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 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 39 */ 40 41 /* 42 * External virtual filesystem routines 43 */ 44 45 #include <sys/cdefs.h> 46 __FBSDID("$FreeBSD$"); 47 48 #include "opt_ddb.h" 49 #include "opt_mac.h" 50 51 #include <sys/param.h> 52 #include <sys/systm.h> 53 #include <sys/bio.h> 54 #include <sys/buf.h> 55 #include <sys/conf.h> 56 #include <sys/eventhandler.h> 57 #include <sys/extattr.h> 58 #include <sys/fcntl.h> 59 #include <sys/kernel.h> 60 #include <sys/kthread.h> 61 #include <sys/mac.h> 62 #include <sys/malloc.h> 63 #include <sys/mount.h> 64 #include <sys/namei.h> 65 #include <sys/stat.h> 66 #include <sys/sysctl.h> 67 #include <sys/syslog.h> 68 #include <sys/vmmeter.h> 69 #include <sys/vnode.h> 70 71 #include <vm/vm.h> 72 #include <vm/vm_object.h> 73 #include <vm/vm_extern.h> 74 #include <vm/pmap.h> 75 #include <vm/vm_map.h> 76 #include <vm/vm_page.h> 77 #include <vm/vm_kern.h> 78 #include <vm/uma.h> 79 80 static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure"); 81 82 static void addalias(struct vnode *vp, dev_t nvp_rdev); 83 static void insmntque(struct vnode *vp, struct mount *mp); 84 static void vclean(struct vnode *vp, int flags, struct thread *td); 85 static void vlruvp(struct vnode *vp); 86 static int flushbuflist(struct buf *blist, int flags, struct vnode *vp, 87 int slpflag, int slptimeo, int *errorp); 88 static int vtryrecycle(struct vnode *vp); 89 static void vx_lock(struct vnode *vp); 90 static void vx_unlock(struct vnode *vp); 91 static void vgonechrl(struct vnode *vp, struct thread *td); 92 93 94 /* 95 * Number of vnodes in existence. Increased whenever getnewvnode() 96 * allocates a new vnode, never decreased. 97 */ 98 static unsigned long numvnodes; 99 100 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); 101 102 /* 103 * Conversion tables for conversion from vnode types to inode formats 104 * and back. 105 */ 106 enum vtype iftovt_tab[16] = { 107 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 108 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 109 }; 110 int vttoif_tab[9] = { 111 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 112 S_IFSOCK, S_IFIFO, S_IFMT, 113 }; 114 115 /* 116 * List of vnodes that are ready for recycling. 117 */ 118 static TAILQ_HEAD(freelst, vnode) vnode_free_list; 119 120 /* 121 * Minimum number of free vnodes. If there are fewer than this free vnodes, 122 * getnewvnode() will return a newly allocated vnode. 123 */ 124 static u_long wantfreevnodes = 25; 125 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 126 /* Number of vnodes in the free list. */ 127 static u_long freevnodes; 128 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); 129 130 /* 131 * Various variables used for debugging the new implementation of 132 * reassignbuf(). 133 * XXX these are probably of (very) limited utility now. 134 */ 135 static int reassignbufcalls; 136 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); 137 static int nameileafonly; 138 SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, ""); 139 140 /* 141 * Cache for the mount type id assigned to NFS. This is used for 142 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 143 */ 144 int nfs_mount_type = -1; 145 146 /* To keep more than one thread at a time from running vfs_getnewfsid */ 147 static struct mtx mntid_mtx; 148 149 /* 150 * Lock for any access to the following: 151 * vnode_free_list 152 * numvnodes 153 * freevnodes 154 */ 155 static struct mtx vnode_free_list_mtx; 156 157 /* 158 * For any iteration/modification of dev->si_hlist (linked through 159 * v_specnext) 160 */ 161 static struct mtx spechash_mtx; 162 163 /* Publicly exported FS */ 164 struct nfs_public nfs_pub; 165 166 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 167 static uma_zone_t vnode_zone; 168 static uma_zone_t vnodepoll_zone; 169 170 /* Set to 1 to print out reclaim of active vnodes */ 171 int prtactive; 172 173 /* 174 * The workitem queue. 175 * 176 * It is useful to delay writes of file data and filesystem metadata 177 * for tens of seconds so that quickly created and deleted files need 178 * not waste disk bandwidth being created and removed. To realize this, 179 * we append vnodes to a "workitem" queue. When running with a soft 180 * updates implementation, most pending metadata dependencies should 181 * not wait for more than a few seconds. Thus, mounted on block devices 182 * are delayed only about a half the time that file data is delayed. 183 * Similarly, directory updates are more critical, so are only delayed 184 * about a third the time that file data is delayed. Thus, there are 185 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 186 * one each second (driven off the filesystem syncer process). The 187 * syncer_delayno variable indicates the next queue that is to be processed. 188 * Items that need to be processed soon are placed in this queue: 189 * 190 * syncer_workitem_pending[syncer_delayno] 191 * 192 * A delay of fifteen seconds is done by placing the request fifteen 193 * entries later in the queue: 194 * 195 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 196 * 197 */ 198 static int syncer_delayno; 199 static long syncer_mask; 200 LIST_HEAD(synclist, vnode); 201 static struct synclist *syncer_workitem_pending; 202 /* 203 * The sync_mtx protects: 204 * vp->v_synclist 205 * syncer_delayno 206 * syncer_workitem_pending 207 * rushjob 208 */ 209 static struct mtx sync_mtx; 210 211 #define SYNCER_MAXDELAY 32 212 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 213 static int syncdelay = 30; /* max time to delay syncing data */ 214 static int filedelay = 30; /* time to delay syncing files */ 215 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); 216 static int dirdelay = 29; /* time to delay syncing directories */ 217 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); 218 static int metadelay = 28; /* time to delay syncing metadata */ 219 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); 220 static int rushjob; /* number of slots to run ASAP */ 221 static int stat_rush_requests; /* number of times I/O speeded up */ 222 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); 223 224 /* 225 * Number of vnodes we want to exist at any one time. This is mostly used 226 * to size hash tables in vnode-related code. It is normally not used in 227 * getnewvnode(), as wantfreevnodes is normally nonzero.) 228 * 229 * XXX desiredvnodes is historical cruft and should not exist. 230 */ 231 int desiredvnodes; 232 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 233 &desiredvnodes, 0, "Maximum number of vnodes"); 234 static int minvnodes; 235 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 236 &minvnodes, 0, "Minimum number of vnodes"); 237 static int vnlru_nowhere; 238 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 239 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 240 241 /* Hook for calling soft updates. */ 242 int (*softdep_process_worklist_hook)(struct mount *); 243 244 /* 245 * Initialize the vnode management data structures. 246 */ 247 static void 248 vntblinit(void *dummy __unused) 249 { 250 251 /* 252 * Desiredvnodes is a function of the physical memory size and 253 * the kernel's heap size. Specifically, desiredvnodes scales 254 * in proportion to the physical memory size until two fifths 255 * of the kernel's heap size is consumed by vnodes and vm 256 * objects. 257 */ 258 desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size / 259 (5 * (sizeof(struct vm_object) + sizeof(struct vnode)))); 260 minvnodes = desiredvnodes / 4; 261 mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF); 262 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 263 mtx_init(&spechash_mtx, "spechash", NULL, MTX_DEF); 264 TAILQ_INIT(&vnode_free_list); 265 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 266 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 267 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 268 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 269 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 270 /* 271 * Initialize the filesystem syncer. 272 */ 273 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 274 &syncer_mask); 275 syncer_maxdelay = syncer_mask + 1; 276 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 277 } 278 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) 279 280 281 /* 282 * Mark a mount point as busy. Used to synchronize access and to delay 283 * unmounting. Interlock is not released on failure. 284 */ 285 int 286 vfs_busy(mp, flags, interlkp, td) 287 struct mount *mp; 288 int flags; 289 struct mtx *interlkp; 290 struct thread *td; 291 { 292 int lkflags; 293 294 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 295 if (flags & LK_NOWAIT) 296 return (ENOENT); 297 mp->mnt_kern_flag |= MNTK_MWAIT; 298 /* 299 * Since all busy locks are shared except the exclusive 300 * lock granted when unmounting, the only place that a 301 * wakeup needs to be done is at the release of the 302 * exclusive lock at the end of dounmount. 303 */ 304 msleep(mp, interlkp, PVFS, "vfs_busy", 0); 305 return (ENOENT); 306 } 307 lkflags = LK_SHARED | LK_NOPAUSE; 308 if (interlkp) 309 lkflags |= LK_INTERLOCK; 310 if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td)) 311 panic("vfs_busy: unexpected lock failure"); 312 return (0); 313 } 314 315 /* 316 * Free a busy filesystem. 317 */ 318 void 319 vfs_unbusy(mp, td) 320 struct mount *mp; 321 struct thread *td; 322 { 323 324 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); 325 } 326 327 /* 328 * Lookup a mount point by filesystem identifier. 329 */ 330 struct mount * 331 vfs_getvfs(fsid) 332 fsid_t *fsid; 333 { 334 register struct mount *mp; 335 336 mtx_lock(&mountlist_mtx); 337 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 338 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 339 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 340 mtx_unlock(&mountlist_mtx); 341 return (mp); 342 } 343 } 344 mtx_unlock(&mountlist_mtx); 345 return ((struct mount *) 0); 346 } 347 348 /* 349 * Get a new unique fsid. Try to make its val[0] unique, since this value 350 * will be used to create fake device numbers for stat(). Also try (but 351 * not so hard) make its val[0] unique mod 2^16, since some emulators only 352 * support 16-bit device numbers. We end up with unique val[0]'s for the 353 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 354 * 355 * Keep in mind that several mounts may be running in parallel. Starting 356 * the search one past where the previous search terminated is both a 357 * micro-optimization and a defense against returning the same fsid to 358 * different mounts. 359 */ 360 void 361 vfs_getnewfsid(mp) 362 struct mount *mp; 363 { 364 static u_int16_t mntid_base; 365 fsid_t tfsid; 366 int mtype; 367 368 mtx_lock(&mntid_mtx); 369 mtype = mp->mnt_vfc->vfc_typenum; 370 tfsid.val[1] = mtype; 371 mtype = (mtype & 0xFF) << 24; 372 for (;;) { 373 tfsid.val[0] = makeudev(255, 374 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 375 mntid_base++; 376 if (vfs_getvfs(&tfsid) == NULL) 377 break; 378 } 379 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 380 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 381 mtx_unlock(&mntid_mtx); 382 } 383 384 /* 385 * Knob to control the precision of file timestamps: 386 * 387 * 0 = seconds only; nanoseconds zeroed. 388 * 1 = seconds and nanoseconds, accurate within 1/HZ. 389 * 2 = seconds and nanoseconds, truncated to microseconds. 390 * >=3 = seconds and nanoseconds, maximum precision. 391 */ 392 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 393 394 static int timestamp_precision = TSP_SEC; 395 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 396 ×tamp_precision, 0, ""); 397 398 /* 399 * Get a current timestamp. 400 */ 401 void 402 vfs_timestamp(tsp) 403 struct timespec *tsp; 404 { 405 struct timeval tv; 406 407 switch (timestamp_precision) { 408 case TSP_SEC: 409 tsp->tv_sec = time_second; 410 tsp->tv_nsec = 0; 411 break; 412 case TSP_HZ: 413 getnanotime(tsp); 414 break; 415 case TSP_USEC: 416 microtime(&tv); 417 TIMEVAL_TO_TIMESPEC(&tv, tsp); 418 break; 419 case TSP_NSEC: 420 default: 421 nanotime(tsp); 422 break; 423 } 424 } 425 426 /* 427 * Set vnode attributes to VNOVAL 428 */ 429 void 430 vattr_null(vap) 431 register struct vattr *vap; 432 { 433 434 vap->va_type = VNON; 435 vap->va_size = VNOVAL; 436 vap->va_bytes = VNOVAL; 437 vap->va_mode = VNOVAL; 438 vap->va_nlink = VNOVAL; 439 vap->va_uid = VNOVAL; 440 vap->va_gid = VNOVAL; 441 vap->va_fsid = VNOVAL; 442 vap->va_fileid = VNOVAL; 443 vap->va_blocksize = VNOVAL; 444 vap->va_rdev = VNOVAL; 445 vap->va_atime.tv_sec = VNOVAL; 446 vap->va_atime.tv_nsec = VNOVAL; 447 vap->va_mtime.tv_sec = VNOVAL; 448 vap->va_mtime.tv_nsec = VNOVAL; 449 vap->va_ctime.tv_sec = VNOVAL; 450 vap->va_ctime.tv_nsec = VNOVAL; 451 vap->va_birthtime.tv_sec = VNOVAL; 452 vap->va_birthtime.tv_nsec = VNOVAL; 453 vap->va_flags = VNOVAL; 454 vap->va_gen = VNOVAL; 455 vap->va_vaflags = 0; 456 } 457 458 /* 459 * This routine is called when we have too many vnodes. It attempts 460 * to free <count> vnodes and will potentially free vnodes that still 461 * have VM backing store (VM backing store is typically the cause 462 * of a vnode blowout so we want to do this). Therefore, this operation 463 * is not considered cheap. 464 * 465 * A number of conditions may prevent a vnode from being reclaimed. 466 * the buffer cache may have references on the vnode, a directory 467 * vnode may still have references due to the namei cache representing 468 * underlying files, or the vnode may be in active use. It is not 469 * desireable to reuse such vnodes. These conditions may cause the 470 * number of vnodes to reach some minimum value regardless of what 471 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 472 */ 473 static int 474 vlrureclaim(struct mount *mp) 475 { 476 struct vnode *vp; 477 int done; 478 int trigger; 479 int usevnodes; 480 int count; 481 482 /* 483 * Calculate the trigger point, don't allow user 484 * screwups to blow us up. This prevents us from 485 * recycling vnodes with lots of resident pages. We 486 * aren't trying to free memory, we are trying to 487 * free vnodes. 488 */ 489 usevnodes = desiredvnodes; 490 if (usevnodes <= 0) 491 usevnodes = 1; 492 trigger = cnt.v_page_count * 2 / usevnodes; 493 494 done = 0; 495 MNT_ILOCK(mp); 496 count = mp->mnt_nvnodelistsize / 10 + 1; 497 while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) { 498 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 499 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 500 501 if (vp->v_type != VNON && 502 vp->v_type != VBAD && 503 VI_TRYLOCK(vp)) { 504 if (VMIGHTFREE(vp) && /* critical path opt */ 505 (vp->v_object == NULL || 506 vp->v_object->resident_page_count < trigger)) { 507 MNT_IUNLOCK(mp); 508 vgonel(vp, curthread); 509 done++; 510 MNT_ILOCK(mp); 511 } else 512 VI_UNLOCK(vp); 513 } 514 --count; 515 } 516 MNT_IUNLOCK(mp); 517 return done; 518 } 519 520 /* 521 * Attempt to recycle vnodes in a context that is always safe to block. 522 * Calling vlrurecycle() from the bowels of filesystem code has some 523 * interesting deadlock problems. 524 */ 525 static struct proc *vnlruproc; 526 static int vnlruproc_sig; 527 528 static void 529 vnlru_proc(void) 530 { 531 struct mount *mp, *nmp; 532 int done; 533 struct proc *p = vnlruproc; 534 struct thread *td = FIRST_THREAD_IN_PROC(p); /* XXXKSE */ 535 536 mtx_lock(&Giant); 537 538 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 539 SHUTDOWN_PRI_FIRST); 540 541 for (;;) { 542 kthread_suspend_check(p); 543 mtx_lock(&vnode_free_list_mtx); 544 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { 545 mtx_unlock(&vnode_free_list_mtx); 546 vnlruproc_sig = 0; 547 wakeup(&vnlruproc_sig); 548 tsleep(vnlruproc, PVFS, "vlruwt", hz); 549 continue; 550 } 551 mtx_unlock(&vnode_free_list_mtx); 552 done = 0; 553 mtx_lock(&mountlist_mtx); 554 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 555 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 556 nmp = TAILQ_NEXT(mp, mnt_list); 557 continue; 558 } 559 done += vlrureclaim(mp); 560 mtx_lock(&mountlist_mtx); 561 nmp = TAILQ_NEXT(mp, mnt_list); 562 vfs_unbusy(mp, td); 563 } 564 mtx_unlock(&mountlist_mtx); 565 if (done == 0) { 566 #if 0 567 /* These messages are temporary debugging aids */ 568 if (vnlru_nowhere < 5) 569 printf("vnlru process getting nowhere..\n"); 570 else if (vnlru_nowhere == 5) 571 printf("vnlru process messages stopped.\n"); 572 #endif 573 vnlru_nowhere++; 574 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 575 } 576 } 577 } 578 579 static struct kproc_desc vnlru_kp = { 580 "vnlru", 581 vnlru_proc, 582 &vnlruproc 583 }; 584 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 585 586 587 /* 588 * Routines having to do with the management of the vnode table. 589 */ 590 591 /* 592 * Check to see if a free vnode can be recycled. If it can, 593 * recycle it and return it with the vnode interlock held. 594 */ 595 static int 596 vtryrecycle(struct vnode *vp) 597 { 598 struct thread *td = curthread; 599 vm_object_t object; 600 struct mount *vnmp; 601 int error; 602 603 /* Don't recycle if we can't get the interlock */ 604 if (!VI_TRYLOCK(vp)) 605 return (EWOULDBLOCK); 606 /* 607 * This vnode may found and locked via some other list, if so we 608 * can't recycle it yet. 609 */ 610 if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 611 return (EWOULDBLOCK); 612 /* 613 * Don't recycle if its filesystem is being suspended. 614 */ 615 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 616 error = EBUSY; 617 goto done; 618 } 619 620 /* 621 * Don't recycle if we still have cached pages. 622 */ 623 if (VOP_GETVOBJECT(vp, &object) == 0) { 624 VM_OBJECT_LOCK(object); 625 if (object->resident_page_count || 626 object->ref_count) { 627 VM_OBJECT_UNLOCK(object); 628 error = EBUSY; 629 goto done; 630 } 631 VM_OBJECT_UNLOCK(object); 632 } 633 if (LIST_FIRST(&vp->v_cache_src)) { 634 /* 635 * note: nameileafonly sysctl is temporary, 636 * for debugging only, and will eventually be 637 * removed. 638 */ 639 if (nameileafonly > 0) { 640 /* 641 * Do not reuse namei-cached directory 642 * vnodes that have cached 643 * subdirectories. 644 */ 645 if (cache_leaf_test(vp) < 0) { 646 error = EISDIR; 647 goto done; 648 } 649 } else if (nameileafonly < 0 || 650 vmiodirenable == 0) { 651 /* 652 * Do not reuse namei-cached directory 653 * vnodes if nameileafonly is -1 or 654 * if VMIO backing for directories is 655 * turned off (otherwise we reuse them 656 * too quickly). 657 */ 658 error = EBUSY; 659 goto done; 660 } 661 } 662 /* 663 * If we got this far, we need to acquire the interlock and see if 664 * anyone picked up this vnode from another list. If not, we will 665 * mark it with XLOCK via vgonel() so that anyone who does find it 666 * will skip over it. 667 */ 668 VI_LOCK(vp); 669 if (VSHOULDBUSY(vp) && (vp->v_iflag & VI_XLOCK) == 0) { 670 VI_UNLOCK(vp); 671 error = EBUSY; 672 goto done; 673 } 674 mtx_lock(&vnode_free_list_mtx); 675 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 676 vp->v_iflag &= ~VI_FREE; 677 mtx_unlock(&vnode_free_list_mtx); 678 vp->v_iflag |= VI_DOOMED; 679 if (vp->v_type != VBAD) { 680 VOP_UNLOCK(vp, 0, td); 681 vgonel(vp, td); 682 VI_LOCK(vp); 683 } else 684 VOP_UNLOCK(vp, 0, td); 685 vn_finished_write(vnmp); 686 return (0); 687 done: 688 VOP_UNLOCK(vp, 0, td); 689 return (error); 690 } 691 692 /* 693 * Return the next vnode from the free list. 694 */ 695 int 696 getnewvnode(tag, mp, vops, vpp) 697 const char *tag; 698 struct mount *mp; 699 vop_t **vops; 700 struct vnode **vpp; 701 { 702 struct vnode *vp = NULL; 703 struct vpollinfo *pollinfo = NULL; 704 705 mtx_lock(&vnode_free_list_mtx); 706 707 /* 708 * Try to reuse vnodes if we hit the max. This situation only 709 * occurs in certain large-memory (2G+) situations. We cannot 710 * attempt to directly reclaim vnodes due to nasty recursion 711 * problems. 712 */ 713 while (numvnodes - freevnodes > desiredvnodes) { 714 if (vnlruproc_sig == 0) { 715 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 716 wakeup(vnlruproc); 717 } 718 mtx_unlock(&vnode_free_list_mtx); 719 tsleep(&vnlruproc_sig, PVFS, "vlruwk", hz); 720 mtx_lock(&vnode_free_list_mtx); 721 } 722 723 /* 724 * Attempt to reuse a vnode already on the free list, allocating 725 * a new vnode if we can't find one or if we have not reached a 726 * good minimum for good LRU performance. 727 */ 728 729 if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) { 730 int error; 731 int count; 732 733 for (count = 0; count < freevnodes; count++) { 734 vp = TAILQ_FIRST(&vnode_free_list); 735 736 KASSERT(vp->v_usecount == 0 && 737 (vp->v_iflag & VI_DOINGINACT) == 0, 738 ("getnewvnode: free vnode isn't")); 739 740 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 741 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 742 mtx_unlock(&vnode_free_list_mtx); 743 error = vtryrecycle(vp); 744 mtx_lock(&vnode_free_list_mtx); 745 if (error == 0) 746 break; 747 vp = NULL; 748 } 749 } 750 if (vp) { 751 freevnodes--; 752 mtx_unlock(&vnode_free_list_mtx); 753 754 #ifdef INVARIANTS 755 { 756 if (vp->v_data) 757 panic("cleaned vnode isn't"); 758 if (vp->v_numoutput) 759 panic("Clean vnode has pending I/O's"); 760 if (vp->v_writecount != 0) 761 panic("Non-zero write count"); 762 } 763 #endif 764 if ((pollinfo = vp->v_pollinfo) != NULL) { 765 /* 766 * To avoid lock order reversals, the call to 767 * uma_zfree() must be delayed until the vnode 768 * interlock is released. 769 */ 770 vp->v_pollinfo = NULL; 771 } 772 #ifdef MAC 773 mac_destroy_vnode(vp); 774 #endif 775 vp->v_iflag = 0; 776 vp->v_vflag = 0; 777 vp->v_lastw = 0; 778 vp->v_lasta = 0; 779 vp->v_cstart = 0; 780 vp->v_clen = 0; 781 vp->v_socket = 0; 782 lockdestroy(vp->v_vnlock); 783 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 784 KASSERT(vp->v_cleanbufcnt == 0, ("cleanbufcnt not 0")); 785 KASSERT(vp->v_cleanblkroot == NULL, ("cleanblkroot not NULL")); 786 KASSERT(vp->v_dirtybufcnt == 0, ("dirtybufcnt not 0")); 787 KASSERT(vp->v_dirtyblkroot == NULL, ("dirtyblkroot not NULL")); 788 } else { 789 numvnodes++; 790 mtx_unlock(&vnode_free_list_mtx); 791 792 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); 793 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 794 VI_LOCK(vp); 795 vp->v_dd = vp; 796 vp->v_vnlock = &vp->v_lock; 797 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 798 cache_purge(vp); /* Sets up v_id. */ 799 LIST_INIT(&vp->v_cache_src); 800 TAILQ_INIT(&vp->v_cache_dst); 801 } 802 803 TAILQ_INIT(&vp->v_cleanblkhd); 804 TAILQ_INIT(&vp->v_dirtyblkhd); 805 vp->v_type = VNON; 806 vp->v_tag = tag; 807 vp->v_op = vops; 808 *vpp = vp; 809 vp->v_usecount = 1; 810 vp->v_data = 0; 811 vp->v_cachedid = -1; 812 VI_UNLOCK(vp); 813 if (pollinfo != NULL) { 814 mtx_destroy(&pollinfo->vpi_lock); 815 uma_zfree(vnodepoll_zone, pollinfo); 816 } 817 #ifdef MAC 818 mac_init_vnode(vp); 819 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 820 mac_associate_vnode_singlelabel(mp, vp); 821 #endif 822 insmntque(vp, mp); 823 824 return (0); 825 } 826 827 /* 828 * Move a vnode from one mount queue to another. 829 */ 830 static void 831 insmntque(vp, mp) 832 register struct vnode *vp; 833 register struct mount *mp; 834 { 835 836 /* 837 * Delete from old mount point vnode list, if on one. 838 */ 839 if (vp->v_mount != NULL) { 840 MNT_ILOCK(vp->v_mount); 841 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0, 842 ("bad mount point vnode list size")); 843 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes); 844 vp->v_mount->mnt_nvnodelistsize--; 845 MNT_IUNLOCK(vp->v_mount); 846 } 847 /* 848 * Insert into list of vnodes for the new mount point, if available. 849 */ 850 if ((vp->v_mount = mp) != NULL) { 851 MNT_ILOCK(vp->v_mount); 852 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 853 mp->mnt_nvnodelistsize++; 854 MNT_IUNLOCK(vp->v_mount); 855 } 856 } 857 858 /* 859 * Update outstanding I/O count and do wakeup if requested. 860 */ 861 void 862 vwakeup(bp) 863 register struct buf *bp; 864 { 865 register struct vnode *vp; 866 867 bp->b_flags &= ~B_WRITEINPROG; 868 if ((vp = bp->b_vp)) { 869 VI_LOCK(vp); 870 vp->v_numoutput--; 871 if (vp->v_numoutput < 0) 872 panic("vwakeup: neg numoutput"); 873 if ((vp->v_numoutput == 0) && (vp->v_iflag & VI_BWAIT)) { 874 vp->v_iflag &= ~VI_BWAIT; 875 wakeup(&vp->v_numoutput); 876 } 877 VI_UNLOCK(vp); 878 } 879 } 880 881 /* 882 * Flush out and invalidate all buffers associated with a vnode. 883 * Called with the underlying object locked. 884 */ 885 int 886 vinvalbuf(vp, flags, cred, td, slpflag, slptimeo) 887 struct vnode *vp; 888 int flags; 889 struct ucred *cred; 890 struct thread *td; 891 int slpflag, slptimeo; 892 { 893 struct buf *blist; 894 int error; 895 vm_object_t object; 896 897 GIANT_REQUIRED; 898 899 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 900 901 VI_LOCK(vp); 902 if (flags & V_SAVE) { 903 while (vp->v_numoutput) { 904 vp->v_iflag |= VI_BWAIT; 905 error = msleep(&vp->v_numoutput, VI_MTX(vp), 906 slpflag | (PRIBIO + 1), "vinvlbuf", slptimeo); 907 if (error) { 908 VI_UNLOCK(vp); 909 return (error); 910 } 911 } 912 if (!TAILQ_EMPTY(&vp->v_dirtyblkhd)) { 913 VI_UNLOCK(vp); 914 if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0) 915 return (error); 916 /* 917 * XXX We could save a lock/unlock if this was only 918 * enabled under INVARIANTS 919 */ 920 VI_LOCK(vp); 921 if (vp->v_numoutput > 0 || 922 !TAILQ_EMPTY(&vp->v_dirtyblkhd)) 923 panic("vinvalbuf: dirty bufs"); 924 } 925 } 926 /* 927 * If you alter this loop please notice that interlock is dropped and 928 * reacquired in flushbuflist. Special care is needed to ensure that 929 * no race conditions occur from this. 930 */ 931 for (error = 0;;) { 932 if ((blist = TAILQ_FIRST(&vp->v_cleanblkhd)) != 0 && 933 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 934 if (error) 935 break; 936 continue; 937 } 938 if ((blist = TAILQ_FIRST(&vp->v_dirtyblkhd)) != 0 && 939 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 940 if (error) 941 break; 942 continue; 943 } 944 break; 945 } 946 if (error) { 947 VI_UNLOCK(vp); 948 return (error); 949 } 950 951 /* 952 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 953 * have write I/O in-progress but if there is a VM object then the 954 * VM object can also have read-I/O in-progress. 955 */ 956 do { 957 while (vp->v_numoutput > 0) { 958 vp->v_iflag |= VI_BWAIT; 959 msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vnvlbv", 0); 960 } 961 VI_UNLOCK(vp); 962 if (VOP_GETVOBJECT(vp, &object) == 0) { 963 VM_OBJECT_LOCK(object); 964 vm_object_pip_wait(object, "vnvlbx"); 965 VM_OBJECT_UNLOCK(object); 966 } 967 VI_LOCK(vp); 968 } while (vp->v_numoutput > 0); 969 VI_UNLOCK(vp); 970 971 /* 972 * Destroy the copy in the VM cache, too. 973 */ 974 if (VOP_GETVOBJECT(vp, &object) == 0) { 975 VM_OBJECT_LOCK(object); 976 vm_object_page_remove(object, 0, 0, 977 (flags & V_SAVE) ? TRUE : FALSE); 978 VM_OBJECT_UNLOCK(object); 979 } 980 981 #ifdef INVARIANTS 982 VI_LOCK(vp); 983 if ((flags & (V_ALT | V_NORMAL)) == 0 && 984 (!TAILQ_EMPTY(&vp->v_dirtyblkhd) || 985 !TAILQ_EMPTY(&vp->v_cleanblkhd))) 986 panic("vinvalbuf: flush failed"); 987 VI_UNLOCK(vp); 988 #endif 989 return (0); 990 } 991 992 /* 993 * Flush out buffers on the specified list. 994 * 995 */ 996 static int 997 flushbuflist(blist, flags, vp, slpflag, slptimeo, errorp) 998 struct buf *blist; 999 int flags; 1000 struct vnode *vp; 1001 int slpflag, slptimeo; 1002 int *errorp; 1003 { 1004 struct buf *bp, *nbp; 1005 int found, error; 1006 1007 ASSERT_VI_LOCKED(vp, "flushbuflist"); 1008 1009 for (found = 0, bp = blist; bp; bp = nbp) { 1010 nbp = TAILQ_NEXT(bp, b_vnbufs); 1011 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1012 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1013 continue; 1014 } 1015 found += 1; 1016 error = BUF_TIMELOCK(bp, 1017 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp), 1018 "flushbuf", slpflag, slptimeo); 1019 if (error) { 1020 if (error != ENOLCK) 1021 *errorp = error; 1022 goto done; 1023 } 1024 /* 1025 * XXX Since there are no node locks for NFS, I 1026 * believe there is a slight chance that a delayed 1027 * write will occur while sleeping just above, so 1028 * check for it. Note that vfs_bio_awrite expects 1029 * buffers to reside on a queue, while BUF_WRITE and 1030 * brelse do not. 1031 */ 1032 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1033 (flags & V_SAVE)) { 1034 1035 if (bp->b_vp == vp) { 1036 if (bp->b_flags & B_CLUSTEROK) { 1037 vfs_bio_awrite(bp); 1038 } else { 1039 bremfree(bp); 1040 bp->b_flags |= B_ASYNC; 1041 BUF_WRITE(bp); 1042 } 1043 } else { 1044 bremfree(bp); 1045 (void) BUF_WRITE(bp); 1046 } 1047 goto done; 1048 } 1049 bremfree(bp); 1050 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 1051 bp->b_flags &= ~B_ASYNC; 1052 brelse(bp); 1053 VI_LOCK(vp); 1054 } 1055 return (found); 1056 done: 1057 VI_LOCK(vp); 1058 return (found); 1059 } 1060 1061 /* 1062 * Truncate a file's buffer and pages to a specified length. This 1063 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1064 * sync activity. 1065 */ 1066 int 1067 vtruncbuf(vp, cred, td, length, blksize) 1068 register struct vnode *vp; 1069 struct ucred *cred; 1070 struct thread *td; 1071 off_t length; 1072 int blksize; 1073 { 1074 register struct buf *bp; 1075 struct buf *nbp; 1076 int anyfreed; 1077 int trunclbn; 1078 1079 /* 1080 * Round up to the *next* lbn. 1081 */ 1082 trunclbn = (length + blksize - 1) / blksize; 1083 1084 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1085 restart: 1086 VI_LOCK(vp); 1087 anyfreed = 1; 1088 for (;anyfreed;) { 1089 anyfreed = 0; 1090 for (bp = TAILQ_FIRST(&vp->v_cleanblkhd); bp; bp = nbp) { 1091 nbp = TAILQ_NEXT(bp, b_vnbufs); 1092 if (bp->b_lblkno >= trunclbn) { 1093 if (BUF_LOCK(bp, 1094 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1095 VI_MTX(vp)) == ENOLCK) 1096 goto restart; 1097 1098 bremfree(bp); 1099 bp->b_flags |= (B_INVAL | B_RELBUF); 1100 bp->b_flags &= ~B_ASYNC; 1101 brelse(bp); 1102 anyfreed = 1; 1103 1104 if (nbp && 1105 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1106 (nbp->b_vp != vp) || 1107 (nbp->b_flags & B_DELWRI))) { 1108 goto restart; 1109 } 1110 VI_LOCK(vp); 1111 } 1112 } 1113 1114 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1115 nbp = TAILQ_NEXT(bp, b_vnbufs); 1116 if (bp->b_lblkno >= trunclbn) { 1117 if (BUF_LOCK(bp, 1118 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1119 VI_MTX(vp)) == ENOLCK) 1120 goto restart; 1121 bremfree(bp); 1122 bp->b_flags |= (B_INVAL | B_RELBUF); 1123 bp->b_flags &= ~B_ASYNC; 1124 brelse(bp); 1125 anyfreed = 1; 1126 if (nbp && 1127 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1128 (nbp->b_vp != vp) || 1129 (nbp->b_flags & B_DELWRI) == 0)) { 1130 goto restart; 1131 } 1132 VI_LOCK(vp); 1133 } 1134 } 1135 } 1136 1137 if (length > 0) { 1138 restartsync: 1139 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1140 nbp = TAILQ_NEXT(bp, b_vnbufs); 1141 if (bp->b_lblkno > 0) 1142 continue; 1143 /* 1144 * Since we hold the vnode lock this should only 1145 * fail if we're racing with the buf daemon. 1146 */ 1147 if (BUF_LOCK(bp, 1148 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1149 VI_MTX(vp)) == ENOLCK) { 1150 goto restart; 1151 } 1152 KASSERT((bp->b_flags & B_DELWRI), 1153 ("buf(%p) on dirty queue without DELWRI", bp)); 1154 1155 bremfree(bp); 1156 bawrite(bp); 1157 VI_LOCK(vp); 1158 goto restartsync; 1159 } 1160 } 1161 1162 while (vp->v_numoutput > 0) { 1163 vp->v_iflag |= VI_BWAIT; 1164 msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vbtrunc", 0); 1165 } 1166 VI_UNLOCK(vp); 1167 vnode_pager_setsize(vp, length); 1168 1169 return (0); 1170 } 1171 1172 /* 1173 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1174 * a vnode. 1175 * 1176 * NOTE: We have to deal with the special case of a background bitmap 1177 * buffer, a situation where two buffers will have the same logical 1178 * block offset. We want (1) only the foreground buffer to be accessed 1179 * in a lookup and (2) must differentiate between the foreground and 1180 * background buffer in the splay tree algorithm because the splay 1181 * tree cannot normally handle multiple entities with the same 'index'. 1182 * We accomplish this by adding differentiating flags to the splay tree's 1183 * numerical domain. 1184 */ 1185 static 1186 struct buf * 1187 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1188 { 1189 struct buf dummy; 1190 struct buf *lefttreemax, *righttreemin, *y; 1191 1192 if (root == NULL) 1193 return (NULL); 1194 lefttreemax = righttreemin = &dummy; 1195 for (;;) { 1196 if (lblkno < root->b_lblkno || 1197 (lblkno == root->b_lblkno && 1198 (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1199 if ((y = root->b_left) == NULL) 1200 break; 1201 if (lblkno < y->b_lblkno) { 1202 /* Rotate right. */ 1203 root->b_left = y->b_right; 1204 y->b_right = root; 1205 root = y; 1206 if ((y = root->b_left) == NULL) 1207 break; 1208 } 1209 /* Link into the new root's right tree. */ 1210 righttreemin->b_left = root; 1211 righttreemin = root; 1212 } else if (lblkno > root->b_lblkno || 1213 (lblkno == root->b_lblkno && 1214 (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { 1215 if ((y = root->b_right) == NULL) 1216 break; 1217 if (lblkno > y->b_lblkno) { 1218 /* Rotate left. */ 1219 root->b_right = y->b_left; 1220 y->b_left = root; 1221 root = y; 1222 if ((y = root->b_right) == NULL) 1223 break; 1224 } 1225 /* Link into the new root's left tree. */ 1226 lefttreemax->b_right = root; 1227 lefttreemax = root; 1228 } else { 1229 break; 1230 } 1231 root = y; 1232 } 1233 /* Assemble the new root. */ 1234 lefttreemax->b_right = root->b_left; 1235 righttreemin->b_left = root->b_right; 1236 root->b_left = dummy.b_right; 1237 root->b_right = dummy.b_left; 1238 return (root); 1239 } 1240 1241 static 1242 void 1243 buf_vlist_remove(struct buf *bp) 1244 { 1245 struct vnode *vp = bp->b_vp; 1246 struct buf *root; 1247 1248 ASSERT_VI_LOCKED(vp, "buf_vlist_remove"); 1249 if (bp->b_xflags & BX_VNDIRTY) { 1250 if (bp != vp->v_dirtyblkroot) { 1251 root = buf_splay(bp->b_lblkno, bp->b_xflags, 1252 vp->v_dirtyblkroot); 1253 KASSERT(root == bp, 1254 ("splay lookup failed during dirty remove")); 1255 } 1256 if (bp->b_left == NULL) { 1257 root = bp->b_right; 1258 } else { 1259 root = buf_splay(bp->b_lblkno, bp->b_xflags, 1260 bp->b_left); 1261 root->b_right = bp->b_right; 1262 } 1263 vp->v_dirtyblkroot = root; 1264 TAILQ_REMOVE(&vp->v_dirtyblkhd, bp, b_vnbufs); 1265 vp->v_dirtybufcnt--; 1266 } else { 1267 /* KASSERT(bp->b_xflags & BX_VNCLEAN, ("bp wasn't clean")); */ 1268 if (bp != vp->v_cleanblkroot) { 1269 root = buf_splay(bp->b_lblkno, bp->b_xflags, 1270 vp->v_cleanblkroot); 1271 KASSERT(root == bp, 1272 ("splay lookup failed during clean remove")); 1273 } 1274 if (bp->b_left == NULL) { 1275 root = bp->b_right; 1276 } else { 1277 root = buf_splay(bp->b_lblkno, bp->b_xflags, 1278 bp->b_left); 1279 root->b_right = bp->b_right; 1280 } 1281 vp->v_cleanblkroot = root; 1282 TAILQ_REMOVE(&vp->v_cleanblkhd, bp, b_vnbufs); 1283 vp->v_cleanbufcnt--; 1284 } 1285 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1286 } 1287 1288 /* 1289 * Add the buffer to the sorted clean or dirty block list using a 1290 * splay tree algorithm. 1291 * 1292 * NOTE: xflags is passed as a constant, optimizing this inline function! 1293 */ 1294 static 1295 void 1296 buf_vlist_add(struct buf *bp, struct vnode *vp, b_xflags_t xflags) 1297 { 1298 struct buf *root; 1299 1300 ASSERT_VI_LOCKED(vp, "buf_vlist_add"); 1301 bp->b_xflags |= xflags; 1302 if (xflags & BX_VNDIRTY) { 1303 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot); 1304 if (root == NULL) { 1305 bp->b_left = NULL; 1306 bp->b_right = NULL; 1307 TAILQ_INSERT_TAIL(&vp->v_dirtyblkhd, bp, b_vnbufs); 1308 } else if (bp->b_lblkno < root->b_lblkno || 1309 (bp->b_lblkno == root->b_lblkno && 1310 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1311 bp->b_left = root->b_left; 1312 bp->b_right = root; 1313 root->b_left = NULL; 1314 TAILQ_INSERT_BEFORE(root, bp, b_vnbufs); 1315 } else { 1316 bp->b_right = root->b_right; 1317 bp->b_left = root; 1318 root->b_right = NULL; 1319 TAILQ_INSERT_AFTER(&vp->v_dirtyblkhd, 1320 root, bp, b_vnbufs); 1321 } 1322 vp->v_dirtybufcnt++; 1323 vp->v_dirtyblkroot = bp; 1324 } else { 1325 /* KASSERT(xflags & BX_VNCLEAN, ("xflags not clean")); */ 1326 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot); 1327 if (root == NULL) { 1328 bp->b_left = NULL; 1329 bp->b_right = NULL; 1330 TAILQ_INSERT_TAIL(&vp->v_cleanblkhd, bp, b_vnbufs); 1331 } else if (bp->b_lblkno < root->b_lblkno || 1332 (bp->b_lblkno == root->b_lblkno && 1333 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1334 bp->b_left = root->b_left; 1335 bp->b_right = root; 1336 root->b_left = NULL; 1337 TAILQ_INSERT_BEFORE(root, bp, b_vnbufs); 1338 } else { 1339 bp->b_right = root->b_right; 1340 bp->b_left = root; 1341 root->b_right = NULL; 1342 TAILQ_INSERT_AFTER(&vp->v_cleanblkhd, 1343 root, bp, b_vnbufs); 1344 } 1345 vp->v_cleanbufcnt++; 1346 vp->v_cleanblkroot = bp; 1347 } 1348 } 1349 1350 /* 1351 * Lookup a buffer using the splay tree. Note that we specifically avoid 1352 * shadow buffers used in background bitmap writes. 1353 * 1354 * This code isn't quite efficient as it could be because we are maintaining 1355 * two sorted lists and do not know which list the block resides in. 1356 * 1357 * During a "make buildworld" the desired buffer is found at one of 1358 * the roots more than 60% of the time. Thus, checking both roots 1359 * before performing either splay eliminates unnecessary splays on the 1360 * first tree splayed. 1361 */ 1362 struct buf * 1363 gbincore(struct vnode *vp, daddr_t lblkno) 1364 { 1365 struct buf *bp; 1366 1367 GIANT_REQUIRED; 1368 1369 ASSERT_VI_LOCKED(vp, "gbincore"); 1370 if ((bp = vp->v_cleanblkroot) != NULL && 1371 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1372 return (bp); 1373 if ((bp = vp->v_dirtyblkroot) != NULL && 1374 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1375 return (bp); 1376 if ((bp = vp->v_cleanblkroot) != NULL) { 1377 vp->v_cleanblkroot = bp = buf_splay(lblkno, 0, bp); 1378 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1379 return (bp); 1380 } 1381 if ((bp = vp->v_dirtyblkroot) != NULL) { 1382 vp->v_dirtyblkroot = bp = buf_splay(lblkno, 0, bp); 1383 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1384 return (bp); 1385 } 1386 return (NULL); 1387 } 1388 1389 /* 1390 * Associate a buffer with a vnode. 1391 */ 1392 void 1393 bgetvp(vp, bp) 1394 register struct vnode *vp; 1395 register struct buf *bp; 1396 { 1397 1398 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 1399 1400 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1401 ("bgetvp: bp already attached! %p", bp)); 1402 1403 ASSERT_VI_LOCKED(vp, "bgetvp"); 1404 vholdl(vp); 1405 bp->b_vp = vp; 1406 bp->b_dev = vn_todev(vp); 1407 /* 1408 * Insert onto list for new vnode. 1409 */ 1410 buf_vlist_add(bp, vp, BX_VNCLEAN); 1411 } 1412 1413 /* 1414 * Disassociate a buffer from a vnode. 1415 */ 1416 void 1417 brelvp(bp) 1418 register struct buf *bp; 1419 { 1420 struct vnode *vp; 1421 1422 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1423 1424 /* 1425 * Delete from old vnode list, if on one. 1426 */ 1427 vp = bp->b_vp; 1428 VI_LOCK(vp); 1429 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1430 buf_vlist_remove(bp); 1431 if ((vp->v_iflag & VI_ONWORKLST) && TAILQ_EMPTY(&vp->v_dirtyblkhd)) { 1432 vp->v_iflag &= ~VI_ONWORKLST; 1433 mtx_lock(&sync_mtx); 1434 LIST_REMOVE(vp, v_synclist); 1435 mtx_unlock(&sync_mtx); 1436 } 1437 vdropl(vp); 1438 bp->b_vp = (struct vnode *) 0; 1439 if (bp->b_object) 1440 bp->b_object = NULL; 1441 VI_UNLOCK(vp); 1442 } 1443 1444 /* 1445 * Add an item to the syncer work queue. 1446 */ 1447 static void 1448 vn_syncer_add_to_worklist(struct vnode *vp, int delay) 1449 { 1450 int slot; 1451 1452 ASSERT_VI_LOCKED(vp, "vn_syncer_add_to_worklist"); 1453 1454 mtx_lock(&sync_mtx); 1455 if (vp->v_iflag & VI_ONWORKLST) 1456 LIST_REMOVE(vp, v_synclist); 1457 else 1458 vp->v_iflag |= VI_ONWORKLST; 1459 1460 if (delay > syncer_maxdelay - 2) 1461 delay = syncer_maxdelay - 2; 1462 slot = (syncer_delayno + delay) & syncer_mask; 1463 1464 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist); 1465 mtx_unlock(&sync_mtx); 1466 } 1467 1468 struct proc *updateproc; 1469 static void sched_sync(void); 1470 static struct kproc_desc up_kp = { 1471 "syncer", 1472 sched_sync, 1473 &updateproc 1474 }; 1475 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 1476 1477 /* 1478 * System filesystem synchronizer daemon. 1479 */ 1480 static void 1481 sched_sync(void) 1482 { 1483 struct synclist *next; 1484 struct synclist *slp; 1485 struct vnode *vp; 1486 struct mount *mp; 1487 long starttime; 1488 struct thread *td = FIRST_THREAD_IN_PROC(updateproc); /* XXXKSE */ 1489 1490 mtx_lock(&Giant); 1491 1492 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, td->td_proc, 1493 SHUTDOWN_PRI_LAST); 1494 1495 for (;;) { 1496 kthread_suspend_check(td->td_proc); 1497 1498 starttime = time_second; 1499 1500 /* 1501 * Push files whose dirty time has expired. Be careful 1502 * of interrupt race on slp queue. 1503 */ 1504 mtx_lock(&sync_mtx); 1505 slp = &syncer_workitem_pending[syncer_delayno]; 1506 syncer_delayno += 1; 1507 if (syncer_delayno == syncer_maxdelay) 1508 syncer_delayno = 0; 1509 next = &syncer_workitem_pending[syncer_delayno]; 1510 1511 while ((vp = LIST_FIRST(slp)) != NULL) { 1512 if (VOP_ISLOCKED(vp, NULL) != 0 || 1513 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1514 LIST_REMOVE(vp, v_synclist); 1515 LIST_INSERT_HEAD(next, vp, v_synclist); 1516 continue; 1517 } 1518 if (VI_TRYLOCK(vp) == 0) { 1519 LIST_REMOVE(vp, v_synclist); 1520 LIST_INSERT_HEAD(next, vp, v_synclist); 1521 vn_finished_write(mp); 1522 continue; 1523 } 1524 /* 1525 * We use vhold in case the vnode does not 1526 * successfully sync. vhold prevents the vnode from 1527 * going away when we unlock the sync_mtx so that 1528 * we can acquire the vnode interlock. 1529 */ 1530 vholdl(vp); 1531 mtx_unlock(&sync_mtx); 1532 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, td); 1533 (void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td); 1534 VOP_UNLOCK(vp, 0, td); 1535 vn_finished_write(mp); 1536 VI_LOCK(vp); 1537 if ((vp->v_iflag & VI_ONWORKLST) != 0) { 1538 /* 1539 * Put us back on the worklist. The worklist 1540 * routine will remove us from our current 1541 * position and then add us back in at a later 1542 * position. 1543 */ 1544 vn_syncer_add_to_worklist(vp, syncdelay); 1545 } 1546 vdropl(vp); 1547 VI_UNLOCK(vp); 1548 mtx_lock(&sync_mtx); 1549 } 1550 mtx_unlock(&sync_mtx); 1551 1552 /* 1553 * Do soft update processing. 1554 */ 1555 if (softdep_process_worklist_hook != NULL) 1556 (*softdep_process_worklist_hook)(NULL); 1557 1558 /* 1559 * The variable rushjob allows the kernel to speed up the 1560 * processing of the filesystem syncer process. A rushjob 1561 * value of N tells the filesystem syncer to process the next 1562 * N seconds worth of work on its queue ASAP. Currently rushjob 1563 * is used by the soft update code to speed up the filesystem 1564 * syncer process when the incore state is getting so far 1565 * ahead of the disk that the kernel memory pool is being 1566 * threatened with exhaustion. 1567 */ 1568 mtx_lock(&sync_mtx); 1569 if (rushjob > 0) { 1570 rushjob -= 1; 1571 mtx_unlock(&sync_mtx); 1572 continue; 1573 } 1574 mtx_unlock(&sync_mtx); 1575 /* 1576 * If it has taken us less than a second to process the 1577 * current work, then wait. Otherwise start right over 1578 * again. We can still lose time if any single round 1579 * takes more than two seconds, but it does not really 1580 * matter as we are just trying to generally pace the 1581 * filesystem activity. 1582 */ 1583 if (time_second == starttime) 1584 tsleep(&lbolt, PPAUSE, "syncer", 0); 1585 } 1586 } 1587 1588 /* 1589 * Request the syncer daemon to speed up its work. 1590 * We never push it to speed up more than half of its 1591 * normal turn time, otherwise it could take over the cpu. 1592 * XXXKSE only one update? 1593 */ 1594 int 1595 speedup_syncer() 1596 { 1597 struct thread *td; 1598 int ret = 0; 1599 1600 td = FIRST_THREAD_IN_PROC(updateproc); 1601 mtx_lock_spin(&sched_lock); 1602 if (td->td_wchan == &lbolt) { 1603 unsleep(td); 1604 TD_CLR_SLEEPING(td); 1605 setrunnable(td); 1606 } 1607 mtx_unlock_spin(&sched_lock); 1608 mtx_lock(&sync_mtx); 1609 if (rushjob < syncdelay / 2) { 1610 rushjob += 1; 1611 stat_rush_requests += 1; 1612 ret = 1; 1613 } 1614 mtx_unlock(&sync_mtx); 1615 return (ret); 1616 } 1617 1618 /* 1619 * Associate a p-buffer with a vnode. 1620 * 1621 * Also sets B_PAGING flag to indicate that vnode is not fully associated 1622 * with the buffer. i.e. the bp has not been linked into the vnode or 1623 * ref-counted. 1624 */ 1625 void 1626 pbgetvp(vp, bp) 1627 register struct vnode *vp; 1628 register struct buf *bp; 1629 { 1630 1631 KASSERT(bp->b_vp == NULL, ("pbgetvp: not free")); 1632 1633 bp->b_vp = vp; 1634 bp->b_object = vp->v_object; 1635 bp->b_flags |= B_PAGING; 1636 bp->b_dev = vn_todev(vp); 1637 } 1638 1639 /* 1640 * Disassociate a p-buffer from a vnode. 1641 */ 1642 void 1643 pbrelvp(bp) 1644 register struct buf *bp; 1645 { 1646 1647 KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL")); 1648 1649 /* XXX REMOVE ME */ 1650 VI_LOCK(bp->b_vp); 1651 if (TAILQ_NEXT(bp, b_vnbufs) != NULL) { 1652 panic( 1653 "relpbuf(): b_vp was probably reassignbuf()d %p %x", 1654 bp, 1655 (int)bp->b_flags 1656 ); 1657 } 1658 VI_UNLOCK(bp->b_vp); 1659 bp->b_vp = (struct vnode *) 0; 1660 bp->b_object = NULL; 1661 bp->b_flags &= ~B_PAGING; 1662 } 1663 1664 /* 1665 * Reassign a buffer from one vnode to another. 1666 * Used to assign file specific control information 1667 * (indirect blocks) to the vnode to which they belong. 1668 */ 1669 void 1670 reassignbuf(bp, newvp) 1671 register struct buf *bp; 1672 register struct vnode *newvp; 1673 { 1674 struct vnode *vp; 1675 int delay; 1676 1677 if (newvp == NULL) { 1678 printf("reassignbuf: NULL"); 1679 return; 1680 } 1681 vp = bp->b_vp; 1682 ++reassignbufcalls; 1683 1684 /* 1685 * B_PAGING flagged buffers cannot be reassigned because their vp 1686 * is not fully linked in. 1687 */ 1688 if (bp->b_flags & B_PAGING) 1689 panic("cannot reassign paging buffer"); 1690 1691 /* 1692 * Delete from old vnode list, if on one. 1693 */ 1694 VI_LOCK(vp); 1695 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) { 1696 buf_vlist_remove(bp); 1697 if (vp != newvp) { 1698 vdropl(bp->b_vp); 1699 bp->b_vp = NULL; /* for clarification */ 1700 } 1701 } 1702 if (vp != newvp) { 1703 VI_UNLOCK(vp); 1704 VI_LOCK(newvp); 1705 } 1706 /* 1707 * If dirty, put on list of dirty buffers; otherwise insert onto list 1708 * of clean buffers. 1709 */ 1710 if (bp->b_flags & B_DELWRI) { 1711 if ((newvp->v_iflag & VI_ONWORKLST) == 0) { 1712 switch (newvp->v_type) { 1713 case VDIR: 1714 delay = dirdelay; 1715 break; 1716 case VCHR: 1717 if (newvp->v_rdev->si_mountpoint != NULL) { 1718 delay = metadelay; 1719 break; 1720 } 1721 /* FALLTHROUGH */ 1722 default: 1723 delay = filedelay; 1724 } 1725 vn_syncer_add_to_worklist(newvp, delay); 1726 } 1727 buf_vlist_add(bp, newvp, BX_VNDIRTY); 1728 } else { 1729 buf_vlist_add(bp, newvp, BX_VNCLEAN); 1730 1731 if ((newvp->v_iflag & VI_ONWORKLST) && 1732 TAILQ_EMPTY(&newvp->v_dirtyblkhd)) { 1733 mtx_lock(&sync_mtx); 1734 LIST_REMOVE(newvp, v_synclist); 1735 mtx_unlock(&sync_mtx); 1736 newvp->v_iflag &= ~VI_ONWORKLST; 1737 } 1738 } 1739 if (bp->b_vp != newvp) { 1740 bp->b_vp = newvp; 1741 vholdl(bp->b_vp); 1742 } 1743 VI_UNLOCK(newvp); 1744 } 1745 1746 /* 1747 * Create a vnode for a device. 1748 * Used for mounting the root filesystem. 1749 */ 1750 int 1751 bdevvp(dev, vpp) 1752 dev_t dev; 1753 struct vnode **vpp; 1754 { 1755 register struct vnode *vp; 1756 struct vnode *nvp; 1757 int error; 1758 1759 if (dev == NODEV) { 1760 *vpp = NULLVP; 1761 return (ENXIO); 1762 } 1763 if (vfinddev(dev, VCHR, vpp)) 1764 return (0); 1765 error = getnewvnode("none", (struct mount *)0, spec_vnodeop_p, &nvp); 1766 if (error) { 1767 *vpp = NULLVP; 1768 return (error); 1769 } 1770 vp = nvp; 1771 vp->v_type = VCHR; 1772 addalias(vp, dev); 1773 *vpp = vp; 1774 return (0); 1775 } 1776 1777 static void 1778 v_incr_usecount(struct vnode *vp, int delta) 1779 { 1780 1781 vp->v_usecount += delta; 1782 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1783 mtx_lock(&spechash_mtx); 1784 vp->v_rdev->si_usecount += delta; 1785 mtx_unlock(&spechash_mtx); 1786 } 1787 } 1788 1789 /* 1790 * Add vnode to the alias list hung off the dev_t. 1791 * 1792 * The reason for this gunk is that multiple vnodes can reference 1793 * the same physical device, so checking vp->v_usecount to see 1794 * how many users there are is inadequate; the v_usecount for 1795 * the vnodes need to be accumulated. vcount() does that. 1796 */ 1797 struct vnode * 1798 addaliasu(nvp, nvp_rdev) 1799 struct vnode *nvp; 1800 udev_t nvp_rdev; 1801 { 1802 struct vnode *ovp; 1803 vop_t **ops; 1804 dev_t dev; 1805 1806 if (nvp->v_type == VBLK) 1807 return (nvp); 1808 if (nvp->v_type != VCHR) 1809 panic("addaliasu on non-special vnode"); 1810 dev = udev2dev(nvp_rdev, 0); 1811 /* 1812 * Check to see if we have a bdevvp vnode with no associated 1813 * filesystem. If so, we want to associate the filesystem of 1814 * the new newly instigated vnode with the bdevvp vnode and 1815 * discard the newly created vnode rather than leaving the 1816 * bdevvp vnode lying around with no associated filesystem. 1817 */ 1818 if (vfinddev(dev, nvp->v_type, &ovp) == 0 || ovp->v_data != NULL) { 1819 addalias(nvp, dev); 1820 return (nvp); 1821 } 1822 /* 1823 * Discard unneeded vnode, but save its node specific data. 1824 * Note that if there is a lock, it is carried over in the 1825 * node specific data to the replacement vnode. 1826 */ 1827 vref(ovp); 1828 ovp->v_data = nvp->v_data; 1829 ovp->v_tag = nvp->v_tag; 1830 nvp->v_data = NULL; 1831 lockdestroy(ovp->v_vnlock); 1832 lockinit(ovp->v_vnlock, PVFS, nvp->v_vnlock->lk_wmesg, 1833 nvp->v_vnlock->lk_timo, nvp->v_vnlock->lk_flags & LK_EXTFLG_MASK); 1834 ops = ovp->v_op; 1835 ovp->v_op = nvp->v_op; 1836 if (VOP_ISLOCKED(nvp, curthread)) { 1837 VOP_UNLOCK(nvp, 0, curthread); 1838 vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread); 1839 } 1840 nvp->v_op = ops; 1841 insmntque(ovp, nvp->v_mount); 1842 vrele(nvp); 1843 vgone(nvp); 1844 return (ovp); 1845 } 1846 1847 /* This is a local helper function that do the same as addaliasu, but for a 1848 * dev_t instead of an udev_t. */ 1849 static void 1850 addalias(nvp, dev) 1851 struct vnode *nvp; 1852 dev_t dev; 1853 { 1854 1855 KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode")); 1856 nvp->v_rdev = dev; 1857 VI_LOCK(nvp); 1858 mtx_lock(&spechash_mtx); 1859 SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext); 1860 dev->si_usecount += nvp->v_usecount; 1861 mtx_unlock(&spechash_mtx); 1862 VI_UNLOCK(nvp); 1863 } 1864 1865 /* 1866 * Grab a particular vnode from the free list, increment its 1867 * reference count and lock it. The vnode lock bit is set if the 1868 * vnode is being eliminated in vgone. The process is awakened 1869 * when the transition is completed, and an error returned to 1870 * indicate that the vnode is no longer usable (possibly having 1871 * been changed to a new filesystem type). 1872 */ 1873 int 1874 vget(vp, flags, td) 1875 register struct vnode *vp; 1876 int flags; 1877 struct thread *td; 1878 { 1879 int error; 1880 1881 /* 1882 * If the vnode is in the process of being cleaned out for 1883 * another use, we wait for the cleaning to finish and then 1884 * return failure. Cleaning is determined by checking that 1885 * the VI_XLOCK flag is set. 1886 */ 1887 if ((flags & LK_INTERLOCK) == 0) 1888 VI_LOCK(vp); 1889 if (vp->v_iflag & VI_XLOCK && vp->v_vxthread != curthread) { 1890 if ((flags & LK_NOWAIT) == 0) { 1891 vp->v_iflag |= VI_XWANT; 1892 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0); 1893 return (ENOENT); 1894 } 1895 VI_UNLOCK(vp); 1896 return (EBUSY); 1897 } 1898 1899 v_incr_usecount(vp, 1); 1900 1901 if (VSHOULDBUSY(vp)) 1902 vbusy(vp); 1903 if (flags & LK_TYPE_MASK) { 1904 if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { 1905 /* 1906 * must expand vrele here because we do not want 1907 * to call VOP_INACTIVE if the reference count 1908 * drops back to zero since it was never really 1909 * active. We must remove it from the free list 1910 * before sleeping so that multiple processes do 1911 * not try to recycle it. 1912 */ 1913 VI_LOCK(vp); 1914 v_incr_usecount(vp, -1); 1915 if (VSHOULDFREE(vp)) 1916 vfree(vp); 1917 else 1918 vlruvp(vp); 1919 VI_UNLOCK(vp); 1920 } 1921 return (error); 1922 } 1923 VI_UNLOCK(vp); 1924 return (0); 1925 } 1926 1927 /* 1928 * Increase the reference count of a vnode. 1929 */ 1930 void 1931 vref(struct vnode *vp) 1932 { 1933 1934 VI_LOCK(vp); 1935 v_incr_usecount(vp, 1); 1936 VI_UNLOCK(vp); 1937 } 1938 1939 /* 1940 * Return reference count of a vnode. 1941 * 1942 * The results of this call are only guaranteed when some mechanism other 1943 * than the VI lock is used to stop other processes from gaining references 1944 * to the vnode. This may be the case if the caller holds the only reference. 1945 * This is also useful when stale data is acceptable as race conditions may 1946 * be accounted for by some other means. 1947 */ 1948 int 1949 vrefcnt(struct vnode *vp) 1950 { 1951 int usecnt; 1952 1953 VI_LOCK(vp); 1954 usecnt = vp->v_usecount; 1955 VI_UNLOCK(vp); 1956 1957 return (usecnt); 1958 } 1959 1960 1961 /* 1962 * Vnode put/release. 1963 * If count drops to zero, call inactive routine and return to freelist. 1964 */ 1965 void 1966 vrele(vp) 1967 struct vnode *vp; 1968 { 1969 struct thread *td = curthread; /* XXX */ 1970 1971 KASSERT(vp != NULL, ("vrele: null vp")); 1972 1973 VI_LOCK(vp); 1974 1975 /* Skip this v_writecount check if we're going to panic below. */ 1976 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 1977 ("vrele: missed vn_close")); 1978 1979 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 1980 vp->v_usecount == 1)) { 1981 v_incr_usecount(vp, -1); 1982 VI_UNLOCK(vp); 1983 1984 return; 1985 } 1986 1987 if (vp->v_usecount == 1) { 1988 v_incr_usecount(vp, -1); 1989 /* 1990 * We must call VOP_INACTIVE with the node locked. Mark 1991 * as VI_DOINGINACT to avoid recursion. 1992 */ 1993 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) { 1994 VI_LOCK(vp); 1995 vp->v_iflag |= VI_DOINGINACT; 1996 VI_UNLOCK(vp); 1997 VOP_INACTIVE(vp, td); 1998 VI_LOCK(vp); 1999 KASSERT(vp->v_iflag & VI_DOINGINACT, 2000 ("vrele: lost VI_DOINGINACT")); 2001 vp->v_iflag &= ~VI_DOINGINACT; 2002 } else 2003 VI_LOCK(vp); 2004 if (VSHOULDFREE(vp)) 2005 vfree(vp); 2006 else 2007 vlruvp(vp); 2008 VI_UNLOCK(vp); 2009 2010 } else { 2011 #ifdef DIAGNOSTIC 2012 vprint("vrele: negative ref count", vp); 2013 #endif 2014 VI_UNLOCK(vp); 2015 panic("vrele: negative ref cnt"); 2016 } 2017 } 2018 2019 /* 2020 * Release an already locked vnode. This give the same effects as 2021 * unlock+vrele(), but takes less time and avoids releasing and 2022 * re-aquiring the lock (as vrele() aquires the lock internally.) 2023 */ 2024 void 2025 vput(vp) 2026 struct vnode *vp; 2027 { 2028 struct thread *td = curthread; /* XXX */ 2029 2030 GIANT_REQUIRED; 2031 2032 KASSERT(vp != NULL, ("vput: null vp")); 2033 VI_LOCK(vp); 2034 /* Skip this v_writecount check if we're going to panic below. */ 2035 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2036 ("vput: missed vn_close")); 2037 2038 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2039 vp->v_usecount == 1)) { 2040 v_incr_usecount(vp, -1); 2041 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2042 return; 2043 } 2044 2045 if (vp->v_usecount == 1) { 2046 v_incr_usecount(vp, -1); 2047 /* 2048 * We must call VOP_INACTIVE with the node locked, so 2049 * we just need to release the vnode mutex. Mark as 2050 * as VI_DOINGINACT to avoid recursion. 2051 */ 2052 vp->v_iflag |= VI_DOINGINACT; 2053 VI_UNLOCK(vp); 2054 VOP_INACTIVE(vp, td); 2055 VI_LOCK(vp); 2056 KASSERT(vp->v_iflag & VI_DOINGINACT, 2057 ("vput: lost VI_DOINGINACT")); 2058 vp->v_iflag &= ~VI_DOINGINACT; 2059 if (VSHOULDFREE(vp)) 2060 vfree(vp); 2061 else 2062 vlruvp(vp); 2063 VI_UNLOCK(vp); 2064 2065 } else { 2066 #ifdef DIAGNOSTIC 2067 vprint("vput: negative ref count", vp); 2068 #endif 2069 panic("vput: negative ref cnt"); 2070 } 2071 } 2072 2073 /* 2074 * Somebody doesn't want the vnode recycled. 2075 */ 2076 void 2077 vhold(struct vnode *vp) 2078 { 2079 2080 VI_LOCK(vp); 2081 vholdl(vp); 2082 VI_UNLOCK(vp); 2083 } 2084 2085 void 2086 vholdl(vp) 2087 register struct vnode *vp; 2088 { 2089 2090 vp->v_holdcnt++; 2091 if (VSHOULDBUSY(vp)) 2092 vbusy(vp); 2093 } 2094 2095 /* 2096 * Note that there is one less who cares about this vnode. vdrop() is the 2097 * opposite of vhold(). 2098 */ 2099 void 2100 vdrop(struct vnode *vp) 2101 { 2102 2103 VI_LOCK(vp); 2104 vdropl(vp); 2105 VI_UNLOCK(vp); 2106 } 2107 2108 void 2109 vdropl(vp) 2110 register struct vnode *vp; 2111 { 2112 2113 if (vp->v_holdcnt <= 0) 2114 panic("vdrop: holdcnt"); 2115 vp->v_holdcnt--; 2116 if (VSHOULDFREE(vp)) 2117 vfree(vp); 2118 else 2119 vlruvp(vp); 2120 } 2121 2122 /* 2123 * Remove any vnodes in the vnode table belonging to mount point mp. 2124 * 2125 * If FORCECLOSE is not specified, there should not be any active ones, 2126 * return error if any are found (nb: this is a user error, not a 2127 * system error). If FORCECLOSE is specified, detach any active vnodes 2128 * that are found. 2129 * 2130 * If WRITECLOSE is set, only flush out regular file vnodes open for 2131 * writing. 2132 * 2133 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2134 * 2135 * `rootrefs' specifies the base reference count for the root vnode 2136 * of this filesystem. The root vnode is considered busy if its 2137 * v_usecount exceeds this value. On a successful return, vflush() 2138 * will call vrele() on the root vnode exactly rootrefs times. 2139 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2140 * be zero. 2141 */ 2142 #ifdef DIAGNOSTIC 2143 static int busyprt = 0; /* print out busy vnodes */ 2144 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 2145 #endif 2146 2147 int 2148 vflush(mp, rootrefs, flags) 2149 struct mount *mp; 2150 int rootrefs; 2151 int flags; 2152 { 2153 struct thread *td = curthread; /* XXX */ 2154 struct vnode *vp, *nvp, *rootvp = NULL; 2155 struct vattr vattr; 2156 int busy = 0, error; 2157 2158 if (rootrefs > 0) { 2159 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2160 ("vflush: bad args")); 2161 /* 2162 * Get the filesystem root vnode. We can vput() it 2163 * immediately, since with rootrefs > 0, it won't go away. 2164 */ 2165 if ((error = VFS_ROOT(mp, &rootvp)) != 0) 2166 return (error); 2167 vput(rootvp); 2168 2169 } 2170 MNT_ILOCK(mp); 2171 loop: 2172 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp; vp = nvp) { 2173 /* 2174 * Make sure this vnode wasn't reclaimed in getnewvnode(). 2175 * Start over if it has (it won't be on the list anymore). 2176 */ 2177 if (vp->v_mount != mp) 2178 goto loop; 2179 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 2180 2181 VI_LOCK(vp); 2182 MNT_IUNLOCK(mp); 2183 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td); 2184 if (error) { 2185 MNT_ILOCK(mp); 2186 goto loop; 2187 } 2188 /* 2189 * Skip over a vnodes marked VV_SYSTEM. 2190 */ 2191 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2192 VOP_UNLOCK(vp, 0, td); 2193 MNT_ILOCK(mp); 2194 continue; 2195 } 2196 /* 2197 * If WRITECLOSE is set, flush out unlinked but still open 2198 * files (even if open only for reading) and regular file 2199 * vnodes open for writing. 2200 */ 2201 if (flags & WRITECLOSE) { 2202 error = VOP_GETATTR(vp, &vattr, td->td_ucred, td); 2203 VI_LOCK(vp); 2204 2205 if ((vp->v_type == VNON || 2206 (error == 0 && vattr.va_nlink > 0)) && 2207 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2208 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2209 MNT_ILOCK(mp); 2210 continue; 2211 } 2212 } else 2213 VI_LOCK(vp); 2214 2215 VOP_UNLOCK(vp, 0, td); 2216 2217 /* 2218 * With v_usecount == 0, all we need to do is clear out the 2219 * vnode data structures and we are done. 2220 */ 2221 if (vp->v_usecount == 0) { 2222 vgonel(vp, td); 2223 MNT_ILOCK(mp); 2224 continue; 2225 } 2226 2227 /* 2228 * If FORCECLOSE is set, forcibly close the vnode. For block 2229 * or character devices, revert to an anonymous device. For 2230 * all other files, just kill them. 2231 */ 2232 if (flags & FORCECLOSE) { 2233 if (vp->v_type != VCHR) 2234 vgonel(vp, td); 2235 else 2236 vgonechrl(vp, td); 2237 MNT_ILOCK(mp); 2238 continue; 2239 } 2240 #ifdef DIAGNOSTIC 2241 if (busyprt) 2242 vprint("vflush: busy vnode", vp); 2243 #endif 2244 VI_UNLOCK(vp); 2245 MNT_ILOCK(mp); 2246 busy++; 2247 } 2248 MNT_IUNLOCK(mp); 2249 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2250 /* 2251 * If just the root vnode is busy, and if its refcount 2252 * is equal to `rootrefs', then go ahead and kill it. 2253 */ 2254 VI_LOCK(rootvp); 2255 KASSERT(busy > 0, ("vflush: not busy")); 2256 KASSERT(rootvp->v_usecount >= rootrefs, ("vflush: rootrefs")); 2257 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2258 vgonel(rootvp, td); 2259 busy = 0; 2260 } else 2261 VI_UNLOCK(rootvp); 2262 } 2263 if (busy) 2264 return (EBUSY); 2265 for (; rootrefs > 0; rootrefs--) 2266 vrele(rootvp); 2267 return (0); 2268 } 2269 2270 /* 2271 * This moves a now (likely recyclable) vnode to the end of the 2272 * mountlist. XXX However, it is temporarily disabled until we 2273 * can clean up ffs_sync() and friends, which have loop restart 2274 * conditions which this code causes to operate O(N^2). 2275 */ 2276 static void 2277 vlruvp(struct vnode *vp) 2278 { 2279 #if 0 2280 struct mount *mp; 2281 2282 if ((mp = vp->v_mount) != NULL) { 2283 MNT_ILOCK(mp); 2284 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2285 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2286 MNT_IUNLOCK(mp); 2287 } 2288 #endif 2289 } 2290 2291 static void 2292 vx_lock(struct vnode *vp) 2293 { 2294 2295 ASSERT_VI_LOCKED(vp, "vx_lock"); 2296 2297 /* 2298 * Prevent the vnode from being recycled or brought into use while we 2299 * clean it out. 2300 */ 2301 if (vp->v_iflag & VI_XLOCK) 2302 panic("vclean: deadlock"); 2303 vp->v_iflag |= VI_XLOCK; 2304 vp->v_vxthread = curthread; 2305 } 2306 2307 static void 2308 vx_unlock(struct vnode *vp) 2309 { 2310 ASSERT_VI_LOCKED(vp, "vx_unlock"); 2311 vp->v_iflag &= ~VI_XLOCK; 2312 vp->v_vxthread = NULL; 2313 if (vp->v_iflag & VI_XWANT) { 2314 vp->v_iflag &= ~VI_XWANT; 2315 wakeup(vp); 2316 } 2317 } 2318 2319 /* 2320 * Disassociate the underlying filesystem from a vnode. 2321 */ 2322 static void 2323 vclean(vp, flags, td) 2324 struct vnode *vp; 2325 int flags; 2326 struct thread *td; 2327 { 2328 int active; 2329 2330 ASSERT_VI_LOCKED(vp, "vclean"); 2331 /* 2332 * Check to see if the vnode is in use. If so we have to reference it 2333 * before we clean it out so that its count cannot fall to zero and 2334 * generate a race against ourselves to recycle it. 2335 */ 2336 if ((active = vp->v_usecount)) 2337 v_incr_usecount(vp, 1); 2338 2339 /* 2340 * Even if the count is zero, the VOP_INACTIVE routine may still 2341 * have the object locked while it cleans it out. The VOP_LOCK 2342 * ensures that the VOP_INACTIVE routine is done with its work. 2343 * For active vnodes, it ensures that no other activity can 2344 * occur while the underlying object is being cleaned out. 2345 */ 2346 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2347 2348 /* 2349 * Clean out any buffers associated with the vnode. 2350 * If the flush fails, just toss the buffers. 2351 */ 2352 if (flags & DOCLOSE) { 2353 struct buf *bp; 2354 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 2355 if (bp != NULL) 2356 (void) vn_write_suspend_wait(vp, NULL, V_WAIT); 2357 if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0) 2358 vinvalbuf(vp, 0, NOCRED, td, 0, 0); 2359 } 2360 2361 VOP_DESTROYVOBJECT(vp); 2362 2363 /* 2364 * Any other processes trying to obtain this lock must first 2365 * wait for VXLOCK to clear, then call the new lock operation. 2366 */ 2367 VOP_UNLOCK(vp, 0, td); 2368 2369 /* 2370 * If purging an active vnode, it must be closed and 2371 * deactivated before being reclaimed. Note that the 2372 * VOP_INACTIVE will unlock the vnode. 2373 */ 2374 if (active) { 2375 if (flags & DOCLOSE) 2376 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2377 VI_LOCK(vp); 2378 if ((vp->v_iflag & VI_DOINGINACT) == 0) { 2379 vp->v_iflag |= VI_DOINGINACT; 2380 VI_UNLOCK(vp); 2381 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 2382 panic("vclean: cannot relock."); 2383 VOP_INACTIVE(vp, td); 2384 VI_LOCK(vp); 2385 KASSERT(vp->v_iflag & VI_DOINGINACT, 2386 ("vclean: lost VI_DOINGINACT")); 2387 vp->v_iflag &= ~VI_DOINGINACT; 2388 } 2389 VI_UNLOCK(vp); 2390 } 2391 /* 2392 * Reclaim the vnode. 2393 */ 2394 if (VOP_RECLAIM(vp, td)) 2395 panic("vclean: cannot reclaim"); 2396 2397 if (active) { 2398 /* 2399 * Inline copy of vrele() since VOP_INACTIVE 2400 * has already been called. 2401 */ 2402 VI_LOCK(vp); 2403 v_incr_usecount(vp, -1); 2404 if (vp->v_usecount <= 0) { 2405 #ifdef INVARIANTS 2406 if (vp->v_usecount < 0 || vp->v_writecount != 0) { 2407 vprint("vclean: bad ref count", vp); 2408 panic("vclean: ref cnt"); 2409 } 2410 #endif 2411 if (VSHOULDFREE(vp)) 2412 vfree(vp); 2413 } 2414 VI_UNLOCK(vp); 2415 } 2416 /* 2417 * Delete from old mount point vnode list. 2418 */ 2419 if (vp->v_mount != NULL) 2420 insmntque(vp, (struct mount *)0); 2421 cache_purge(vp); 2422 VI_LOCK(vp); 2423 if (VSHOULDFREE(vp)) 2424 vfree(vp); 2425 2426 /* 2427 * Done with purge, reset to the standard lock and 2428 * notify sleepers of the grim news. 2429 */ 2430 vp->v_vnlock = &vp->v_lock; 2431 vp->v_op = dead_vnodeop_p; 2432 if (vp->v_pollinfo != NULL) 2433 vn_pollgone(vp); 2434 vp->v_tag = "none"; 2435 } 2436 2437 /* 2438 * Eliminate all activity associated with the requested vnode 2439 * and with all vnodes aliased to the requested vnode. 2440 */ 2441 int 2442 vop_revoke(ap) 2443 struct vop_revoke_args /* { 2444 struct vnode *a_vp; 2445 int a_flags; 2446 } */ *ap; 2447 { 2448 struct vnode *vp, *vq; 2449 dev_t dev; 2450 2451 KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke")); 2452 vp = ap->a_vp; 2453 KASSERT((vp->v_type == VCHR), ("vop_revoke: not VCHR")); 2454 2455 VI_LOCK(vp); 2456 /* 2457 * If a vgone (or vclean) is already in progress, 2458 * wait until it is done and return. 2459 */ 2460 if (vp->v_iflag & VI_XLOCK) { 2461 vp->v_iflag |= VI_XWANT; 2462 msleep(vp, VI_MTX(vp), PINOD | PDROP, 2463 "vop_revokeall", 0); 2464 return (0); 2465 } 2466 VI_UNLOCK(vp); 2467 dev = vp->v_rdev; 2468 for (;;) { 2469 mtx_lock(&spechash_mtx); 2470 vq = SLIST_FIRST(&dev->si_hlist); 2471 mtx_unlock(&spechash_mtx); 2472 if (!vq) 2473 break; 2474 vgone(vq); 2475 } 2476 return (0); 2477 } 2478 2479 /* 2480 * Recycle an unused vnode to the front of the free list. 2481 * Release the passed interlock if the vnode will be recycled. 2482 */ 2483 int 2484 vrecycle(vp, inter_lkp, td) 2485 struct vnode *vp; 2486 struct mtx *inter_lkp; 2487 struct thread *td; 2488 { 2489 2490 VI_LOCK(vp); 2491 if (vp->v_usecount == 0) { 2492 if (inter_lkp) { 2493 mtx_unlock(inter_lkp); 2494 } 2495 vgonel(vp, td); 2496 return (1); 2497 } 2498 VI_UNLOCK(vp); 2499 return (0); 2500 } 2501 2502 /* 2503 * Eliminate all activity associated with a vnode 2504 * in preparation for reuse. 2505 */ 2506 void 2507 vgone(vp) 2508 register struct vnode *vp; 2509 { 2510 struct thread *td = curthread; /* XXX */ 2511 2512 VI_LOCK(vp); 2513 vgonel(vp, td); 2514 } 2515 2516 /* 2517 * Disassociate a character device from the its underlying filesystem and 2518 * attach it to spec. This is for use when the chr device is still active 2519 * and the filesystem is going away. 2520 */ 2521 static void 2522 vgonechrl(struct vnode *vp, struct thread *td) 2523 { 2524 ASSERT_VI_LOCKED(vp, "vgonechrl"); 2525 vx_lock(vp); 2526 /* 2527 * This is a custom version of vclean() which does not tearm down 2528 * the bufs or vm objects held by this vnode. This allows filesystems 2529 * to continue using devices which were discovered via another 2530 * filesystem that has been unmounted. 2531 */ 2532 if (vp->v_usecount != 0) { 2533 v_incr_usecount(vp, 1); 2534 /* 2535 * Ensure that no other activity can occur while the 2536 * underlying object is being cleaned out. 2537 */ 2538 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2539 /* 2540 * Any other processes trying to obtain this lock must first 2541 * wait for VXLOCK to clear, then call the new lock operation. 2542 */ 2543 VOP_UNLOCK(vp, 0, td); 2544 vp->v_vnlock = &vp->v_lock; 2545 vp->v_tag = "orphanchr"; 2546 vp->v_op = spec_vnodeop_p; 2547 if (vp->v_mount != NULL) 2548 insmntque(vp, (struct mount *)0); 2549 cache_purge(vp); 2550 vrele(vp); 2551 VI_LOCK(vp); 2552 } else 2553 vclean(vp, 0, td); 2554 vp->v_op = spec_vnodeop_p; 2555 vx_unlock(vp); 2556 VI_UNLOCK(vp); 2557 } 2558 2559 /* 2560 * vgone, with the vp interlock held. 2561 */ 2562 void 2563 vgonel(vp, td) 2564 struct vnode *vp; 2565 struct thread *td; 2566 { 2567 /* 2568 * If a vgone (or vclean) is already in progress, 2569 * wait until it is done and return. 2570 */ 2571 ASSERT_VI_LOCKED(vp, "vgonel"); 2572 if (vp->v_iflag & VI_XLOCK) { 2573 vp->v_iflag |= VI_XWANT; 2574 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0); 2575 return; 2576 } 2577 vx_lock(vp); 2578 2579 /* 2580 * Clean out the filesystem specific data. 2581 */ 2582 vclean(vp, DOCLOSE, td); 2583 VI_UNLOCK(vp); 2584 2585 /* 2586 * If special device, remove it from special device alias list 2587 * if it is on one. 2588 */ 2589 VI_LOCK(vp); 2590 if (vp->v_type == VCHR && vp->v_rdev != NULL && vp->v_rdev != NODEV) { 2591 mtx_lock(&spechash_mtx); 2592 SLIST_REMOVE(&vp->v_rdev->si_hlist, vp, vnode, v_specnext); 2593 vp->v_rdev->si_usecount -= vp->v_usecount; 2594 mtx_unlock(&spechash_mtx); 2595 vp->v_rdev = NULL; 2596 } 2597 2598 /* 2599 * If it is on the freelist and not already at the head, 2600 * move it to the head of the list. The test of the 2601 * VDOOMED flag and the reference count of zero is because 2602 * it will be removed from the free list by getnewvnode, 2603 * but will not have its reference count incremented until 2604 * after calling vgone. If the reference count were 2605 * incremented first, vgone would (incorrectly) try to 2606 * close the previous instance of the underlying object. 2607 */ 2608 if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) { 2609 mtx_lock(&vnode_free_list_mtx); 2610 if (vp->v_iflag & VI_FREE) { 2611 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2612 } else { 2613 vp->v_iflag |= VI_FREE; 2614 freevnodes++; 2615 } 2616 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2617 mtx_unlock(&vnode_free_list_mtx); 2618 } 2619 2620 vp->v_type = VBAD; 2621 vx_unlock(vp); 2622 VI_UNLOCK(vp); 2623 } 2624 2625 /* 2626 * Lookup a vnode by device number. 2627 */ 2628 int 2629 vfinddev(dev, type, vpp) 2630 dev_t dev; 2631 enum vtype type; 2632 struct vnode **vpp; 2633 { 2634 struct vnode *vp; 2635 2636 mtx_lock(&spechash_mtx); 2637 SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) { 2638 if (type == vp->v_type) { 2639 *vpp = vp; 2640 mtx_unlock(&spechash_mtx); 2641 return (1); 2642 } 2643 } 2644 mtx_unlock(&spechash_mtx); 2645 return (0); 2646 } 2647 2648 /* 2649 * Calculate the total number of references to a special device. 2650 */ 2651 int 2652 vcount(vp) 2653 struct vnode *vp; 2654 { 2655 int count; 2656 2657 mtx_lock(&spechash_mtx); 2658 count = vp->v_rdev->si_usecount; 2659 mtx_unlock(&spechash_mtx); 2660 return (count); 2661 } 2662 2663 /* 2664 * Same as above, but using the dev_t as argument 2665 */ 2666 int 2667 count_dev(dev) 2668 dev_t dev; 2669 { 2670 int count; 2671 2672 mtx_lock(&spechash_mtx); 2673 count = dev->si_usecount; 2674 mtx_unlock(&spechash_mtx); 2675 return(count); 2676 } 2677 2678 /* 2679 * Print out a description of a vnode. 2680 */ 2681 static char *typename[] = 2682 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 2683 2684 void 2685 vprint(label, vp) 2686 char *label; 2687 struct vnode *vp; 2688 { 2689 char buf[96]; 2690 2691 if (label != NULL) 2692 printf("%s: %p: ", label, (void *)vp); 2693 else 2694 printf("%p: ", (void *)vp); 2695 printf("tag %s, type %s, usecount %d, writecount %d, refcount %d,", 2696 vp->v_tag, typename[vp->v_type], vp->v_usecount, 2697 vp->v_writecount, vp->v_holdcnt); 2698 buf[0] = '\0'; 2699 if (vp->v_vflag & VV_ROOT) 2700 strcat(buf, "|VV_ROOT"); 2701 if (vp->v_vflag & VV_TEXT) 2702 strcat(buf, "|VV_TEXT"); 2703 if (vp->v_vflag & VV_SYSTEM) 2704 strcat(buf, "|VV_SYSTEM"); 2705 if (vp->v_iflag & VI_XLOCK) 2706 strcat(buf, "|VI_XLOCK"); 2707 if (vp->v_iflag & VI_XWANT) 2708 strcat(buf, "|VI_XWANT"); 2709 if (vp->v_iflag & VI_BWAIT) 2710 strcat(buf, "|VI_BWAIT"); 2711 if (vp->v_iflag & VI_DOOMED) 2712 strcat(buf, "|VI_DOOMED"); 2713 if (vp->v_iflag & VI_FREE) 2714 strcat(buf, "|VI_FREE"); 2715 if (vp->v_vflag & VV_OBJBUF) 2716 strcat(buf, "|VV_OBJBUF"); 2717 if (buf[0] != '\0') 2718 printf(" flags (%s),", &buf[1]); 2719 lockmgr_printinfo(vp->v_vnlock); 2720 printf("\n"); 2721 if (vp->v_data != NULL) 2722 VOP_PRINT(vp); 2723 } 2724 2725 #ifdef DDB 2726 #include <ddb/ddb.h> 2727 /* 2728 * List all of the locked vnodes in the system. 2729 * Called when debugging the kernel. 2730 */ 2731 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 2732 { 2733 struct mount *mp, *nmp; 2734 struct vnode *vp; 2735 2736 /* 2737 * Note: because this is DDB, we can't obey the locking semantics 2738 * for these structures, which means we could catch an inconsistent 2739 * state and dereference a nasty pointer. Not much to be done 2740 * about that. 2741 */ 2742 printf("Locked vnodes\n"); 2743 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2744 nmp = TAILQ_NEXT(mp, mnt_list); 2745 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2746 if (VOP_ISLOCKED(vp, NULL)) 2747 vprint(NULL, vp); 2748 } 2749 nmp = TAILQ_NEXT(mp, mnt_list); 2750 } 2751 } 2752 #endif 2753 2754 /* 2755 * Fill in a struct xvfsconf based on a struct vfsconf. 2756 */ 2757 static void 2758 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) 2759 { 2760 2761 strcpy(xvfsp->vfc_name, vfsp->vfc_name); 2762 xvfsp->vfc_typenum = vfsp->vfc_typenum; 2763 xvfsp->vfc_refcount = vfsp->vfc_refcount; 2764 xvfsp->vfc_flags = vfsp->vfc_flags; 2765 /* 2766 * These are unused in userland, we keep them 2767 * to not break binary compatibility. 2768 */ 2769 xvfsp->vfc_vfsops = NULL; 2770 xvfsp->vfc_next = NULL; 2771 } 2772 2773 static int 2774 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 2775 { 2776 struct vfsconf *vfsp; 2777 struct xvfsconf *xvfsp; 2778 int cnt, error, i; 2779 2780 cnt = 0; 2781 for (vfsp = vfsconf; vfsp != NULL; vfsp = vfsp->vfc_next) 2782 cnt++; 2783 xvfsp = malloc(sizeof(struct xvfsconf) * cnt, M_TEMP, M_WAITOK); 2784 /* 2785 * Handle the race that we will have here when struct vfsconf 2786 * will be locked down by using both cnt and checking vfc_next 2787 * against NULL to determine the end of the loop. The race will 2788 * happen because we will have to unlock before calling malloc(). 2789 * We are protected by Giant for now. 2790 */ 2791 i = 0; 2792 for (vfsp = vfsconf; vfsp != NULL && i < cnt; vfsp = vfsp->vfc_next) { 2793 vfsconf2x(vfsp, xvfsp + i); 2794 i++; 2795 } 2796 error = SYSCTL_OUT(req, xvfsp, sizeof(struct xvfsconf) * i); 2797 free(xvfsp, M_TEMP); 2798 return (error); 2799 } 2800 2801 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, 2802 "S,xvfsconf", "List of all configured filesystems"); 2803 2804 /* 2805 * Top level filesystem related information gathering. 2806 */ 2807 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2808 2809 static int 2810 vfs_sysctl(SYSCTL_HANDLER_ARGS) 2811 { 2812 int *name = (int *)arg1 - 1; /* XXX */ 2813 u_int namelen = arg2 + 1; /* XXX */ 2814 struct vfsconf *vfsp; 2815 struct xvfsconf xvfsp; 2816 2817 printf("WARNING: userland calling deprecated sysctl, " 2818 "please rebuild world\n"); 2819 2820 #if 1 || defined(COMPAT_PRELITE2) 2821 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2822 if (namelen == 1) 2823 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2824 #endif 2825 2826 switch (name[1]) { 2827 case VFS_MAXTYPENUM: 2828 if (namelen != 2) 2829 return (ENOTDIR); 2830 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2831 case VFS_CONF: 2832 if (namelen != 3) 2833 return (ENOTDIR); /* overloaded */ 2834 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) 2835 if (vfsp->vfc_typenum == name[2]) 2836 break; 2837 if (vfsp == NULL) 2838 return (EOPNOTSUPP); 2839 vfsconf2x(vfsp, &xvfsp); 2840 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 2841 } 2842 return (EOPNOTSUPP); 2843 } 2844 2845 SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl, 2846 "Generic filesystem"); 2847 2848 #if 1 || defined(COMPAT_PRELITE2) 2849 2850 static int 2851 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 2852 { 2853 int error; 2854 struct vfsconf *vfsp; 2855 struct ovfsconf ovfs; 2856 2857 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) { 2858 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 2859 strcpy(ovfs.vfc_name, vfsp->vfc_name); 2860 ovfs.vfc_index = vfsp->vfc_typenum; 2861 ovfs.vfc_refcount = vfsp->vfc_refcount; 2862 ovfs.vfc_flags = vfsp->vfc_flags; 2863 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 2864 if (error) 2865 return error; 2866 } 2867 return 0; 2868 } 2869 2870 #endif /* 1 || COMPAT_PRELITE2 */ 2871 2872 #define KINFO_VNODESLOP 10 2873 #ifdef notyet 2874 /* 2875 * Dump vnode list (via sysctl). 2876 */ 2877 /* ARGSUSED */ 2878 static int 2879 sysctl_vnode(SYSCTL_HANDLER_ARGS) 2880 { 2881 struct xvnode *xvn; 2882 struct thread *td = req->td; 2883 struct mount *mp; 2884 struct vnode *vp; 2885 int error, len, n; 2886 2887 /* 2888 * Stale numvnodes access is not fatal here. 2889 */ 2890 req->lock = 0; 2891 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 2892 if (!req->oldptr) 2893 /* Make an estimate */ 2894 return (SYSCTL_OUT(req, 0, len)); 2895 2896 sysctl_wire_old_buffer(req, 0); 2897 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 2898 n = 0; 2899 mtx_lock(&mountlist_mtx); 2900 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 2901 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 2902 continue; 2903 MNT_ILOCK(mp); 2904 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2905 if (n == len) 2906 break; 2907 vref(vp); 2908 xvn[n].xv_size = sizeof *xvn; 2909 xvn[n].xv_vnode = vp; 2910 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 2911 XV_COPY(usecount); 2912 XV_COPY(writecount); 2913 XV_COPY(holdcnt); 2914 XV_COPY(id); 2915 XV_COPY(mount); 2916 XV_COPY(numoutput); 2917 XV_COPY(type); 2918 #undef XV_COPY 2919 xvn[n].xv_flag = vp->v_vflag; 2920 2921 switch (vp->v_type) { 2922 case VREG: 2923 case VDIR: 2924 case VLNK: 2925 xvn[n].xv_dev = vp->v_cachedfs; 2926 xvn[n].xv_ino = vp->v_cachedid; 2927 break; 2928 case VBLK: 2929 case VCHR: 2930 if (vp->v_rdev == NULL) { 2931 vrele(vp); 2932 continue; 2933 } 2934 xvn[n].xv_dev = dev2udev(vp->v_rdev); 2935 break; 2936 case VSOCK: 2937 xvn[n].xv_socket = vp->v_socket; 2938 break; 2939 case VFIFO: 2940 xvn[n].xv_fifo = vp->v_fifoinfo; 2941 break; 2942 case VNON: 2943 case VBAD: 2944 default: 2945 /* shouldn't happen? */ 2946 vrele(vp); 2947 continue; 2948 } 2949 vrele(vp); 2950 ++n; 2951 } 2952 MNT_IUNLOCK(mp); 2953 mtx_lock(&mountlist_mtx); 2954 vfs_unbusy(mp, td); 2955 if (n == len) 2956 break; 2957 } 2958 mtx_unlock(&mountlist_mtx); 2959 2960 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 2961 free(xvn, M_TEMP); 2962 return (error); 2963 } 2964 2965 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 2966 0, 0, sysctl_vnode, "S,xvnode", ""); 2967 #endif 2968 2969 /* 2970 * Check to see if a filesystem is mounted on a block device. 2971 */ 2972 int 2973 vfs_mountedon(vp) 2974 struct vnode *vp; 2975 { 2976 2977 if (vp->v_rdev->si_mountpoint != NULL) 2978 return (EBUSY); 2979 return (0); 2980 } 2981 2982 /* 2983 * Unmount all filesystems. The list is traversed in reverse order 2984 * of mounting to avoid dependencies. 2985 */ 2986 void 2987 vfs_unmountall() 2988 { 2989 struct mount *mp; 2990 struct thread *td; 2991 int error; 2992 2993 if (curthread != NULL) 2994 td = curthread; 2995 else 2996 td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */ 2997 /* 2998 * Since this only runs when rebooting, it is not interlocked. 2999 */ 3000 while(!TAILQ_EMPTY(&mountlist)) { 3001 mp = TAILQ_LAST(&mountlist, mntlist); 3002 error = dounmount(mp, MNT_FORCE, td); 3003 if (error) { 3004 TAILQ_REMOVE(&mountlist, mp, mnt_list); 3005 printf("unmount of %s failed (", 3006 mp->mnt_stat.f_mntonname); 3007 if (error == EBUSY) 3008 printf("BUSY)\n"); 3009 else 3010 printf("%d)\n", error); 3011 } else { 3012 /* The unmount has removed mp from the mountlist */ 3013 } 3014 } 3015 } 3016 3017 /* 3018 * perform msync on all vnodes under a mount point 3019 * the mount point must be locked. 3020 */ 3021 void 3022 vfs_msync(struct mount *mp, int flags) 3023 { 3024 struct vnode *vp, *nvp; 3025 struct vm_object *obj; 3026 int tries; 3027 3028 GIANT_REQUIRED; 3029 3030 tries = 5; 3031 MNT_ILOCK(mp); 3032 loop: 3033 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp != NULL; vp = nvp) { 3034 if (vp->v_mount != mp) { 3035 if (--tries > 0) 3036 goto loop; 3037 break; 3038 } 3039 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 3040 3041 VI_LOCK(vp); 3042 if (vp->v_iflag & VI_XLOCK) { 3043 VI_UNLOCK(vp); 3044 continue; 3045 } 3046 3047 if ((vp->v_iflag & VI_OBJDIRTY) && 3048 (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { 3049 MNT_IUNLOCK(mp); 3050 if (!vget(vp, 3051 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 3052 curthread)) { 3053 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 3054 vput(vp); 3055 MNT_ILOCK(mp); 3056 continue; 3057 } 3058 3059 if (VOP_GETVOBJECT(vp, &obj) == 0) { 3060 VM_OBJECT_LOCK(obj); 3061 vm_object_page_clean(obj, 0, 0, 3062 flags == MNT_WAIT ? 3063 OBJPC_SYNC : OBJPC_NOSYNC); 3064 VM_OBJECT_UNLOCK(obj); 3065 } 3066 vput(vp); 3067 } 3068 MNT_ILOCK(mp); 3069 if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) { 3070 if (--tries > 0) 3071 goto loop; 3072 break; 3073 } 3074 } else 3075 VI_UNLOCK(vp); 3076 } 3077 MNT_IUNLOCK(mp); 3078 } 3079 3080 /* 3081 * Create the VM object needed for VMIO and mmap support. This 3082 * is done for all VREG files in the system. Some filesystems might 3083 * afford the additional metadata buffering capability of the 3084 * VMIO code by making the device node be VMIO mode also. 3085 * 3086 * vp must be locked when vfs_object_create is called. 3087 */ 3088 int 3089 vfs_object_create(vp, td, cred) 3090 struct vnode *vp; 3091 struct thread *td; 3092 struct ucred *cred; 3093 { 3094 3095 GIANT_REQUIRED; 3096 return (VOP_CREATEVOBJECT(vp, cred, td)); 3097 } 3098 3099 /* 3100 * Mark a vnode as free, putting it up for recycling. 3101 */ 3102 void 3103 vfree(vp) 3104 struct vnode *vp; 3105 { 3106 3107 ASSERT_VI_LOCKED(vp, "vfree"); 3108 mtx_lock(&vnode_free_list_mtx); 3109 KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free")); 3110 if (vp->v_iflag & VI_AGE) { 3111 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 3112 } else { 3113 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 3114 } 3115 freevnodes++; 3116 mtx_unlock(&vnode_free_list_mtx); 3117 vp->v_iflag &= ~VI_AGE; 3118 vp->v_iflag |= VI_FREE; 3119 } 3120 3121 /* 3122 * Opposite of vfree() - mark a vnode as in use. 3123 */ 3124 void 3125 vbusy(vp) 3126 struct vnode *vp; 3127 { 3128 3129 ASSERT_VI_LOCKED(vp, "vbusy"); 3130 KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free")); 3131 3132 mtx_lock(&vnode_free_list_mtx); 3133 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 3134 freevnodes--; 3135 mtx_unlock(&vnode_free_list_mtx); 3136 3137 vp->v_iflag &= ~(VI_FREE|VI_AGE); 3138 } 3139 3140 /* 3141 * Initalize per-vnode helper structure to hold poll-related state. 3142 */ 3143 void 3144 v_addpollinfo(struct vnode *vp) 3145 { 3146 3147 vp->v_pollinfo = uma_zalloc(vnodepoll_zone, M_WAITOK); 3148 mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 3149 } 3150 3151 /* 3152 * Record a process's interest in events which might happen to 3153 * a vnode. Because poll uses the historic select-style interface 3154 * internally, this routine serves as both the ``check for any 3155 * pending events'' and the ``record my interest in future events'' 3156 * functions. (These are done together, while the lock is held, 3157 * to avoid race conditions.) 3158 */ 3159 int 3160 vn_pollrecord(vp, td, events) 3161 struct vnode *vp; 3162 struct thread *td; 3163 short events; 3164 { 3165 3166 if (vp->v_pollinfo == NULL) 3167 v_addpollinfo(vp); 3168 mtx_lock(&vp->v_pollinfo->vpi_lock); 3169 if (vp->v_pollinfo->vpi_revents & events) { 3170 /* 3171 * This leaves events we are not interested 3172 * in available for the other process which 3173 * which presumably had requested them 3174 * (otherwise they would never have been 3175 * recorded). 3176 */ 3177 events &= vp->v_pollinfo->vpi_revents; 3178 vp->v_pollinfo->vpi_revents &= ~events; 3179 3180 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3181 return events; 3182 } 3183 vp->v_pollinfo->vpi_events |= events; 3184 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 3185 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3186 return 0; 3187 } 3188 3189 /* 3190 * Note the occurrence of an event. If the VN_POLLEVENT macro is used, 3191 * it is possible for us to miss an event due to race conditions, but 3192 * that condition is expected to be rare, so for the moment it is the 3193 * preferred interface. 3194 */ 3195 void 3196 vn_pollevent(vp, events) 3197 struct vnode *vp; 3198 short events; 3199 { 3200 3201 if (vp->v_pollinfo == NULL) 3202 v_addpollinfo(vp); 3203 mtx_lock(&vp->v_pollinfo->vpi_lock); 3204 if (vp->v_pollinfo->vpi_events & events) { 3205 /* 3206 * We clear vpi_events so that we don't 3207 * call selwakeup() twice if two events are 3208 * posted before the polling process(es) is 3209 * awakened. This also ensures that we take at 3210 * most one selwakeup() if the polling process 3211 * is no longer interested. However, it does 3212 * mean that only one event can be noticed at 3213 * a time. (Perhaps we should only clear those 3214 * event bits which we note?) XXX 3215 */ 3216 vp->v_pollinfo->vpi_events = 0; /* &= ~events ??? */ 3217 vp->v_pollinfo->vpi_revents |= events; 3218 selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO); 3219 } 3220 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3221 } 3222 3223 /* 3224 * Wake up anyone polling on vp because it is being revoked. 3225 * This depends on dead_poll() returning POLLHUP for correct 3226 * behavior. 3227 */ 3228 void 3229 vn_pollgone(vp) 3230 struct vnode *vp; 3231 { 3232 3233 mtx_lock(&vp->v_pollinfo->vpi_lock); 3234 VN_KNOTE(vp, NOTE_REVOKE); 3235 if (vp->v_pollinfo->vpi_events) { 3236 vp->v_pollinfo->vpi_events = 0; 3237 selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO); 3238 } 3239 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3240 } 3241 3242 3243 3244 /* 3245 * Routine to create and manage a filesystem syncer vnode. 3246 */ 3247 #define sync_close ((int (*)(struct vop_close_args *))nullop) 3248 static int sync_fsync(struct vop_fsync_args *); 3249 static int sync_inactive(struct vop_inactive_args *); 3250 static int sync_reclaim(struct vop_reclaim_args *); 3251 3252 static vop_t **sync_vnodeop_p; 3253 static struct vnodeopv_entry_desc sync_vnodeop_entries[] = { 3254 { &vop_default_desc, (vop_t *) vop_eopnotsupp }, 3255 { &vop_close_desc, (vop_t *) sync_close }, /* close */ 3256 { &vop_fsync_desc, (vop_t *) sync_fsync }, /* fsync */ 3257 { &vop_inactive_desc, (vop_t *) sync_inactive }, /* inactive */ 3258 { &vop_reclaim_desc, (vop_t *) sync_reclaim }, /* reclaim */ 3259 { &vop_lock_desc, (vop_t *) vop_stdlock }, /* lock */ 3260 { &vop_unlock_desc, (vop_t *) vop_stdunlock }, /* unlock */ 3261 { &vop_islocked_desc, (vop_t *) vop_stdislocked }, /* islocked */ 3262 { NULL, NULL } 3263 }; 3264 static struct vnodeopv_desc sync_vnodeop_opv_desc = 3265 { &sync_vnodeop_p, sync_vnodeop_entries }; 3266 3267 VNODEOP_SET(sync_vnodeop_opv_desc); 3268 3269 /* 3270 * Create a new filesystem syncer vnode for the specified mount point. 3271 */ 3272 int 3273 vfs_allocate_syncvnode(mp) 3274 struct mount *mp; 3275 { 3276 struct vnode *vp; 3277 static long start, incr, next; 3278 int error; 3279 3280 /* Allocate a new vnode */ 3281 if ((error = getnewvnode("syncer", mp, sync_vnodeop_p, &vp)) != 0) { 3282 mp->mnt_syncer = NULL; 3283 return (error); 3284 } 3285 vp->v_type = VNON; 3286 /* 3287 * Place the vnode onto the syncer worklist. We attempt to 3288 * scatter them about on the list so that they will go off 3289 * at evenly distributed times even if all the filesystems 3290 * are mounted at once. 3291 */ 3292 next += incr; 3293 if (next == 0 || next > syncer_maxdelay) { 3294 start /= 2; 3295 incr /= 2; 3296 if (start == 0) { 3297 start = syncer_maxdelay / 2; 3298 incr = syncer_maxdelay; 3299 } 3300 next = start; 3301 } 3302 VI_LOCK(vp); 3303 vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0); 3304 VI_UNLOCK(vp); 3305 mp->mnt_syncer = vp; 3306 return (0); 3307 } 3308 3309 /* 3310 * Do a lazy sync of the filesystem. 3311 */ 3312 static int 3313 sync_fsync(ap) 3314 struct vop_fsync_args /* { 3315 struct vnode *a_vp; 3316 struct ucred *a_cred; 3317 int a_waitfor; 3318 struct thread *a_td; 3319 } */ *ap; 3320 { 3321 struct vnode *syncvp = ap->a_vp; 3322 struct mount *mp = syncvp->v_mount; 3323 struct thread *td = ap->a_td; 3324 int error, asyncflag; 3325 3326 /* 3327 * We only need to do something if this is a lazy evaluation. 3328 */ 3329 if (ap->a_waitfor != MNT_LAZY) 3330 return (0); 3331 3332 /* 3333 * Move ourselves to the back of the sync list. 3334 */ 3335 VI_LOCK(syncvp); 3336 vn_syncer_add_to_worklist(syncvp, syncdelay); 3337 VI_UNLOCK(syncvp); 3338 3339 /* 3340 * Walk the list of vnodes pushing all that are dirty and 3341 * not already on the sync list. 3342 */ 3343 mtx_lock(&mountlist_mtx); 3344 if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { 3345 mtx_unlock(&mountlist_mtx); 3346 return (0); 3347 } 3348 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 3349 vfs_unbusy(mp, td); 3350 return (0); 3351 } 3352 asyncflag = mp->mnt_flag & MNT_ASYNC; 3353 mp->mnt_flag &= ~MNT_ASYNC; 3354 vfs_msync(mp, MNT_NOWAIT); 3355 error = VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td); 3356 if (asyncflag) 3357 mp->mnt_flag |= MNT_ASYNC; 3358 vn_finished_write(mp); 3359 vfs_unbusy(mp, td); 3360 return (error); 3361 } 3362 3363 /* 3364 * The syncer vnode is no referenced. 3365 */ 3366 static int 3367 sync_inactive(ap) 3368 struct vop_inactive_args /* { 3369 struct vnode *a_vp; 3370 struct thread *a_td; 3371 } */ *ap; 3372 { 3373 3374 VOP_UNLOCK(ap->a_vp, 0, ap->a_td); 3375 vgone(ap->a_vp); 3376 return (0); 3377 } 3378 3379 /* 3380 * The syncer vnode is no longer needed and is being decommissioned. 3381 * 3382 * Modifications to the worklist must be protected by sync_mtx. 3383 */ 3384 static int 3385 sync_reclaim(ap) 3386 struct vop_reclaim_args /* { 3387 struct vnode *a_vp; 3388 } */ *ap; 3389 { 3390 struct vnode *vp = ap->a_vp; 3391 3392 VI_LOCK(vp); 3393 vp->v_mount->mnt_syncer = NULL; 3394 if (vp->v_iflag & VI_ONWORKLST) { 3395 mtx_lock(&sync_mtx); 3396 LIST_REMOVE(vp, v_synclist); 3397 mtx_unlock(&sync_mtx); 3398 vp->v_iflag &= ~VI_ONWORKLST; 3399 } 3400 VI_UNLOCK(vp); 3401 3402 return (0); 3403 } 3404 3405 /* 3406 * extract the dev_t from a VCHR 3407 */ 3408 dev_t 3409 vn_todev(vp) 3410 struct vnode *vp; 3411 { 3412 3413 if (vp->v_type != VCHR) 3414 return (NODEV); 3415 return (vp->v_rdev); 3416 } 3417 3418 /* 3419 * Check if vnode represents a disk device 3420 */ 3421 int 3422 vn_isdisk(vp, errp) 3423 struct vnode *vp; 3424 int *errp; 3425 { 3426 int error; 3427 3428 error = 0; 3429 if (vp->v_type != VCHR) 3430 error = ENOTBLK; 3431 else if (vp->v_rdev == NULL) 3432 error = ENXIO; 3433 else if (!(devsw(vp->v_rdev)->d_flags & D_DISK)) 3434 error = ENOTBLK; 3435 if (errp != NULL) 3436 *errp = error; 3437 return (error == 0); 3438 } 3439 3440 /* 3441 * Free data allocated by namei(); see namei(9) for details. 3442 */ 3443 void 3444 NDFREE(ndp, flags) 3445 struct nameidata *ndp; 3446 const u_int flags; 3447 { 3448 3449 if (!(flags & NDF_NO_FREE_PNBUF) && 3450 (ndp->ni_cnd.cn_flags & HASBUF)) { 3451 uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf); 3452 ndp->ni_cnd.cn_flags &= ~HASBUF; 3453 } 3454 if (!(flags & NDF_NO_DVP_UNLOCK) && 3455 (ndp->ni_cnd.cn_flags & LOCKPARENT) && 3456 ndp->ni_dvp != ndp->ni_vp) 3457 VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread); 3458 if (!(flags & NDF_NO_DVP_RELE) && 3459 (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) { 3460 vrele(ndp->ni_dvp); 3461 ndp->ni_dvp = NULL; 3462 } 3463 if (!(flags & NDF_NO_VP_UNLOCK) && 3464 (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp) 3465 VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread); 3466 if (!(flags & NDF_NO_VP_RELE) && 3467 ndp->ni_vp) { 3468 vrele(ndp->ni_vp); 3469 ndp->ni_vp = NULL; 3470 } 3471 if (!(flags & NDF_NO_STARTDIR_RELE) && 3472 (ndp->ni_cnd.cn_flags & SAVESTART)) { 3473 vrele(ndp->ni_startdir); 3474 ndp->ni_startdir = NULL; 3475 } 3476 } 3477 3478 /* 3479 * Common filesystem object access control check routine. Accepts a 3480 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3481 * and optional call-by-reference privused argument allowing vaccess() 3482 * to indicate to the caller whether privilege was used to satisfy the 3483 * request (obsoleted). Returns 0 on success, or an errno on failure. 3484 */ 3485 int 3486 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused) 3487 enum vtype type; 3488 mode_t file_mode; 3489 uid_t file_uid; 3490 gid_t file_gid; 3491 mode_t acc_mode; 3492 struct ucred *cred; 3493 int *privused; 3494 { 3495 mode_t dac_granted; 3496 #ifdef CAPABILITIES 3497 mode_t cap_granted; 3498 #endif 3499 3500 /* 3501 * Look for a normal, non-privileged way to access the file/directory 3502 * as requested. If it exists, go with that. 3503 */ 3504 3505 if (privused != NULL) 3506 *privused = 0; 3507 3508 dac_granted = 0; 3509 3510 /* Check the owner. */ 3511 if (cred->cr_uid == file_uid) { 3512 dac_granted |= VADMIN; 3513 if (file_mode & S_IXUSR) 3514 dac_granted |= VEXEC; 3515 if (file_mode & S_IRUSR) 3516 dac_granted |= VREAD; 3517 if (file_mode & S_IWUSR) 3518 dac_granted |= (VWRITE | VAPPEND); 3519 3520 if ((acc_mode & dac_granted) == acc_mode) 3521 return (0); 3522 3523 goto privcheck; 3524 } 3525 3526 /* Otherwise, check the groups (first match) */ 3527 if (groupmember(file_gid, cred)) { 3528 if (file_mode & S_IXGRP) 3529 dac_granted |= VEXEC; 3530 if (file_mode & S_IRGRP) 3531 dac_granted |= VREAD; 3532 if (file_mode & S_IWGRP) 3533 dac_granted |= (VWRITE | VAPPEND); 3534 3535 if ((acc_mode & dac_granted) == acc_mode) 3536 return (0); 3537 3538 goto privcheck; 3539 } 3540 3541 /* Otherwise, check everyone else. */ 3542 if (file_mode & S_IXOTH) 3543 dac_granted |= VEXEC; 3544 if (file_mode & S_IROTH) 3545 dac_granted |= VREAD; 3546 if (file_mode & S_IWOTH) 3547 dac_granted |= (VWRITE | VAPPEND); 3548 if ((acc_mode & dac_granted) == acc_mode) 3549 return (0); 3550 3551 privcheck: 3552 if (!suser_cred(cred, PRISON_ROOT)) { 3553 /* XXX audit: privilege used */ 3554 if (privused != NULL) 3555 *privused = 1; 3556 return (0); 3557 } 3558 3559 #ifdef CAPABILITIES 3560 /* 3561 * Build a capability mask to determine if the set of capabilities 3562 * satisfies the requirements when combined with the granted mask 3563 * from above. 3564 * For each capability, if the capability is required, bitwise 3565 * or the request type onto the cap_granted mask. 3566 */ 3567 cap_granted = 0; 3568 3569 if (type == VDIR) { 3570 /* 3571 * For directories, use CAP_DAC_READ_SEARCH to satisfy 3572 * VEXEC requests, instead of CAP_DAC_EXECUTE. 3573 */ 3574 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3575 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) 3576 cap_granted |= VEXEC; 3577 } else { 3578 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3579 !cap_check(cred, NULL, CAP_DAC_EXECUTE, PRISON_ROOT)) 3580 cap_granted |= VEXEC; 3581 } 3582 3583 if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && 3584 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) 3585 cap_granted |= VREAD; 3586 3587 if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3588 !cap_check(cred, NULL, CAP_DAC_WRITE, PRISON_ROOT)) 3589 cap_granted |= (VWRITE | VAPPEND); 3590 3591 if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3592 !cap_check(cred, NULL, CAP_FOWNER, PRISON_ROOT)) 3593 cap_granted |= VADMIN; 3594 3595 if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { 3596 /* XXX audit: privilege used */ 3597 if (privused != NULL) 3598 *privused = 1; 3599 return (0); 3600 } 3601 #endif 3602 3603 return ((acc_mode & VADMIN) ? EPERM : EACCES); 3604 } 3605 3606 /* 3607 * Credential check based on process requesting service, and per-attribute 3608 * permissions. 3609 */ 3610 int 3611 extattr_check_cred(struct vnode *vp, int attrnamespace, 3612 struct ucred *cred, struct thread *td, int access) 3613 { 3614 3615 /* 3616 * Kernel-invoked always succeeds. 3617 */ 3618 if (cred == NOCRED) 3619 return (0); 3620 3621 /* 3622 * Do not allow privileged processes in jail to directly 3623 * manipulate system attributes. 3624 * 3625 * XXX What capability should apply here? 3626 * Probably CAP_SYS_SETFFLAG. 3627 */ 3628 switch (attrnamespace) { 3629 case EXTATTR_NAMESPACE_SYSTEM: 3630 /* Potentially should be: return (EPERM); */ 3631 return (suser_cred(cred, 0)); 3632 case EXTATTR_NAMESPACE_USER: 3633 return (VOP_ACCESS(vp, access, cred, td)); 3634 default: 3635 return (EPERM); 3636 } 3637 } 3638 3639 #ifdef DEBUG_VFS_LOCKS 3640 /* 3641 * This only exists to supress warnings from unlocked specfs accesses. It is 3642 * no longer ok to have an unlocked VFS. 3643 */ 3644 #define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD) 3645 3646 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 3647 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 3648 int vfs_badlock_print = 1; /* Print lock violations. */ 3649 3650 static void 3651 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 3652 { 3653 3654 if (vfs_badlock_print) 3655 printf("%s: %p %s\n", str, (void *)vp, msg); 3656 if (vfs_badlock_ddb) 3657 Debugger("lock violation"); 3658 } 3659 3660 void 3661 assert_vi_locked(struct vnode *vp, const char *str) 3662 { 3663 3664 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 3665 vfs_badlock("interlock is not locked but should be", str, vp); 3666 } 3667 3668 void 3669 assert_vi_unlocked(struct vnode *vp, const char *str) 3670 { 3671 3672 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 3673 vfs_badlock("interlock is locked but should not be", str, vp); 3674 } 3675 3676 void 3677 assert_vop_locked(struct vnode *vp, const char *str) 3678 { 3679 3680 if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0) 3681 vfs_badlock("is not locked but should be", str, vp); 3682 } 3683 3684 void 3685 assert_vop_unlocked(struct vnode *vp, const char *str) 3686 { 3687 3688 if (vp && !IGNORE_LOCK(vp) && 3689 VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE) 3690 vfs_badlock("is locked but should not be", str, vp); 3691 } 3692 3693 #if 0 3694 void 3695 assert_vop_elocked(struct vnode *vp, const char *str) 3696 { 3697 3698 if (vp && !IGNORE_LOCK(vp) && 3699 VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE) 3700 vfs_badlock("is not exclusive locked but should be", str, vp); 3701 } 3702 3703 void 3704 assert_vop_elocked_other(struct vnode *vp, const char *str) 3705 { 3706 3707 if (vp && !IGNORE_LOCK(vp) && 3708 VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER) 3709 vfs_badlock("is not exclusive locked by another thread", 3710 str, vp); 3711 } 3712 3713 void 3714 assert_vop_slocked(struct vnode *vp, const char *str) 3715 { 3716 3717 if (vp && !IGNORE_LOCK(vp) && 3718 VOP_ISLOCKED(vp, curthread) != LK_SHARED) 3719 vfs_badlock("is not locked shared but should be", str, vp); 3720 } 3721 #endif /* 0 */ 3722 3723 void 3724 vop_rename_pre(void *ap) 3725 { 3726 struct vop_rename_args *a = ap; 3727 3728 if (a->a_tvp) 3729 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 3730 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 3731 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 3732 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 3733 3734 /* Check the source (from). */ 3735 if (a->a_tdvp != a->a_fdvp) 3736 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 3737 if (a->a_tvp != a->a_fvp) 3738 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked"); 3739 3740 /* Check the target. */ 3741 if (a->a_tvp) 3742 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 3743 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 3744 } 3745 3746 void 3747 vop_strategy_pre(void *ap) 3748 { 3749 struct vop_strategy_args *a; 3750 struct buf *bp; 3751 3752 a = ap; 3753 bp = a->a_bp; 3754 3755 /* 3756 * Cluster ops lock their component buffers but not the IO container. 3757 */ 3758 if ((bp->b_flags & B_CLUSTER) != 0) 3759 return; 3760 3761 if (BUF_REFCNT(bp) < 1) { 3762 if (vfs_badlock_print) 3763 printf( 3764 "VOP_STRATEGY: bp is not locked but should be\n"); 3765 if (vfs_badlock_ddb) 3766 Debugger("lock violation"); 3767 } 3768 } 3769 3770 void 3771 vop_lookup_pre(void *ap) 3772 { 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 } 3781 3782 void 3783 vop_lookup_post(void *ap, int rc) 3784 { 3785 struct vop_lookup_args *a; 3786 struct componentname *cnp; 3787 struct vnode *dvp; 3788 struct vnode *vp; 3789 int flags; 3790 3791 a = ap; 3792 dvp = a->a_dvp; 3793 cnp = a->a_cnp; 3794 vp = *(a->a_vpp); 3795 flags = cnp->cn_flags; 3796 3797 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3798 3799 /* 3800 * If this is the last path component for this lookup and LOCKPARENT 3801 * is set, OR if there is an error the directory has to be locked. 3802 */ 3803 if ((flags & LOCKPARENT) && (flags & ISLASTCN)) 3804 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)"); 3805 else if (rc != 0) 3806 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)"); 3807 else if (dvp != vp) 3808 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)"); 3809 if (flags & PDIRUNLOCK) 3810 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)"); 3811 } 3812 3813 void 3814 vop_lock_pre(void *ap) 3815 { 3816 struct vop_lock_args *a = ap; 3817 3818 if ((a->a_flags & LK_INTERLOCK) == 0) 3819 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3820 else 3821 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 3822 } 3823 3824 void 3825 vop_lock_post(void *ap, int rc) 3826 { 3827 struct vop_lock_args *a = ap; 3828 3829 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3830 if (rc == 0) 3831 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 3832 } 3833 3834 void 3835 vop_unlock_pre(void *ap) 3836 { 3837 struct vop_unlock_args *a = ap; 3838 3839 if (a->a_flags & LK_INTERLOCK) 3840 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 3841 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 3842 } 3843 3844 void 3845 vop_unlock_post(void *ap, int rc) 3846 { 3847 struct vop_unlock_args *a = ap; 3848 3849 if (a->a_flags & LK_INTERLOCK) 3850 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 3851 } 3852 #endif /* DEBUG_VFS_LOCKS */ 3853