1 /* 2 * Copyright (c) 2004 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * Copyright (c) 1989, 1993 35 * The Regents of the University of California. All rights reserved. 36 * (c) UNIX System Laboratories, Inc. 37 * All or some portions of this file are derived from material licensed 38 * to the University of California by American Telephone and Telegraph 39 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 40 * the permission of UNIX System Laboratories, Inc. 41 * 42 * Redistribution and use in source and binary forms, with or without 43 * modification, are permitted provided that the following conditions 44 * are met: 45 * 1. Redistributions of source code must retain the above copyright 46 * notice, this list of conditions and the following disclaimer. 47 * 2. Redistributions in binary form must reproduce the above copyright 48 * notice, this list of conditions and the following disclaimer in the 49 * documentation and/or other materials provided with the distribution. 50 * 3. All advertising materials mentioning features or use of this software 51 * must display the following acknowledgement: 52 * This product includes software developed by the University of 53 * California, Berkeley and its contributors. 54 * 4. Neither the name of the University nor the names of its contributors 55 * may be used to endorse or promote products derived from this software 56 * without specific prior written permission. 57 * 58 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 59 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 60 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 61 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 62 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 63 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 64 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 65 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 66 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 67 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 68 * SUCH DAMAGE. 69 * 70 * $DragonFly: src/sys/kern/vfs_mount.c,v 1.11 2005/06/06 15:02:28 dillon Exp $ 71 */ 72 73 /* 74 * External virtual filesystem routines 75 */ 76 #include "opt_ddb.h" 77 78 #include <sys/param.h> 79 #include <sys/systm.h> 80 #include <sys/kernel.h> 81 #include <sys/malloc.h> 82 #include <sys/mount.h> 83 #include <sys/proc.h> 84 #include <sys/vnode.h> 85 #include <sys/buf.h> 86 #include <sys/eventhandler.h> 87 #include <sys/kthread.h> 88 #include <sys/sysctl.h> 89 90 #include <machine/limits.h> 91 92 #include <sys/buf2.h> 93 #include <sys/thread2.h> 94 95 #include <vm/vm.h> 96 #include <vm/vm_object.h> 97 98 struct mountscan_info { 99 TAILQ_ENTRY(mountscan_info) msi_entry; 100 int msi_how; 101 struct mount *msi_node; 102 }; 103 104 struct vmntvnodescan_info { 105 TAILQ_ENTRY(vmntvnodescan_info) entry; 106 struct vnode *vp; 107 }; 108 109 static int vnlru_nowhere = 0; 110 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RD, 111 &vnlru_nowhere, 0, 112 "Number of times the vnlru process ran without success"); 113 114 115 static struct lwkt_token mntid_token; 116 117 static struct mntlist mountlist = TAILQ_HEAD_INITIALIZER(mountlist); 118 static TAILQ_HEAD(,mountscan_info) mountscan_list; 119 static struct lwkt_token mountlist_token; 120 static TAILQ_HEAD(,vmntvnodescan_info) mntvnodescan_list; 121 struct lwkt_token mntvnode_token; 122 123 /* 124 * Called from vfsinit() 125 */ 126 void 127 vfs_mount_init(void) 128 { 129 lwkt_token_init(&mountlist_token); 130 lwkt_token_init(&mntvnode_token); 131 lwkt_token_init(&mntid_token); 132 TAILQ_INIT(&mountscan_list); 133 TAILQ_INIT(&mntvnodescan_list); 134 } 135 136 /* 137 * Support function called with mntvnode_token held to remove a vnode 138 * from the mountlist. We must update any list scans which are in progress. 139 */ 140 static void 141 vremovevnodemnt(struct vnode *vp) 142 { 143 struct vmntvnodescan_info *info; 144 145 TAILQ_FOREACH(info, &mntvnodescan_list, entry) { 146 if (info->vp == vp) 147 info->vp = TAILQ_NEXT(vp, v_nmntvnodes); 148 } 149 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes); 150 } 151 152 /* 153 * Support function called with mntvnode_token held to move a vnode to 154 * the end of the list. 155 */ 156 static void 157 vmovevnodetoend(struct mount *mp, struct vnode *vp) 158 { 159 vremovevnodemnt(vp); 160 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 161 } 162 163 164 /* 165 * Allocate a new vnode and associate it with a tag, mount point, and 166 * operations vector. 167 * 168 * A VX locked and refd vnode is returned. The caller should setup the 169 * remaining fields and vx_put() or, if he wishes to leave a vref, 170 * vx_unlock() the vnode. 171 */ 172 int 173 getnewvnode(enum vtagtype tag, struct mount *mp, 174 struct vnode **vpp, int lktimeout, int lkflags) 175 { 176 struct vnode *vp; 177 178 KKASSERT(mp != NULL); 179 180 vp = allocvnode(lktimeout, lkflags); 181 vp->v_tag = tag; 182 vp->v_data = NULL; 183 184 /* 185 * By default the vnode is assigned the mount point's normal 186 * operations vector. 187 */ 188 vp->v_ops = &mp->mnt_vn_use_ops; 189 190 /* 191 * Placing the vnode on the mount point's queue makes it visible. 192 * VNON prevents it from being messed with, however. 193 */ 194 insmntque(vp, mp); 195 vfs_object_create(vp, curthread); 196 197 /* 198 * A VX locked & refd vnode is returned. 199 */ 200 *vpp = vp; 201 return (0); 202 } 203 204 /* 205 * This function creates vnodes with special operations vectors. The 206 * mount point is optional. 207 * 208 * This routine is being phased out. 209 */ 210 int 211 getspecialvnode(enum vtagtype tag, struct mount *mp, 212 struct vop_ops **ops_pp, 213 struct vnode **vpp, int lktimeout, int lkflags) 214 { 215 struct vnode *vp; 216 217 vp = allocvnode(lktimeout, lkflags); 218 vp->v_tag = tag; 219 vp->v_data = NULL; 220 vp->v_ops = ops_pp; 221 222 /* 223 * Placing the vnode on the mount point's queue makes it visible. 224 * VNON prevents it from being messed with, however. 225 */ 226 insmntque(vp, mp); 227 vfs_object_create(vp, curthread); 228 229 /* 230 * A VX locked & refd vnode is returned. 231 */ 232 *vpp = vp; 233 return (0); 234 } 235 236 /* 237 * Interlock against an unmount, return 0 on success, non-zero on failure. 238 * 239 * The passed flag may be 0 or LK_NOWAIT and is only used if an unmount 240 * is in-progress. 241 * 242 * If no unmount is in-progress LK_NOWAIT is ignored. No other flag bits 243 * are used. A shared locked will be obtained and the filesystem will not 244 * be unmountable until the lock is released. 245 */ 246 int 247 vfs_busy(struct mount *mp, int flags, struct thread *td) 248 { 249 int lkflags; 250 251 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 252 if (flags & LK_NOWAIT) 253 return (ENOENT); 254 /* XXX not MP safe */ 255 mp->mnt_kern_flag |= MNTK_MWAIT; 256 /* 257 * Since all busy locks are shared except the exclusive 258 * lock granted when unmounting, the only place that a 259 * wakeup needs to be done is at the release of the 260 * exclusive lock at the end of dounmount. 261 */ 262 tsleep((caddr_t)mp, 0, "vfs_busy", 0); 263 return (ENOENT); 264 } 265 lkflags = LK_SHARED | LK_NOPAUSE; 266 if (lockmgr(&mp->mnt_lock, lkflags, NULL, td)) 267 panic("vfs_busy: unexpected lock failure"); 268 return (0); 269 } 270 271 /* 272 * Free a busy filesystem. 273 */ 274 void 275 vfs_unbusy(struct mount *mp, struct thread *td) 276 { 277 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td); 278 } 279 280 /* 281 * Lookup a filesystem type, and if found allocate and initialize 282 * a mount structure for it. 283 * 284 * Devname is usually updated by mount(8) after booting. 285 */ 286 int 287 vfs_rootmountalloc(char *fstypename, char *devname, struct mount **mpp) 288 { 289 struct thread *td = curthread; /* XXX */ 290 struct vfsconf *vfsp; 291 struct mount *mp; 292 293 if (fstypename == NULL) 294 return (ENODEV); 295 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) { 296 if (!strcmp(vfsp->vfc_name, fstypename)) 297 break; 298 } 299 if (vfsp == NULL) 300 return (ENODEV); 301 mp = malloc(sizeof(struct mount), M_MOUNT, M_WAITOK); 302 bzero((char *)mp, (u_long)sizeof(struct mount)); 303 lockinit(&mp->mnt_lock, 0, "vfslock", VLKTIMEOUT, LK_NOPAUSE); 304 vfs_busy(mp, LK_NOWAIT, td); 305 TAILQ_INIT(&mp->mnt_nvnodelist); 306 TAILQ_INIT(&mp->mnt_reservedvnlist); 307 TAILQ_INIT(&mp->mnt_jlist); 308 mp->mnt_nvnodelistsize = 0; 309 mp->mnt_vfc = vfsp; 310 mp->mnt_op = vfsp->vfc_vfsops; 311 mp->mnt_flag = MNT_RDONLY; 312 mp->mnt_vnodecovered = NULLVP; 313 vfsp->vfc_refcount++; 314 mp->mnt_iosize_max = DFLTPHYS; 315 mp->mnt_stat.f_type = vfsp->vfc_typenum; 316 mp->mnt_flag |= vfsp->vfc_flags & MNT_VISFLAGMASK; 317 strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN); 318 copystr(devname, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, 0); 319 *mpp = mp; 320 return (0); 321 } 322 323 /* 324 * Lookup a mount point by filesystem identifier. 325 */ 326 struct mount * 327 vfs_getvfs(fsid_t *fsid) 328 { 329 struct mount *mp; 330 lwkt_tokref ilock; 331 332 lwkt_gettoken(&ilock, &mountlist_token); 333 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 334 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 335 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 336 break; 337 } 338 } 339 lwkt_reltoken(&ilock); 340 return (mp); 341 } 342 343 /* 344 * Get a new unique fsid. Try to make its val[0] unique, since this value 345 * will be used to create fake device numbers for stat(). Also try (but 346 * not so hard) make its val[0] unique mod 2^16, since some emulators only 347 * support 16-bit device numbers. We end up with unique val[0]'s for the 348 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 349 * 350 * Keep in mind that several mounts may be running in parallel. Starting 351 * the search one past where the previous search terminated is both a 352 * micro-optimization and a defense against returning the same fsid to 353 * different mounts. 354 */ 355 void 356 vfs_getnewfsid(struct mount *mp) 357 { 358 static u_int16_t mntid_base; 359 lwkt_tokref ilock; 360 fsid_t tfsid; 361 int mtype; 362 363 lwkt_gettoken(&ilock, &mntid_token); 364 mtype = mp->mnt_vfc->vfc_typenum; 365 tfsid.val[1] = mtype; 366 mtype = (mtype & 0xFF) << 24; 367 for (;;) { 368 tfsid.val[0] = makeudev(255, 369 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 370 mntid_base++; 371 if (vfs_getvfs(&tfsid) == NULL) 372 break; 373 } 374 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 375 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 376 lwkt_reltoken(&ilock); 377 } 378 379 /* 380 * This routine is called when we have too many vnodes. It attempts 381 * to free <count> vnodes and will potentially free vnodes that still 382 * have VM backing store (VM backing store is typically the cause 383 * of a vnode blowout so we want to do this). Therefore, this operation 384 * is not considered cheap. 385 * 386 * A number of conditions may prevent a vnode from being reclaimed. 387 * the buffer cache may have references on the vnode, a directory 388 * vnode may still have references due to the namei cache representing 389 * underlying files, or the vnode may be in active use. It is not 390 * desireable to reuse such vnodes. These conditions may cause the 391 * number of vnodes to reach some minimum value regardless of what 392 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 393 */ 394 395 /* 396 * This is a quick non-blocking check to determine if the vnode is a good 397 * candidate for being (eventually) vgone()'d. Returns 0 if the vnode is 398 * not a good candidate, 1 if it is. 399 * 400 * vnodes marked VFREE are already on the free list, but may still need 401 * to be recycled due to eating namecache resources and potentially blocking 402 * the namecache directory chain and related vnodes from being freed. 403 */ 404 static __inline int 405 vmightfree(struct vnode *vp, int page_count) 406 { 407 if (vp->v_flag & VRECLAIMED) 408 return (0); 409 if ((vp->v_flag & VFREE) && TAILQ_EMPTY(&vp->v_namecache)) 410 return (0); 411 if (vp->v_usecount != 0) 412 return (0); 413 if (vp->v_object && vp->v_object->resident_page_count >= page_count) 414 return (0); 415 return (1); 416 } 417 418 /* 419 * The vnode was found to be possibly vgone()able and the caller has locked it 420 * (thus the usecount should be 1 now). Determine if the vnode is actually 421 * vgone()able, doing some cleanups in the process. Returns 1 if the vnode 422 * can be vgone()'d, 0 otherwise. 423 * 424 * Note that v_holdcnt may be non-zero because (A) this vnode is not a leaf 425 * in the namecache topology and (B) this vnode has buffer cache bufs. 426 * We cannot remove vnodes with non-leaf namecache associations. We do a 427 * tentitive leaf check prior to attempting to flush out any buffers but the 428 * 'real' test when all is said in done is that v_holdcnt must become 0 for 429 * the vnode to be freeable. 430 * 431 * We could theoretically just unconditionally flush when v_holdcnt != 0, 432 * but flushing data associated with non-leaf nodes (which are always 433 * directories), just throws it away for no benefit. It is the buffer 434 * cache's responsibility to choose buffers to recycle from the cached 435 * data point of view. 436 */ 437 static int 438 visleaf(struct vnode *vp) 439 { 440 struct namecache *ncp; 441 442 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 443 if (!TAILQ_EMPTY(&ncp->nc_list)) 444 return(0); 445 } 446 return(1); 447 } 448 449 /* 450 * Try to clean up the vnode to the point where it can be vgone()'d, returning 451 * 0 if it cannot be vgone()'d (or already has been), 1 if it can. Unlike 452 * vmightfree() this routine may flush the vnode and block. Vnodes marked 453 * VFREE are still candidates for vgone()ing because they may hold namecache 454 * resources and could be blocking the namecache directory hierarchy (and 455 * related vnodes) from being freed. 456 */ 457 static int 458 vtrytomakegoneable(struct vnode *vp, int page_count) 459 { 460 if (vp->v_flag & VRECLAIMED) 461 return (0); 462 if (vp->v_usecount != 1) 463 return (0); 464 if (vp->v_object && vp->v_object->resident_page_count >= page_count) 465 return (0); 466 if (vp->v_holdcnt && visleaf(vp)) { 467 vinvalbuf(vp, V_SAVE, NULL, 0, 0); 468 #if 0 /* DEBUG */ 469 printf((vp->v_holdcnt ? "vrecycle: vp %p failed: %s\n" : 470 "vrecycle: vp %p succeeded: %s\n"), vp, 471 (TAILQ_FIRST(&vp->v_namecache) ? 472 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?")); 473 #endif 474 } 475 return(vp->v_usecount == 1 && vp->v_holdcnt == 0); 476 } 477 478 /* 479 * Reclaim up to 1/10 of the vnodes associated with a mount point. Try 480 * to avoid vnodes which have lots of resident pages (we are trying to free 481 * vnodes, not memory). 482 * 483 * This routine is a callback from the mountlist scan. The mount point 484 * in question will be busied. 485 */ 486 static int 487 vlrureclaim(struct mount *mp, void *data) 488 { 489 struct vnode *vp; 490 lwkt_tokref ilock; 491 int done; 492 int trigger; 493 int usevnodes; 494 int count; 495 int trigger_mult = vnlru_nowhere; 496 497 /* 498 * Calculate the trigger point for the resident pages check. The 499 * minimum trigger value is approximately the number of pages in 500 * the system divded by the number of vnodes. However, due to 501 * various other system memory overheads unrelated to data caching 502 * it is a good idea to double the trigger (at least). 503 * 504 * trigger_mult starts at 0. If the recycler is having problems 505 * finding enough freeable vnodes it will increase trigger_mult. 506 * This should not happen in normal operation, even on machines with 507 * low amounts of memory, but extraordinary memory use by the system 508 * verses the amount of cached data can trigger it. 509 */ 510 usevnodes = desiredvnodes; 511 if (usevnodes <= 0) 512 usevnodes = 1; 513 trigger = vmstats.v_page_count * (trigger_mult + 2) / usevnodes; 514 515 done = 0; 516 lwkt_gettoken(&ilock, &mntvnode_token); 517 count = mp->mnt_nvnodelistsize / 10 + 1; 518 while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) { 519 /* 520 * __VNODESCAN__ 521 * 522 * The VP will stick around while we hold mntvnode_token, 523 * at least until we block, so we can safely do an initial 524 * check, and then must check again after we lock the vnode. 525 */ 526 if (vp->v_type == VNON || /* XXX */ 527 vp->v_type == VBAD || /* XXX */ 528 !vmightfree(vp, trigger) /* critical path opt */ 529 ) { 530 vmovevnodetoend(mp, vp); 531 --count; 532 continue; 533 } 534 535 /* 536 * VX get the candidate vnode. If the VX get fails the 537 * vnode might still be on the mountlist. Our loop depends 538 * on us at least cycling the vnode to the end of the 539 * mountlist. 540 */ 541 if (vx_get_nonblock(vp) != 0) { 542 if (vp->v_mount == mp) 543 vmovevnodetoend(mp, vp); 544 --count; 545 continue; 546 } 547 548 /* 549 * Since we blocked locking the vp, make sure it is still 550 * a candidate for reclamation. That is, it has not already 551 * been reclaimed and only has our VX reference associated 552 * with it. 553 */ 554 if (vp->v_type == VNON || /* XXX */ 555 vp->v_type == VBAD || /* XXX */ 556 (vp->v_flag & VRECLAIMED) || 557 vp->v_mount != mp || 558 !vtrytomakegoneable(vp, trigger) /* critical path opt */ 559 ) { 560 if (vp->v_mount == mp) 561 vmovevnodetoend(mp, vp); 562 --count; 563 vx_put(vp); 564 continue; 565 } 566 567 /* 568 * All right, we are good, move the vp to the end of the 569 * mountlist and clean it out. The vget will have returned 570 * an error if the vnode was destroyed (VRECLAIMED set), so we 571 * do not have to check again. The vput() will move the 572 * vnode to the free list if the vgone() was successful. 573 */ 574 KKASSERT(vp->v_mount == mp); 575 vmovevnodetoend(mp, vp); 576 vgone(vp); 577 vx_put(vp); 578 ++done; 579 --count; 580 } 581 lwkt_reltoken(&ilock); 582 return (done); 583 } 584 585 /* 586 * Attempt to recycle vnodes in a context that is always safe to block. 587 * Calling vlrurecycle() from the bowels of file system code has some 588 * interesting deadlock problems. 589 */ 590 static struct thread *vnlruthread; 591 static int vnlruproc_sig; 592 593 void 594 vnlru_proc_wait(void) 595 { 596 if (vnlruproc_sig == 0) { 597 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 598 wakeup(vnlruthread); 599 } 600 tsleep(&vnlruproc_sig, 0, "vlruwk", hz); 601 } 602 603 static void 604 vnlru_proc(void) 605 { 606 struct thread *td = curthread; 607 int done; 608 609 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td, 610 SHUTDOWN_PRI_FIRST); 611 612 crit_enter(); 613 for (;;) { 614 kproc_suspend_loop(); 615 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { 616 vnlruproc_sig = 0; 617 wakeup(&vnlruproc_sig); 618 tsleep(td, 0, "vlruwt", hz); 619 continue; 620 } 621 cache_cleanneg(0); 622 done = mountlist_scan(vlrureclaim, NULL, MNTSCAN_FORWARD); 623 624 /* 625 * The vlrureclaim() call only processes 1/10 of the vnodes 626 * on each mount. If we couldn't find any repeat the loop 627 * at least enough times to cover all available vnodes before 628 * we start sleeping. Complain if the failure extends past 629 * 30 second, every 30 seconds. 630 */ 631 if (done == 0) { 632 ++vnlru_nowhere; 633 if (vnlru_nowhere % 10 == 0) 634 tsleep(td, 0, "vlrup", hz * 3); 635 if (vnlru_nowhere % 100 == 0) 636 printf("vnlru_proc: vnode recycler stopped working!\n"); 637 if (vnlru_nowhere == 1000) 638 vnlru_nowhere = 900; 639 } else { 640 vnlru_nowhere = 0; 641 } 642 } 643 crit_exit(); 644 } 645 646 /* 647 * MOUNTLIST FUNCTIONS 648 */ 649 650 /* 651 * mountlist_insert (MP SAFE) 652 * 653 * Add a new mount point to the mount list. 654 */ 655 void 656 mountlist_insert(struct mount *mp, int how) 657 { 658 lwkt_tokref ilock; 659 660 lwkt_gettoken(&ilock, &mountlist_token); 661 if (how == MNTINS_FIRST) 662 TAILQ_INSERT_HEAD(&mountlist, mp, mnt_list); 663 else 664 TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list); 665 lwkt_reltoken(&ilock); 666 } 667 668 /* 669 * mountlist_interlock (MP SAFE) 670 * 671 * Execute the specified interlock function with the mountlist token 672 * held. The function will be called in a serialized fashion verses 673 * other functions called through this mechanism. 674 */ 675 int 676 mountlist_interlock(int (*callback)(struct mount *), struct mount *mp) 677 { 678 lwkt_tokref ilock; 679 int error; 680 681 lwkt_gettoken(&ilock, &mountlist_token); 682 error = callback(mp); 683 lwkt_reltoken(&ilock); 684 return (error); 685 } 686 687 /* 688 * mountlist_boot_getfirst (DURING BOOT ONLY) 689 * 690 * This function returns the first mount on the mountlist, which is 691 * expected to be the root mount. Since no interlocks are obtained 692 * this function is only safe to use during booting. 693 */ 694 695 struct mount * 696 mountlist_boot_getfirst(void) 697 { 698 return(TAILQ_FIRST(&mountlist)); 699 } 700 701 /* 702 * mountlist_remove (MP SAFE) 703 * 704 * Remove a node from the mountlist. If this node is the next scan node 705 * for any active mountlist scans, the active mountlist scan will be 706 * adjusted to skip the node, thus allowing removals during mountlist 707 * scans. 708 */ 709 void 710 mountlist_remove(struct mount *mp) 711 { 712 struct mountscan_info *msi; 713 lwkt_tokref ilock; 714 715 lwkt_gettoken(&ilock, &mountlist_token); 716 TAILQ_FOREACH(msi, &mountscan_list, msi_entry) { 717 if (msi->msi_node == mp) { 718 if (msi->msi_how & MNTSCAN_FORWARD) 719 msi->msi_node = TAILQ_NEXT(mp, mnt_list); 720 else 721 msi->msi_node = TAILQ_PREV(mp, mntlist, mnt_list); 722 } 723 } 724 TAILQ_REMOVE(&mountlist, mp, mnt_list); 725 lwkt_reltoken(&ilock); 726 } 727 728 /* 729 * mountlist_scan (MP SAFE) 730 * 731 * Safely scan the mount points on the mount list. Unless otherwise 732 * specified each mount point will be busied prior to the callback and 733 * unbusied afterwords. The callback may safely remove any mount point 734 * without interfering with the scan. If the current callback 735 * mount is removed the scanner will not attempt to unbusy it. 736 * 737 * If a mount node cannot be busied it is silently skipped. 738 * 739 * The callback return value is aggregated and a total is returned. A return 740 * value of < 0 is not aggregated and will terminate the scan. 741 * 742 * MNTSCAN_FORWARD - the mountlist is scanned in the forward direction 743 * MNTSCAN_REVERSE - the mountlist is scanned in reverse 744 * MNTSCAN_NOBUSY - the scanner will make the callback without busying 745 * the mount node. 746 */ 747 int 748 mountlist_scan(int (*callback)(struct mount *, void *), void *data, int how) 749 { 750 struct mountscan_info info; 751 lwkt_tokref ilock; 752 struct mount *mp; 753 thread_t td; 754 int count; 755 int res; 756 757 lwkt_gettoken(&ilock, &mountlist_token); 758 759 info.msi_how = how; 760 info.msi_node = NULL; /* paranoia */ 761 TAILQ_INSERT_TAIL(&mountscan_list, &info, msi_entry); 762 763 res = 0; 764 td = curthread; 765 766 if (how & MNTSCAN_FORWARD) { 767 info.msi_node = TAILQ_FIRST(&mountlist); 768 while ((mp = info.msi_node) != NULL) { 769 if (how & MNTSCAN_NOBUSY) { 770 count = callback(mp, data); 771 } else if (vfs_busy(mp, LK_NOWAIT, td) == 0) { 772 count = callback(mp, data); 773 if (mp == info.msi_node) 774 vfs_unbusy(mp, td); 775 } else { 776 count = 0; 777 } 778 if (count < 0) 779 break; 780 res += count; 781 if (mp == info.msi_node) 782 info.msi_node = TAILQ_NEXT(mp, mnt_list); 783 } 784 } else if (how & MNTSCAN_REVERSE) { 785 info.msi_node = TAILQ_LAST(&mountlist, mntlist); 786 while ((mp = info.msi_node) != NULL) { 787 if (how & MNTSCAN_NOBUSY) { 788 count = callback(mp, data); 789 } else if (vfs_busy(mp, LK_NOWAIT, td) == 0) { 790 count = callback(mp, data); 791 if (mp == info.msi_node) 792 vfs_unbusy(mp, td); 793 } else { 794 count = 0; 795 } 796 if (count < 0) 797 break; 798 res += count; 799 if (mp == info.msi_node) 800 info.msi_node = TAILQ_PREV(mp, mntlist, mnt_list); 801 } 802 } 803 TAILQ_REMOVE(&mountscan_list, &info, msi_entry); 804 lwkt_reltoken(&ilock); 805 return(res); 806 } 807 808 /* 809 * MOUNT RELATED VNODE FUNCTIONS 810 */ 811 812 static struct kproc_desc vnlru_kp = { 813 "vnlru", 814 vnlru_proc, 815 &vnlruthread 816 }; 817 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 818 819 /* 820 * Move a vnode from one mount queue to another. 821 */ 822 void 823 insmntque(struct vnode *vp, struct mount *mp) 824 { 825 lwkt_tokref ilock; 826 827 lwkt_gettoken(&ilock, &mntvnode_token); 828 /* 829 * Delete from old mount point vnode list, if on one. 830 */ 831 if (vp->v_mount != NULL) { 832 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0, 833 ("bad mount point vnode list size")); 834 vremovevnodemnt(vp); 835 vp->v_mount->mnt_nvnodelistsize--; 836 } 837 /* 838 * Insert into list of vnodes for the new mount point, if available. 839 */ 840 if ((vp->v_mount = mp) == NULL) { 841 lwkt_reltoken(&ilock); 842 return; 843 } 844 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 845 mp->mnt_nvnodelistsize++; 846 lwkt_reltoken(&ilock); 847 } 848 849 850 /* 851 * Scan the vnodes under a mount point and issue appropriate callbacks. 852 * 853 * The fastfunc() callback is called with just the mountlist token held 854 * (no vnode lock). It may not block and the vnode may be undergoing 855 * modifications while the caller is processing it. The vnode will 856 * not be entirely destroyed, however, due to the fact that the mountlist 857 * token is held. A return value < 0 skips to the next vnode without calling 858 * the slowfunc(), a return value > 0 terminates the loop. 859 * 860 * The slowfunc() callback is called after the vnode has been successfully 861 * locked based on passed flags. The vnode is skipped if it gets rearranged 862 * or destroyed while blocking on the lock. A non-zero return value from 863 * the slow function terminates the loop. The slow function is allowed to 864 * arbitrarily block. The scanning code guarentees consistency of operation 865 * even if the slow function deletes or moves the node, or blocks and some 866 * other thread deletes or moves the node. 867 */ 868 int 869 vmntvnodescan( 870 struct mount *mp, 871 int flags, 872 int (*fastfunc)(struct mount *mp, struct vnode *vp, void *data), 873 int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data), 874 void *data 875 ) { 876 struct vmntvnodescan_info info; 877 lwkt_tokref ilock; 878 struct vnode *vp; 879 int r = 0; 880 int maxcount = 1000000; 881 882 lwkt_gettoken(&ilock, &mntvnode_token); 883 884 info.vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 885 TAILQ_INSERT_TAIL(&mntvnodescan_list, &info, entry); 886 while ((vp = info.vp) != NULL) { 887 if (--maxcount == 0) 888 panic("maxcount reached during vmntvnodescan"); 889 890 if (vp->v_type == VNON) /* visible but not ready */ 891 goto next; 892 KKASSERT(vp->v_mount == mp); 893 894 /* 895 * Quick test. A negative return continues the loop without 896 * calling the slow test. 0 continues onto the slow test. 897 * A positive number aborts the loop. 898 */ 899 if (fastfunc) { 900 if ((r = fastfunc(mp, vp, data)) < 0) 901 goto next; 902 if (r) 903 break; 904 } 905 906 /* 907 * Get a vxlock on the vnode, retry if it has moved or isn't 908 * in the mountlist where we expect it. 909 */ 910 if (slowfunc) { 911 int error; 912 913 switch(flags) { 914 case VMSC_GETVP: 915 error = vget(vp, LK_EXCLUSIVE, curthread); 916 break; 917 case VMSC_GETVP|VMSC_NOWAIT: 918 error = vget(vp, LK_EXCLUSIVE|LK_NOWAIT, 919 curthread); 920 break; 921 case VMSC_GETVX: 922 error = vx_get(vp); 923 break; 924 case VMSC_REFVP: 925 vref(vp); 926 /* fall through */ 927 default: 928 error = 0; 929 break; 930 } 931 if (error) 932 goto next; 933 /* 934 * Do not call the slow function if the vnode is 935 * invalid or if it was ripped out from under us 936 * while we (potentially) blocked. 937 */ 938 if (info.vp == vp && vp->v_type != VNON) 939 r = slowfunc(mp, vp, data); 940 941 /* 942 * Cleanup 943 */ 944 switch(flags) { 945 case VMSC_GETVP: 946 case VMSC_GETVP|VMSC_NOWAIT: 947 vput(vp); 948 break; 949 case VMSC_GETVX: 950 vx_put(vp); 951 break; 952 case VMSC_REFVP: 953 vrele(vp); 954 /* fall through */ 955 default: 956 break; 957 } 958 if (r != 0) 959 break; 960 } 961 962 /* 963 * Iterate. If the vnode was ripped out from under us 964 * info.vp will already point to the next vnode, otherwise 965 * we have to obtain the next valid vnode ourselves. 966 */ 967 next: 968 if (info.vp == vp) 969 info.vp = TAILQ_NEXT(vp, v_nmntvnodes); 970 } 971 TAILQ_REMOVE(&mntvnodescan_list, &info, entry); 972 lwkt_reltoken(&ilock); 973 return(r); 974 } 975 976 /* 977 * Remove any vnodes in the vnode table belonging to mount point mp. 978 * 979 * If FORCECLOSE is not specified, there should not be any active ones, 980 * return error if any are found (nb: this is a user error, not a 981 * system error). If FORCECLOSE is specified, detach any active vnodes 982 * that are found. 983 * 984 * If WRITECLOSE is set, only flush out regular file vnodes open for 985 * writing. 986 * 987 * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped. 988 * 989 * `rootrefs' specifies the base reference count for the root vnode 990 * of this filesystem. The root vnode is considered busy if its 991 * v_usecount exceeds this value. On a successful return, vflush() 992 * will call vrele() on the root vnode exactly rootrefs times. 993 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 994 * be zero. 995 */ 996 #ifdef DIAGNOSTIC 997 static int busyprt = 0; /* print out busy vnodes */ 998 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 999 #endif 1000 1001 static int vflush_scan(struct mount *mp, struct vnode *vp, void *data); 1002 1003 struct vflush_info { 1004 int flags; 1005 int busy; 1006 thread_t td; 1007 }; 1008 1009 int 1010 vflush(struct mount *mp, int rootrefs, int flags) 1011 { 1012 struct thread *td = curthread; /* XXX */ 1013 struct vnode *rootvp = NULL; 1014 int error; 1015 struct vflush_info vflush_info; 1016 1017 if (rootrefs > 0) { 1018 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 1019 ("vflush: bad args")); 1020 /* 1021 * Get the filesystem root vnode. We can vput() it 1022 * immediately, since with rootrefs > 0, it won't go away. 1023 */ 1024 if ((error = VFS_ROOT(mp, &rootvp)) != 0) 1025 return (error); 1026 vput(rootvp); 1027 } 1028 1029 vflush_info.busy = 0; 1030 vflush_info.flags = flags; 1031 vflush_info.td = td; 1032 vmntvnodescan(mp, VMSC_GETVX, NULL, vflush_scan, &vflush_info); 1033 1034 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 1035 /* 1036 * If just the root vnode is busy, and if its refcount 1037 * is equal to `rootrefs', then go ahead and kill it. 1038 */ 1039 KASSERT(vflush_info.busy > 0, ("vflush: not busy")); 1040 KASSERT(rootvp->v_usecount >= rootrefs, ("vflush: rootrefs")); 1041 if (vflush_info.busy == 1 && rootvp->v_usecount == rootrefs) { 1042 if (vx_lock(rootvp) == 0) { 1043 vgone(rootvp); 1044 vx_unlock(rootvp); 1045 vflush_info.busy = 0; 1046 } 1047 } 1048 } 1049 if (vflush_info.busy) 1050 return (EBUSY); 1051 for (; rootrefs > 0; rootrefs--) 1052 vrele(rootvp); 1053 return (0); 1054 } 1055 1056 /* 1057 * The scan callback is made with an VX locked vnode. 1058 */ 1059 static int 1060 vflush_scan(struct mount *mp, struct vnode *vp, void *data) 1061 { 1062 struct vflush_info *info = data; 1063 struct vattr vattr; 1064 1065 /* 1066 * Skip over a vnodes marked VSYSTEM. 1067 */ 1068 if ((info->flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) { 1069 return(0); 1070 } 1071 1072 /* 1073 * If WRITECLOSE is set, flush out unlinked but still open 1074 * files (even if open only for reading) and regular file 1075 * vnodes open for writing. 1076 */ 1077 if ((info->flags & WRITECLOSE) && 1078 (vp->v_type == VNON || 1079 (VOP_GETATTR(vp, &vattr, info->td) == 0 && 1080 vattr.va_nlink > 0)) && 1081 (vp->v_writecount == 0 || vp->v_type != VREG)) { 1082 return(0); 1083 } 1084 1085 /* 1086 * With v_usecount == 0, all we need to do is clear out the 1087 * vnode data structures and we are done. 1088 */ 1089 if (vp->v_usecount == 1) { 1090 vgone(vp); 1091 return(0); 1092 } 1093 1094 /* 1095 * If FORCECLOSE is set, forcibly close the vnode. For block 1096 * or character devices, revert to an anonymous device. For 1097 * all other files, just kill them. 1098 */ 1099 if (info->flags & FORCECLOSE) { 1100 if (vp->v_type != VBLK && vp->v_type != VCHR) { 1101 vgone(vp); 1102 } else { 1103 vclean(vp, 0, info->td); 1104 vp->v_ops = &spec_vnode_vops; 1105 insmntque(vp, NULL); 1106 } 1107 return(0); 1108 } 1109 #ifdef DIAGNOSTIC 1110 if (busyprt) 1111 vprint("vflush: busy vnode", vp); 1112 #endif 1113 ++info->busy; 1114 return(0); 1115 } 1116 1117