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 1.36 2008/07/14 22:16:35 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 #include <sys/sysref2.h> 95 96 #include <vm/vm.h> 97 #include <vm/vm_object.h> 98 99 struct mountscan_info { 100 TAILQ_ENTRY(mountscan_info) msi_entry; 101 int msi_how; 102 struct mount *msi_node; 103 }; 104 105 struct vmntvnodescan_info { 106 TAILQ_ENTRY(vmntvnodescan_info) entry; 107 struct vnode *vp; 108 }; 109 110 static int vnlru_nowhere = 0; 111 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RD, 112 &vnlru_nowhere, 0, 113 "Number of times the vnlru process ran without success"); 114 115 116 static struct lwkt_token mntid_token; 117 118 /* note: mountlist exported to pstat */ 119 struct mntlist mountlist = TAILQ_HEAD_INITIALIZER(mountlist); 120 static TAILQ_HEAD(,mountscan_info) mountscan_list; 121 static struct lwkt_token mountlist_token; 122 static TAILQ_HEAD(,vmntvnodescan_info) mntvnodescan_list; 123 struct lwkt_token mntvnode_token; 124 125 static TAILQ_HEAD(,bio_ops) bio_ops_list = TAILQ_HEAD_INITIALIZER(bio_ops_list); 126 127 /* 128 * Called from vfsinit() 129 */ 130 void 131 vfs_mount_init(void) 132 { 133 lwkt_token_init(&mountlist_token); 134 lwkt_token_init(&mntvnode_token); 135 lwkt_token_init(&mntid_token); 136 TAILQ_INIT(&mountscan_list); 137 TAILQ_INIT(&mntvnodescan_list); 138 } 139 140 /* 141 * Support function called with mntvnode_token held to remove a vnode 142 * from the mountlist. We must update any list scans which are in progress. 143 */ 144 static void 145 vremovevnodemnt(struct vnode *vp) 146 { 147 struct vmntvnodescan_info *info; 148 149 TAILQ_FOREACH(info, &mntvnodescan_list, entry) { 150 if (info->vp == vp) 151 info->vp = TAILQ_NEXT(vp, v_nmntvnodes); 152 } 153 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes); 154 } 155 156 /* 157 * Allocate a new vnode and associate it with a tag, mount point, and 158 * operations vector. 159 * 160 * A VX locked and refd vnode is returned. The caller should setup the 161 * remaining fields and vx_put() or, if he wishes to leave a vref, 162 * vx_unlock() the vnode. 163 */ 164 int 165 getnewvnode(enum vtagtype tag, struct mount *mp, 166 struct vnode **vpp, int lktimeout, int lkflags) 167 { 168 struct vnode *vp; 169 170 KKASSERT(mp != NULL); 171 172 vp = allocvnode(lktimeout, lkflags); 173 vp->v_tag = tag; 174 vp->v_data = NULL; 175 176 /* 177 * By default the vnode is assigned the mount point's normal 178 * operations vector. 179 */ 180 vp->v_ops = &mp->mnt_vn_use_ops; 181 182 /* 183 * Placing the vnode on the mount point's queue makes it visible. 184 * VNON prevents it from being messed with, however. 185 */ 186 insmntque(vp, mp); 187 188 /* 189 * A VX locked & refd vnode is returned. 190 */ 191 *vpp = vp; 192 return (0); 193 } 194 195 /* 196 * This function creates vnodes with special operations vectors. The 197 * mount point is optional. 198 * 199 * This routine is being phased out. 200 */ 201 int 202 getspecialvnode(enum vtagtype tag, struct mount *mp, 203 struct vop_ops **ops, 204 struct vnode **vpp, int lktimeout, int lkflags) 205 { 206 struct vnode *vp; 207 208 vp = allocvnode(lktimeout, lkflags); 209 vp->v_tag = tag; 210 vp->v_data = NULL; 211 vp->v_ops = ops; 212 213 /* 214 * Placing the vnode on the mount point's queue makes it visible. 215 * VNON prevents it from being messed with, however. 216 */ 217 insmntque(vp, mp); 218 219 /* 220 * A VX locked & refd vnode is returned. 221 */ 222 *vpp = vp; 223 return (0); 224 } 225 226 /* 227 * Interlock against an unmount, return 0 on success, non-zero on failure. 228 * 229 * The passed flag may be 0 or LK_NOWAIT and is only used if an unmount 230 * is in-progress. 231 * 232 * If no unmount is in-progress LK_NOWAIT is ignored. No other flag bits 233 * are used. A shared locked will be obtained and the filesystem will not 234 * be unmountable until the lock is released. 235 */ 236 int 237 vfs_busy(struct mount *mp, int flags) 238 { 239 int lkflags; 240 241 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 242 if (flags & LK_NOWAIT) 243 return (ENOENT); 244 /* XXX not MP safe */ 245 mp->mnt_kern_flag |= MNTK_MWAIT; 246 /* 247 * Since all busy locks are shared except the exclusive 248 * lock granted when unmounting, the only place that a 249 * wakeup needs to be done is at the release of the 250 * exclusive lock at the end of dounmount. 251 */ 252 tsleep((caddr_t)mp, 0, "vfs_busy", 0); 253 return (ENOENT); 254 } 255 lkflags = LK_SHARED; 256 if (lockmgr(&mp->mnt_lock, lkflags)) 257 panic("vfs_busy: unexpected lock failure"); 258 return (0); 259 } 260 261 /* 262 * Free a busy filesystem. 263 */ 264 void 265 vfs_unbusy(struct mount *mp) 266 { 267 lockmgr(&mp->mnt_lock, LK_RELEASE); 268 } 269 270 /* 271 * Lookup a filesystem type, and if found allocate and initialize 272 * a mount structure for it. 273 * 274 * Devname is usually updated by mount(8) after booting. 275 */ 276 int 277 vfs_rootmountalloc(char *fstypename, char *devname, struct mount **mpp) 278 { 279 struct vfsconf *vfsp; 280 struct mount *mp; 281 282 if (fstypename == NULL) 283 return (ENODEV); 284 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) { 285 if (!strcmp(vfsp->vfc_name, fstypename)) 286 break; 287 } 288 if (vfsp == NULL) 289 return (ENODEV); 290 mp = kmalloc(sizeof(struct mount), M_MOUNT, M_WAITOK | M_ZERO); 291 lockinit(&mp->mnt_lock, "vfslock", VLKTIMEOUT, 0); 292 vfs_busy(mp, LK_NOWAIT); 293 TAILQ_INIT(&mp->mnt_nvnodelist); 294 TAILQ_INIT(&mp->mnt_reservedvnlist); 295 TAILQ_INIT(&mp->mnt_jlist); 296 mp->mnt_nvnodelistsize = 0; 297 mp->mnt_vfc = vfsp; 298 mp->mnt_op = vfsp->vfc_vfsops; 299 mp->mnt_flag = MNT_RDONLY; 300 vfsp->vfc_refcount++; 301 mp->mnt_iosize_max = DFLTPHYS; 302 mp->mnt_stat.f_type = vfsp->vfc_typenum; 303 mp->mnt_flag |= vfsp->vfc_flags & MNT_VISFLAGMASK; 304 strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN); 305 copystr(devname, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, 0); 306 *mpp = mp; 307 return (0); 308 } 309 310 /* 311 * Lookup a mount point by filesystem identifier. 312 */ 313 struct mount * 314 vfs_getvfs(fsid_t *fsid) 315 { 316 struct mount *mp; 317 lwkt_tokref ilock; 318 319 lwkt_gettoken(&ilock, &mountlist_token); 320 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 321 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 322 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 323 break; 324 } 325 } 326 lwkt_reltoken(&ilock); 327 return (mp); 328 } 329 330 /* 331 * Get a new unique fsid. Try to make its val[0] unique, since this value 332 * will be used to create fake device numbers for stat(). Also try (but 333 * not so hard) make its val[0] unique mod 2^16, since some emulators only 334 * support 16-bit device numbers. We end up with unique val[0]'s for the 335 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 336 * 337 * Keep in mind that several mounts may be running in parallel. Starting 338 * the search one past where the previous search terminated is both a 339 * micro-optimization and a defense against returning the same fsid to 340 * different mounts. 341 */ 342 void 343 vfs_getnewfsid(struct mount *mp) 344 { 345 static u_int16_t mntid_base; 346 lwkt_tokref ilock; 347 fsid_t tfsid; 348 int mtype; 349 350 lwkt_gettoken(&ilock, &mntid_token); 351 mtype = mp->mnt_vfc->vfc_typenum; 352 tfsid.val[1] = mtype; 353 mtype = (mtype & 0xFF) << 24; 354 for (;;) { 355 tfsid.val[0] = makeudev(255, 356 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 357 mntid_base++; 358 if (vfs_getvfs(&tfsid) == NULL) 359 break; 360 } 361 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 362 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 363 lwkt_reltoken(&ilock); 364 } 365 366 /* 367 * This routine is called when we have too many vnodes. It attempts 368 * to free <count> vnodes and will potentially free vnodes that still 369 * have VM backing store (VM backing store is typically the cause 370 * of a vnode blowout so we want to do this). Therefore, this operation 371 * is not considered cheap. 372 * 373 * A number of conditions may prevent a vnode from being reclaimed. 374 * the buffer cache may have references on the vnode, a directory 375 * vnode may still have references due to the namei cache representing 376 * underlying files, or the vnode may be in active use. It is not 377 * desireable to reuse such vnodes. These conditions may cause the 378 * number of vnodes to reach some minimum value regardless of what 379 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 380 */ 381 382 /* 383 * This is a quick non-blocking check to determine if the vnode is a good 384 * candidate for being (eventually) vgone()'d. Returns 0 if the vnode is 385 * not a good candidate, 1 if it is. 386 */ 387 static __inline int 388 vmightfree(struct vnode *vp, int page_count) 389 { 390 if (vp->v_flag & VRECLAIMED) 391 return (0); 392 #if 0 393 if ((vp->v_flag & VFREE) && TAILQ_EMPTY(&vp->v_namecache)) 394 return (0); 395 #endif 396 if (sysref_isactive(&vp->v_sysref)) 397 return (0); 398 if (vp->v_object && vp->v_object->resident_page_count >= page_count) 399 return (0); 400 return (1); 401 } 402 403 /* 404 * The vnode was found to be possibly vgone()able and the caller has locked it 405 * (thus the usecount should be 1 now). Determine if the vnode is actually 406 * vgone()able, doing some cleanups in the process. Returns 1 if the vnode 407 * can be vgone()'d, 0 otherwise. 408 * 409 * Note that v_auxrefs may be non-zero because (A) this vnode is not a leaf 410 * in the namecache topology and (B) this vnode has buffer cache bufs. 411 * We cannot remove vnodes with non-leaf namecache associations. We do a 412 * tentitive leaf check prior to attempting to flush out any buffers but the 413 * 'real' test when all is said in done is that v_auxrefs must become 0 for 414 * the vnode to be freeable. 415 * 416 * We could theoretically just unconditionally flush when v_auxrefs != 0, 417 * but flushing data associated with non-leaf nodes (which are always 418 * directories), just throws it away for no benefit. It is the buffer 419 * cache's responsibility to choose buffers to recycle from the cached 420 * data point of view. 421 */ 422 static int 423 visleaf(struct vnode *vp) 424 { 425 struct namecache *ncp; 426 427 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 428 if (!TAILQ_EMPTY(&ncp->nc_list)) 429 return(0); 430 } 431 return(1); 432 } 433 434 /* 435 * Try to clean up the vnode to the point where it can be vgone()'d, returning 436 * 0 if it cannot be vgone()'d (or already has been), 1 if it can. Unlike 437 * vmightfree() this routine may flush the vnode and block. Vnodes marked 438 * VFREE are still candidates for vgone()ing because they may hold namecache 439 * resources and could be blocking the namecache directory hierarchy (and 440 * related vnodes) from being freed. 441 */ 442 static int 443 vtrytomakegoneable(struct vnode *vp, int page_count) 444 { 445 if (vp->v_flag & VRECLAIMED) 446 return (0); 447 if (vp->v_sysref.refcnt > 1) 448 return (0); 449 if (vp->v_object && vp->v_object->resident_page_count >= page_count) 450 return (0); 451 if (vp->v_auxrefs && visleaf(vp)) { 452 vinvalbuf(vp, V_SAVE, 0, 0); 453 #if 0 /* DEBUG */ 454 kprintf((vp->v_auxrefs ? "vrecycle: vp %p failed: %s\n" : 455 "vrecycle: vp %p succeeded: %s\n"), vp, 456 (TAILQ_FIRST(&vp->v_namecache) ? 457 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?")); 458 #endif 459 } 460 461 /* 462 * This sequence may seem a little strange, but we need to optimize 463 * the critical path a bit. We can't recycle vnodes with other 464 * references and because we are trying to recycle an otherwise 465 * perfectly fine vnode we have to invalidate the namecache in a 466 * way that avoids possible deadlocks (since the vnode lock is being 467 * held here). Finally, we have to check for other references one 468 * last time in case something snuck in during the inval. 469 */ 470 if (vp->v_sysref.refcnt > 1 || vp->v_auxrefs != 0) 471 return (0); 472 if (cache_inval_vp_nonblock(vp)) 473 return (0); 474 return (vp->v_sysref.refcnt <= 1 && vp->v_auxrefs == 0); 475 } 476 477 /* 478 * Reclaim up to 1/10 of the vnodes associated with a mount point. Try 479 * to avoid vnodes which have lots of resident pages (we are trying to free 480 * vnodes, not memory). 481 * 482 * This routine is a callback from the mountlist scan. The mount point 483 * in question will be busied. 484 */ 485 static int 486 vlrureclaim(struct mount *mp, void *data) 487 { 488 struct vnode *vp; 489 lwkt_tokref ilock; 490 int done; 491 int trigger; 492 int usevnodes; 493 int count; 494 int trigger_mult = vnlru_nowhere; 495 496 /* 497 * Calculate the trigger point for the resident pages check. The 498 * minimum trigger value is approximately the number of pages in 499 * the system divded by the number of vnodes. However, due to 500 * various other system memory overheads unrelated to data caching 501 * it is a good idea to double the trigger (at least). 502 * 503 * trigger_mult starts at 0. If the recycler is having problems 504 * finding enough freeable vnodes it will increase trigger_mult. 505 * This should not happen in normal operation, even on machines with 506 * low amounts of memory, but extraordinary memory use by the system 507 * verses the amount of cached data can trigger it. 508 */ 509 usevnodes = desiredvnodes; 510 if (usevnodes <= 0) 511 usevnodes = 1; 512 trigger = vmstats.v_page_count * (trigger_mult + 2) / usevnodes; 513 514 done = 0; 515 lwkt_gettoken(&ilock, &mntvnode_token); 516 count = mp->mnt_nvnodelistsize / 10 + 1; 517 while (count && mp->mnt_syncer) { 518 /* 519 * Next vnode. Use the special syncer vnode to placemark 520 * the LRU. This way the LRU code does not interfere with 521 * vmntvnodescan(). 522 */ 523 vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes); 524 TAILQ_REMOVE(&mp->mnt_nvnodelist, mp->mnt_syncer, v_nmntvnodes); 525 if (vp) { 526 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, 527 mp->mnt_syncer, v_nmntvnodes); 528 } else { 529 TAILQ_INSERT_HEAD(&mp->mnt_nvnodelist, mp->mnt_syncer, 530 v_nmntvnodes); 531 vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes); 532 if (vp == NULL) 533 break; 534 } 535 536 /* 537 * __VNODESCAN__ 538 * 539 * The VP will stick around while we hold mntvnode_token, 540 * at least until we block, so we can safely do an initial 541 * check, and then must check again after we lock the vnode. 542 */ 543 if (vp->v_type == VNON || /* syncer or indeterminant */ 544 !vmightfree(vp, trigger) /* critical path opt */ 545 ) { 546 --count; 547 continue; 548 } 549 550 /* 551 * VX get the candidate vnode. If the VX get fails the 552 * vnode might still be on the mountlist. Our loop depends 553 * on us at least cycling the vnode to the end of the 554 * mountlist. 555 */ 556 if (vx_get_nonblock(vp) != 0) { 557 --count; 558 continue; 559 } 560 561 /* 562 * Since we blocked locking the vp, make sure it is still 563 * a candidate for reclamation. That is, it has not already 564 * been reclaimed and only has our VX reference associated 565 * with it. 566 */ 567 if (vp->v_type == VNON || /* syncer or indeterminant */ 568 (vp->v_flag & VRECLAIMED) || 569 vp->v_mount != mp || 570 !vtrytomakegoneable(vp, trigger) /* critical path opt */ 571 ) { 572 --count; 573 vx_put(vp); 574 continue; 575 } 576 577 /* 578 * All right, we are good, move the vp to the end of the 579 * mountlist and clean it out. The vget will have returned 580 * an error if the vnode was destroyed (VRECLAIMED set), so we 581 * do not have to check again. The vput() will move the 582 * vnode to the free list if the vgone() was successful. 583 */ 584 KKASSERT(vp->v_mount == mp); 585 vgone_vxlocked(vp); 586 vx_put(vp); 587 ++done; 588 --count; 589 } 590 lwkt_reltoken(&ilock); 591 return (done); 592 } 593 594 /* 595 * Attempt to recycle vnodes in a context that is always safe to block. 596 * Calling vlrurecycle() from the bowels of file system code has some 597 * interesting deadlock problems. 598 */ 599 static struct thread *vnlruthread; 600 static int vnlruproc_sig; 601 602 void 603 vnlru_proc_wait(void) 604 { 605 if (vnlruproc_sig == 0) { 606 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 607 wakeup(vnlruthread); 608 } 609 tsleep(&vnlruproc_sig, 0, "vlruwk", hz); 610 } 611 612 static void 613 vnlru_proc(void) 614 { 615 struct thread *td = curthread; 616 int done; 617 618 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td, 619 SHUTDOWN_PRI_FIRST); 620 621 crit_enter(); 622 for (;;) { 623 kproc_suspend_loop(); 624 625 /* 626 * Try to free some vnodes if we have too many 627 */ 628 if (numvnodes > desiredvnodes && 629 freevnodes > desiredvnodes * 2 / 10) { 630 int count = numvnodes - desiredvnodes; 631 632 if (count > freevnodes / 100) 633 count = freevnodes / 100; 634 if (count < 5) 635 count = 5; 636 freesomevnodes(count); 637 } 638 639 /* 640 * Nothing to do if most of our vnodes are already on 641 * the free list. 642 */ 643 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { 644 vnlruproc_sig = 0; 645 wakeup(&vnlruproc_sig); 646 tsleep(td, 0, "vlruwt", hz); 647 continue; 648 } 649 cache_cleanneg(0); 650 done = mountlist_scan(vlrureclaim, NULL, MNTSCAN_FORWARD); 651 652 /* 653 * The vlrureclaim() call only processes 1/10 of the vnodes 654 * on each mount. If we couldn't find any repeat the loop 655 * at least enough times to cover all available vnodes before 656 * we start sleeping. Complain if the failure extends past 657 * 30 second, every 30 seconds. 658 */ 659 if (done == 0) { 660 ++vnlru_nowhere; 661 if (vnlru_nowhere % 10 == 0) 662 tsleep(td, 0, "vlrup", hz * 3); 663 if (vnlru_nowhere % 100 == 0) 664 kprintf("vnlru_proc: vnode recycler stopped working!\n"); 665 if (vnlru_nowhere == 1000) 666 vnlru_nowhere = 900; 667 } else { 668 vnlru_nowhere = 0; 669 } 670 } 671 crit_exit(); 672 } 673 674 /* 675 * MOUNTLIST FUNCTIONS 676 */ 677 678 /* 679 * mountlist_insert (MP SAFE) 680 * 681 * Add a new mount point to the mount list. 682 */ 683 void 684 mountlist_insert(struct mount *mp, int how) 685 { 686 lwkt_tokref ilock; 687 688 lwkt_gettoken(&ilock, &mountlist_token); 689 if (how == MNTINS_FIRST) 690 TAILQ_INSERT_HEAD(&mountlist, mp, mnt_list); 691 else 692 TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list); 693 lwkt_reltoken(&ilock); 694 } 695 696 /* 697 * mountlist_interlock (MP SAFE) 698 * 699 * Execute the specified interlock function with the mountlist token 700 * held. The function will be called in a serialized fashion verses 701 * other functions called through this mechanism. 702 */ 703 int 704 mountlist_interlock(int (*callback)(struct mount *), struct mount *mp) 705 { 706 lwkt_tokref ilock; 707 int error; 708 709 lwkt_gettoken(&ilock, &mountlist_token); 710 error = callback(mp); 711 lwkt_reltoken(&ilock); 712 return (error); 713 } 714 715 /* 716 * mountlist_boot_getfirst (DURING BOOT ONLY) 717 * 718 * This function returns the first mount on the mountlist, which is 719 * expected to be the root mount. Since no interlocks are obtained 720 * this function is only safe to use during booting. 721 */ 722 723 struct mount * 724 mountlist_boot_getfirst(void) 725 { 726 return(TAILQ_FIRST(&mountlist)); 727 } 728 729 /* 730 * mountlist_remove (MP SAFE) 731 * 732 * Remove a node from the mountlist. If this node is the next scan node 733 * for any active mountlist scans, the active mountlist scan will be 734 * adjusted to skip the node, thus allowing removals during mountlist 735 * scans. 736 */ 737 void 738 mountlist_remove(struct mount *mp) 739 { 740 struct mountscan_info *msi; 741 lwkt_tokref ilock; 742 743 lwkt_gettoken(&ilock, &mountlist_token); 744 TAILQ_FOREACH(msi, &mountscan_list, msi_entry) { 745 if (msi->msi_node == mp) { 746 if (msi->msi_how & MNTSCAN_FORWARD) 747 msi->msi_node = TAILQ_NEXT(mp, mnt_list); 748 else 749 msi->msi_node = TAILQ_PREV(mp, mntlist, mnt_list); 750 } 751 } 752 TAILQ_REMOVE(&mountlist, mp, mnt_list); 753 lwkt_reltoken(&ilock); 754 } 755 756 /* 757 * mountlist_scan (MP SAFE) 758 * 759 * Safely scan the mount points on the mount list. Unless otherwise 760 * specified each mount point will be busied prior to the callback and 761 * unbusied afterwords. The callback may safely remove any mount point 762 * without interfering with the scan. If the current callback 763 * mount is removed the scanner will not attempt to unbusy it. 764 * 765 * If a mount node cannot be busied it is silently skipped. 766 * 767 * The callback return value is aggregated and a total is returned. A return 768 * value of < 0 is not aggregated and will terminate the scan. 769 * 770 * MNTSCAN_FORWARD - the mountlist is scanned in the forward direction 771 * MNTSCAN_REVERSE - the mountlist is scanned in reverse 772 * MNTSCAN_NOBUSY - the scanner will make the callback without busying 773 * the mount node. 774 */ 775 int 776 mountlist_scan(int (*callback)(struct mount *, void *), void *data, int how) 777 { 778 struct mountscan_info info; 779 lwkt_tokref ilock; 780 struct mount *mp; 781 thread_t td; 782 int count; 783 int res; 784 785 lwkt_gettoken(&ilock, &mountlist_token); 786 787 info.msi_how = how; 788 info.msi_node = NULL; /* paranoia */ 789 TAILQ_INSERT_TAIL(&mountscan_list, &info, msi_entry); 790 791 res = 0; 792 td = curthread; 793 794 if (how & MNTSCAN_FORWARD) { 795 info.msi_node = TAILQ_FIRST(&mountlist); 796 while ((mp = info.msi_node) != NULL) { 797 if (how & MNTSCAN_NOBUSY) { 798 count = callback(mp, data); 799 } else if (vfs_busy(mp, LK_NOWAIT) == 0) { 800 count = callback(mp, data); 801 if (mp == info.msi_node) 802 vfs_unbusy(mp); 803 } else { 804 count = 0; 805 } 806 if (count < 0) 807 break; 808 res += count; 809 if (mp == info.msi_node) 810 info.msi_node = TAILQ_NEXT(mp, mnt_list); 811 } 812 } else if (how & MNTSCAN_REVERSE) { 813 info.msi_node = TAILQ_LAST(&mountlist, mntlist); 814 while ((mp = info.msi_node) != NULL) { 815 if (how & MNTSCAN_NOBUSY) { 816 count = callback(mp, data); 817 } else if (vfs_busy(mp, LK_NOWAIT) == 0) { 818 count = callback(mp, data); 819 if (mp == info.msi_node) 820 vfs_unbusy(mp); 821 } else { 822 count = 0; 823 } 824 if (count < 0) 825 break; 826 res += count; 827 if (mp == info.msi_node) 828 info.msi_node = TAILQ_PREV(mp, mntlist, mnt_list); 829 } 830 } 831 TAILQ_REMOVE(&mountscan_list, &info, msi_entry); 832 lwkt_reltoken(&ilock); 833 return(res); 834 } 835 836 /* 837 * MOUNT RELATED VNODE FUNCTIONS 838 */ 839 840 static struct kproc_desc vnlru_kp = { 841 "vnlru", 842 vnlru_proc, 843 &vnlruthread 844 }; 845 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 846 847 /* 848 * Move a vnode from one mount queue to another. 849 */ 850 void 851 insmntque(struct vnode *vp, struct mount *mp) 852 { 853 lwkt_tokref ilock; 854 855 lwkt_gettoken(&ilock, &mntvnode_token); 856 /* 857 * Delete from old mount point vnode list, if on one. 858 */ 859 if (vp->v_mount != NULL) { 860 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0, 861 ("bad mount point vnode list size")); 862 vremovevnodemnt(vp); 863 vp->v_mount->mnt_nvnodelistsize--; 864 } 865 /* 866 * Insert into list of vnodes for the new mount point, if available. 867 * The 'end' of the LRU list is the vnode prior to mp->mnt_syncer. 868 */ 869 if ((vp->v_mount = mp) == NULL) { 870 lwkt_reltoken(&ilock); 871 return; 872 } 873 if (mp->mnt_syncer) { 874 TAILQ_INSERT_BEFORE(mp->mnt_syncer, vp, v_nmntvnodes); 875 } else { 876 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 877 } 878 mp->mnt_nvnodelistsize++; 879 lwkt_reltoken(&ilock); 880 } 881 882 883 /* 884 * Scan the vnodes under a mount point and issue appropriate callbacks. 885 * 886 * The fastfunc() callback is called with just the mountlist token held 887 * (no vnode lock). It may not block and the vnode may be undergoing 888 * modifications while the caller is processing it. The vnode will 889 * not be entirely destroyed, however, due to the fact that the mountlist 890 * token is held. A return value < 0 skips to the next vnode without calling 891 * the slowfunc(), a return value > 0 terminates the loop. 892 * 893 * The slowfunc() callback is called after the vnode has been successfully 894 * locked based on passed flags. The vnode is skipped if it gets rearranged 895 * or destroyed while blocking on the lock. A non-zero return value from 896 * the slow function terminates the loop. The slow function is allowed to 897 * arbitrarily block. The scanning code guarentees consistency of operation 898 * even if the slow function deletes or moves the node, or blocks and some 899 * other thread deletes or moves the node. 900 */ 901 int 902 vmntvnodescan( 903 struct mount *mp, 904 int flags, 905 int (*fastfunc)(struct mount *mp, struct vnode *vp, void *data), 906 int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data), 907 void *data 908 ) { 909 struct vmntvnodescan_info info; 910 lwkt_tokref ilock; 911 struct vnode *vp; 912 int r = 0; 913 int maxcount = 1000000; 914 int stopcount = 0; 915 int count = 0; 916 917 lwkt_gettoken(&ilock, &mntvnode_token); 918 919 /* 920 * If asked to do one pass stop after iterating available vnodes. 921 * Under heavy loads new vnodes can be added while we are scanning, 922 * so this isn't perfect. Create a slop factor of 2x. 923 */ 924 if (flags & VMSC_ONEPASS) 925 stopcount = mp->mnt_nvnodelistsize * 2; 926 927 info.vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 928 TAILQ_INSERT_TAIL(&mntvnodescan_list, &info, entry); 929 while ((vp = info.vp) != NULL) { 930 if (--maxcount == 0) 931 panic("maxcount reached during vmntvnodescan"); 932 933 /* 934 * Skip if visible but not ready, or special (e.g. 935 * mp->mnt_syncer) 936 */ 937 if (vp->v_type == VNON) 938 goto next; 939 KKASSERT(vp->v_mount == mp); 940 941 /* 942 * Quick test. A negative return continues the loop without 943 * calling the slow test. 0 continues onto the slow test. 944 * A positive number aborts the loop. 945 */ 946 if (fastfunc) { 947 if ((r = fastfunc(mp, vp, data)) < 0) { 948 r = 0; 949 goto next; 950 } 951 if (r) 952 break; 953 } 954 955 /* 956 * Get a vxlock on the vnode, retry if it has moved or isn't 957 * in the mountlist where we expect it. 958 */ 959 if (slowfunc) { 960 int error; 961 962 switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) { 963 case VMSC_GETVP: 964 error = vget(vp, LK_EXCLUSIVE); 965 break; 966 case VMSC_GETVP|VMSC_NOWAIT: 967 error = vget(vp, LK_EXCLUSIVE|LK_NOWAIT); 968 break; 969 case VMSC_GETVX: 970 vx_get(vp); 971 error = 0; 972 break; 973 default: 974 error = 0; 975 break; 976 } 977 if (error) 978 goto next; 979 /* 980 * Do not call the slow function if the vnode is 981 * invalid or if it was ripped out from under us 982 * while we (potentially) blocked. 983 */ 984 if (info.vp == vp && vp->v_type != VNON) 985 r = slowfunc(mp, vp, data); 986 987 /* 988 * Cleanup 989 */ 990 switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) { 991 case VMSC_GETVP: 992 case VMSC_GETVP|VMSC_NOWAIT: 993 vput(vp); 994 break; 995 case VMSC_GETVX: 996 vx_put(vp); 997 break; 998 default: 999 break; 1000 } 1001 if (r != 0) 1002 break; 1003 } 1004 1005 next: 1006 /* 1007 * Yield after some processing. Depending on the number 1008 * of vnodes, we might wind up running for a long time. 1009 * Because threads are not preemptable, time critical 1010 * userland processes might starve. Give them a chance 1011 * now and then. 1012 */ 1013 if (++count == 10000) { 1014 /* We really want to yield a bit, so we simply sleep a tick */ 1015 tsleep(mp, 0, "vnodescn", 1); 1016 count = 0; 1017 } 1018 1019 /* 1020 * If doing one pass this decrements to zero. If it starts 1021 * at zero it is effectively unlimited for the purposes of 1022 * this loop. 1023 */ 1024 if (--stopcount == 0) 1025 break; 1026 1027 /* 1028 * Iterate. If the vnode was ripped out from under us 1029 * info.vp will already point to the next vnode, otherwise 1030 * we have to obtain the next valid vnode ourselves. 1031 */ 1032 if (info.vp == vp) 1033 info.vp = TAILQ_NEXT(vp, v_nmntvnodes); 1034 } 1035 TAILQ_REMOVE(&mntvnodescan_list, &info, entry); 1036 lwkt_reltoken(&ilock); 1037 return(r); 1038 } 1039 1040 /* 1041 * Remove any vnodes in the vnode table belonging to mount point mp. 1042 * 1043 * If FORCECLOSE is not specified, there should not be any active ones, 1044 * return error if any are found (nb: this is a user error, not a 1045 * system error). If FORCECLOSE is specified, detach any active vnodes 1046 * that are found. 1047 * 1048 * If WRITECLOSE is set, only flush out regular file vnodes open for 1049 * writing. 1050 * 1051 * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped. 1052 * 1053 * `rootrefs' specifies the base reference count for the root vnode 1054 * of this filesystem. The root vnode is considered busy if its 1055 * v_sysref.refcnt exceeds this value. On a successful return, vflush() 1056 * will call vrele() on the root vnode exactly rootrefs times. 1057 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 1058 * be zero. 1059 */ 1060 #ifdef DIAGNOSTIC 1061 static int busyprt = 0; /* print out busy vnodes */ 1062 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 1063 #endif 1064 1065 static int vflush_scan(struct mount *mp, struct vnode *vp, void *data); 1066 1067 struct vflush_info { 1068 int flags; 1069 int busy; 1070 thread_t td; 1071 }; 1072 1073 int 1074 vflush(struct mount *mp, int rootrefs, int flags) 1075 { 1076 struct thread *td = curthread; /* XXX */ 1077 struct vnode *rootvp = NULL; 1078 int error; 1079 struct vflush_info vflush_info; 1080 1081 if (rootrefs > 0) { 1082 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 1083 ("vflush: bad args")); 1084 /* 1085 * Get the filesystem root vnode. We can vput() it 1086 * immediately, since with rootrefs > 0, it won't go away. 1087 */ 1088 if ((error = VFS_ROOT(mp, &rootvp)) != 0) 1089 return (error); 1090 vput(rootvp); 1091 } 1092 1093 vflush_info.busy = 0; 1094 vflush_info.flags = flags; 1095 vflush_info.td = td; 1096 vmntvnodescan(mp, VMSC_GETVX, NULL, vflush_scan, &vflush_info); 1097 1098 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 1099 /* 1100 * If just the root vnode is busy, and if its refcount 1101 * is equal to `rootrefs', then go ahead and kill it. 1102 */ 1103 KASSERT(vflush_info.busy > 0, ("vflush: not busy")); 1104 KASSERT(rootvp->v_sysref.refcnt >= rootrefs, ("vflush: rootrefs")); 1105 if (vflush_info.busy == 1 && rootvp->v_sysref.refcnt == rootrefs) { 1106 vx_lock(rootvp); 1107 vgone_vxlocked(rootvp); 1108 vx_unlock(rootvp); 1109 vflush_info.busy = 0; 1110 } 1111 } 1112 if (vflush_info.busy) 1113 return (EBUSY); 1114 for (; rootrefs > 0; rootrefs--) 1115 vrele(rootvp); 1116 return (0); 1117 } 1118 1119 /* 1120 * The scan callback is made with an VX locked vnode. 1121 */ 1122 static int 1123 vflush_scan(struct mount *mp, struct vnode *vp, void *data) 1124 { 1125 struct vflush_info *info = data; 1126 struct vattr vattr; 1127 1128 /* 1129 * Skip over a vnodes marked VSYSTEM. 1130 */ 1131 if ((info->flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) { 1132 return(0); 1133 } 1134 1135 /* 1136 * If WRITECLOSE is set, flush out unlinked but still open 1137 * files (even if open only for reading) and regular file 1138 * vnodes open for writing. 1139 */ 1140 if ((info->flags & WRITECLOSE) && 1141 (vp->v_type == VNON || 1142 (VOP_GETATTR(vp, &vattr) == 0 && 1143 vattr.va_nlink > 0)) && 1144 (vp->v_writecount == 0 || vp->v_type != VREG)) { 1145 return(0); 1146 } 1147 1148 /* 1149 * If we are the only holder (refcnt of 1) or the vnode is in 1150 * termination (refcnt < 0), we can vgone the vnode. 1151 */ 1152 if (vp->v_sysref.refcnt <= 1) { 1153 vgone_vxlocked(vp); 1154 return(0); 1155 } 1156 1157 /* 1158 * If FORCECLOSE is set, forcibly close the vnode. For block 1159 * or character devices, revert to an anonymous device. For 1160 * all other files, just kill them. 1161 */ 1162 if (info->flags & FORCECLOSE) { 1163 if (vp->v_type != VBLK && vp->v_type != VCHR) { 1164 vgone_vxlocked(vp); 1165 } else { 1166 vclean_vxlocked(vp, 0); 1167 vp->v_ops = &spec_vnode_vops_p; 1168 insmntque(vp, NULL); 1169 } 1170 return(0); 1171 } 1172 #ifdef DIAGNOSTIC 1173 if (busyprt) 1174 vprint("vflush: busy vnode", vp); 1175 #endif 1176 ++info->busy; 1177 return(0); 1178 } 1179 1180 void 1181 add_bio_ops(struct bio_ops *ops) 1182 { 1183 TAILQ_INSERT_TAIL(&bio_ops_list, ops, entry); 1184 } 1185 1186 void 1187 rem_bio_ops(struct bio_ops *ops) 1188 { 1189 TAILQ_REMOVE(&bio_ops_list, ops, entry); 1190 } 1191 1192 /* 1193 * This calls the bio_ops io_sync function either for a mount point 1194 * or generally. 1195 * 1196 * WARNING: softdeps is weirdly coded and just isn't happy unless 1197 * io_sync is called with a NULL mount from the general syncing code. 1198 */ 1199 void 1200 bio_ops_sync(struct mount *mp) 1201 { 1202 struct bio_ops *ops; 1203 1204 if (mp) { 1205 if ((ops = mp->mnt_bioops) != NULL) 1206 ops->io_sync(mp); 1207 } else { 1208 TAILQ_FOREACH(ops, &bio_ops_list, entry) { 1209 ops->io_sync(NULL); 1210 } 1211 } 1212 } 1213 1214