3788 struct vnode *dvp; 3789 struct vnode *vp; 3790 3791 a = ap; 3792 dvp = a->a_dvp; 3793 vp = *(a->a_vpp); 3794 3795 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3796 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 3797 3798 if (!rc) 3799 ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)"); 3800 #endif 3801 } 3802 3803 void 3804 vop_lock_pre(void *ap) 3805 { 3806 #ifdef DEBUG_VFS_LOCKS 3807 struct vop_lock1_args *a = ap; 3808 3809 if ((a->a_flags & LK_INTERLOCK) == 0) 3810 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3811 else 3812 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 3813 #endif 3814 } 3815 3816 void 3817 vop_lock_post(void *ap, int rc) 3818 { 3819 #ifdef DEBUG_VFS_LOCKS 3820 struct vop_lock1_args *a = ap; 3821 3822 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3823 if (rc == 0) 3824 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 3825 #endif 3826 } 3827 3828 void 3829 vop_unlock_pre(void *ap) 3830 { 3831 #ifdef DEBUG_VFS_LOCKS 3832 struct vop_unlock_args *a = ap; 3833 3834 if (a->a_flags & LK_INTERLOCK) 3835 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 3836 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 3837 #endif 3838 } 3839 3840 void 3841 vop_unlock_post(void *ap, int rc) 3842 { 3843 #ifdef DEBUG_VFS_LOCKS 3844 struct vop_unlock_args *a = ap; 3845 3846 if (a->a_flags & LK_INTERLOCK) 3847 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 3848 #endif 3849 } 3850 3851 void 3852 vop_create_post(void *ap, int rc) 3853 { 3854 struct vop_create_args *a = ap; 3855 3856 if (!rc) 3857 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 3858 } 3859 3860 void 3861 vop_link_post(void *ap, int rc) 3862 { 3863 struct vop_link_args *a = ap; 3864 3865 if (!rc) { 3866 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK); 3867 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE); 3868 } 3869 } 3870 3871 void 3872 vop_mkdir_post(void *ap, int rc) 3873 { 3874 struct vop_mkdir_args *a = ap; 3875 3876 if (!rc) 3877 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 3878 } 3879 3880 void 3881 vop_mknod_post(void *ap, int rc) 3882 { 3883 struct vop_mknod_args *a = ap; 3884 3885 if (!rc) 3886 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 3887 } 3888 3889 void 3890 vop_remove_post(void *ap, int rc) 3891 { 3892 struct vop_remove_args *a = ap; 3893 3894 if (!rc) { 3895 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 3896 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 3897 } 3898 } 3899 3900 void 3901 vop_rename_post(void *ap, int rc) 3902 { 3903 struct vop_rename_args *a = ap; 3904 3905 if (!rc) { 3906 VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE); 3907 VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE); 3908 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); 3909 if (a->a_tvp) 3910 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); 3911 } 3912 if (a->a_tdvp != a->a_fdvp) 3913 vdrop(a->a_fdvp); 3914 if (a->a_tvp != a->a_fvp) 3915 vdrop(a->a_fvp); 3916 vdrop(a->a_tdvp); 3917 if (a->a_tvp) 3918 vdrop(a->a_tvp); 3919 } 3920 3921 void 3922 vop_rmdir_post(void *ap, int rc) 3923 { 3924 struct vop_rmdir_args *a = ap; 3925 3926 if (!rc) { 3927 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); 3928 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); 3929 } 3930 } 3931 3932 void 3933 vop_setattr_post(void *ap, int rc) 3934 { 3935 struct vop_setattr_args *a = ap; 3936 3937 if (!rc) 3938 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); 3939 } 3940 3941 void 3942 vop_symlink_post(void *ap, int rc) 3943 { 3944 struct vop_symlink_args *a = ap; 3945 3946 if (!rc) 3947 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); 3948 } 3949 3950 static struct knlist fs_knlist; 3951 3952 static void 3953 vfs_event_init(void *arg) 3954 { 3955 knlist_init(&fs_knlist, NULL, NULL, NULL, NULL); 3956 } 3957 /* XXX - correct order? */ 3958 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); 3959 3960 void 3961 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused) 3962 { 3963 3964 KNOTE_UNLOCKED(&fs_knlist, event); 3965 } 3966 3967 static int filt_fsattach(struct knote *kn); 3968 static void filt_fsdetach(struct knote *kn); 3969 static int filt_fsevent(struct knote *kn, long hint); 3970 3971 struct filterops fs_filtops = 3972 { 0, filt_fsattach, filt_fsdetach, filt_fsevent }; 3973 3974 static int 3975 filt_fsattach(struct knote *kn) 3976 { 3977 3978 kn->kn_flags |= EV_CLEAR; 3979 knlist_add(&fs_knlist, kn, 0); 3980 return (0); 3981 } 3982 3983 static void 3984 filt_fsdetach(struct knote *kn) 3985 { 3986 3987 knlist_remove(&fs_knlist, kn, 0); 3988 } 3989 3990 static int 3991 filt_fsevent(struct knote *kn, long hint) 3992 { 3993 3994 kn->kn_fflags |= hint; 3995 return (kn->kn_fflags != 0); 3996 } 3997 3998 static int 3999 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) 4000 { 4001 struct vfsidctl vc; 4002 int error; 4003 struct mount *mp; 4004 4005 error = SYSCTL_IN(req, &vc, sizeof(vc)); 4006 if (error) 4007 return (error); 4008 if (vc.vc_vers != VFS_CTL_VERS1) 4009 return (EINVAL); 4010 mp = vfs_getvfs(&vc.vc_fsid); 4011 if (mp == NULL) 4012 return (ENOENT); 4013 /* ensure that a specific sysctl goes to the right filesystem. */ 4014 if (strcmp(vc.vc_fstypename, "*") != 0 && 4015 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { 4016 vfs_rel(mp); 4017 return (EINVAL); 4018 } 4019 VCTLTOREQ(&vc, req); 4020 error = VFS_SYSCTL(mp, vc.vc_op, req); 4021 vfs_rel(mp); 4022 return (error); 4023 } 4024 4025 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR, NULL, 0, sysctl_vfs_ctl, "", 4026 "Sysctl by fsid"); 4027 4028 /* 4029 * Function to initialize a va_filerev field sensibly. 4030 * XXX: Wouldn't a random number make a lot more sense ?? 4031 */ 4032 u_quad_t 4033 init_va_filerev(void) 4034 { 4035 struct bintime bt; 4036 4037 getbinuptime(&bt); 4038 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); 4039 } 4040 4041 static int filt_vfsread(struct knote *kn, long hint); 4042 static int filt_vfswrite(struct knote *kn, long hint); 4043 static int filt_vfsvnode(struct knote *kn, long hint); 4044 static void filt_vfsdetach(struct knote *kn); 4045 static struct filterops vfsread_filtops = 4046 { 1, NULL, filt_vfsdetach, filt_vfsread }; 4047 static struct filterops vfswrite_filtops = 4048 { 1, NULL, filt_vfsdetach, filt_vfswrite }; 4049 static struct filterops vfsvnode_filtops = 4050 { 1, NULL, filt_vfsdetach, filt_vfsvnode }; 4051 4052 static void 4053 vfs_knllock(void *arg) 4054 { 4055 struct vnode *vp = arg; 4056 4057 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4058 } 4059 4060 static void 4061 vfs_knlunlock(void *arg) 4062 { 4063 struct vnode *vp = arg; 4064 4065 VOP_UNLOCK(vp, 0); 4066 } 4067 4068 static int 4069 vfs_knllocked(void *arg) 4070 { 4071 struct vnode *vp = arg; 4072 4073 return (VOP_ISLOCKED(vp) == LK_EXCLUSIVE); 4074 } 4075 4076 int 4077 vfs_kqfilter(struct vop_kqfilter_args *ap) 4078 { 4079 struct vnode *vp = ap->a_vp; 4080 struct knote *kn = ap->a_kn; 4081 struct knlist *knl; 4082 4083 switch (kn->kn_filter) { 4084 case EVFILT_READ: 4085 kn->kn_fop = &vfsread_filtops; 4086 break; 4087 case EVFILT_WRITE: 4088 kn->kn_fop = &vfswrite_filtops; 4089 break; 4090 case EVFILT_VNODE: 4091 kn->kn_fop = &vfsvnode_filtops; 4092 break; 4093 default: 4094 return (EINVAL); 4095 } 4096 4097 kn->kn_hook = (caddr_t)vp; 4098 4099 v_addpollinfo(vp); 4100 if (vp->v_pollinfo == NULL) 4101 return (ENOMEM); 4102 knl = &vp->v_pollinfo->vpi_selinfo.si_note; 4103 knlist_add(knl, kn, 0); 4104 4105 return (0); 4106 } 4107 4108 /* 4109 * Detach knote from vnode 4110 */ 4111 static void 4112 filt_vfsdetach(struct knote *kn) 4113 { 4114 struct vnode *vp = (struct vnode *)kn->kn_hook; 4115 4116 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); 4117 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); 4118 } 4119 4120 /*ARGSUSED*/ 4121 static int 4122 filt_vfsread(struct knote *kn, long hint) 4123 { 4124 struct vnode *vp = (struct vnode *)kn->kn_hook; 4125 struct vattr va; 4126 4127 /* 4128 * filesystem is gone, so set the EOF flag and schedule 4129 * the knote for deletion. 4130 */ 4131 if (hint == NOTE_REVOKE) { 4132 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 4133 return (1); 4134 } 4135 4136 if (VOP_GETATTR(vp, &va, curthread->td_ucred)) 4137 return (0); 4138 4139 kn->kn_data = va.va_size - kn->kn_fp->f_offset; 4140 return (kn->kn_data != 0); 4141 } 4142 4143 /*ARGSUSED*/ 4144 static int 4145 filt_vfswrite(struct knote *kn, long hint) 4146 { 4147 /* 4148 * filesystem is gone, so set the EOF flag and schedule 4149 * the knote for deletion. 4150 */ 4151 if (hint == NOTE_REVOKE) 4152 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 4153 4154 kn->kn_data = 0; 4155 return (1); 4156 } 4157 4158 static int 4159 filt_vfsvnode(struct knote *kn, long hint) 4160 { 4161 if (kn->kn_sfflags & hint) 4162 kn->kn_fflags |= hint; 4163 if (hint == NOTE_REVOKE) { 4164 kn->kn_flags |= EV_EOF; 4165 return (1); 4166 } 4167 return (kn->kn_fflags != 0); 4168 } 4169 4170 int 4171 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) 4172 { 4173 int error; 4174 4175 if (dp->d_reclen > ap->a_uio->uio_resid) 4176 return (ENAMETOOLONG); 4177 error = uiomove(dp, dp->d_reclen, ap->a_uio); 4178 if (error) { 4179 if (ap->a_ncookies != NULL) { 4180 if (ap->a_cookies != NULL) 4181 free(ap->a_cookies, M_TEMP); 4182 ap->a_cookies = NULL; 4183 *ap->a_ncookies = 0; 4184 } 4185 return (error); 4186 } 4187 if (ap->a_ncookies == NULL) 4188 return (0); 4189 4190 KASSERT(ap->a_cookies, 4191 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); 4192 4193 *ap->a_cookies = realloc(*ap->a_cookies, 4194 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); 4195 (*ap->a_cookies)[*ap->a_ncookies] = off; 4196 return (0); 4197 } 4198 4199 /* 4200 * Mark for update the access time of the file if the filesystem 4201 * supports VOP_MARKATIME. This functionality is used by execve and 4202 * mmap, so we want to avoid the I/O implied by directly setting 4203 * va_atime for the sake of efficiency. 4204 */ 4205 void 4206 vfs_mark_atime(struct vnode *vp, struct ucred *cred) 4207 { 4208 4209 if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) 4210 (void)VOP_MARKATIME(vp); 4211 } 4212