1 /* 2 * Copyright (c) 1989, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 4. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 35 */ 36 37 /* 38 * External virtual filesystem routines 39 */ 40 41 #include <sys/cdefs.h> 42 __FBSDID("$FreeBSD$"); 43 44 #include "opt_ddb.h" 45 #include "opt_mac.h" 46 47 #include <sys/param.h> 48 #include <sys/systm.h> 49 #include <sys/bio.h> 50 #include <sys/buf.h> 51 #include <sys/conf.h> 52 #include <sys/eventhandler.h> 53 #include <sys/extattr.h> 54 #include <sys/fcntl.h> 55 #include <sys/kernel.h> 56 #include <sys/kthread.h> 57 #include <sys/mac.h> 58 #include <sys/malloc.h> 59 #include <sys/mount.h> 60 #include <sys/namei.h> 61 #include <sys/sleepqueue.h> 62 #include <sys/stat.h> 63 #include <sys/sysctl.h> 64 #include <sys/syslog.h> 65 #include <sys/vmmeter.h> 66 #include <sys/vnode.h> 67 68 #include <vm/vm.h> 69 #include <vm/vm_object.h> 70 #include <vm/vm_extern.h> 71 #include <vm/pmap.h> 72 #include <vm/vm_map.h> 73 #include <vm/vm_page.h> 74 #include <vm/vm_kern.h> 75 #include <vm/uma.h> 76 77 static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure"); 78 79 static void addalias(struct vnode *vp, struct cdev *nvp_rdev); 80 static void insmntque(struct vnode *vp, struct mount *mp); 81 static void vclean(struct vnode *vp, int flags, struct thread *td); 82 static void vlruvp(struct vnode *vp); 83 static int flushbuflist(struct buf *blist, int flags, struct vnode *vp, 84 int slpflag, int slptimeo, int *errorp); 85 static void syncer_shutdown(void *arg, int howto); 86 static int vtryrecycle(struct vnode *vp); 87 static void vx_lock(struct vnode *vp); 88 static void vx_unlock(struct vnode *vp); 89 static void vgonechrl(struct vnode *vp, struct thread *td); 90 91 92 /* 93 * Number of vnodes in existence. Increased whenever getnewvnode() 94 * allocates a new vnode, never decreased. 95 */ 96 static unsigned long numvnodes; 97 98 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, ""); 99 100 /* 101 * Conversion tables for conversion from vnode types to inode formats 102 * and back. 103 */ 104 enum vtype iftovt_tab[16] = { 105 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, 106 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD, 107 }; 108 int vttoif_tab[9] = { 109 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, 110 S_IFSOCK, S_IFIFO, S_IFMT, 111 }; 112 113 /* 114 * List of vnodes that are ready for recycling. 115 */ 116 static TAILQ_HEAD(freelst, vnode) vnode_free_list; 117 118 /* 119 * Minimum number of free vnodes. If there are fewer than this free vnodes, 120 * getnewvnode() will return a newly allocated vnode. 121 */ 122 static u_long wantfreevnodes = 25; 123 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, ""); 124 /* Number of vnodes in the free list. */ 125 static u_long freevnodes; 126 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, ""); 127 128 /* 129 * Various variables used for debugging the new implementation of 130 * reassignbuf(). 131 * XXX these are probably of (very) limited utility now. 132 */ 133 static int reassignbufcalls; 134 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, ""); 135 static int nameileafonly; 136 SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, ""); 137 138 /* 139 * Cache for the mount type id assigned to NFS. This is used for 140 * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c. 141 */ 142 int nfs_mount_type = -1; 143 144 /* To keep more than one thread at a time from running vfs_getnewfsid */ 145 static struct mtx mntid_mtx; 146 147 /* 148 * Lock for any access to the following: 149 * vnode_free_list 150 * numvnodes 151 * freevnodes 152 */ 153 static struct mtx vnode_free_list_mtx; 154 155 /* 156 * For any iteration/modification of dev->si_hlist (linked through 157 * v_specnext) 158 */ 159 static struct mtx spechash_mtx; 160 161 /* Publicly exported FS */ 162 struct nfs_public nfs_pub; 163 164 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ 165 static uma_zone_t vnode_zone; 166 static uma_zone_t vnodepoll_zone; 167 168 /* Set to 1 to print out reclaim of active vnodes */ 169 int prtactive; 170 171 /* 172 * The workitem queue. 173 * 174 * It is useful to delay writes of file data and filesystem metadata 175 * for tens of seconds so that quickly created and deleted files need 176 * not waste disk bandwidth being created and removed. To realize this, 177 * we append vnodes to a "workitem" queue. When running with a soft 178 * updates implementation, most pending metadata dependencies should 179 * not wait for more than a few seconds. Thus, mounted on block devices 180 * are delayed only about a half the time that file data is delayed. 181 * Similarly, directory updates are more critical, so are only delayed 182 * about a third the time that file data is delayed. Thus, there are 183 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 184 * one each second (driven off the filesystem syncer process). The 185 * syncer_delayno variable indicates the next queue that is to be processed. 186 * Items that need to be processed soon are placed in this queue: 187 * 188 * syncer_workitem_pending[syncer_delayno] 189 * 190 * A delay of fifteen seconds is done by placing the request fifteen 191 * entries later in the queue: 192 * 193 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 194 * 195 */ 196 static int syncer_delayno; 197 static long syncer_mask; 198 LIST_HEAD(synclist, vnode); 199 static struct synclist *syncer_workitem_pending; 200 /* 201 * The sync_mtx protects: 202 * vp->v_synclist 203 * sync_vnode_count 204 * syncer_delayno 205 * syncer_shutdown_iter 206 * syncer_workitem_pending 207 * syncer_worklist_len 208 * rushjob 209 */ 210 static struct mtx sync_mtx; 211 212 #define SYNCER_MAXDELAY 32 213 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 214 static int syncdelay = 30; /* max time to delay syncing data */ 215 static int filedelay = 30; /* time to delay syncing files */ 216 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, ""); 217 static int dirdelay = 29; /* time to delay syncing directories */ 218 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, ""); 219 static int metadelay = 28; /* time to delay syncing metadata */ 220 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, ""); 221 static int rushjob; /* number of slots to run ASAP */ 222 static int stat_rush_requests; /* number of times I/O speeded up */ 223 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, ""); 224 225 /* 226 * Tell the syncer to make three passes through the work list before 227 * shutting down (unless it runs out of work and shuts down sooner). 228 * 229 * Run at 8 times normal speed when shutting down the syncer. With 230 * the default settings, the syncer will take approximately 12 231 * seconds to shut down, which is less than the default 60 timeout 232 * in kproc_shutdown(). 233 */ 234 #define SYNCER_SHUTDOWN_ITER_LIMIT (3*SYNCER_MAXDELAY) 235 #define SYNCER_SHUTDOWN_SPEEDUP 7 236 static int sync_vnode_count; 237 static int syncer_shutdown_iter; 238 static int syncer_worklist_len; 239 240 /* 241 * Number of vnodes we want to exist at any one time. This is mostly used 242 * to size hash tables in vnode-related code. It is normally not used in 243 * getnewvnode(), as wantfreevnodes is normally nonzero.) 244 * 245 * XXX desiredvnodes is historical cruft and should not exist. 246 */ 247 int desiredvnodes; 248 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW, 249 &desiredvnodes, 0, "Maximum number of vnodes"); 250 static int minvnodes; 251 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, 252 &minvnodes, 0, "Minimum number of vnodes"); 253 static int vnlru_nowhere; 254 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, 255 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); 256 257 /* Hook for calling soft updates. */ 258 int (*softdep_process_worklist_hook)(struct mount *); 259 260 /* 261 * Initialize the vnode management data structures. 262 */ 263 static void 264 vntblinit(void *dummy __unused) 265 { 266 267 /* 268 * Desiredvnodes is a function of the physical memory size and 269 * the kernel's heap size. Specifically, desiredvnodes scales 270 * in proportion to the physical memory size until two fifths 271 * of the kernel's heap size is consumed by vnodes and vm 272 * objects. 273 */ 274 desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size / 275 (5 * (sizeof(struct vm_object) + sizeof(struct vnode)))); 276 minvnodes = desiredvnodes / 4; 277 mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF); 278 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); 279 mtx_init(&spechash_mtx, "spechash", NULL, MTX_DEF); 280 TAILQ_INIT(&vnode_free_list); 281 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF); 282 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, 283 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 284 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), 285 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 286 /* 287 * Initialize the filesystem syncer. 288 */ 289 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, 290 &syncer_mask); 291 syncer_maxdelay = syncer_mask + 1; 292 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); 293 } 294 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL) 295 296 297 /* 298 * Mark a mount point as busy. Used to synchronize access and to delay 299 * unmounting. Interlock is not released on failure. 300 */ 301 int 302 vfs_busy(mp, flags, interlkp, td) 303 struct mount *mp; 304 int flags; 305 struct mtx *interlkp; 306 struct thread *td; 307 { 308 int lkflags; 309 310 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 311 if (flags & LK_NOWAIT) 312 return (ENOENT); 313 mp->mnt_kern_flag |= MNTK_MWAIT; 314 /* 315 * Since all busy locks are shared except the exclusive 316 * lock granted when unmounting, the only place that a 317 * wakeup needs to be done is at the release of the 318 * exclusive lock at the end of dounmount. 319 */ 320 msleep(mp, interlkp, PVFS, "vfs_busy", 0); 321 return (ENOENT); 322 } 323 lkflags = LK_SHARED | LK_NOPAUSE; 324 if (interlkp) 325 lkflags |= LK_INTERLOCK; 326 if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td)) 327 panic("vfs_busy: unexpected lock failure"); 328 return (0); 329 } 330 331 /* 332 * Free a busy filesystem. 333 */ 334 void 335 vfs_unbusy(mp, td) 336 struct mount *mp; 337 struct thread *td; 338 { 339 340 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); 341 } 342 343 /* 344 * Lookup a mount point by filesystem identifier. 345 */ 346 struct mount * 347 vfs_getvfs(fsid) 348 fsid_t *fsid; 349 { 350 register struct mount *mp; 351 352 mtx_lock(&mountlist_mtx); 353 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 354 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 355 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 356 mtx_unlock(&mountlist_mtx); 357 return (mp); 358 } 359 } 360 mtx_unlock(&mountlist_mtx); 361 return ((struct mount *) 0); 362 } 363 364 /* 365 * Get a new unique fsid. Try to make its val[0] unique, since this value 366 * will be used to create fake device numbers for stat(). Also try (but 367 * not so hard) make its val[0] unique mod 2^16, since some emulators only 368 * support 16-bit device numbers. We end up with unique val[0]'s for the 369 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 370 * 371 * Keep in mind that several mounts may be running in parallel. Starting 372 * the search one past where the previous search terminated is both a 373 * micro-optimization and a defense against returning the same fsid to 374 * different mounts. 375 */ 376 void 377 vfs_getnewfsid(mp) 378 struct mount *mp; 379 { 380 static u_int16_t mntid_base; 381 fsid_t tfsid; 382 int mtype; 383 384 mtx_lock(&mntid_mtx); 385 mtype = mp->mnt_vfc->vfc_typenum; 386 tfsid.val[1] = mtype; 387 mtype = (mtype & 0xFF) << 24; 388 for (;;) { 389 tfsid.val[0] = makedev(255, 390 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 391 mntid_base++; 392 if (vfs_getvfs(&tfsid) == NULL) 393 break; 394 } 395 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 396 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 397 mtx_unlock(&mntid_mtx); 398 } 399 400 /* 401 * Knob to control the precision of file timestamps: 402 * 403 * 0 = seconds only; nanoseconds zeroed. 404 * 1 = seconds and nanoseconds, accurate within 1/HZ. 405 * 2 = seconds and nanoseconds, truncated to microseconds. 406 * >=3 = seconds and nanoseconds, maximum precision. 407 */ 408 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; 409 410 static int timestamp_precision = TSP_SEC; 411 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, 412 ×tamp_precision, 0, ""); 413 414 /* 415 * Get a current timestamp. 416 */ 417 void 418 vfs_timestamp(tsp) 419 struct timespec *tsp; 420 { 421 struct timeval tv; 422 423 switch (timestamp_precision) { 424 case TSP_SEC: 425 tsp->tv_sec = time_second; 426 tsp->tv_nsec = 0; 427 break; 428 case TSP_HZ: 429 getnanotime(tsp); 430 break; 431 case TSP_USEC: 432 microtime(&tv); 433 TIMEVAL_TO_TIMESPEC(&tv, tsp); 434 break; 435 case TSP_NSEC: 436 default: 437 nanotime(tsp); 438 break; 439 } 440 } 441 442 /* 443 * Set vnode attributes to VNOVAL 444 */ 445 void 446 vattr_null(vap) 447 register struct vattr *vap; 448 { 449 450 vap->va_type = VNON; 451 vap->va_size = VNOVAL; 452 vap->va_bytes = VNOVAL; 453 vap->va_mode = VNOVAL; 454 vap->va_nlink = VNOVAL; 455 vap->va_uid = VNOVAL; 456 vap->va_gid = VNOVAL; 457 vap->va_fsid = VNOVAL; 458 vap->va_fileid = VNOVAL; 459 vap->va_blocksize = VNOVAL; 460 vap->va_rdev = VNOVAL; 461 vap->va_atime.tv_sec = VNOVAL; 462 vap->va_atime.tv_nsec = VNOVAL; 463 vap->va_mtime.tv_sec = VNOVAL; 464 vap->va_mtime.tv_nsec = VNOVAL; 465 vap->va_ctime.tv_sec = VNOVAL; 466 vap->va_ctime.tv_nsec = VNOVAL; 467 vap->va_birthtime.tv_sec = VNOVAL; 468 vap->va_birthtime.tv_nsec = VNOVAL; 469 vap->va_flags = VNOVAL; 470 vap->va_gen = VNOVAL; 471 vap->va_vaflags = 0; 472 } 473 474 /* 475 * This routine is called when we have too many vnodes. It attempts 476 * to free <count> vnodes and will potentially free vnodes that still 477 * have VM backing store (VM backing store is typically the cause 478 * of a vnode blowout so we want to do this). Therefore, this operation 479 * is not considered cheap. 480 * 481 * A number of conditions may prevent a vnode from being reclaimed. 482 * the buffer cache may have references on the vnode, a directory 483 * vnode may still have references due to the namei cache representing 484 * underlying files, or the vnode may be in active use. It is not 485 * desireable to reuse such vnodes. These conditions may cause the 486 * number of vnodes to reach some minimum value regardless of what 487 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 488 */ 489 static int 490 vlrureclaim(struct mount *mp) 491 { 492 struct vnode *vp; 493 int done; 494 int trigger; 495 int usevnodes; 496 int count; 497 498 /* 499 * Calculate the trigger point, don't allow user 500 * screwups to blow us up. This prevents us from 501 * recycling vnodes with lots of resident pages. We 502 * aren't trying to free memory, we are trying to 503 * free vnodes. 504 */ 505 usevnodes = desiredvnodes; 506 if (usevnodes <= 0) 507 usevnodes = 1; 508 trigger = cnt.v_page_count * 2 / usevnodes; 509 510 done = 0; 511 MNT_ILOCK(mp); 512 count = mp->mnt_nvnodelistsize / 10 + 1; 513 while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) { 514 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 515 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 516 517 if (vp->v_type != VNON && 518 vp->v_type != VBAD && 519 VI_TRYLOCK(vp)) { 520 if (VMIGHTFREE(vp) && /* critical path opt */ 521 (vp->v_object == NULL || 522 vp->v_object->resident_page_count < trigger)) { 523 MNT_IUNLOCK(mp); 524 vgonel(vp, curthread); 525 done++; 526 MNT_ILOCK(mp); 527 } else 528 VI_UNLOCK(vp); 529 } 530 --count; 531 } 532 MNT_IUNLOCK(mp); 533 return done; 534 } 535 536 /* 537 * Attempt to recycle vnodes in a context that is always safe to block. 538 * Calling vlrurecycle() from the bowels of filesystem code has some 539 * interesting deadlock problems. 540 */ 541 static struct proc *vnlruproc; 542 static int vnlruproc_sig; 543 544 static void 545 vnlru_proc(void) 546 { 547 struct mount *mp, *nmp; 548 int done; 549 struct proc *p = vnlruproc; 550 struct thread *td = FIRST_THREAD_IN_PROC(p); 551 552 mtx_lock(&Giant); 553 554 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p, 555 SHUTDOWN_PRI_FIRST); 556 557 for (;;) { 558 kthread_suspend_check(p); 559 mtx_lock(&vnode_free_list_mtx); 560 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { 561 mtx_unlock(&vnode_free_list_mtx); 562 vnlruproc_sig = 0; 563 wakeup(&vnlruproc_sig); 564 tsleep(vnlruproc, PVFS, "vlruwt", hz); 565 continue; 566 } 567 mtx_unlock(&vnode_free_list_mtx); 568 done = 0; 569 mtx_lock(&mountlist_mtx); 570 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 571 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) { 572 nmp = TAILQ_NEXT(mp, mnt_list); 573 continue; 574 } 575 done += vlrureclaim(mp); 576 mtx_lock(&mountlist_mtx); 577 nmp = TAILQ_NEXT(mp, mnt_list); 578 vfs_unbusy(mp, td); 579 } 580 mtx_unlock(&mountlist_mtx); 581 if (done == 0) { 582 #if 0 583 /* These messages are temporary debugging aids */ 584 if (vnlru_nowhere < 5) 585 printf("vnlru process getting nowhere..\n"); 586 else if (vnlru_nowhere == 5) 587 printf("vnlru process messages stopped.\n"); 588 #endif 589 vnlru_nowhere++; 590 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); 591 } 592 } 593 } 594 595 static struct kproc_desc vnlru_kp = { 596 "vnlru", 597 vnlru_proc, 598 &vnlruproc 599 }; 600 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 601 602 603 /* 604 * Routines having to do with the management of the vnode table. 605 */ 606 607 /* 608 * Check to see if a free vnode can be recycled. If it can, 609 * recycle it and return it with the vnode interlock held. 610 */ 611 static int 612 vtryrecycle(struct vnode *vp) 613 { 614 struct thread *td = curthread; 615 vm_object_t object; 616 struct mount *vnmp; 617 int error; 618 619 /* Don't recycle if we can't get the interlock */ 620 if (!VI_TRYLOCK(vp)) 621 return (EWOULDBLOCK); 622 /* 623 * This vnode may found and locked via some other list, if so we 624 * can't recycle it yet. 625 */ 626 if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 627 return (EWOULDBLOCK); 628 /* 629 * Don't recycle if its filesystem is being suspended. 630 */ 631 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { 632 VOP_UNLOCK(vp, 0, td); 633 return (EBUSY); 634 } 635 636 /* 637 * Don't recycle if we still have cached pages. 638 */ 639 if (VOP_GETVOBJECT(vp, &object) == 0) { 640 VM_OBJECT_LOCK(object); 641 if (object->resident_page_count || 642 object->ref_count) { 643 VM_OBJECT_UNLOCK(object); 644 error = EBUSY; 645 goto done; 646 } 647 VM_OBJECT_UNLOCK(object); 648 } 649 if (LIST_FIRST(&vp->v_cache_src)) { 650 /* 651 * note: nameileafonly sysctl is temporary, 652 * for debugging only, and will eventually be 653 * removed. 654 */ 655 if (nameileafonly > 0) { 656 /* 657 * Do not reuse namei-cached directory 658 * vnodes that have cached 659 * subdirectories. 660 */ 661 if (cache_leaf_test(vp) < 0) { 662 error = EISDIR; 663 goto done; 664 } 665 } else if (nameileafonly < 0 || 666 vmiodirenable == 0) { 667 /* 668 * Do not reuse namei-cached directory 669 * vnodes if nameileafonly is -1 or 670 * if VMIO backing for directories is 671 * turned off (otherwise we reuse them 672 * too quickly). 673 */ 674 error = EBUSY; 675 goto done; 676 } 677 } 678 /* 679 * If we got this far, we need to acquire the interlock and see if 680 * anyone picked up this vnode from another list. If not, we will 681 * mark it with XLOCK via vgonel() so that anyone who does find it 682 * will skip over it. 683 */ 684 VI_LOCK(vp); 685 if (VSHOULDBUSY(vp) && (vp->v_iflag & VI_XLOCK) == 0) { 686 VI_UNLOCK(vp); 687 error = EBUSY; 688 goto done; 689 } 690 mtx_lock(&vnode_free_list_mtx); 691 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 692 vp->v_iflag &= ~VI_FREE; 693 mtx_unlock(&vnode_free_list_mtx); 694 vp->v_iflag |= VI_DOOMED; 695 if (vp->v_type != VBAD) { 696 VOP_UNLOCK(vp, 0, td); 697 vgonel(vp, td); 698 VI_LOCK(vp); 699 } else 700 VOP_UNLOCK(vp, 0, td); 701 vn_finished_write(vnmp); 702 return (0); 703 done: 704 VOP_UNLOCK(vp, 0, td); 705 vn_finished_write(vnmp); 706 return (error); 707 } 708 709 /* 710 * Return the next vnode from the free list. 711 */ 712 int 713 getnewvnode(tag, mp, vops, vpp) 714 const char *tag; 715 struct mount *mp; 716 vop_t **vops; 717 struct vnode **vpp; 718 { 719 struct vnode *vp = NULL; 720 struct vpollinfo *pollinfo = NULL; 721 722 mtx_lock(&vnode_free_list_mtx); 723 724 /* 725 * Try to reuse vnodes if we hit the max. This situation only 726 * occurs in certain large-memory (2G+) situations. We cannot 727 * attempt to directly reclaim vnodes due to nasty recursion 728 * problems. 729 */ 730 while (numvnodes - freevnodes > desiredvnodes) { 731 if (vnlruproc_sig == 0) { 732 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 733 wakeup(vnlruproc); 734 } 735 mtx_unlock(&vnode_free_list_mtx); 736 tsleep(&vnlruproc_sig, PVFS, "vlruwk", hz); 737 mtx_lock(&vnode_free_list_mtx); 738 } 739 740 /* 741 * Attempt to reuse a vnode already on the free list, allocating 742 * a new vnode if we can't find one or if we have not reached a 743 * good minimum for good LRU performance. 744 */ 745 746 if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) { 747 int error; 748 int count; 749 750 for (count = 0; count < freevnodes; count++) { 751 vp = TAILQ_FIRST(&vnode_free_list); 752 753 KASSERT(vp->v_usecount == 0 && 754 (vp->v_iflag & VI_DOINGINACT) == 0, 755 ("getnewvnode: free vnode isn't")); 756 757 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 758 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 759 mtx_unlock(&vnode_free_list_mtx); 760 error = vtryrecycle(vp); 761 mtx_lock(&vnode_free_list_mtx); 762 if (error == 0) 763 break; 764 vp = NULL; 765 } 766 } 767 if (vp) { 768 freevnodes--; 769 mtx_unlock(&vnode_free_list_mtx); 770 771 #ifdef INVARIANTS 772 { 773 if (vp->v_data) 774 panic("cleaned vnode isn't"); 775 if (vp->v_numoutput) 776 panic("Clean vnode has pending I/O's"); 777 if (vp->v_writecount != 0) 778 panic("Non-zero write count"); 779 } 780 #endif 781 if ((pollinfo = vp->v_pollinfo) != NULL) { 782 /* 783 * To avoid lock order reversals, the call to 784 * uma_zfree() must be delayed until the vnode 785 * interlock is released. 786 */ 787 vp->v_pollinfo = NULL; 788 } 789 #ifdef MAC 790 mac_destroy_vnode(vp); 791 #endif 792 vp->v_iflag = 0; 793 vp->v_vflag = 0; 794 vp->v_lastw = 0; 795 vp->v_lasta = 0; 796 vp->v_cstart = 0; 797 vp->v_clen = 0; 798 vp->v_socket = 0; 799 lockdestroy(vp->v_vnlock); 800 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 801 KASSERT(vp->v_cleanbufcnt == 0, ("cleanbufcnt not 0")); 802 KASSERT(vp->v_cleanblkroot == NULL, ("cleanblkroot not NULL")); 803 KASSERT(vp->v_dirtybufcnt == 0, ("dirtybufcnt not 0")); 804 KASSERT(vp->v_dirtyblkroot == NULL, ("dirtyblkroot not NULL")); 805 } else { 806 numvnodes++; 807 mtx_unlock(&vnode_free_list_mtx); 808 809 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO); 810 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); 811 VI_LOCK(vp); 812 vp->v_dd = vp; 813 vp->v_vnlock = &vp->v_lock; 814 lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE); 815 cache_purge(vp); /* Sets up v_id. */ 816 LIST_INIT(&vp->v_cache_src); 817 TAILQ_INIT(&vp->v_cache_dst); 818 } 819 820 TAILQ_INIT(&vp->v_cleanblkhd); 821 TAILQ_INIT(&vp->v_dirtyblkhd); 822 vp->v_type = VNON; 823 vp->v_tag = tag; 824 vp->v_op = vops; 825 *vpp = vp; 826 vp->v_usecount = 1; 827 vp->v_data = 0; 828 vp->v_cachedid = -1; 829 VI_UNLOCK(vp); 830 if (pollinfo != NULL) { 831 mtx_destroy(&pollinfo->vpi_lock); 832 uma_zfree(vnodepoll_zone, pollinfo); 833 } 834 #ifdef MAC 835 mac_init_vnode(vp); 836 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) 837 mac_associate_vnode_singlelabel(mp, vp); 838 #endif 839 insmntque(vp, mp); 840 841 return (0); 842 } 843 844 /* 845 * Move a vnode from one mount queue to another. 846 */ 847 static void 848 insmntque(vp, mp) 849 register struct vnode *vp; 850 register struct mount *mp; 851 { 852 853 /* 854 * Delete from old mount point vnode list, if on one. 855 */ 856 if (vp->v_mount != NULL) { 857 MNT_ILOCK(vp->v_mount); 858 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0, 859 ("bad mount point vnode list size")); 860 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes); 861 vp->v_mount->mnt_nvnodelistsize--; 862 MNT_IUNLOCK(vp->v_mount); 863 } 864 /* 865 * Insert into list of vnodes for the new mount point, if available. 866 */ 867 if ((vp->v_mount = mp) != NULL) { 868 MNT_ILOCK(vp->v_mount); 869 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 870 mp->mnt_nvnodelistsize++; 871 MNT_IUNLOCK(vp->v_mount); 872 } 873 } 874 875 /* 876 * Update outstanding I/O count and do wakeup if requested. 877 */ 878 void 879 vwakeup(bp) 880 register struct buf *bp; 881 { 882 register struct vnode *vp; 883 884 bp->b_flags &= ~B_WRITEINPROG; 885 if ((vp = bp->b_vp)) { 886 VI_LOCK(vp); 887 vp->v_numoutput--; 888 if (vp->v_numoutput < 0) 889 panic("vwakeup: neg numoutput"); 890 if ((vp->v_numoutput == 0) && (vp->v_iflag & VI_BWAIT)) { 891 vp->v_iflag &= ~VI_BWAIT; 892 wakeup(&vp->v_numoutput); 893 } 894 VI_UNLOCK(vp); 895 } 896 } 897 898 /* 899 * Flush out and invalidate all buffers associated with a vnode. 900 * Called with the underlying object locked. 901 */ 902 int 903 vinvalbuf(vp, flags, cred, td, slpflag, slptimeo) 904 struct vnode *vp; 905 int flags; 906 struct ucred *cred; 907 struct thread *td; 908 int slpflag, slptimeo; 909 { 910 struct buf *blist; 911 int error; 912 vm_object_t object; 913 914 GIANT_REQUIRED; 915 916 ASSERT_VOP_LOCKED(vp, "vinvalbuf"); 917 918 VI_LOCK(vp); 919 if (flags & V_SAVE) { 920 while (vp->v_numoutput) { 921 vp->v_iflag |= VI_BWAIT; 922 error = msleep(&vp->v_numoutput, VI_MTX(vp), 923 slpflag | (PRIBIO + 1), "vinvlbuf", slptimeo); 924 if (error) { 925 VI_UNLOCK(vp); 926 return (error); 927 } 928 } 929 if (!TAILQ_EMPTY(&vp->v_dirtyblkhd)) { 930 VI_UNLOCK(vp); 931 if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0) 932 return (error); 933 /* 934 * XXX We could save a lock/unlock if this was only 935 * enabled under INVARIANTS 936 */ 937 VI_LOCK(vp); 938 if (vp->v_numoutput > 0 || 939 !TAILQ_EMPTY(&vp->v_dirtyblkhd)) 940 panic("vinvalbuf: dirty bufs"); 941 } 942 } 943 /* 944 * If you alter this loop please notice that interlock is dropped and 945 * reacquired in flushbuflist. Special care is needed to ensure that 946 * no race conditions occur from this. 947 */ 948 for (error = 0;;) { 949 if ((blist = TAILQ_FIRST(&vp->v_cleanblkhd)) != 0 && 950 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 951 if (error) 952 break; 953 continue; 954 } 955 if ((blist = TAILQ_FIRST(&vp->v_dirtyblkhd)) != 0 && 956 flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) { 957 if (error) 958 break; 959 continue; 960 } 961 break; 962 } 963 if (error) { 964 VI_UNLOCK(vp); 965 return (error); 966 } 967 968 /* 969 * Wait for I/O to complete. XXX needs cleaning up. The vnode can 970 * have write I/O in-progress but if there is a VM object then the 971 * VM object can also have read-I/O in-progress. 972 */ 973 do { 974 while (vp->v_numoutput > 0) { 975 vp->v_iflag |= VI_BWAIT; 976 msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vnvlbv", 0); 977 } 978 VI_UNLOCK(vp); 979 if (VOP_GETVOBJECT(vp, &object) == 0) { 980 VM_OBJECT_LOCK(object); 981 vm_object_pip_wait(object, "vnvlbx"); 982 VM_OBJECT_UNLOCK(object); 983 } 984 VI_LOCK(vp); 985 } while (vp->v_numoutput > 0); 986 VI_UNLOCK(vp); 987 988 /* 989 * Destroy the copy in the VM cache, too. 990 */ 991 if (VOP_GETVOBJECT(vp, &object) == 0) { 992 VM_OBJECT_LOCK(object); 993 vm_object_page_remove(object, 0, 0, 994 (flags & V_SAVE) ? TRUE : FALSE); 995 VM_OBJECT_UNLOCK(object); 996 } 997 998 #ifdef INVARIANTS 999 VI_LOCK(vp); 1000 if ((flags & (V_ALT | V_NORMAL)) == 0 && 1001 (!TAILQ_EMPTY(&vp->v_dirtyblkhd) || 1002 !TAILQ_EMPTY(&vp->v_cleanblkhd))) 1003 panic("vinvalbuf: flush failed"); 1004 VI_UNLOCK(vp); 1005 #endif 1006 return (0); 1007 } 1008 1009 /* 1010 * Flush out buffers on the specified list. 1011 * 1012 */ 1013 static int 1014 flushbuflist(blist, flags, vp, slpflag, slptimeo, errorp) 1015 struct buf *blist; 1016 int flags; 1017 struct vnode *vp; 1018 int slpflag, slptimeo; 1019 int *errorp; 1020 { 1021 struct buf *bp, *nbp; 1022 int found, error; 1023 1024 ASSERT_VI_LOCKED(vp, "flushbuflist"); 1025 1026 for (found = 0, bp = blist; bp; bp = nbp) { 1027 nbp = TAILQ_NEXT(bp, b_vnbufs); 1028 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) || 1029 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) { 1030 continue; 1031 } 1032 found += 1; 1033 error = BUF_TIMELOCK(bp, 1034 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp), 1035 "flushbuf", slpflag, slptimeo); 1036 if (error) { 1037 if (error != ENOLCK) 1038 *errorp = error; 1039 goto done; 1040 } 1041 /* 1042 * XXX Since there are no node locks for NFS, I 1043 * believe there is a slight chance that a delayed 1044 * write will occur while sleeping just above, so 1045 * check for it. Note that vfs_bio_awrite expects 1046 * buffers to reside on a queue, while bwrite and 1047 * brelse do not. 1048 */ 1049 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && 1050 (flags & V_SAVE)) { 1051 1052 if (bp->b_vp == vp) { 1053 if (bp->b_flags & B_CLUSTEROK) { 1054 vfs_bio_awrite(bp); 1055 } else { 1056 bremfree(bp); 1057 bp->b_flags |= B_ASYNC; 1058 bwrite(bp); 1059 } 1060 } else { 1061 bremfree(bp); 1062 (void) bwrite(bp); 1063 } 1064 goto done; 1065 } 1066 bremfree(bp); 1067 bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF); 1068 bp->b_flags &= ~B_ASYNC; 1069 brelse(bp); 1070 VI_LOCK(vp); 1071 } 1072 return (found); 1073 done: 1074 VI_LOCK(vp); 1075 return (found); 1076 } 1077 1078 /* 1079 * Truncate a file's buffer and pages to a specified length. This 1080 * is in lieu of the old vinvalbuf mechanism, which performed unneeded 1081 * sync activity. 1082 */ 1083 int 1084 vtruncbuf(vp, cred, td, length, blksize) 1085 register struct vnode *vp; 1086 struct ucred *cred; 1087 struct thread *td; 1088 off_t length; 1089 int blksize; 1090 { 1091 register struct buf *bp; 1092 struct buf *nbp; 1093 int anyfreed; 1094 int trunclbn; 1095 1096 /* 1097 * Round up to the *next* lbn. 1098 */ 1099 trunclbn = (length + blksize - 1) / blksize; 1100 1101 ASSERT_VOP_LOCKED(vp, "vtruncbuf"); 1102 restart: 1103 VI_LOCK(vp); 1104 anyfreed = 1; 1105 for (;anyfreed;) { 1106 anyfreed = 0; 1107 for (bp = TAILQ_FIRST(&vp->v_cleanblkhd); bp; bp = nbp) { 1108 nbp = TAILQ_NEXT(bp, b_vnbufs); 1109 if (bp->b_lblkno >= trunclbn) { 1110 if (BUF_LOCK(bp, 1111 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1112 VI_MTX(vp)) == ENOLCK) 1113 goto restart; 1114 1115 bremfree(bp); 1116 bp->b_flags |= (B_INVAL | B_RELBUF); 1117 bp->b_flags &= ~B_ASYNC; 1118 brelse(bp); 1119 anyfreed = 1; 1120 1121 if (nbp && 1122 (((nbp->b_xflags & BX_VNCLEAN) == 0) || 1123 (nbp->b_vp != vp) || 1124 (nbp->b_flags & B_DELWRI))) { 1125 goto restart; 1126 } 1127 VI_LOCK(vp); 1128 } 1129 } 1130 1131 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1132 nbp = TAILQ_NEXT(bp, b_vnbufs); 1133 if (bp->b_lblkno >= trunclbn) { 1134 if (BUF_LOCK(bp, 1135 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1136 VI_MTX(vp)) == ENOLCK) 1137 goto restart; 1138 bremfree(bp); 1139 bp->b_flags |= (B_INVAL | B_RELBUF); 1140 bp->b_flags &= ~B_ASYNC; 1141 brelse(bp); 1142 anyfreed = 1; 1143 if (nbp && 1144 (((nbp->b_xflags & BX_VNDIRTY) == 0) || 1145 (nbp->b_vp != vp) || 1146 (nbp->b_flags & B_DELWRI) == 0)) { 1147 goto restart; 1148 } 1149 VI_LOCK(vp); 1150 } 1151 } 1152 } 1153 1154 if (length > 0) { 1155 restartsync: 1156 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1157 nbp = TAILQ_NEXT(bp, b_vnbufs); 1158 if (bp->b_lblkno > 0) 1159 continue; 1160 /* 1161 * Since we hold the vnode lock this should only 1162 * fail if we're racing with the buf daemon. 1163 */ 1164 if (BUF_LOCK(bp, 1165 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 1166 VI_MTX(vp)) == ENOLCK) { 1167 goto restart; 1168 } 1169 KASSERT((bp->b_flags & B_DELWRI), 1170 ("buf(%p) on dirty queue without DELWRI", bp)); 1171 1172 bremfree(bp); 1173 bawrite(bp); 1174 VI_LOCK(vp); 1175 goto restartsync; 1176 } 1177 } 1178 1179 while (vp->v_numoutput > 0) { 1180 vp->v_iflag |= VI_BWAIT; 1181 msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vbtrunc", 0); 1182 } 1183 VI_UNLOCK(vp); 1184 vnode_pager_setsize(vp, length); 1185 1186 return (0); 1187 } 1188 1189 /* 1190 * buf_splay() - splay tree core for the clean/dirty list of buffers in 1191 * a vnode. 1192 * 1193 * NOTE: We have to deal with the special case of a background bitmap 1194 * buffer, a situation where two buffers will have the same logical 1195 * block offset. We want (1) only the foreground buffer to be accessed 1196 * in a lookup and (2) must differentiate between the foreground and 1197 * background buffer in the splay tree algorithm because the splay 1198 * tree cannot normally handle multiple entities with the same 'index'. 1199 * We accomplish this by adding differentiating flags to the splay tree's 1200 * numerical domain. 1201 */ 1202 static 1203 struct buf * 1204 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root) 1205 { 1206 struct buf dummy; 1207 struct buf *lefttreemax, *righttreemin, *y; 1208 1209 if (root == NULL) 1210 return (NULL); 1211 lefttreemax = righttreemin = &dummy; 1212 for (;;) { 1213 if (lblkno < root->b_lblkno || 1214 (lblkno == root->b_lblkno && 1215 (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1216 if ((y = root->b_left) == NULL) 1217 break; 1218 if (lblkno < y->b_lblkno) { 1219 /* Rotate right. */ 1220 root->b_left = y->b_right; 1221 y->b_right = root; 1222 root = y; 1223 if ((y = root->b_left) == NULL) 1224 break; 1225 } 1226 /* Link into the new root's right tree. */ 1227 righttreemin->b_left = root; 1228 righttreemin = root; 1229 } else if (lblkno > root->b_lblkno || 1230 (lblkno == root->b_lblkno && 1231 (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) { 1232 if ((y = root->b_right) == NULL) 1233 break; 1234 if (lblkno > y->b_lblkno) { 1235 /* Rotate left. */ 1236 root->b_right = y->b_left; 1237 y->b_left = root; 1238 root = y; 1239 if ((y = root->b_right) == NULL) 1240 break; 1241 } 1242 /* Link into the new root's left tree. */ 1243 lefttreemax->b_right = root; 1244 lefttreemax = root; 1245 } else { 1246 break; 1247 } 1248 root = y; 1249 } 1250 /* Assemble the new root. */ 1251 lefttreemax->b_right = root->b_left; 1252 righttreemin->b_left = root->b_right; 1253 root->b_left = dummy.b_right; 1254 root->b_right = dummy.b_left; 1255 return (root); 1256 } 1257 1258 static 1259 void 1260 buf_vlist_remove(struct buf *bp) 1261 { 1262 struct vnode *vp = bp->b_vp; 1263 struct buf *root; 1264 1265 ASSERT_VI_LOCKED(vp, "buf_vlist_remove"); 1266 if (bp->b_xflags & BX_VNDIRTY) { 1267 if (bp != vp->v_dirtyblkroot) { 1268 root = buf_splay(bp->b_lblkno, bp->b_xflags, 1269 vp->v_dirtyblkroot); 1270 KASSERT(root == bp, 1271 ("splay lookup failed during dirty remove")); 1272 } 1273 if (bp->b_left == NULL) { 1274 root = bp->b_right; 1275 } else { 1276 root = buf_splay(bp->b_lblkno, bp->b_xflags, 1277 bp->b_left); 1278 root->b_right = bp->b_right; 1279 } 1280 vp->v_dirtyblkroot = root; 1281 TAILQ_REMOVE(&vp->v_dirtyblkhd, bp, b_vnbufs); 1282 vp->v_dirtybufcnt--; 1283 } else { 1284 /* KASSERT(bp->b_xflags & BX_VNCLEAN, ("bp wasn't clean")); */ 1285 if (bp != vp->v_cleanblkroot) { 1286 root = buf_splay(bp->b_lblkno, bp->b_xflags, 1287 vp->v_cleanblkroot); 1288 KASSERT(root == bp, 1289 ("splay lookup failed during clean remove")); 1290 } 1291 if (bp->b_left == NULL) { 1292 root = bp->b_right; 1293 } else { 1294 root = buf_splay(bp->b_lblkno, bp->b_xflags, 1295 bp->b_left); 1296 root->b_right = bp->b_right; 1297 } 1298 vp->v_cleanblkroot = root; 1299 TAILQ_REMOVE(&vp->v_cleanblkhd, bp, b_vnbufs); 1300 vp->v_cleanbufcnt--; 1301 } 1302 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); 1303 } 1304 1305 /* 1306 * Add the buffer to the sorted clean or dirty block list using a 1307 * splay tree algorithm. 1308 * 1309 * NOTE: xflags is passed as a constant, optimizing this inline function! 1310 */ 1311 static 1312 void 1313 buf_vlist_add(struct buf *bp, struct vnode *vp, b_xflags_t xflags) 1314 { 1315 struct buf *root; 1316 1317 ASSERT_VI_LOCKED(vp, "buf_vlist_add"); 1318 bp->b_xflags |= xflags; 1319 if (xflags & BX_VNDIRTY) { 1320 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot); 1321 if (root == NULL) { 1322 bp->b_left = NULL; 1323 bp->b_right = NULL; 1324 TAILQ_INSERT_TAIL(&vp->v_dirtyblkhd, bp, b_vnbufs); 1325 } else if (bp->b_lblkno < root->b_lblkno || 1326 (bp->b_lblkno == root->b_lblkno && 1327 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1328 bp->b_left = root->b_left; 1329 bp->b_right = root; 1330 root->b_left = NULL; 1331 TAILQ_INSERT_BEFORE(root, bp, b_vnbufs); 1332 } else { 1333 bp->b_right = root->b_right; 1334 bp->b_left = root; 1335 root->b_right = NULL; 1336 TAILQ_INSERT_AFTER(&vp->v_dirtyblkhd, 1337 root, bp, b_vnbufs); 1338 } 1339 vp->v_dirtybufcnt++; 1340 vp->v_dirtyblkroot = bp; 1341 } else { 1342 /* KASSERT(xflags & BX_VNCLEAN, ("xflags not clean")); */ 1343 root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot); 1344 if (root == NULL) { 1345 bp->b_left = NULL; 1346 bp->b_right = NULL; 1347 TAILQ_INSERT_TAIL(&vp->v_cleanblkhd, bp, b_vnbufs); 1348 } else if (bp->b_lblkno < root->b_lblkno || 1349 (bp->b_lblkno == root->b_lblkno && 1350 (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) { 1351 bp->b_left = root->b_left; 1352 bp->b_right = root; 1353 root->b_left = NULL; 1354 TAILQ_INSERT_BEFORE(root, bp, b_vnbufs); 1355 } else { 1356 bp->b_right = root->b_right; 1357 bp->b_left = root; 1358 root->b_right = NULL; 1359 TAILQ_INSERT_AFTER(&vp->v_cleanblkhd, 1360 root, bp, b_vnbufs); 1361 } 1362 vp->v_cleanbufcnt++; 1363 vp->v_cleanblkroot = bp; 1364 } 1365 } 1366 1367 /* 1368 * Lookup a buffer using the splay tree. Note that we specifically avoid 1369 * shadow buffers used in background bitmap writes. 1370 * 1371 * This code isn't quite efficient as it could be because we are maintaining 1372 * two sorted lists and do not know which list the block resides in. 1373 * 1374 * During a "make buildworld" the desired buffer is found at one of 1375 * the roots more than 60% of the time. Thus, checking both roots 1376 * before performing either splay eliminates unnecessary splays on the 1377 * first tree splayed. 1378 */ 1379 struct buf * 1380 gbincore(struct vnode *vp, daddr_t lblkno) 1381 { 1382 struct buf *bp; 1383 1384 GIANT_REQUIRED; 1385 1386 ASSERT_VI_LOCKED(vp, "gbincore"); 1387 if ((bp = vp->v_cleanblkroot) != NULL && 1388 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1389 return (bp); 1390 if ((bp = vp->v_dirtyblkroot) != NULL && 1391 bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1392 return (bp); 1393 if ((bp = vp->v_cleanblkroot) != NULL) { 1394 vp->v_cleanblkroot = bp = buf_splay(lblkno, 0, bp); 1395 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1396 return (bp); 1397 } 1398 if ((bp = vp->v_dirtyblkroot) != NULL) { 1399 vp->v_dirtyblkroot = bp = buf_splay(lblkno, 0, bp); 1400 if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER)) 1401 return (bp); 1402 } 1403 return (NULL); 1404 } 1405 1406 /* 1407 * Associate a buffer with a vnode. 1408 */ 1409 void 1410 bgetvp(vp, bp) 1411 register struct vnode *vp; 1412 register struct buf *bp; 1413 { 1414 1415 KASSERT(bp->b_vp == NULL, ("bgetvp: not free")); 1416 1417 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, 1418 ("bgetvp: bp already attached! %p", bp)); 1419 1420 ASSERT_VI_LOCKED(vp, "bgetvp"); 1421 vholdl(vp); 1422 bp->b_vp = vp; 1423 bp->b_dev = vn_todev(vp); 1424 /* 1425 * Insert onto list for new vnode. 1426 */ 1427 buf_vlist_add(bp, vp, BX_VNCLEAN); 1428 } 1429 1430 /* 1431 * Disassociate a buffer from a vnode. 1432 */ 1433 void 1434 brelvp(bp) 1435 register struct buf *bp; 1436 { 1437 struct vnode *vp; 1438 1439 KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); 1440 1441 /* 1442 * Delete from old vnode list, if on one. 1443 */ 1444 vp = bp->b_vp; 1445 VI_LOCK(vp); 1446 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) 1447 buf_vlist_remove(bp); 1448 if ((vp->v_iflag & VI_ONWORKLST) && TAILQ_EMPTY(&vp->v_dirtyblkhd)) { 1449 vp->v_iflag &= ~VI_ONWORKLST; 1450 mtx_lock(&sync_mtx); 1451 LIST_REMOVE(vp, v_synclist); 1452 syncer_worklist_len--; 1453 mtx_unlock(&sync_mtx); 1454 } 1455 vdropl(vp); 1456 bp->b_vp = (struct vnode *) 0; 1457 if (bp->b_object) 1458 bp->b_object = NULL; 1459 VI_UNLOCK(vp); 1460 } 1461 1462 /* 1463 * Add an item to the syncer work queue. 1464 */ 1465 static void 1466 vn_syncer_add_to_worklist(struct vnode *vp, int delay) 1467 { 1468 int slot; 1469 1470 ASSERT_VI_LOCKED(vp, "vn_syncer_add_to_worklist"); 1471 1472 mtx_lock(&sync_mtx); 1473 if (vp->v_iflag & VI_ONWORKLST) 1474 LIST_REMOVE(vp, v_synclist); 1475 else { 1476 vp->v_iflag |= VI_ONWORKLST; 1477 syncer_worklist_len++; 1478 } 1479 1480 if (delay > syncer_maxdelay - 2) 1481 delay = syncer_maxdelay - 2; 1482 slot = (syncer_delayno + delay) & syncer_mask; 1483 1484 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist); 1485 mtx_unlock(&sync_mtx); 1486 } 1487 1488 struct proc *updateproc; 1489 static void sched_sync(void); 1490 static struct kproc_desc up_kp = { 1491 "syncer", 1492 sched_sync, 1493 &updateproc 1494 }; 1495 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 1496 1497 /* 1498 * System filesystem synchronizer daemon. 1499 */ 1500 static void 1501 sched_sync(void) 1502 { 1503 struct synclist *next; 1504 struct synclist *slp; 1505 struct vnode *vp; 1506 struct mount *mp; 1507 long starttime; 1508 struct thread *td = FIRST_THREAD_IN_PROC(updateproc); 1509 1510 mtx_lock(&Giant); 1511 1512 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, 1513 SHUTDOWN_PRI_LAST); 1514 1515 for (;;) { 1516 mtx_lock(&sync_mtx); 1517 /* 1518 * Make one more full pass through the work list after 1519 * the only vnodes remaining on the work list are the 1520 * syncer vnodes. 1521 */ 1522 if (syncer_shutdown_iter > SYNCER_MAXDELAY && 1523 syncer_worklist_len == sync_vnode_count) 1524 syncer_shutdown_iter = SYNCER_MAXDELAY; 1525 if (syncer_shutdown_iter == 0) { 1526 mtx_unlock(&sync_mtx); 1527 kthread_suspend_check(td->td_proc); 1528 mtx_lock(&sync_mtx); 1529 } 1530 starttime = time_second; 1531 1532 /* 1533 * Push files whose dirty time has expired. Be careful 1534 * of interrupt race on slp queue. 1535 */ 1536 slp = &syncer_workitem_pending[syncer_delayno]; 1537 syncer_delayno += 1; 1538 if (syncer_delayno == syncer_maxdelay) 1539 syncer_delayno = 0; 1540 next = &syncer_workitem_pending[syncer_delayno]; 1541 1542 while ((vp = LIST_FIRST(slp)) != NULL) { 1543 if (VOP_ISLOCKED(vp, NULL) != 0 || 1544 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1545 LIST_REMOVE(vp, v_synclist); 1546 LIST_INSERT_HEAD(next, vp, v_synclist); 1547 continue; 1548 } 1549 if (VI_TRYLOCK(vp) == 0) { 1550 LIST_REMOVE(vp, v_synclist); 1551 LIST_INSERT_HEAD(next, vp, v_synclist); 1552 vn_finished_write(mp); 1553 continue; 1554 } 1555 /* 1556 * We use vhold in case the vnode does not 1557 * successfully sync. vhold prevents the vnode from 1558 * going away when we unlock the sync_mtx so that 1559 * we can acquire the vnode interlock. 1560 */ 1561 vholdl(vp); 1562 mtx_unlock(&sync_mtx); 1563 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, td); 1564 (void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td); 1565 VOP_UNLOCK(vp, 0, td); 1566 vn_finished_write(mp); 1567 VI_LOCK(vp); 1568 if ((vp->v_iflag & VI_ONWORKLST) != 0) { 1569 /* 1570 * Put us back on the worklist. The worklist 1571 * routine will remove us from our current 1572 * position and then add us back in at a later 1573 * position. 1574 */ 1575 vn_syncer_add_to_worklist(vp, syncdelay); 1576 } 1577 vdropl(vp); 1578 VI_UNLOCK(vp); 1579 mtx_lock(&sync_mtx); 1580 } 1581 if (syncer_shutdown_iter > 0) 1582 syncer_shutdown_iter--; 1583 mtx_unlock(&sync_mtx); 1584 1585 /* 1586 * Do soft update processing. 1587 */ 1588 if (softdep_process_worklist_hook != NULL) 1589 (*softdep_process_worklist_hook)(NULL); 1590 1591 /* 1592 * The variable rushjob allows the kernel to speed up the 1593 * processing of the filesystem syncer process. A rushjob 1594 * value of N tells the filesystem syncer to process the next 1595 * N seconds worth of work on its queue ASAP. Currently rushjob 1596 * is used by the soft update code to speed up the filesystem 1597 * syncer process when the incore state is getting so far 1598 * ahead of the disk that the kernel memory pool is being 1599 * threatened with exhaustion. 1600 */ 1601 mtx_lock(&sync_mtx); 1602 if (rushjob > 0) { 1603 rushjob -= 1; 1604 mtx_unlock(&sync_mtx); 1605 continue; 1606 } else if (syncer_shutdown_iter > 0) 1607 rushjob = SYNCER_SHUTDOWN_SPEEDUP; 1608 mtx_unlock(&sync_mtx); 1609 /* 1610 * If it has taken us less than a second to process the 1611 * current work, then wait. Otherwise start right over 1612 * again. We can still lose time if any single round 1613 * takes more than two seconds, but it does not really 1614 * matter as we are just trying to generally pace the 1615 * filesystem activity. 1616 */ 1617 if (time_second == starttime) 1618 tsleep(&lbolt, PPAUSE, "syncer", 0); 1619 } 1620 } 1621 1622 /* 1623 * Request the syncer daemon to speed up its work. 1624 * We never push it to speed up more than half of its 1625 * normal turn time, otherwise it could take over the cpu. 1626 */ 1627 int 1628 speedup_syncer() 1629 { 1630 struct thread *td; 1631 int ret = 0; 1632 1633 td = FIRST_THREAD_IN_PROC(updateproc); 1634 sleepq_remove(td, &lbolt); 1635 mtx_lock(&sync_mtx); 1636 if (rushjob < syncdelay / 2) { 1637 rushjob += 1; 1638 stat_rush_requests += 1; 1639 ret = 1; 1640 } 1641 mtx_unlock(&sync_mtx); 1642 return (ret); 1643 } 1644 1645 /* 1646 * Tell the syncer to speed up its work and run though its work 1647 * list several times, then tell it to shut down. 1648 */ 1649 static void 1650 syncer_shutdown(void *arg, int howto) 1651 { 1652 struct thread *td; 1653 1654 td = FIRST_THREAD_IN_PROC(updateproc); 1655 sleepq_remove(td, &lbolt); 1656 mtx_lock(&sync_mtx); 1657 if (rushjob < SYNCER_SHUTDOWN_SPEEDUP) 1658 rushjob = SYNCER_SHUTDOWN_SPEEDUP; 1659 syncer_shutdown_iter = SYNCER_SHUTDOWN_ITER_LIMIT; 1660 mtx_unlock(&sync_mtx); 1661 kproc_shutdown(arg, howto); 1662 } 1663 1664 /* 1665 * Associate a p-buffer with a vnode. 1666 * 1667 * Also sets B_PAGING flag to indicate that vnode is not fully associated 1668 * with the buffer. i.e. the bp has not been linked into the vnode or 1669 * ref-counted. 1670 */ 1671 void 1672 pbgetvp(vp, bp) 1673 register struct vnode *vp; 1674 register struct buf *bp; 1675 { 1676 1677 KASSERT(bp->b_vp == NULL, ("pbgetvp: not free")); 1678 1679 bp->b_vp = vp; 1680 bp->b_object = vp->v_object; 1681 bp->b_flags |= B_PAGING; 1682 bp->b_dev = vn_todev(vp); 1683 } 1684 1685 /* 1686 * Disassociate a p-buffer from a vnode. 1687 */ 1688 void 1689 pbrelvp(bp) 1690 register struct buf *bp; 1691 { 1692 1693 KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL")); 1694 1695 /* XXX REMOVE ME */ 1696 VI_LOCK(bp->b_vp); 1697 if (TAILQ_NEXT(bp, b_vnbufs) != NULL) { 1698 panic( 1699 "relpbuf(): b_vp was probably reassignbuf()d %p %x", 1700 bp, 1701 (int)bp->b_flags 1702 ); 1703 } 1704 VI_UNLOCK(bp->b_vp); 1705 bp->b_vp = (struct vnode *) 0; 1706 bp->b_object = NULL; 1707 bp->b_flags &= ~B_PAGING; 1708 } 1709 1710 /* 1711 * Reassign a buffer from one vnode to another. 1712 * Used to assign file specific control information 1713 * (indirect blocks) to the vnode to which they belong. 1714 */ 1715 void 1716 reassignbuf(bp, newvp) 1717 register struct buf *bp; 1718 register struct vnode *newvp; 1719 { 1720 struct vnode *vp; 1721 int delay; 1722 1723 if (newvp == NULL) { 1724 printf("reassignbuf: NULL"); 1725 return; 1726 } 1727 vp = bp->b_vp; 1728 ++reassignbufcalls; 1729 1730 /* 1731 * B_PAGING flagged buffers cannot be reassigned because their vp 1732 * is not fully linked in. 1733 */ 1734 if (bp->b_flags & B_PAGING) 1735 panic("cannot reassign paging buffer"); 1736 1737 /* 1738 * Delete from old vnode list, if on one. 1739 */ 1740 VI_LOCK(vp); 1741 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) { 1742 buf_vlist_remove(bp); 1743 if (vp != newvp) { 1744 vdropl(bp->b_vp); 1745 bp->b_vp = NULL; /* for clarification */ 1746 } 1747 } 1748 if (vp != newvp) { 1749 VI_UNLOCK(vp); 1750 VI_LOCK(newvp); 1751 } 1752 /* 1753 * If dirty, put on list of dirty buffers; otherwise insert onto list 1754 * of clean buffers. 1755 */ 1756 if (bp->b_flags & B_DELWRI) { 1757 if ((newvp->v_iflag & VI_ONWORKLST) == 0) { 1758 switch (newvp->v_type) { 1759 case VDIR: 1760 delay = dirdelay; 1761 break; 1762 case VCHR: 1763 delay = metadelay; 1764 break; 1765 default: 1766 delay = filedelay; 1767 } 1768 vn_syncer_add_to_worklist(newvp, delay); 1769 } 1770 buf_vlist_add(bp, newvp, BX_VNDIRTY); 1771 } else { 1772 buf_vlist_add(bp, newvp, BX_VNCLEAN); 1773 1774 if ((newvp->v_iflag & VI_ONWORKLST) && 1775 TAILQ_EMPTY(&newvp->v_dirtyblkhd)) { 1776 mtx_lock(&sync_mtx); 1777 LIST_REMOVE(newvp, v_synclist); 1778 syncer_worklist_len--; 1779 mtx_unlock(&sync_mtx); 1780 newvp->v_iflag &= ~VI_ONWORKLST; 1781 } 1782 } 1783 if (bp->b_vp != newvp) { 1784 bp->b_vp = newvp; 1785 vholdl(bp->b_vp); 1786 } 1787 VI_UNLOCK(newvp); 1788 } 1789 1790 /* 1791 * Create a vnode for a device. 1792 * Used for mounting the root filesystem. 1793 */ 1794 int 1795 bdevvp(dev, vpp) 1796 struct cdev *dev; 1797 struct vnode **vpp; 1798 { 1799 register struct vnode *vp; 1800 struct vnode *nvp; 1801 int error; 1802 1803 if (dev == NULL) { 1804 *vpp = NULLVP; 1805 return (ENXIO); 1806 } 1807 if (vfinddev(dev, vpp)) 1808 return (0); 1809 1810 error = getnewvnode("none", (struct mount *)0, spec_vnodeop_p, &nvp); 1811 if (error) { 1812 *vpp = NULLVP; 1813 return (error); 1814 } 1815 vp = nvp; 1816 vp->v_type = VCHR; 1817 addalias(vp, dev); 1818 *vpp = vp; 1819 return (0); 1820 } 1821 1822 static void 1823 v_incr_usecount(struct vnode *vp, int delta) 1824 { 1825 1826 vp->v_usecount += delta; 1827 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 1828 mtx_lock(&spechash_mtx); 1829 vp->v_rdev->si_usecount += delta; 1830 mtx_unlock(&spechash_mtx); 1831 } 1832 } 1833 1834 /* 1835 * Add vnode to the alias list hung off the struct cdev *. 1836 * 1837 * The reason for this gunk is that multiple vnodes can reference 1838 * the same physical device, so checking vp->v_usecount to see 1839 * how many users there are is inadequate; the v_usecount for 1840 * the vnodes need to be accumulated. vcount() does that. 1841 */ 1842 struct vnode * 1843 addaliasu(nvp, nvp_rdev) 1844 struct vnode *nvp; 1845 dev_t nvp_rdev; 1846 { 1847 struct vnode *ovp; 1848 vop_t **ops; 1849 struct cdev *dev; 1850 1851 if (nvp->v_type == VBLK) 1852 return (nvp); 1853 if (nvp->v_type != VCHR) 1854 panic("addaliasu on non-special vnode"); 1855 dev = findcdev(nvp_rdev); 1856 if (dev == NULL) 1857 return (nvp); 1858 /* 1859 * Check to see if we have a bdevvp vnode with no associated 1860 * filesystem. If so, we want to associate the filesystem of 1861 * the new newly instigated vnode with the bdevvp vnode and 1862 * discard the newly created vnode rather than leaving the 1863 * bdevvp vnode lying around with no associated filesystem. 1864 */ 1865 if (vfinddev(dev, &ovp) == 0 || ovp->v_data != NULL) { 1866 addalias(nvp, dev); 1867 return (nvp); 1868 } 1869 /* 1870 * Discard unneeded vnode, but save its node specific data. 1871 * Note that if there is a lock, it is carried over in the 1872 * node specific data to the replacement vnode. 1873 */ 1874 vref(ovp); 1875 ovp->v_data = nvp->v_data; 1876 ovp->v_tag = nvp->v_tag; 1877 nvp->v_data = NULL; 1878 lockdestroy(ovp->v_vnlock); 1879 lockinit(ovp->v_vnlock, PVFS, nvp->v_vnlock->lk_wmesg, 1880 nvp->v_vnlock->lk_timo, nvp->v_vnlock->lk_flags & LK_EXTFLG_MASK); 1881 ops = ovp->v_op; 1882 ovp->v_op = nvp->v_op; 1883 if (VOP_ISLOCKED(nvp, curthread)) { 1884 VOP_UNLOCK(nvp, 0, curthread); 1885 vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread); 1886 } 1887 nvp->v_op = ops; 1888 insmntque(ovp, nvp->v_mount); 1889 vrele(nvp); 1890 vgone(nvp); 1891 return (ovp); 1892 } 1893 1894 /* This is a local helper function that do the same as addaliasu, but for a 1895 * struct cdev *instead of an dev_t. */ 1896 static void 1897 addalias(nvp, dev) 1898 struct vnode *nvp; 1899 struct cdev *dev; 1900 { 1901 1902 KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode")); 1903 dev_ref(dev); 1904 nvp->v_rdev = dev; 1905 VI_LOCK(nvp); 1906 mtx_lock(&spechash_mtx); 1907 SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext); 1908 dev->si_usecount += nvp->v_usecount; 1909 mtx_unlock(&spechash_mtx); 1910 VI_UNLOCK(nvp); 1911 } 1912 1913 /* 1914 * Grab a particular vnode from the free list, increment its 1915 * reference count and lock it. The vnode lock bit is set if the 1916 * vnode is being eliminated in vgone. The process is awakened 1917 * when the transition is completed, and an error returned to 1918 * indicate that the vnode is no longer usable (possibly having 1919 * been changed to a new filesystem type). 1920 */ 1921 int 1922 vget(vp, flags, td) 1923 register struct vnode *vp; 1924 int flags; 1925 struct thread *td; 1926 { 1927 int error; 1928 1929 /* 1930 * If the vnode is in the process of being cleaned out for 1931 * another use, we wait for the cleaning to finish and then 1932 * return failure. Cleaning is determined by checking that 1933 * the VI_XLOCK flag is set. 1934 */ 1935 if ((flags & LK_INTERLOCK) == 0) 1936 VI_LOCK(vp); 1937 if (vp->v_iflag & VI_XLOCK && vp->v_vxthread != curthread) { 1938 if ((flags & LK_NOWAIT) == 0) { 1939 vp->v_iflag |= VI_XWANT; 1940 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0); 1941 return (ENOENT); 1942 } 1943 VI_UNLOCK(vp); 1944 return (EBUSY); 1945 } 1946 1947 v_incr_usecount(vp, 1); 1948 1949 if (VSHOULDBUSY(vp)) 1950 vbusy(vp); 1951 if (flags & LK_TYPE_MASK) { 1952 if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) { 1953 /* 1954 * must expand vrele here because we do not want 1955 * to call VOP_INACTIVE if the reference count 1956 * drops back to zero since it was never really 1957 * active. We must remove it from the free list 1958 * before sleeping so that multiple processes do 1959 * not try to recycle it. 1960 */ 1961 VI_LOCK(vp); 1962 v_incr_usecount(vp, -1); 1963 if (VSHOULDFREE(vp)) 1964 vfree(vp); 1965 else 1966 vlruvp(vp); 1967 VI_UNLOCK(vp); 1968 } 1969 return (error); 1970 } 1971 VI_UNLOCK(vp); 1972 return (0); 1973 } 1974 1975 /* 1976 * Increase the reference count of a vnode. 1977 */ 1978 void 1979 vref(struct vnode *vp) 1980 { 1981 1982 VI_LOCK(vp); 1983 v_incr_usecount(vp, 1); 1984 VI_UNLOCK(vp); 1985 } 1986 1987 /* 1988 * Return reference count of a vnode. 1989 * 1990 * The results of this call are only guaranteed when some mechanism other 1991 * than the VI lock is used to stop other processes from gaining references 1992 * to the vnode. This may be the case if the caller holds the only reference. 1993 * This is also useful when stale data is acceptable as race conditions may 1994 * be accounted for by some other means. 1995 */ 1996 int 1997 vrefcnt(struct vnode *vp) 1998 { 1999 int usecnt; 2000 2001 VI_LOCK(vp); 2002 usecnt = vp->v_usecount; 2003 VI_UNLOCK(vp); 2004 2005 return (usecnt); 2006 } 2007 2008 2009 /* 2010 * Vnode put/release. 2011 * If count drops to zero, call inactive routine and return to freelist. 2012 */ 2013 void 2014 vrele(vp) 2015 struct vnode *vp; 2016 { 2017 struct thread *td = curthread; /* XXX */ 2018 2019 GIANT_REQUIRED; 2020 2021 KASSERT(vp != NULL, ("vrele: null vp")); 2022 2023 VI_LOCK(vp); 2024 2025 /* Skip this v_writecount check if we're going to panic below. */ 2026 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2027 ("vrele: missed vn_close")); 2028 2029 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2030 vp->v_usecount == 1)) { 2031 v_incr_usecount(vp, -1); 2032 VI_UNLOCK(vp); 2033 2034 return; 2035 } 2036 2037 if (vp->v_usecount == 1) { 2038 v_incr_usecount(vp, -1); 2039 /* 2040 * We must call VOP_INACTIVE with the node locked. Mark 2041 * as VI_DOINGINACT to avoid recursion. 2042 */ 2043 if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) { 2044 VI_LOCK(vp); 2045 vp->v_iflag |= VI_DOINGINACT; 2046 VI_UNLOCK(vp); 2047 VOP_INACTIVE(vp, td); 2048 VI_LOCK(vp); 2049 KASSERT(vp->v_iflag & VI_DOINGINACT, 2050 ("vrele: lost VI_DOINGINACT")); 2051 vp->v_iflag &= ~VI_DOINGINACT; 2052 } else 2053 VI_LOCK(vp); 2054 if (VSHOULDFREE(vp)) 2055 vfree(vp); 2056 else 2057 vlruvp(vp); 2058 VI_UNLOCK(vp); 2059 2060 } else { 2061 #ifdef DIAGNOSTIC 2062 vprint("vrele: negative ref count", vp); 2063 #endif 2064 VI_UNLOCK(vp); 2065 panic("vrele: negative ref cnt"); 2066 } 2067 } 2068 2069 /* 2070 * Release an already locked vnode. This give the same effects as 2071 * unlock+vrele(), but takes less time and avoids releasing and 2072 * re-aquiring the lock (as vrele() aquires the lock internally.) 2073 */ 2074 void 2075 vput(vp) 2076 struct vnode *vp; 2077 { 2078 struct thread *td = curthread; /* XXX */ 2079 2080 GIANT_REQUIRED; 2081 2082 KASSERT(vp != NULL, ("vput: null vp")); 2083 VI_LOCK(vp); 2084 /* Skip this v_writecount check if we're going to panic below. */ 2085 KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, 2086 ("vput: missed vn_close")); 2087 2088 if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) && 2089 vp->v_usecount == 1)) { 2090 v_incr_usecount(vp, -1); 2091 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2092 return; 2093 } 2094 2095 if (vp->v_usecount == 1) { 2096 v_incr_usecount(vp, -1); 2097 /* 2098 * We must call VOP_INACTIVE with the node locked, so 2099 * we just need to release the vnode mutex. Mark as 2100 * as VI_DOINGINACT to avoid recursion. 2101 */ 2102 vp->v_iflag |= VI_DOINGINACT; 2103 VI_UNLOCK(vp); 2104 VOP_INACTIVE(vp, td); 2105 VI_LOCK(vp); 2106 KASSERT(vp->v_iflag & VI_DOINGINACT, 2107 ("vput: lost VI_DOINGINACT")); 2108 vp->v_iflag &= ~VI_DOINGINACT; 2109 if (VSHOULDFREE(vp)) 2110 vfree(vp); 2111 else 2112 vlruvp(vp); 2113 VI_UNLOCK(vp); 2114 2115 } else { 2116 #ifdef DIAGNOSTIC 2117 vprint("vput: negative ref count", vp); 2118 #endif 2119 panic("vput: negative ref cnt"); 2120 } 2121 } 2122 2123 /* 2124 * Somebody doesn't want the vnode recycled. 2125 */ 2126 void 2127 vhold(struct vnode *vp) 2128 { 2129 2130 VI_LOCK(vp); 2131 vholdl(vp); 2132 VI_UNLOCK(vp); 2133 } 2134 2135 void 2136 vholdl(vp) 2137 register struct vnode *vp; 2138 { 2139 2140 vp->v_holdcnt++; 2141 if (VSHOULDBUSY(vp)) 2142 vbusy(vp); 2143 } 2144 2145 /* 2146 * Note that there is one less who cares about this vnode. vdrop() is the 2147 * opposite of vhold(). 2148 */ 2149 void 2150 vdrop(struct vnode *vp) 2151 { 2152 2153 VI_LOCK(vp); 2154 vdropl(vp); 2155 VI_UNLOCK(vp); 2156 } 2157 2158 void 2159 vdropl(vp) 2160 register struct vnode *vp; 2161 { 2162 2163 if (vp->v_holdcnt <= 0) 2164 panic("vdrop: holdcnt"); 2165 vp->v_holdcnt--; 2166 if (VSHOULDFREE(vp)) 2167 vfree(vp); 2168 else 2169 vlruvp(vp); 2170 } 2171 2172 /* 2173 * Remove any vnodes in the vnode table belonging to mount point mp. 2174 * 2175 * If FORCECLOSE is not specified, there should not be any active ones, 2176 * return error if any are found (nb: this is a user error, not a 2177 * system error). If FORCECLOSE is specified, detach any active vnodes 2178 * that are found. 2179 * 2180 * If WRITECLOSE is set, only flush out regular file vnodes open for 2181 * writing. 2182 * 2183 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. 2184 * 2185 * `rootrefs' specifies the base reference count for the root vnode 2186 * of this filesystem. The root vnode is considered busy if its 2187 * v_usecount exceeds this value. On a successful return, vflush() 2188 * will call vrele() on the root vnode exactly rootrefs times. 2189 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 2190 * be zero. 2191 */ 2192 #ifdef DIAGNOSTIC 2193 static int busyprt = 0; /* print out busy vnodes */ 2194 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 2195 #endif 2196 2197 int 2198 vflush(mp, rootrefs, flags) 2199 struct mount *mp; 2200 int rootrefs; 2201 int flags; 2202 { 2203 struct thread *td = curthread; /* XXX */ 2204 struct vnode *vp, *nvp, *rootvp = NULL; 2205 struct vattr vattr; 2206 int busy = 0, error; 2207 2208 if (rootrefs > 0) { 2209 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 2210 ("vflush: bad args")); 2211 /* 2212 * Get the filesystem root vnode. We can vput() it 2213 * immediately, since with rootrefs > 0, it won't go away. 2214 */ 2215 if ((error = VFS_ROOT(mp, &rootvp)) != 0) 2216 return (error); 2217 vput(rootvp); 2218 2219 } 2220 MNT_ILOCK(mp); 2221 loop: 2222 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp; vp = nvp) { 2223 /* 2224 * Make sure this vnode wasn't reclaimed in getnewvnode(). 2225 * Start over if it has (it won't be on the list anymore). 2226 */ 2227 if (vp->v_mount != mp) 2228 goto loop; 2229 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 2230 2231 VI_LOCK(vp); 2232 MNT_IUNLOCK(mp); 2233 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td); 2234 if (error) { 2235 MNT_ILOCK(mp); 2236 goto loop; 2237 } 2238 /* 2239 * Skip over a vnodes marked VV_SYSTEM. 2240 */ 2241 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { 2242 VOP_UNLOCK(vp, 0, td); 2243 MNT_ILOCK(mp); 2244 continue; 2245 } 2246 /* 2247 * If WRITECLOSE is set, flush out unlinked but still open 2248 * files (even if open only for reading) and regular file 2249 * vnodes open for writing. 2250 */ 2251 if (flags & WRITECLOSE) { 2252 error = VOP_GETATTR(vp, &vattr, td->td_ucred, td); 2253 VI_LOCK(vp); 2254 2255 if ((vp->v_type == VNON || 2256 (error == 0 && vattr.va_nlink > 0)) && 2257 (vp->v_writecount == 0 || vp->v_type != VREG)) { 2258 VOP_UNLOCK(vp, LK_INTERLOCK, td); 2259 MNT_ILOCK(mp); 2260 continue; 2261 } 2262 } else 2263 VI_LOCK(vp); 2264 2265 VOP_UNLOCK(vp, 0, td); 2266 2267 /* 2268 * With v_usecount == 0, all we need to do is clear out the 2269 * vnode data structures and we are done. 2270 */ 2271 if (vp->v_usecount == 0) { 2272 vgonel(vp, td); 2273 MNT_ILOCK(mp); 2274 continue; 2275 } 2276 2277 /* 2278 * If FORCECLOSE is set, forcibly close the vnode. For block 2279 * or character devices, revert to an anonymous device. For 2280 * all other files, just kill them. 2281 */ 2282 if (flags & FORCECLOSE) { 2283 if (vp->v_type != VCHR) 2284 vgonel(vp, td); 2285 else 2286 vgonechrl(vp, td); 2287 MNT_ILOCK(mp); 2288 continue; 2289 } 2290 #ifdef DIAGNOSTIC 2291 if (busyprt) 2292 vprint("vflush: busy vnode", vp); 2293 #endif 2294 VI_UNLOCK(vp); 2295 MNT_ILOCK(mp); 2296 busy++; 2297 } 2298 MNT_IUNLOCK(mp); 2299 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 2300 /* 2301 * If just the root vnode is busy, and if its refcount 2302 * is equal to `rootrefs', then go ahead and kill it. 2303 */ 2304 VI_LOCK(rootvp); 2305 KASSERT(busy > 0, ("vflush: not busy")); 2306 KASSERT(rootvp->v_usecount >= rootrefs, ("vflush: rootrefs")); 2307 if (busy == 1 && rootvp->v_usecount == rootrefs) { 2308 vgonel(rootvp, td); 2309 busy = 0; 2310 } else 2311 VI_UNLOCK(rootvp); 2312 } 2313 if (busy) 2314 return (EBUSY); 2315 for (; rootrefs > 0; rootrefs--) 2316 vrele(rootvp); 2317 return (0); 2318 } 2319 2320 /* 2321 * This moves a now (likely recyclable) vnode to the end of the 2322 * mountlist. XXX However, it is temporarily disabled until we 2323 * can clean up ffs_sync() and friends, which have loop restart 2324 * conditions which this code causes to operate O(N^2). 2325 */ 2326 static void 2327 vlruvp(struct vnode *vp) 2328 { 2329 #if 0 2330 struct mount *mp; 2331 2332 if ((mp = vp->v_mount) != NULL) { 2333 MNT_ILOCK(mp); 2334 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2335 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 2336 MNT_IUNLOCK(mp); 2337 } 2338 #endif 2339 } 2340 2341 static void 2342 vx_lock(struct vnode *vp) 2343 { 2344 2345 ASSERT_VI_LOCKED(vp, "vx_lock"); 2346 2347 /* 2348 * Prevent the vnode from being recycled or brought into use while we 2349 * clean it out. 2350 */ 2351 if (vp->v_iflag & VI_XLOCK) 2352 panic("vclean: deadlock"); 2353 vp->v_iflag |= VI_XLOCK; 2354 vp->v_vxthread = curthread; 2355 } 2356 2357 static void 2358 vx_unlock(struct vnode *vp) 2359 { 2360 ASSERT_VI_LOCKED(vp, "vx_unlock"); 2361 vp->v_iflag &= ~VI_XLOCK; 2362 vp->v_vxthread = NULL; 2363 if (vp->v_iflag & VI_XWANT) { 2364 vp->v_iflag &= ~VI_XWANT; 2365 wakeup(vp); 2366 } 2367 } 2368 2369 /* 2370 * Disassociate the underlying filesystem from a vnode. 2371 */ 2372 static void 2373 vclean(vp, flags, td) 2374 struct vnode *vp; 2375 int flags; 2376 struct thread *td; 2377 { 2378 int active; 2379 2380 ASSERT_VI_LOCKED(vp, "vclean"); 2381 /* 2382 * Check to see if the vnode is in use. If so we have to reference it 2383 * before we clean it out so that its count cannot fall to zero and 2384 * generate a race against ourselves to recycle it. 2385 */ 2386 if ((active = vp->v_usecount)) 2387 v_incr_usecount(vp, 1); 2388 2389 /* 2390 * Even if the count is zero, the VOP_INACTIVE routine may still 2391 * have the object locked while it cleans it out. The VOP_LOCK 2392 * ensures that the VOP_INACTIVE routine is done with its work. 2393 * For active vnodes, it ensures that no other activity can 2394 * occur while the underlying object is being cleaned out. 2395 */ 2396 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2397 2398 /* 2399 * Clean out any buffers associated with the vnode. 2400 * If the flush fails, just toss the buffers. 2401 */ 2402 if (flags & DOCLOSE) { 2403 struct buf *bp; 2404 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 2405 if (bp != NULL) 2406 (void) vn_write_suspend_wait(vp, NULL, V_WAIT); 2407 if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0) 2408 vinvalbuf(vp, 0, NOCRED, td, 0, 0); 2409 } 2410 2411 VOP_DESTROYVOBJECT(vp); 2412 2413 /* 2414 * Any other processes trying to obtain this lock must first 2415 * wait for VXLOCK to clear, then call the new lock operation. 2416 */ 2417 VOP_UNLOCK(vp, 0, td); 2418 2419 /* 2420 * If purging an active vnode, it must be closed and 2421 * deactivated before being reclaimed. Note that the 2422 * VOP_INACTIVE will unlock the vnode. 2423 */ 2424 if (active) { 2425 if (flags & DOCLOSE) 2426 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); 2427 VI_LOCK(vp); 2428 if ((vp->v_iflag & VI_DOINGINACT) == 0) { 2429 vp->v_iflag |= VI_DOINGINACT; 2430 VI_UNLOCK(vp); 2431 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0) 2432 panic("vclean: cannot relock."); 2433 VOP_INACTIVE(vp, td); 2434 VI_LOCK(vp); 2435 KASSERT(vp->v_iflag & VI_DOINGINACT, 2436 ("vclean: lost VI_DOINGINACT")); 2437 vp->v_iflag &= ~VI_DOINGINACT; 2438 } 2439 VI_UNLOCK(vp); 2440 } 2441 /* 2442 * Reclaim the vnode. 2443 */ 2444 if (VOP_RECLAIM(vp, td)) 2445 panic("vclean: cannot reclaim"); 2446 2447 if (active) { 2448 /* 2449 * Inline copy of vrele() since VOP_INACTIVE 2450 * has already been called. 2451 */ 2452 VI_LOCK(vp); 2453 v_incr_usecount(vp, -1); 2454 if (vp->v_usecount <= 0) { 2455 #ifdef INVARIANTS 2456 if (vp->v_usecount < 0 || vp->v_writecount != 0) { 2457 vprint("vclean: bad ref count", vp); 2458 panic("vclean: ref cnt"); 2459 } 2460 #endif 2461 if (VSHOULDFREE(vp)) 2462 vfree(vp); 2463 } 2464 VI_UNLOCK(vp); 2465 } 2466 /* 2467 * Delete from old mount point vnode list. 2468 */ 2469 if (vp->v_mount != NULL) 2470 insmntque(vp, (struct mount *)0); 2471 cache_purge(vp); 2472 VI_LOCK(vp); 2473 if (VSHOULDFREE(vp)) 2474 vfree(vp); 2475 2476 /* 2477 * Done with purge, reset to the standard lock and 2478 * notify sleepers of the grim news. 2479 */ 2480 vp->v_vnlock = &vp->v_lock; 2481 vp->v_op = dead_vnodeop_p; 2482 if (vp->v_pollinfo != NULL) 2483 vn_pollgone(vp); 2484 vp->v_tag = "none"; 2485 } 2486 2487 /* 2488 * Eliminate all activity associated with the requested vnode 2489 * and with all vnodes aliased to the requested vnode. 2490 */ 2491 int 2492 vop_revoke(ap) 2493 struct vop_revoke_args /* { 2494 struct vnode *a_vp; 2495 int a_flags; 2496 } */ *ap; 2497 { 2498 struct vnode *vp, *vq; 2499 struct cdev *dev; 2500 2501 KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke")); 2502 vp = ap->a_vp; 2503 KASSERT((vp->v_type == VCHR), ("vop_revoke: not VCHR")); 2504 2505 VI_LOCK(vp); 2506 /* 2507 * If a vgone (or vclean) is already in progress, 2508 * wait until it is done and return. 2509 */ 2510 if (vp->v_iflag & VI_XLOCK) { 2511 vp->v_iflag |= VI_XWANT; 2512 msleep(vp, VI_MTX(vp), PINOD | PDROP, 2513 "vop_revokeall", 0); 2514 return (0); 2515 } 2516 VI_UNLOCK(vp); 2517 dev = vp->v_rdev; 2518 for (;;) { 2519 mtx_lock(&spechash_mtx); 2520 vq = SLIST_FIRST(&dev->si_hlist); 2521 mtx_unlock(&spechash_mtx); 2522 if (vq == NULL) 2523 break; 2524 vgone(vq); 2525 } 2526 return (0); 2527 } 2528 2529 /* 2530 * Recycle an unused vnode to the front of the free list. 2531 * Release the passed interlock if the vnode will be recycled. 2532 */ 2533 int 2534 vrecycle(vp, inter_lkp, td) 2535 struct vnode *vp; 2536 struct mtx *inter_lkp; 2537 struct thread *td; 2538 { 2539 2540 VI_LOCK(vp); 2541 if (vp->v_usecount == 0) { 2542 if (inter_lkp) { 2543 mtx_unlock(inter_lkp); 2544 } 2545 vgonel(vp, td); 2546 return (1); 2547 } 2548 VI_UNLOCK(vp); 2549 return (0); 2550 } 2551 2552 /* 2553 * Eliminate all activity associated with a vnode 2554 * in preparation for reuse. 2555 */ 2556 void 2557 vgone(vp) 2558 register struct vnode *vp; 2559 { 2560 struct thread *td = curthread; /* XXX */ 2561 2562 VI_LOCK(vp); 2563 vgonel(vp, td); 2564 } 2565 2566 /* 2567 * Disassociate a character device from the its underlying filesystem and 2568 * attach it to spec. This is for use when the chr device is still active 2569 * and the filesystem is going away. 2570 */ 2571 static void 2572 vgonechrl(struct vnode *vp, struct thread *td) 2573 { 2574 ASSERT_VI_LOCKED(vp, "vgonechrl"); 2575 vx_lock(vp); 2576 /* 2577 * This is a custom version of vclean() which does not tearm down 2578 * the bufs or vm objects held by this vnode. This allows filesystems 2579 * to continue using devices which were discovered via another 2580 * filesystem that has been unmounted. 2581 */ 2582 if (vp->v_usecount != 0) { 2583 v_incr_usecount(vp, 1); 2584 /* 2585 * Ensure that no other activity can occur while the 2586 * underlying object is being cleaned out. 2587 */ 2588 VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td); 2589 /* 2590 * Any other processes trying to obtain this lock must first 2591 * wait for VXLOCK to clear, then call the new lock operation. 2592 */ 2593 VOP_UNLOCK(vp, 0, td); 2594 vp->v_vnlock = &vp->v_lock; 2595 vp->v_tag = "orphanchr"; 2596 vp->v_op = spec_vnodeop_p; 2597 if (vp->v_mount != NULL) 2598 insmntque(vp, (struct mount *)0); 2599 cache_purge(vp); 2600 vrele(vp); 2601 VI_LOCK(vp); 2602 } else 2603 vclean(vp, 0, td); 2604 vp->v_op = spec_vnodeop_p; 2605 vx_unlock(vp); 2606 VI_UNLOCK(vp); 2607 } 2608 2609 /* 2610 * vgone, with the vp interlock held. 2611 */ 2612 void 2613 vgonel(vp, td) 2614 struct vnode *vp; 2615 struct thread *td; 2616 { 2617 /* 2618 * If a vgone (or vclean) is already in progress, 2619 * wait until it is done and return. 2620 */ 2621 ASSERT_VI_LOCKED(vp, "vgonel"); 2622 if (vp->v_iflag & VI_XLOCK) { 2623 vp->v_iflag |= VI_XWANT; 2624 msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0); 2625 return; 2626 } 2627 vx_lock(vp); 2628 2629 /* 2630 * Clean out the filesystem specific data. 2631 */ 2632 vclean(vp, DOCLOSE, td); 2633 VI_UNLOCK(vp); 2634 2635 /* 2636 * If special device, remove it from special device alias list 2637 * if it is on one. 2638 */ 2639 VI_LOCK(vp); 2640 if (vp->v_type == VCHR && vp->v_rdev != NULL) { 2641 mtx_lock(&spechash_mtx); 2642 SLIST_REMOVE(&vp->v_rdev->si_hlist, vp, vnode, v_specnext); 2643 vp->v_rdev->si_usecount -= vp->v_usecount; 2644 mtx_unlock(&spechash_mtx); 2645 dev_rel(vp->v_rdev); 2646 vp->v_rdev = NULL; 2647 } 2648 2649 /* 2650 * If it is on the freelist and not already at the head, 2651 * move it to the head of the list. The test of the 2652 * VDOOMED flag and the reference count of zero is because 2653 * it will be removed from the free list by getnewvnode, 2654 * but will not have its reference count incremented until 2655 * after calling vgone. If the reference count were 2656 * incremented first, vgone would (incorrectly) try to 2657 * close the previous instance of the underlying object. 2658 */ 2659 if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) { 2660 mtx_lock(&vnode_free_list_mtx); 2661 if (vp->v_iflag & VI_FREE) { 2662 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 2663 } else { 2664 vp->v_iflag |= VI_FREE; 2665 freevnodes++; 2666 } 2667 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2668 mtx_unlock(&vnode_free_list_mtx); 2669 } 2670 2671 vp->v_type = VBAD; 2672 vx_unlock(vp); 2673 VI_UNLOCK(vp); 2674 } 2675 2676 /* 2677 * Lookup a vnode by device number. 2678 */ 2679 int 2680 vfinddev(dev, vpp) 2681 struct cdev *dev; 2682 struct vnode **vpp; 2683 { 2684 struct vnode *vp; 2685 2686 mtx_lock(&spechash_mtx); 2687 SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) { 2688 *vpp = vp; 2689 mtx_unlock(&spechash_mtx); 2690 return (1); 2691 } 2692 mtx_unlock(&spechash_mtx); 2693 return (0); 2694 } 2695 2696 /* 2697 * Calculate the total number of references to a special device. 2698 */ 2699 int 2700 vcount(vp) 2701 struct vnode *vp; 2702 { 2703 int count; 2704 2705 mtx_lock(&spechash_mtx); 2706 count = vp->v_rdev->si_usecount; 2707 mtx_unlock(&spechash_mtx); 2708 return (count); 2709 } 2710 2711 /* 2712 * Same as above, but using the struct cdev *as argument 2713 */ 2714 int 2715 count_dev(dev) 2716 struct cdev *dev; 2717 { 2718 int count; 2719 2720 mtx_lock(&spechash_mtx); 2721 count = dev->si_usecount; 2722 mtx_unlock(&spechash_mtx); 2723 return(count); 2724 } 2725 2726 /* 2727 * Print out a description of a vnode. 2728 */ 2729 static char *typename[] = 2730 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"}; 2731 2732 void 2733 vprint(label, vp) 2734 char *label; 2735 struct vnode *vp; 2736 { 2737 char buf[96]; 2738 2739 if (label != NULL) 2740 printf("%s: %p: ", label, (void *)vp); 2741 else 2742 printf("%p: ", (void *)vp); 2743 printf("tag %s, type %s, usecount %d, writecount %d, refcount %d,", 2744 vp->v_tag, typename[vp->v_type], vp->v_usecount, 2745 vp->v_writecount, vp->v_holdcnt); 2746 buf[0] = '\0'; 2747 if (vp->v_vflag & VV_ROOT) 2748 strcat(buf, "|VV_ROOT"); 2749 if (vp->v_vflag & VV_TEXT) 2750 strcat(buf, "|VV_TEXT"); 2751 if (vp->v_vflag & VV_SYSTEM) 2752 strcat(buf, "|VV_SYSTEM"); 2753 if (vp->v_iflag & VI_XLOCK) 2754 strcat(buf, "|VI_XLOCK"); 2755 if (vp->v_iflag & VI_XWANT) 2756 strcat(buf, "|VI_XWANT"); 2757 if (vp->v_iflag & VI_BWAIT) 2758 strcat(buf, "|VI_BWAIT"); 2759 if (vp->v_iflag & VI_DOOMED) 2760 strcat(buf, "|VI_DOOMED"); 2761 if (vp->v_iflag & VI_FREE) 2762 strcat(buf, "|VI_FREE"); 2763 if (vp->v_vflag & VV_OBJBUF) 2764 strcat(buf, "|VV_OBJBUF"); 2765 if (buf[0] != '\0') 2766 printf(" flags (%s),", &buf[1]); 2767 lockmgr_printinfo(vp->v_vnlock); 2768 printf("\n"); 2769 if (vp->v_data != NULL) 2770 VOP_PRINT(vp); 2771 } 2772 2773 #ifdef DDB 2774 #include <ddb/ddb.h> 2775 /* 2776 * List all of the locked vnodes in the system. 2777 * Called when debugging the kernel. 2778 */ 2779 DB_SHOW_COMMAND(lockedvnods, lockedvnodes) 2780 { 2781 struct mount *mp, *nmp; 2782 struct vnode *vp; 2783 2784 /* 2785 * Note: because this is DDB, we can't obey the locking semantics 2786 * for these structures, which means we could catch an inconsistent 2787 * state and dereference a nasty pointer. Not much to be done 2788 * about that. 2789 */ 2790 printf("Locked vnodes\n"); 2791 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 2792 nmp = TAILQ_NEXT(mp, mnt_list); 2793 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2794 if (VOP_ISLOCKED(vp, NULL)) 2795 vprint(NULL, vp); 2796 } 2797 nmp = TAILQ_NEXT(mp, mnt_list); 2798 } 2799 } 2800 #endif 2801 2802 /* 2803 * Fill in a struct xvfsconf based on a struct vfsconf. 2804 */ 2805 static void 2806 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp) 2807 { 2808 2809 strcpy(xvfsp->vfc_name, vfsp->vfc_name); 2810 xvfsp->vfc_typenum = vfsp->vfc_typenum; 2811 xvfsp->vfc_refcount = vfsp->vfc_refcount; 2812 xvfsp->vfc_flags = vfsp->vfc_flags; 2813 /* 2814 * These are unused in userland, we keep them 2815 * to not break binary compatibility. 2816 */ 2817 xvfsp->vfc_vfsops = NULL; 2818 xvfsp->vfc_next = NULL; 2819 } 2820 2821 /* 2822 * Top level filesystem related information gathering. 2823 */ 2824 static int 2825 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) 2826 { 2827 struct vfsconf *vfsp; 2828 struct xvfsconf *xvfsp; 2829 int cnt, error, i; 2830 2831 cnt = 0; 2832 for (vfsp = vfsconf; vfsp != NULL; vfsp = vfsp->vfc_next) 2833 cnt++; 2834 xvfsp = malloc(sizeof(struct xvfsconf) * cnt, M_TEMP, M_WAITOK); 2835 /* 2836 * Handle the race that we will have here when struct vfsconf 2837 * will be locked down by using both cnt and checking vfc_next 2838 * against NULL to determine the end of the loop. The race will 2839 * happen because we will have to unlock before calling malloc(). 2840 * We are protected by Giant for now. 2841 */ 2842 i = 0; 2843 for (vfsp = vfsconf; vfsp != NULL && i < cnt; vfsp = vfsp->vfc_next) { 2844 vfsconf2x(vfsp, xvfsp + i); 2845 i++; 2846 } 2847 error = SYSCTL_OUT(req, xvfsp, sizeof(struct xvfsconf) * i); 2848 free(xvfsp, M_TEMP); 2849 return (error); 2850 } 2851 2852 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist, 2853 "S,xvfsconf", "List of all configured filesystems"); 2854 2855 #ifndef BURN_BRIDGES 2856 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); 2857 2858 static int 2859 vfs_sysctl(SYSCTL_HANDLER_ARGS) 2860 { 2861 int *name = (int *)arg1 - 1; /* XXX */ 2862 u_int namelen = arg2 + 1; /* XXX */ 2863 struct vfsconf *vfsp; 2864 struct xvfsconf xvfsp; 2865 2866 printf("WARNING: userland calling deprecated sysctl, " 2867 "please rebuild world\n"); 2868 2869 #if 1 || defined(COMPAT_PRELITE2) 2870 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ 2871 if (namelen == 1) 2872 return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); 2873 #endif 2874 2875 switch (name[1]) { 2876 case VFS_MAXTYPENUM: 2877 if (namelen != 2) 2878 return (ENOTDIR); 2879 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); 2880 case VFS_CONF: 2881 if (namelen != 3) 2882 return (ENOTDIR); /* overloaded */ 2883 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) 2884 if (vfsp->vfc_typenum == name[2]) 2885 break; 2886 if (vfsp == NULL) 2887 return (EOPNOTSUPP); 2888 vfsconf2x(vfsp, &xvfsp); 2889 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); 2890 } 2891 return (EOPNOTSUPP); 2892 } 2893 2894 SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl, 2895 "Generic filesystem"); 2896 2897 #if 1 || defined(COMPAT_PRELITE2) 2898 2899 static int 2900 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) 2901 { 2902 int error; 2903 struct vfsconf *vfsp; 2904 struct ovfsconf ovfs; 2905 2906 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) { 2907 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ 2908 strcpy(ovfs.vfc_name, vfsp->vfc_name); 2909 ovfs.vfc_index = vfsp->vfc_typenum; 2910 ovfs.vfc_refcount = vfsp->vfc_refcount; 2911 ovfs.vfc_flags = vfsp->vfc_flags; 2912 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); 2913 if (error) 2914 return error; 2915 } 2916 return 0; 2917 } 2918 2919 #endif /* 1 || COMPAT_PRELITE2 */ 2920 #endif /* !BURN_BRIDGES */ 2921 2922 #define KINFO_VNODESLOP 10 2923 #ifdef notyet 2924 /* 2925 * Dump vnode list (via sysctl). 2926 */ 2927 /* ARGSUSED */ 2928 static int 2929 sysctl_vnode(SYSCTL_HANDLER_ARGS) 2930 { 2931 struct xvnode *xvn; 2932 struct thread *td = req->td; 2933 struct mount *mp; 2934 struct vnode *vp; 2935 int error, len, n; 2936 2937 /* 2938 * Stale numvnodes access is not fatal here. 2939 */ 2940 req->lock = 0; 2941 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; 2942 if (!req->oldptr) 2943 /* Make an estimate */ 2944 return (SYSCTL_OUT(req, 0, len)); 2945 2946 error = sysctl_wire_old_buffer(req, 0); 2947 if (error != 0) 2948 return (error); 2949 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); 2950 n = 0; 2951 mtx_lock(&mountlist_mtx); 2952 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 2953 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 2954 continue; 2955 MNT_ILOCK(mp); 2956 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { 2957 if (n == len) 2958 break; 2959 vref(vp); 2960 xvn[n].xv_size = sizeof *xvn; 2961 xvn[n].xv_vnode = vp; 2962 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field 2963 XV_COPY(usecount); 2964 XV_COPY(writecount); 2965 XV_COPY(holdcnt); 2966 XV_COPY(id); 2967 XV_COPY(mount); 2968 XV_COPY(numoutput); 2969 XV_COPY(type); 2970 #undef XV_COPY 2971 xvn[n].xv_flag = vp->v_vflag; 2972 2973 switch (vp->v_type) { 2974 case VREG: 2975 case VDIR: 2976 case VLNK: 2977 xvn[n].xv_dev = vp->v_cachedfs; 2978 xvn[n].xv_ino = vp->v_cachedid; 2979 break; 2980 case VBLK: 2981 case VCHR: 2982 if (vp->v_rdev == NULL) { 2983 vrele(vp); 2984 continue; 2985 } 2986 xvn[n].xv_dev = dev2udev(vp->v_rdev); 2987 break; 2988 case VSOCK: 2989 xvn[n].xv_socket = vp->v_socket; 2990 break; 2991 case VFIFO: 2992 xvn[n].xv_fifo = vp->v_fifoinfo; 2993 break; 2994 case VNON: 2995 case VBAD: 2996 default: 2997 /* shouldn't happen? */ 2998 vrele(vp); 2999 continue; 3000 } 3001 vrele(vp); 3002 ++n; 3003 } 3004 MNT_IUNLOCK(mp); 3005 mtx_lock(&mountlist_mtx); 3006 vfs_unbusy(mp, td); 3007 if (n == len) 3008 break; 3009 } 3010 mtx_unlock(&mountlist_mtx); 3011 3012 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); 3013 free(xvn, M_TEMP); 3014 return (error); 3015 } 3016 3017 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD, 3018 0, 0, sysctl_vnode, "S,xvnode", ""); 3019 #endif 3020 3021 /* 3022 * Check to see if a filesystem is mounted on a block device. 3023 */ 3024 int 3025 vfs_mountedon(vp) 3026 struct vnode *vp; 3027 { 3028 3029 if (vp->v_rdev->si_mountpoint != NULL) 3030 return (EBUSY); 3031 return (0); 3032 } 3033 3034 /* 3035 * Unmount all filesystems. The list is traversed in reverse order 3036 * of mounting to avoid dependencies. 3037 */ 3038 void 3039 vfs_unmountall() 3040 { 3041 struct mount *mp; 3042 struct thread *td; 3043 int error; 3044 3045 if (curthread != NULL) 3046 td = curthread; 3047 else 3048 td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */ 3049 /* 3050 * Since this only runs when rebooting, it is not interlocked. 3051 */ 3052 while(!TAILQ_EMPTY(&mountlist)) { 3053 mp = TAILQ_LAST(&mountlist, mntlist); 3054 error = dounmount(mp, MNT_FORCE, td); 3055 if (error) { 3056 TAILQ_REMOVE(&mountlist, mp, mnt_list); 3057 printf("unmount of %s failed (", 3058 mp->mnt_stat.f_mntonname); 3059 if (error == EBUSY) 3060 printf("BUSY)\n"); 3061 else 3062 printf("%d)\n", error); 3063 } else { 3064 /* The unmount has removed mp from the mountlist */ 3065 } 3066 } 3067 } 3068 3069 /* 3070 * perform msync on all vnodes under a mount point 3071 * the mount point must be locked. 3072 */ 3073 void 3074 vfs_msync(struct mount *mp, int flags) 3075 { 3076 struct vnode *vp, *nvp; 3077 struct vm_object *obj; 3078 int tries; 3079 3080 GIANT_REQUIRED; 3081 3082 tries = 5; 3083 MNT_ILOCK(mp); 3084 loop: 3085 for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp != NULL; vp = nvp) { 3086 if (vp->v_mount != mp) { 3087 if (--tries > 0) 3088 goto loop; 3089 break; 3090 } 3091 nvp = TAILQ_NEXT(vp, v_nmntvnodes); 3092 3093 VI_LOCK(vp); 3094 if (vp->v_iflag & VI_XLOCK) { 3095 VI_UNLOCK(vp); 3096 continue; 3097 } 3098 3099 if ((vp->v_iflag & VI_OBJDIRTY) && 3100 (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) { 3101 MNT_IUNLOCK(mp); 3102 if (!vget(vp, 3103 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, 3104 curthread)) { 3105 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */ 3106 vput(vp); 3107 MNT_ILOCK(mp); 3108 continue; 3109 } 3110 3111 if (VOP_GETVOBJECT(vp, &obj) == 0) { 3112 VM_OBJECT_LOCK(obj); 3113 vm_object_page_clean(obj, 0, 0, 3114 flags == MNT_WAIT ? 3115 OBJPC_SYNC : OBJPC_NOSYNC); 3116 VM_OBJECT_UNLOCK(obj); 3117 } 3118 vput(vp); 3119 } 3120 MNT_ILOCK(mp); 3121 if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) { 3122 if (--tries > 0) 3123 goto loop; 3124 break; 3125 } 3126 } else 3127 VI_UNLOCK(vp); 3128 } 3129 MNT_IUNLOCK(mp); 3130 } 3131 3132 /* 3133 * Create the VM object needed for VMIO and mmap support. This 3134 * is done for all VREG files in the system. Some filesystems might 3135 * afford the additional metadata buffering capability of the 3136 * VMIO code by making the device node be VMIO mode also. 3137 * 3138 * vp must be locked when vfs_object_create is called. 3139 */ 3140 int 3141 vfs_object_create(vp, td, cred) 3142 struct vnode *vp; 3143 struct thread *td; 3144 struct ucred *cred; 3145 { 3146 3147 GIANT_REQUIRED; 3148 return (VOP_CREATEVOBJECT(vp, cred, td)); 3149 } 3150 3151 /* 3152 * Mark a vnode as free, putting it up for recycling. 3153 */ 3154 void 3155 vfree(vp) 3156 struct vnode *vp; 3157 { 3158 3159 ASSERT_VI_LOCKED(vp, "vfree"); 3160 mtx_lock(&vnode_free_list_mtx); 3161 KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free")); 3162 if (vp->v_iflag & VI_AGE) { 3163 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 3164 } else { 3165 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 3166 } 3167 freevnodes++; 3168 mtx_unlock(&vnode_free_list_mtx); 3169 vp->v_iflag &= ~VI_AGE; 3170 vp->v_iflag |= VI_FREE; 3171 } 3172 3173 /* 3174 * Opposite of vfree() - mark a vnode as in use. 3175 */ 3176 void 3177 vbusy(vp) 3178 struct vnode *vp; 3179 { 3180 3181 ASSERT_VI_LOCKED(vp, "vbusy"); 3182 KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free")); 3183 3184 mtx_lock(&vnode_free_list_mtx); 3185 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 3186 freevnodes--; 3187 mtx_unlock(&vnode_free_list_mtx); 3188 3189 vp->v_iflag &= ~(VI_FREE|VI_AGE); 3190 } 3191 3192 /* 3193 * Initalize per-vnode helper structure to hold poll-related state. 3194 */ 3195 void 3196 v_addpollinfo(struct vnode *vp) 3197 { 3198 3199 vp->v_pollinfo = uma_zalloc(vnodepoll_zone, M_WAITOK); 3200 mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); 3201 } 3202 3203 /* 3204 * Record a process's interest in events which might happen to 3205 * a vnode. Because poll uses the historic select-style interface 3206 * internally, this routine serves as both the ``check for any 3207 * pending events'' and the ``record my interest in future events'' 3208 * functions. (These are done together, while the lock is held, 3209 * to avoid race conditions.) 3210 */ 3211 int 3212 vn_pollrecord(vp, td, events) 3213 struct vnode *vp; 3214 struct thread *td; 3215 short events; 3216 { 3217 3218 if (vp->v_pollinfo == NULL) 3219 v_addpollinfo(vp); 3220 mtx_lock(&vp->v_pollinfo->vpi_lock); 3221 if (vp->v_pollinfo->vpi_revents & events) { 3222 /* 3223 * This leaves events we are not interested 3224 * in available for the other process which 3225 * which presumably had requested them 3226 * (otherwise they would never have been 3227 * recorded). 3228 */ 3229 events &= vp->v_pollinfo->vpi_revents; 3230 vp->v_pollinfo->vpi_revents &= ~events; 3231 3232 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3233 return events; 3234 } 3235 vp->v_pollinfo->vpi_events |= events; 3236 selrecord(td, &vp->v_pollinfo->vpi_selinfo); 3237 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3238 return 0; 3239 } 3240 3241 /* 3242 * Note the occurrence of an event. If the VN_POLLEVENT macro is used, 3243 * it is possible for us to miss an event due to race conditions, but 3244 * that condition is expected to be rare, so for the moment it is the 3245 * preferred interface. 3246 */ 3247 void 3248 vn_pollevent(vp, events) 3249 struct vnode *vp; 3250 short events; 3251 { 3252 3253 if (vp->v_pollinfo == NULL) 3254 v_addpollinfo(vp); 3255 mtx_lock(&vp->v_pollinfo->vpi_lock); 3256 if (vp->v_pollinfo->vpi_events & events) { 3257 /* 3258 * We clear vpi_events so that we don't 3259 * call selwakeup() twice if two events are 3260 * posted before the polling process(es) is 3261 * awakened. This also ensures that we take at 3262 * most one selwakeup() if the polling process 3263 * is no longer interested. However, it does 3264 * mean that only one event can be noticed at 3265 * a time. (Perhaps we should only clear those 3266 * event bits which we note?) XXX 3267 */ 3268 vp->v_pollinfo->vpi_events = 0; /* &= ~events ??? */ 3269 vp->v_pollinfo->vpi_revents |= events; 3270 selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO); 3271 } 3272 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3273 } 3274 3275 /* 3276 * Wake up anyone polling on vp because it is being revoked. 3277 * This depends on dead_poll() returning POLLHUP for correct 3278 * behavior. 3279 */ 3280 void 3281 vn_pollgone(vp) 3282 struct vnode *vp; 3283 { 3284 3285 mtx_lock(&vp->v_pollinfo->vpi_lock); 3286 VN_KNOTE(vp, NOTE_REVOKE); 3287 if (vp->v_pollinfo->vpi_events) { 3288 vp->v_pollinfo->vpi_events = 0; 3289 selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO); 3290 } 3291 mtx_unlock(&vp->v_pollinfo->vpi_lock); 3292 } 3293 3294 3295 3296 /* 3297 * Routine to create and manage a filesystem syncer vnode. 3298 */ 3299 #define sync_close ((int (*)(struct vop_close_args *))nullop) 3300 static int sync_fsync(struct vop_fsync_args *); 3301 static int sync_inactive(struct vop_inactive_args *); 3302 static int sync_reclaim(struct vop_reclaim_args *); 3303 3304 static vop_t **sync_vnodeop_p; 3305 static struct vnodeopv_entry_desc sync_vnodeop_entries[] = { 3306 { &vop_default_desc, (vop_t *) vop_eopnotsupp }, 3307 { &vop_close_desc, (vop_t *) sync_close }, /* close */ 3308 { &vop_fsync_desc, (vop_t *) sync_fsync }, /* fsync */ 3309 { &vop_inactive_desc, (vop_t *) sync_inactive }, /* inactive */ 3310 { &vop_reclaim_desc, (vop_t *) sync_reclaim }, /* reclaim */ 3311 { &vop_lock_desc, (vop_t *) vop_stdlock }, /* lock */ 3312 { &vop_unlock_desc, (vop_t *) vop_stdunlock }, /* unlock */ 3313 { &vop_islocked_desc, (vop_t *) vop_stdislocked }, /* islocked */ 3314 { NULL, NULL } 3315 }; 3316 static struct vnodeopv_desc sync_vnodeop_opv_desc = 3317 { &sync_vnodeop_p, sync_vnodeop_entries }; 3318 3319 VNODEOP_SET(sync_vnodeop_opv_desc); 3320 3321 /* 3322 * Create a new filesystem syncer vnode for the specified mount point. 3323 */ 3324 int 3325 vfs_allocate_syncvnode(mp) 3326 struct mount *mp; 3327 { 3328 struct vnode *vp; 3329 static long start, incr, next; 3330 int error; 3331 3332 /* Allocate a new vnode */ 3333 if ((error = getnewvnode("syncer", mp, sync_vnodeop_p, &vp)) != 0) { 3334 mp->mnt_syncer = NULL; 3335 return (error); 3336 } 3337 vp->v_type = VNON; 3338 /* 3339 * Place the vnode onto the syncer worklist. We attempt to 3340 * scatter them about on the list so that they will go off 3341 * at evenly distributed times even if all the filesystems 3342 * are mounted at once. 3343 */ 3344 next += incr; 3345 if (next == 0 || next > syncer_maxdelay) { 3346 start /= 2; 3347 incr /= 2; 3348 if (start == 0) { 3349 start = syncer_maxdelay / 2; 3350 incr = syncer_maxdelay; 3351 } 3352 next = start; 3353 } 3354 VI_LOCK(vp); 3355 vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0); 3356 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ 3357 mtx_lock(&sync_mtx); 3358 sync_vnode_count++; 3359 mtx_unlock(&sync_mtx); 3360 VI_UNLOCK(vp); 3361 mp->mnt_syncer = vp; 3362 return (0); 3363 } 3364 3365 /* 3366 * Do a lazy sync of the filesystem. 3367 */ 3368 static int 3369 sync_fsync(ap) 3370 struct vop_fsync_args /* { 3371 struct vnode *a_vp; 3372 struct ucred *a_cred; 3373 int a_waitfor; 3374 struct thread *a_td; 3375 } */ *ap; 3376 { 3377 struct vnode *syncvp = ap->a_vp; 3378 struct mount *mp = syncvp->v_mount; 3379 struct thread *td = ap->a_td; 3380 int error, asyncflag; 3381 3382 /* 3383 * We only need to do something if this is a lazy evaluation. 3384 */ 3385 if (ap->a_waitfor != MNT_LAZY) 3386 return (0); 3387 3388 /* 3389 * Move ourselves to the back of the sync list. 3390 */ 3391 VI_LOCK(syncvp); 3392 vn_syncer_add_to_worklist(syncvp, syncdelay); 3393 VI_UNLOCK(syncvp); 3394 3395 /* 3396 * Walk the list of vnodes pushing all that are dirty and 3397 * not already on the sync list. 3398 */ 3399 mtx_lock(&mountlist_mtx); 3400 if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) { 3401 mtx_unlock(&mountlist_mtx); 3402 return (0); 3403 } 3404 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { 3405 vfs_unbusy(mp, td); 3406 return (0); 3407 } 3408 asyncflag = mp->mnt_flag & MNT_ASYNC; 3409 mp->mnt_flag &= ~MNT_ASYNC; 3410 vfs_msync(mp, MNT_NOWAIT); 3411 error = VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td); 3412 if (asyncflag) 3413 mp->mnt_flag |= MNT_ASYNC; 3414 vn_finished_write(mp); 3415 vfs_unbusy(mp, td); 3416 return (error); 3417 } 3418 3419 /* 3420 * The syncer vnode is no referenced. 3421 */ 3422 static int 3423 sync_inactive(ap) 3424 struct vop_inactive_args /* { 3425 struct vnode *a_vp; 3426 struct thread *a_td; 3427 } */ *ap; 3428 { 3429 3430 VOP_UNLOCK(ap->a_vp, 0, ap->a_td); 3431 vgone(ap->a_vp); 3432 return (0); 3433 } 3434 3435 /* 3436 * The syncer vnode is no longer needed and is being decommissioned. 3437 * 3438 * Modifications to the worklist must be protected by sync_mtx. 3439 */ 3440 static int 3441 sync_reclaim(ap) 3442 struct vop_reclaim_args /* { 3443 struct vnode *a_vp; 3444 } */ *ap; 3445 { 3446 struct vnode *vp = ap->a_vp; 3447 3448 VI_LOCK(vp); 3449 vp->v_mount->mnt_syncer = NULL; 3450 if (vp->v_iflag & VI_ONWORKLST) { 3451 mtx_lock(&sync_mtx); 3452 LIST_REMOVE(vp, v_synclist); 3453 syncer_worklist_len--; 3454 sync_vnode_count--; 3455 mtx_unlock(&sync_mtx); 3456 vp->v_iflag &= ~VI_ONWORKLST; 3457 } 3458 VI_UNLOCK(vp); 3459 3460 return (0); 3461 } 3462 3463 /* 3464 * extract the struct cdev *from a VCHR 3465 */ 3466 struct cdev * 3467 vn_todev(vp) 3468 struct vnode *vp; 3469 { 3470 3471 if (vp->v_type != VCHR) 3472 return (NULL); 3473 return (vp->v_rdev); 3474 } 3475 3476 /* 3477 * Check if vnode represents a disk device 3478 */ 3479 int 3480 vn_isdisk(vp, errp) 3481 struct vnode *vp; 3482 int *errp; 3483 { 3484 int error; 3485 3486 error = 0; 3487 if (vp->v_type != VCHR) 3488 error = ENOTBLK; 3489 else if (vp->v_rdev == NULL) 3490 error = ENXIO; 3491 else if (!(devsw(vp->v_rdev)->d_flags & D_DISK)) 3492 error = ENOTBLK; 3493 if (errp != NULL) 3494 *errp = error; 3495 return (error == 0); 3496 } 3497 3498 /* 3499 * Free data allocated by namei(); see namei(9) for details. 3500 */ 3501 void 3502 NDFREE(ndp, flags) 3503 struct nameidata *ndp; 3504 const u_int flags; 3505 { 3506 3507 if (!(flags & NDF_NO_FREE_PNBUF) && 3508 (ndp->ni_cnd.cn_flags & HASBUF)) { 3509 uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf); 3510 ndp->ni_cnd.cn_flags &= ~HASBUF; 3511 } 3512 if (!(flags & NDF_NO_DVP_UNLOCK) && 3513 (ndp->ni_cnd.cn_flags & LOCKPARENT) && 3514 ndp->ni_dvp != ndp->ni_vp) 3515 VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread); 3516 if (!(flags & NDF_NO_DVP_RELE) && 3517 (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) { 3518 vrele(ndp->ni_dvp); 3519 ndp->ni_dvp = NULL; 3520 } 3521 if (!(flags & NDF_NO_VP_UNLOCK) && 3522 (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp) 3523 VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread); 3524 if (!(flags & NDF_NO_VP_RELE) && 3525 ndp->ni_vp) { 3526 vrele(ndp->ni_vp); 3527 ndp->ni_vp = NULL; 3528 } 3529 if (!(flags & NDF_NO_STARTDIR_RELE) && 3530 (ndp->ni_cnd.cn_flags & SAVESTART)) { 3531 vrele(ndp->ni_startdir); 3532 ndp->ni_startdir = NULL; 3533 } 3534 } 3535 3536 /* 3537 * Common filesystem object access control check routine. Accepts a 3538 * vnode's type, "mode", uid and gid, requested access mode, credentials, 3539 * and optional call-by-reference privused argument allowing vaccess() 3540 * to indicate to the caller whether privilege was used to satisfy the 3541 * request (obsoleted). Returns 0 on success, or an errno on failure. 3542 */ 3543 int 3544 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused) 3545 enum vtype type; 3546 mode_t file_mode; 3547 uid_t file_uid; 3548 gid_t file_gid; 3549 mode_t acc_mode; 3550 struct ucred *cred; 3551 int *privused; 3552 { 3553 mode_t dac_granted; 3554 #ifdef CAPABILITIES 3555 mode_t cap_granted; 3556 #endif 3557 3558 /* 3559 * Look for a normal, non-privileged way to access the file/directory 3560 * as requested. If it exists, go with that. 3561 */ 3562 3563 if (privused != NULL) 3564 *privused = 0; 3565 3566 dac_granted = 0; 3567 3568 /* Check the owner. */ 3569 if (cred->cr_uid == file_uid) { 3570 dac_granted |= VADMIN; 3571 if (file_mode & S_IXUSR) 3572 dac_granted |= VEXEC; 3573 if (file_mode & S_IRUSR) 3574 dac_granted |= VREAD; 3575 if (file_mode & S_IWUSR) 3576 dac_granted |= (VWRITE | VAPPEND); 3577 3578 if ((acc_mode & dac_granted) == acc_mode) 3579 return (0); 3580 3581 goto privcheck; 3582 } 3583 3584 /* Otherwise, check the groups (first match) */ 3585 if (groupmember(file_gid, cred)) { 3586 if (file_mode & S_IXGRP) 3587 dac_granted |= VEXEC; 3588 if (file_mode & S_IRGRP) 3589 dac_granted |= VREAD; 3590 if (file_mode & S_IWGRP) 3591 dac_granted |= (VWRITE | VAPPEND); 3592 3593 if ((acc_mode & dac_granted) == acc_mode) 3594 return (0); 3595 3596 goto privcheck; 3597 } 3598 3599 /* Otherwise, check everyone else. */ 3600 if (file_mode & S_IXOTH) 3601 dac_granted |= VEXEC; 3602 if (file_mode & S_IROTH) 3603 dac_granted |= VREAD; 3604 if (file_mode & S_IWOTH) 3605 dac_granted |= (VWRITE | VAPPEND); 3606 if ((acc_mode & dac_granted) == acc_mode) 3607 return (0); 3608 3609 privcheck: 3610 if (!suser_cred(cred, PRISON_ROOT)) { 3611 /* XXX audit: privilege used */ 3612 if (privused != NULL) 3613 *privused = 1; 3614 return (0); 3615 } 3616 3617 #ifdef CAPABILITIES 3618 /* 3619 * Build a capability mask to determine if the set of capabilities 3620 * satisfies the requirements when combined with the granted mask 3621 * from above. 3622 * For each capability, if the capability is required, bitwise 3623 * or the request type onto the cap_granted mask. 3624 */ 3625 cap_granted = 0; 3626 3627 if (type == VDIR) { 3628 /* 3629 * For directories, use CAP_DAC_READ_SEARCH to satisfy 3630 * VEXEC requests, instead of CAP_DAC_EXECUTE. 3631 */ 3632 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3633 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) 3634 cap_granted |= VEXEC; 3635 } else { 3636 if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) && 3637 !cap_check(cred, NULL, CAP_DAC_EXECUTE, PRISON_ROOT)) 3638 cap_granted |= VEXEC; 3639 } 3640 3641 if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) && 3642 !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT)) 3643 cap_granted |= VREAD; 3644 3645 if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) && 3646 !cap_check(cred, NULL, CAP_DAC_WRITE, PRISON_ROOT)) 3647 cap_granted |= (VWRITE | VAPPEND); 3648 3649 if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) && 3650 !cap_check(cred, NULL, CAP_FOWNER, PRISON_ROOT)) 3651 cap_granted |= VADMIN; 3652 3653 if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) { 3654 /* XXX audit: privilege used */ 3655 if (privused != NULL) 3656 *privused = 1; 3657 return (0); 3658 } 3659 #endif 3660 3661 return ((acc_mode & VADMIN) ? EPERM : EACCES); 3662 } 3663 3664 /* 3665 * Credential check based on process requesting service, and per-attribute 3666 * permissions. 3667 */ 3668 int 3669 extattr_check_cred(struct vnode *vp, int attrnamespace, 3670 struct ucred *cred, struct thread *td, int access) 3671 { 3672 3673 /* 3674 * Kernel-invoked always succeeds. 3675 */ 3676 if (cred == NOCRED) 3677 return (0); 3678 3679 /* 3680 * Do not allow privileged processes in jail to directly 3681 * manipulate system attributes. 3682 * 3683 * XXX What capability should apply here? 3684 * Probably CAP_SYS_SETFFLAG. 3685 */ 3686 switch (attrnamespace) { 3687 case EXTATTR_NAMESPACE_SYSTEM: 3688 /* Potentially should be: return (EPERM); */ 3689 return (suser_cred(cred, 0)); 3690 case EXTATTR_NAMESPACE_USER: 3691 return (VOP_ACCESS(vp, access, cred, td)); 3692 default: 3693 return (EPERM); 3694 } 3695 } 3696 3697 #ifdef DEBUG_VFS_LOCKS 3698 /* 3699 * This only exists to supress warnings from unlocked specfs accesses. It is 3700 * no longer ok to have an unlocked VFS. 3701 */ 3702 #define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD) 3703 3704 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ 3705 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ 3706 int vfs_badlock_print = 1; /* Print lock violations. */ 3707 3708 static void 3709 vfs_badlock(const char *msg, const char *str, struct vnode *vp) 3710 { 3711 3712 if (vfs_badlock_print) 3713 printf("%s: %p %s\n", str, (void *)vp, msg); 3714 if (vfs_badlock_ddb) 3715 Debugger("lock violation"); 3716 } 3717 3718 void 3719 assert_vi_locked(struct vnode *vp, const char *str) 3720 { 3721 3722 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) 3723 vfs_badlock("interlock is not locked but should be", str, vp); 3724 } 3725 3726 void 3727 assert_vi_unlocked(struct vnode *vp, const char *str) 3728 { 3729 3730 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) 3731 vfs_badlock("interlock is locked but should not be", str, vp); 3732 } 3733 3734 void 3735 assert_vop_locked(struct vnode *vp, const char *str) 3736 { 3737 3738 if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0) 3739 vfs_badlock("is not locked but should be", str, vp); 3740 } 3741 3742 void 3743 assert_vop_unlocked(struct vnode *vp, const char *str) 3744 { 3745 3746 if (vp && !IGNORE_LOCK(vp) && 3747 VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE) 3748 vfs_badlock("is locked but should not be", str, vp); 3749 } 3750 3751 #if 0 3752 void 3753 assert_vop_elocked(struct vnode *vp, const char *str) 3754 { 3755 3756 if (vp && !IGNORE_LOCK(vp) && 3757 VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE) 3758 vfs_badlock("is not exclusive locked but should be", str, vp); 3759 } 3760 3761 void 3762 assert_vop_elocked_other(struct vnode *vp, const char *str) 3763 { 3764 3765 if (vp && !IGNORE_LOCK(vp) && 3766 VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER) 3767 vfs_badlock("is not exclusive locked by another thread", 3768 str, vp); 3769 } 3770 3771 void 3772 assert_vop_slocked(struct vnode *vp, const char *str) 3773 { 3774 3775 if (vp && !IGNORE_LOCK(vp) && 3776 VOP_ISLOCKED(vp, curthread) != LK_SHARED) 3777 vfs_badlock("is not locked shared but should be", str, vp); 3778 } 3779 #endif /* 0 */ 3780 3781 void 3782 vop_rename_pre(void *ap) 3783 { 3784 struct vop_rename_args *a = ap; 3785 3786 if (a->a_tvp) 3787 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); 3788 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); 3789 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); 3790 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); 3791 3792 /* Check the source (from). */ 3793 if (a->a_tdvp != a->a_fdvp) 3794 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); 3795 if (a->a_tvp != a->a_fvp) 3796 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked"); 3797 3798 /* Check the target. */ 3799 if (a->a_tvp) 3800 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); 3801 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); 3802 } 3803 3804 void 3805 vop_strategy_pre(void *ap) 3806 { 3807 struct vop_strategy_args *a; 3808 struct buf *bp; 3809 3810 a = ap; 3811 bp = a->a_bp; 3812 3813 /* 3814 * Cluster ops lock their component buffers but not the IO container. 3815 */ 3816 if ((bp->b_flags & B_CLUSTER) != 0) 3817 return; 3818 3819 if (BUF_REFCNT(bp) < 1) { 3820 if (vfs_badlock_print) 3821 printf( 3822 "VOP_STRATEGY: bp is not locked but should be\n"); 3823 if (vfs_badlock_ddb) 3824 Debugger("lock violation"); 3825 } 3826 } 3827 3828 void 3829 vop_lookup_pre(void *ap) 3830 { 3831 struct vop_lookup_args *a; 3832 struct vnode *dvp; 3833 3834 a = ap; 3835 dvp = a->a_dvp; 3836 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3837 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP"); 3838 } 3839 3840 void 3841 vop_lookup_post(void *ap, int rc) 3842 { 3843 struct vop_lookup_args *a; 3844 struct componentname *cnp; 3845 struct vnode *dvp; 3846 struct vnode *vp; 3847 int flags; 3848 3849 a = ap; 3850 dvp = a->a_dvp; 3851 cnp = a->a_cnp; 3852 vp = *(a->a_vpp); 3853 flags = cnp->cn_flags; 3854 3855 ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP"); 3856 3857 /* 3858 * If this is the last path component for this lookup and LOCKPARENT 3859 * is set, OR if there is an error the directory has to be locked. 3860 */ 3861 if ((flags & LOCKPARENT) && (flags & ISLASTCN)) 3862 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)"); 3863 else if (rc != 0) 3864 ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)"); 3865 else if (dvp != vp) 3866 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)"); 3867 if (flags & PDIRUNLOCK) 3868 ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)"); 3869 } 3870 3871 void 3872 vop_lock_pre(void *ap) 3873 { 3874 struct vop_lock_args *a = ap; 3875 3876 if ((a->a_flags & LK_INTERLOCK) == 0) 3877 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3878 else 3879 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); 3880 } 3881 3882 void 3883 vop_lock_post(void *ap, int rc) 3884 { 3885 struct vop_lock_args *a = ap; 3886 3887 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); 3888 if (rc == 0) 3889 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); 3890 } 3891 3892 void 3893 vop_unlock_pre(void *ap) 3894 { 3895 struct vop_unlock_args *a = ap; 3896 3897 if (a->a_flags & LK_INTERLOCK) 3898 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK"); 3899 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); 3900 } 3901 3902 void 3903 vop_unlock_post(void *ap, int rc) 3904 { 3905 struct vop_unlock_args *a = ap; 3906 3907 if (a->a_flags & LK_INTERLOCK) 3908 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK"); 3909 } 3910 #endif /* DEBUG_VFS_LOCKS */ 3911