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 1.37 2008/09/17 21:44:18 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 #include <sys/mplock2.h> 96 97 #include <vm/vm.h> 98 #include <vm/vm_object.h> 99 100 struct mountscan_info { 101 TAILQ_ENTRY(mountscan_info) msi_entry; 102 int msi_how; 103 struct mount *msi_node; 104 }; 105 106 struct vmntvnodescan_info { 107 TAILQ_ENTRY(vmntvnodescan_info) entry; 108 struct vnode *vp; 109 }; 110 111 struct vnlru_info { 112 int pass; 113 }; 114 115 static int vnlru_nowhere = 0; 116 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RD, 117 &vnlru_nowhere, 0, 118 "Number of times the vnlru process ran without success"); 119 120 121 static struct lwkt_token mntid_token; 122 static struct mount dummymount; 123 124 /* note: mountlist exported to pstat */ 125 struct mntlist mountlist = TAILQ_HEAD_INITIALIZER(mountlist); 126 static TAILQ_HEAD(,mountscan_info) mountscan_list; 127 static struct lwkt_token mountlist_token; 128 static TAILQ_HEAD(,vmntvnodescan_info) mntvnodescan_list; 129 struct lwkt_token mntvnode_token; 130 131 static TAILQ_HEAD(,bio_ops) bio_ops_list = TAILQ_HEAD_INITIALIZER(bio_ops_list); 132 133 /* 134 * Called from vfsinit() 135 */ 136 void 137 vfs_mount_init(void) 138 { 139 lwkt_token_init(&mountlist_token, 1, "mntlist"); 140 lwkt_token_init(&mntvnode_token, 1, "mntvnode"); 141 lwkt_token_init(&mntid_token, 1, "mntid"); 142 TAILQ_INIT(&mountscan_list); 143 TAILQ_INIT(&mntvnodescan_list); 144 mount_init(&dummymount); 145 dummymount.mnt_flag |= MNT_RDONLY; 146 dummymount.mnt_kern_flag |= MNTK_ALL_MPSAFE; 147 } 148 149 /* 150 * Support function called with mntvnode_token held to remove a vnode 151 * from the mountlist. We must update any list scans which are in progress. 152 */ 153 static void 154 vremovevnodemnt(struct vnode *vp) 155 { 156 struct vmntvnodescan_info *info; 157 158 TAILQ_FOREACH(info, &mntvnodescan_list, entry) { 159 if (info->vp == vp) 160 info->vp = TAILQ_NEXT(vp, v_nmntvnodes); 161 } 162 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes); 163 } 164 165 /* 166 * Allocate a new vnode and associate it with a tag, mount point, and 167 * operations vector. 168 * 169 * A VX locked and refd vnode is returned. The caller should setup the 170 * remaining fields and vx_put() or, if he wishes to leave a vref, 171 * vx_unlock() the vnode. 172 */ 173 int 174 getnewvnode(enum vtagtype tag, struct mount *mp, 175 struct vnode **vpp, int lktimeout, int lkflags) 176 { 177 struct vnode *vp; 178 179 KKASSERT(mp != NULL); 180 181 vp = allocvnode(lktimeout, lkflags); 182 vp->v_tag = tag; 183 vp->v_data = NULL; 184 185 /* 186 * By default the vnode is assigned the mount point's normal 187 * operations vector. 188 */ 189 vp->v_ops = &mp->mnt_vn_use_ops; 190 191 /* 192 * Placing the vnode on the mount point's queue makes it visible. 193 * VNON prevents it from being messed with, however. 194 */ 195 insmntque(vp, mp); 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 but is still used by vfs_conf to 209 * create vnodes for devices prior to the root mount (with mp == NULL). 210 */ 211 int 212 getspecialvnode(enum vtagtype tag, struct mount *mp, 213 struct vop_ops **ops, 214 struct vnode **vpp, int lktimeout, int lkflags) 215 { 216 struct vnode *vp; 217 218 vp = allocvnode(lktimeout, lkflags); 219 vp->v_tag = tag; 220 vp->v_data = NULL; 221 vp->v_ops = ops; 222 223 if (mp == NULL) 224 mp = &dummymount; 225 226 /* 227 * Placing the vnode on the mount point's queue makes it visible. 228 * VNON prevents it from being messed with, however. 229 */ 230 insmntque(vp, mp); 231 232 /* 233 * A VX locked & refd vnode is returned. 234 */ 235 *vpp = vp; 236 return (0); 237 } 238 239 /* 240 * Interlock against an unmount, return 0 on success, non-zero on failure. 241 * 242 * The passed flag may be 0 or LK_NOWAIT and is only used if an unmount 243 * is in-progress. 244 * 245 * If no unmount is in-progress LK_NOWAIT is ignored. No other flag bits 246 * are used. A shared locked will be obtained and the filesystem will not 247 * be unmountable until the lock is released. 248 */ 249 int 250 vfs_busy(struct mount *mp, int flags) 251 { 252 int lkflags; 253 254 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 255 if (flags & LK_NOWAIT) 256 return (ENOENT); 257 /* XXX not MP safe */ 258 mp->mnt_kern_flag |= MNTK_MWAIT; 259 /* 260 * Since all busy locks are shared except the exclusive 261 * lock granted when unmounting, the only place that a 262 * wakeup needs to be done is at the release of the 263 * exclusive lock at the end of dounmount. 264 */ 265 tsleep((caddr_t)mp, 0, "vfs_busy", 0); 266 return (ENOENT); 267 } 268 lkflags = LK_SHARED; 269 if (lockmgr(&mp->mnt_lock, lkflags)) 270 panic("vfs_busy: unexpected lock failure"); 271 return (0); 272 } 273 274 /* 275 * Free a busy filesystem. 276 */ 277 void 278 vfs_unbusy(struct mount *mp) 279 { 280 lockmgr(&mp->mnt_lock, LK_RELEASE); 281 } 282 283 /* 284 * Lookup a filesystem type, and if found allocate and initialize 285 * a mount structure for it. 286 * 287 * Devname is usually updated by mount(8) after booting. 288 */ 289 int 290 vfs_rootmountalloc(char *fstypename, char *devname, struct mount **mpp) 291 { 292 struct vfsconf *vfsp; 293 struct mount *mp; 294 295 if (fstypename == NULL) 296 return (ENODEV); 297 298 vfsp = vfsconf_find_by_name(fstypename); 299 if (vfsp == NULL) 300 return (ENODEV); 301 mp = kmalloc(sizeof(struct mount), M_MOUNT, M_WAITOK | M_ZERO); 302 mount_init(mp); 303 lockinit(&mp->mnt_lock, "vfslock", VLKTIMEOUT, 0); 304 305 vfs_busy(mp, LK_NOWAIT); 306 mp->mnt_vfc = vfsp; 307 mp->mnt_op = vfsp->vfc_vfsops; 308 vfsp->vfc_refcount++; 309 mp->mnt_stat.f_type = vfsp->vfc_typenum; 310 mp->mnt_flag |= MNT_RDONLY; 311 mp->mnt_flag |= vfsp->vfc_flags & MNT_VISFLAGMASK; 312 strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN); 313 copystr(devname, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, 0); 314 *mpp = mp; 315 return (0); 316 } 317 318 /* 319 * Basic mount structure initialization 320 */ 321 void 322 mount_init(struct mount *mp) 323 { 324 lockinit(&mp->mnt_lock, "vfslock", 0, 0); 325 lwkt_token_init(&mp->mnt_token, 1, "permnt"); 326 327 TAILQ_INIT(&mp->mnt_nvnodelist); 328 TAILQ_INIT(&mp->mnt_reservedvnlist); 329 TAILQ_INIT(&mp->mnt_jlist); 330 mp->mnt_nvnodelistsize = 0; 331 mp->mnt_flag = 0; 332 mp->mnt_iosize_max = DFLTPHYS; 333 } 334 335 /* 336 * Lookup a mount point by filesystem identifier. 337 */ 338 struct mount * 339 vfs_getvfs(fsid_t *fsid) 340 { 341 struct mount *mp; 342 343 lwkt_gettoken(&mountlist_token); 344 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 345 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 346 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 347 break; 348 } 349 } 350 lwkt_reltoken(&mountlist_token); 351 return (mp); 352 } 353 354 /* 355 * Get a new unique fsid. Try to make its val[0] unique, since this value 356 * will be used to create fake device numbers for stat(). Also try (but 357 * not so hard) make its val[0] unique mod 2^16, since some emulators only 358 * support 16-bit device numbers. We end up with unique val[0]'s for the 359 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 360 * 361 * Keep in mind that several mounts may be running in parallel. Starting 362 * the search one past where the previous search terminated is both a 363 * micro-optimization and a defense against returning the same fsid to 364 * different mounts. 365 */ 366 void 367 vfs_getnewfsid(struct mount *mp) 368 { 369 static u_int16_t mntid_base; 370 fsid_t tfsid; 371 int mtype; 372 373 lwkt_gettoken(&mntid_token); 374 mtype = mp->mnt_vfc->vfc_typenum; 375 tfsid.val[1] = mtype; 376 mtype = (mtype & 0xFF) << 24; 377 for (;;) { 378 tfsid.val[0] = makeudev(255, 379 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 380 mntid_base++; 381 if (vfs_getvfs(&tfsid) == NULL) 382 break; 383 } 384 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 385 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 386 lwkt_reltoken(&mntid_token); 387 } 388 389 /* 390 * Set the FSID for a new mount point to the template. Adjust 391 * the FSID to avoid collisions. 392 */ 393 int 394 vfs_setfsid(struct mount *mp, fsid_t *template) 395 { 396 int didmunge = 0; 397 398 bzero(&mp->mnt_stat.f_fsid, sizeof(mp->mnt_stat.f_fsid)); 399 for (;;) { 400 if (vfs_getvfs(template) == NULL) 401 break; 402 didmunge = 1; 403 ++template->val[1]; 404 } 405 mp->mnt_stat.f_fsid = *template; 406 return(didmunge); 407 } 408 409 /* 410 * This routine is called when we have too many vnodes. It attempts 411 * to free <count> vnodes and will potentially free vnodes that still 412 * have VM backing store (VM backing store is typically the cause 413 * of a vnode blowout so we want to do this). Therefore, this operation 414 * is not considered cheap. 415 * 416 * A number of conditions may prevent a vnode from being reclaimed. 417 * the buffer cache may have references on the vnode, a directory 418 * vnode may still have references due to the namei cache representing 419 * underlying files, or the vnode may be in active use. It is not 420 * desireable to reuse such vnodes. These conditions may cause the 421 * number of vnodes to reach some minimum value regardless of what 422 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 423 */ 424 425 /* 426 * This is a quick non-blocking check to determine if the vnode is a good 427 * candidate for being (eventually) vgone()'d. Returns 0 if the vnode is 428 * not a good candidate, 1 if it is. 429 */ 430 static __inline int 431 vmightfree(struct vnode *vp, int page_count, int pass) 432 { 433 if (vp->v_flag & VRECLAIMED) 434 return (0); 435 #if 0 436 if ((vp->v_flag & VFREE) && TAILQ_EMPTY(&vp->v_namecache)) 437 return (0); 438 #endif 439 if (sysref_isactive(&vp->v_sysref)) 440 return (0); 441 if (vp->v_object && vp->v_object->resident_page_count >= page_count) 442 return (0); 443 444 /* 445 * XXX horrible hack. Up to four passes will be taken. Each pass 446 * makes a larger set of vnodes eligible. For now what this really 447 * means is that we try to recycle files opened only once before 448 * recycling files opened multiple times. 449 */ 450 switch(vp->v_flag & (VAGE0 | VAGE1)) { 451 case 0: 452 if (pass < 3) 453 return(0); 454 break; 455 case VAGE0: 456 if (pass < 2) 457 return(0); 458 break; 459 case VAGE1: 460 if (pass < 1) 461 return(0); 462 break; 463 case VAGE0 | VAGE1: 464 break; 465 } 466 return (1); 467 } 468 469 /* 470 * The vnode was found to be possibly vgone()able and the caller has locked it 471 * (thus the usecount should be 1 now). Determine if the vnode is actually 472 * vgone()able, doing some cleanups in the process. Returns 1 if the vnode 473 * can be vgone()'d, 0 otherwise. 474 * 475 * Note that v_auxrefs may be non-zero because (A) this vnode is not a leaf 476 * in the namecache topology and (B) this vnode has buffer cache bufs. 477 * We cannot remove vnodes with non-leaf namecache associations. We do a 478 * tentitive leaf check prior to attempting to flush out any buffers but the 479 * 'real' test when all is said in done is that v_auxrefs must become 0 for 480 * the vnode to be freeable. 481 * 482 * We could theoretically just unconditionally flush when v_auxrefs != 0, 483 * but flushing data associated with non-leaf nodes (which are always 484 * directories), just throws it away for no benefit. It is the buffer 485 * cache's responsibility to choose buffers to recycle from the cached 486 * data point of view. 487 */ 488 static int 489 visleaf(struct vnode *vp) 490 { 491 struct namecache *ncp; 492 493 spin_lock(&vp->v_spinlock); 494 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 495 if (!TAILQ_EMPTY(&ncp->nc_list)) { 496 spin_unlock(&vp->v_spinlock); 497 return(0); 498 } 499 } 500 spin_unlock(&vp->v_spinlock); 501 return(1); 502 } 503 504 /* 505 * Try to clean up the vnode to the point where it can be vgone()'d, returning 506 * 0 if it cannot be vgone()'d (or already has been), 1 if it can. Unlike 507 * vmightfree() this routine may flush the vnode and block. Vnodes marked 508 * VFREE are still candidates for vgone()ing because they may hold namecache 509 * resources and could be blocking the namecache directory hierarchy (and 510 * related vnodes) from being freed. 511 */ 512 static int 513 vtrytomakegoneable(struct vnode *vp, int page_count) 514 { 515 if (vp->v_flag & VRECLAIMED) 516 return (0); 517 if (vp->v_sysref.refcnt > 1) 518 return (0); 519 if (vp->v_object && vp->v_object->resident_page_count >= page_count) 520 return (0); 521 if (vp->v_auxrefs && visleaf(vp)) { 522 vinvalbuf(vp, V_SAVE, 0, 0); 523 #if 0 /* DEBUG */ 524 kprintf((vp->v_auxrefs ? "vrecycle: vp %p failed: %s\n" : 525 "vrecycle: vp %p succeeded: %s\n"), vp, 526 (TAILQ_FIRST(&vp->v_namecache) ? 527 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?")); 528 #endif 529 } 530 531 /* 532 * This sequence may seem a little strange, but we need to optimize 533 * the critical path a bit. We can't recycle vnodes with other 534 * references and because we are trying to recycle an otherwise 535 * perfectly fine vnode we have to invalidate the namecache in a 536 * way that avoids possible deadlocks (since the vnode lock is being 537 * held here). Finally, we have to check for other references one 538 * last time in case something snuck in during the inval. 539 */ 540 if (vp->v_sysref.refcnt > 1 || vp->v_auxrefs != 0) 541 return (0); 542 if (cache_inval_vp_nonblock(vp)) 543 return (0); 544 return (vp->v_sysref.refcnt <= 1 && vp->v_auxrefs == 0); 545 } 546 547 /* 548 * Reclaim up to 1/10 of the vnodes associated with a mount point. Try 549 * to avoid vnodes which have lots of resident pages (we are trying to free 550 * vnodes, not memory). 551 * 552 * This routine is a callback from the mountlist scan. The mount point 553 * in question will be busied. 554 * 555 * NOTE: The 1/10 reclamation also ensures that the inactive data set 556 * (the vnodes being recycled by the one-time use) does not degenerate 557 * into too-small a set. This is important because once a vnode is 558 * marked as not being one-time-use (VAGE0/VAGE1 both 0) that vnode 559 * will not be destroyed EXCEPT by this mechanism. VM pages can still 560 * be cleaned/freed by the pageout daemon. 561 */ 562 static int 563 vlrureclaim(struct mount *mp, void *data) 564 { 565 struct vnlru_info *info = data; 566 struct vnode *vp; 567 int done; 568 int trigger; 569 int usevnodes; 570 int count; 571 int trigger_mult = vnlru_nowhere; 572 573 /* 574 * Calculate the trigger point for the resident pages check. The 575 * minimum trigger value is approximately the number of pages in 576 * the system divded by the number of vnodes. However, due to 577 * various other system memory overheads unrelated to data caching 578 * it is a good idea to double the trigger (at least). 579 * 580 * trigger_mult starts at 0. If the recycler is having problems 581 * finding enough freeable vnodes it will increase trigger_mult. 582 * This should not happen in normal operation, even on machines with 583 * low amounts of memory, but extraordinary memory use by the system 584 * verses the amount of cached data can trigger it. 585 */ 586 usevnodes = desiredvnodes; 587 if (usevnodes <= 0) 588 usevnodes = 1; 589 trigger = vmstats.v_page_count * (trigger_mult + 2) / usevnodes; 590 591 done = 0; 592 lwkt_gettoken(&mntvnode_token); 593 count = mp->mnt_nvnodelistsize / 10 + 1; 594 595 while (count && mp->mnt_syncer) { 596 /* 597 * Next vnode. Use the special syncer vnode to placemark 598 * the LRU. This way the LRU code does not interfere with 599 * vmntvnodescan(). 600 */ 601 vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes); 602 TAILQ_REMOVE(&mp->mnt_nvnodelist, mp->mnt_syncer, v_nmntvnodes); 603 if (vp) { 604 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, 605 mp->mnt_syncer, v_nmntvnodes); 606 } else { 607 TAILQ_INSERT_HEAD(&mp->mnt_nvnodelist, mp->mnt_syncer, 608 v_nmntvnodes); 609 vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes); 610 if (vp == NULL) 611 break; 612 } 613 614 /* 615 * __VNODESCAN__ 616 * 617 * The VP will stick around while we hold mntvnode_token, 618 * at least until we block, so we can safely do an initial 619 * check, and then must check again after we lock the vnode. 620 */ 621 if (vp->v_type == VNON || /* syncer or indeterminant */ 622 !vmightfree(vp, trigger, info->pass) /* critical path opt */ 623 ) { 624 --count; 625 continue; 626 } 627 628 /* 629 * VX get the candidate vnode. If the VX get fails the 630 * vnode might still be on the mountlist. Our loop depends 631 * on us at least cycling the vnode to the end of the 632 * mountlist. 633 */ 634 if (vx_get_nonblock(vp) != 0) { 635 --count; 636 continue; 637 } 638 639 /* 640 * Since we blocked locking the vp, make sure it is still 641 * a candidate for reclamation. That is, it has not already 642 * been reclaimed and only has our VX reference associated 643 * with it. 644 */ 645 if (vp->v_type == VNON || /* syncer or indeterminant */ 646 (vp->v_flag & VRECLAIMED) || 647 vp->v_mount != mp || 648 !vtrytomakegoneable(vp, trigger) /* critical path opt */ 649 ) { 650 --count; 651 vx_put(vp); 652 continue; 653 } 654 655 /* 656 * All right, we are good, move the vp to the end of the 657 * mountlist and clean it out. The vget will have returned 658 * an error if the vnode was destroyed (VRECLAIMED set), so we 659 * do not have to check again. The vput() will move the 660 * vnode to the free list if the vgone() was successful. 661 */ 662 KKASSERT(vp->v_mount == mp); 663 vgone_vxlocked(vp); 664 vx_put(vp); 665 ++done; 666 --count; 667 } 668 lwkt_reltoken(&mntvnode_token); 669 return (done); 670 } 671 672 /* 673 * Attempt to recycle vnodes in a context that is always safe to block. 674 * Calling vlrurecycle() from the bowels of file system code has some 675 * interesting deadlock problems. 676 */ 677 static struct thread *vnlruthread; 678 static int vnlruproc_sig; 679 680 void 681 vnlru_proc_wait(void) 682 { 683 if (vnlruproc_sig == 0) { 684 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 685 wakeup(vnlruthread); 686 } 687 tsleep(&vnlruproc_sig, 0, "vlruwk", hz); 688 } 689 690 static void 691 vnlru_proc(void) 692 { 693 struct thread *td = curthread; 694 struct vnlru_info info; 695 int done; 696 697 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td, 698 SHUTDOWN_PRI_FIRST); 699 700 get_mplock(); 701 crit_enter(); 702 703 for (;;) { 704 kproc_suspend_loop(); 705 706 /* 707 * Try to free some vnodes if we have too many 708 */ 709 if (numvnodes > desiredvnodes && 710 freevnodes > desiredvnodes * 2 / 10) { 711 int count = numvnodes - desiredvnodes; 712 713 if (count > freevnodes / 100) 714 count = freevnodes / 100; 715 if (count < 5) 716 count = 5; 717 freesomevnodes(count); 718 } 719 720 /* 721 * Nothing to do if most of our vnodes are already on 722 * the free list. 723 */ 724 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { 725 vnlruproc_sig = 0; 726 wakeup(&vnlruproc_sig); 727 tsleep(td, 0, "vlruwt", hz); 728 continue; 729 } 730 cache_hysteresis(); 731 732 /* 733 * The pass iterates through the four combinations of 734 * VAGE0/VAGE1. We want to get rid of aged small files 735 * first. 736 */ 737 info.pass = 0; 738 done = 0; 739 while (done == 0 && info.pass < 4) { 740 done = mountlist_scan(vlrureclaim, &info, 741 MNTSCAN_FORWARD); 742 ++info.pass; 743 } 744 745 /* 746 * The vlrureclaim() call only processes 1/10 of the vnodes 747 * on each mount. If we couldn't find any repeat the loop 748 * at least enough times to cover all available vnodes before 749 * we start sleeping. Complain if the failure extends past 750 * 30 second, every 30 seconds. 751 */ 752 if (done == 0) { 753 ++vnlru_nowhere; 754 if (vnlru_nowhere % 10 == 0) 755 tsleep(td, 0, "vlrup", hz * 3); 756 if (vnlru_nowhere % 100 == 0) 757 kprintf("vnlru_proc: vnode recycler stopped working!\n"); 758 if (vnlru_nowhere == 1000) 759 vnlru_nowhere = 900; 760 } else { 761 vnlru_nowhere = 0; 762 } 763 } 764 765 crit_exit(); 766 rel_mplock(); 767 } 768 769 /* 770 * MOUNTLIST FUNCTIONS 771 */ 772 773 /* 774 * mountlist_insert (MP SAFE) 775 * 776 * Add a new mount point to the mount list. 777 */ 778 void 779 mountlist_insert(struct mount *mp, int how) 780 { 781 lwkt_gettoken(&mountlist_token); 782 if (how == MNTINS_FIRST) 783 TAILQ_INSERT_HEAD(&mountlist, mp, mnt_list); 784 else 785 TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list); 786 lwkt_reltoken(&mountlist_token); 787 } 788 789 /* 790 * mountlist_interlock (MP SAFE) 791 * 792 * Execute the specified interlock function with the mountlist token 793 * held. The function will be called in a serialized fashion verses 794 * other functions called through this mechanism. 795 */ 796 int 797 mountlist_interlock(int (*callback)(struct mount *), struct mount *mp) 798 { 799 int error; 800 801 lwkt_gettoken(&mountlist_token); 802 error = callback(mp); 803 lwkt_reltoken(&mountlist_token); 804 return (error); 805 } 806 807 /* 808 * mountlist_boot_getfirst (DURING BOOT ONLY) 809 * 810 * This function returns the first mount on the mountlist, which is 811 * expected to be the root mount. Since no interlocks are obtained 812 * this function is only safe to use during booting. 813 */ 814 815 struct mount * 816 mountlist_boot_getfirst(void) 817 { 818 return(TAILQ_FIRST(&mountlist)); 819 } 820 821 /* 822 * mountlist_remove (MP SAFE) 823 * 824 * Remove a node from the mountlist. If this node is the next scan node 825 * for any active mountlist scans, the active mountlist scan will be 826 * adjusted to skip the node, thus allowing removals during mountlist 827 * scans. 828 */ 829 void 830 mountlist_remove(struct mount *mp) 831 { 832 struct mountscan_info *msi; 833 834 lwkt_gettoken(&mountlist_token); 835 TAILQ_FOREACH(msi, &mountscan_list, msi_entry) { 836 if (msi->msi_node == mp) { 837 if (msi->msi_how & MNTSCAN_FORWARD) 838 msi->msi_node = TAILQ_NEXT(mp, mnt_list); 839 else 840 msi->msi_node = TAILQ_PREV(mp, mntlist, mnt_list); 841 } 842 } 843 TAILQ_REMOVE(&mountlist, mp, mnt_list); 844 lwkt_reltoken(&mountlist_token); 845 } 846 847 /* 848 * mountlist_scan (MP SAFE) 849 * 850 * Safely scan the mount points on the mount list. Unless otherwise 851 * specified each mount point will be busied prior to the callback and 852 * unbusied afterwords. The callback may safely remove any mount point 853 * without interfering with the scan. If the current callback 854 * mount is removed the scanner will not attempt to unbusy it. 855 * 856 * If a mount node cannot be busied it is silently skipped. 857 * 858 * The callback return value is aggregated and a total is returned. A return 859 * value of < 0 is not aggregated and will terminate the scan. 860 * 861 * MNTSCAN_FORWARD - the mountlist is scanned in the forward direction 862 * MNTSCAN_REVERSE - the mountlist is scanned in reverse 863 * MNTSCAN_NOBUSY - the scanner will make the callback without busying 864 * the mount node. 865 */ 866 int 867 mountlist_scan(int (*callback)(struct mount *, void *), void *data, int how) 868 { 869 struct mountscan_info info; 870 struct mount *mp; 871 thread_t td; 872 int count; 873 int res; 874 875 lwkt_gettoken(&mountlist_token); 876 877 info.msi_how = how; 878 info.msi_node = NULL; /* paranoia */ 879 TAILQ_INSERT_TAIL(&mountscan_list, &info, msi_entry); 880 881 res = 0; 882 td = curthread; 883 884 if (how & MNTSCAN_FORWARD) { 885 info.msi_node = TAILQ_FIRST(&mountlist); 886 while ((mp = info.msi_node) != NULL) { 887 if (how & MNTSCAN_NOBUSY) { 888 count = callback(mp, data); 889 } else if (vfs_busy(mp, LK_NOWAIT) == 0) { 890 count = callback(mp, data); 891 if (mp == info.msi_node) 892 vfs_unbusy(mp); 893 } else { 894 count = 0; 895 } 896 if (count < 0) 897 break; 898 res += count; 899 if (mp == info.msi_node) 900 info.msi_node = TAILQ_NEXT(mp, mnt_list); 901 } 902 } else if (how & MNTSCAN_REVERSE) { 903 info.msi_node = TAILQ_LAST(&mountlist, mntlist); 904 while ((mp = info.msi_node) != NULL) { 905 if (how & MNTSCAN_NOBUSY) { 906 count = callback(mp, data); 907 } else if (vfs_busy(mp, LK_NOWAIT) == 0) { 908 count = callback(mp, data); 909 if (mp == info.msi_node) 910 vfs_unbusy(mp); 911 } else { 912 count = 0; 913 } 914 if (count < 0) 915 break; 916 res += count; 917 if (mp == info.msi_node) 918 info.msi_node = TAILQ_PREV(mp, mntlist, mnt_list); 919 } 920 } 921 TAILQ_REMOVE(&mountscan_list, &info, msi_entry); 922 lwkt_reltoken(&mountlist_token); 923 return(res); 924 } 925 926 /* 927 * MOUNT RELATED VNODE FUNCTIONS 928 */ 929 930 static struct kproc_desc vnlru_kp = { 931 "vnlru", 932 vnlru_proc, 933 &vnlruthread 934 }; 935 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 936 937 /* 938 * Move a vnode from one mount queue to another. 939 * 940 * MPSAFE 941 */ 942 void 943 insmntque(struct vnode *vp, struct mount *mp) 944 { 945 lwkt_gettoken(&mntvnode_token); 946 /* 947 * Delete from old mount point vnode list, if on one. 948 */ 949 if (vp->v_mount != NULL) { 950 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0, 951 ("bad mount point vnode list size")); 952 vremovevnodemnt(vp); 953 vp->v_mount->mnt_nvnodelistsize--; 954 } 955 /* 956 * Insert into list of vnodes for the new mount point, if available. 957 * The 'end' of the LRU list is the vnode prior to mp->mnt_syncer. 958 */ 959 if ((vp->v_mount = mp) == NULL) { 960 lwkt_reltoken(&mntvnode_token); 961 return; 962 } 963 if (mp->mnt_syncer) { 964 TAILQ_INSERT_BEFORE(mp->mnt_syncer, vp, v_nmntvnodes); 965 } else { 966 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 967 } 968 mp->mnt_nvnodelistsize++; 969 lwkt_reltoken(&mntvnode_token); 970 } 971 972 973 /* 974 * Scan the vnodes under a mount point and issue appropriate callbacks. 975 * 976 * The fastfunc() callback is called with just the mountlist token held 977 * (no vnode lock). It may not block and the vnode may be undergoing 978 * modifications while the caller is processing it. The vnode will 979 * not be entirely destroyed, however, due to the fact that the mountlist 980 * token is held. A return value < 0 skips to the next vnode without calling 981 * the slowfunc(), a return value > 0 terminates the loop. 982 * 983 * The slowfunc() callback is called after the vnode has been successfully 984 * locked based on passed flags. The vnode is skipped if it gets rearranged 985 * or destroyed while blocking on the lock. A non-zero return value from 986 * the slow function terminates the loop. The slow function is allowed to 987 * arbitrarily block. The scanning code guarentees consistency of operation 988 * even if the slow function deletes or moves the node, or blocks and some 989 * other thread deletes or moves the node. 990 * 991 * NOTE: We hold vmobj_token to prevent a VM object from being destroyed 992 * out from under the fastfunc()'s vnode test. It will not prevent 993 * v_object from getting NULL'd out but it will ensure that the 994 * pointer (if we race) will remain stable. 995 */ 996 int 997 vmntvnodescan( 998 struct mount *mp, 999 int flags, 1000 int (*fastfunc)(struct mount *mp, struct vnode *vp, void *data), 1001 int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data), 1002 void *data 1003 ) { 1004 struct vmntvnodescan_info info; 1005 struct vnode *vp; 1006 int r = 0; 1007 int maxcount = 1000000; 1008 int stopcount = 0; 1009 int count = 0; 1010 1011 lwkt_gettoken(&mntvnode_token); 1012 lwkt_gettoken(&vmobj_token); 1013 1014 /* 1015 * If asked to do one pass stop after iterating available vnodes. 1016 * Under heavy loads new vnodes can be added while we are scanning, 1017 * so this isn't perfect. Create a slop factor of 2x. 1018 */ 1019 if (flags & VMSC_ONEPASS) 1020 stopcount = mp->mnt_nvnodelistsize * 2; 1021 1022 info.vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 1023 TAILQ_INSERT_TAIL(&mntvnodescan_list, &info, entry); 1024 while ((vp = info.vp) != NULL) { 1025 if (--maxcount == 0) 1026 panic("maxcount reached during vmntvnodescan"); 1027 1028 /* 1029 * Skip if visible but not ready, or special (e.g. 1030 * mp->mnt_syncer) 1031 */ 1032 if (vp->v_type == VNON) 1033 goto next; 1034 KKASSERT(vp->v_mount == mp); 1035 1036 /* 1037 * Quick test. A negative return continues the loop without 1038 * calling the slow test. 0 continues onto the slow test. 1039 * A positive number aborts the loop. 1040 */ 1041 if (fastfunc) { 1042 if ((r = fastfunc(mp, vp, data)) < 0) { 1043 r = 0; 1044 goto next; 1045 } 1046 if (r) 1047 break; 1048 } 1049 1050 /* 1051 * Get a vxlock on the vnode, retry if it has moved or isn't 1052 * in the mountlist where we expect it. 1053 */ 1054 if (slowfunc) { 1055 int error; 1056 1057 switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) { 1058 case VMSC_GETVP: 1059 error = vget(vp, LK_EXCLUSIVE); 1060 break; 1061 case VMSC_GETVP|VMSC_NOWAIT: 1062 error = vget(vp, LK_EXCLUSIVE|LK_NOWAIT); 1063 break; 1064 case VMSC_GETVX: 1065 vx_get(vp); 1066 error = 0; 1067 break; 1068 default: 1069 error = 0; 1070 break; 1071 } 1072 if (error) 1073 goto next; 1074 /* 1075 * Do not call the slow function if the vnode is 1076 * invalid or if it was ripped out from under us 1077 * while we (potentially) blocked. 1078 */ 1079 if (info.vp == vp && vp->v_type != VNON) 1080 r = slowfunc(mp, vp, data); 1081 1082 /* 1083 * Cleanup 1084 */ 1085 switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) { 1086 case VMSC_GETVP: 1087 case VMSC_GETVP|VMSC_NOWAIT: 1088 vput(vp); 1089 break; 1090 case VMSC_GETVX: 1091 vx_put(vp); 1092 break; 1093 default: 1094 break; 1095 } 1096 if (r != 0) 1097 break; 1098 } 1099 1100 next: 1101 /* 1102 * Yield after some processing. Depending on the number 1103 * of vnodes, we might wind up running for a long time. 1104 * Because threads are not preemptable, time critical 1105 * userland processes might starve. Give them a chance 1106 * now and then. 1107 */ 1108 if (++count == 10000) { 1109 /* We really want to yield a bit, so we simply sleep a tick */ 1110 tsleep(mp, 0, "vnodescn", 1); 1111 count = 0; 1112 } 1113 1114 /* 1115 * If doing one pass this decrements to zero. If it starts 1116 * at zero it is effectively unlimited for the purposes of 1117 * this loop. 1118 */ 1119 if (--stopcount == 0) 1120 break; 1121 1122 /* 1123 * Iterate. If the vnode was ripped out from under us 1124 * info.vp will already point to the next vnode, otherwise 1125 * we have to obtain the next valid vnode ourselves. 1126 */ 1127 if (info.vp == vp) 1128 info.vp = TAILQ_NEXT(vp, v_nmntvnodes); 1129 } 1130 TAILQ_REMOVE(&mntvnodescan_list, &info, entry); 1131 lwkt_reltoken(&vmobj_token); 1132 lwkt_reltoken(&mntvnode_token); 1133 return(r); 1134 } 1135 1136 /* 1137 * Remove any vnodes in the vnode table belonging to mount point mp. 1138 * 1139 * If FORCECLOSE is not specified, there should not be any active ones, 1140 * return error if any are found (nb: this is a user error, not a 1141 * system error). If FORCECLOSE is specified, detach any active vnodes 1142 * that are found. 1143 * 1144 * If WRITECLOSE is set, only flush out regular file vnodes open for 1145 * writing. 1146 * 1147 * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped. 1148 * 1149 * `rootrefs' specifies the base reference count for the root vnode 1150 * of this filesystem. The root vnode is considered busy if its 1151 * v_sysref.refcnt exceeds this value. On a successful return, vflush() 1152 * will call vrele() on the root vnode exactly rootrefs times. 1153 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 1154 * be zero. 1155 */ 1156 #ifdef DIAGNOSTIC 1157 static int busyprt = 0; /* print out busy vnodes */ 1158 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 1159 #endif 1160 1161 static int vflush_scan(struct mount *mp, struct vnode *vp, void *data); 1162 1163 struct vflush_info { 1164 int flags; 1165 int busy; 1166 thread_t td; 1167 }; 1168 1169 int 1170 vflush(struct mount *mp, int rootrefs, int flags) 1171 { 1172 struct thread *td = curthread; /* XXX */ 1173 struct vnode *rootvp = NULL; 1174 int error; 1175 struct vflush_info vflush_info; 1176 1177 if (rootrefs > 0) { 1178 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 1179 ("vflush: bad args")); 1180 /* 1181 * Get the filesystem root vnode. We can vput() it 1182 * immediately, since with rootrefs > 0, it won't go away. 1183 */ 1184 if ((error = VFS_ROOT(mp, &rootvp)) != 0) { 1185 if ((flags & FORCECLOSE) == 0) 1186 return (error); 1187 rootrefs = 0; 1188 /* continue anyway */ 1189 } 1190 if (rootrefs) 1191 vput(rootvp); 1192 } 1193 1194 vflush_info.busy = 0; 1195 vflush_info.flags = flags; 1196 vflush_info.td = td; 1197 vmntvnodescan(mp, VMSC_GETVX, NULL, vflush_scan, &vflush_info); 1198 1199 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 1200 /* 1201 * If just the root vnode is busy, and if its refcount 1202 * is equal to `rootrefs', then go ahead and kill it. 1203 */ 1204 KASSERT(vflush_info.busy > 0, ("vflush: not busy")); 1205 KASSERT(rootvp->v_sysref.refcnt >= rootrefs, ("vflush: rootrefs")); 1206 if (vflush_info.busy == 1 && rootvp->v_sysref.refcnt == rootrefs) { 1207 vx_lock(rootvp); 1208 vgone_vxlocked(rootvp); 1209 vx_unlock(rootvp); 1210 vflush_info.busy = 0; 1211 } 1212 } 1213 if (vflush_info.busy) 1214 return (EBUSY); 1215 for (; rootrefs > 0; rootrefs--) 1216 vrele(rootvp); 1217 return (0); 1218 } 1219 1220 /* 1221 * The scan callback is made with an VX locked vnode. 1222 */ 1223 static int 1224 vflush_scan(struct mount *mp, struct vnode *vp, void *data) 1225 { 1226 struct vflush_info *info = data; 1227 struct vattr vattr; 1228 1229 /* 1230 * Skip over a vnodes marked VSYSTEM. 1231 */ 1232 if ((info->flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) { 1233 return(0); 1234 } 1235 1236 /* 1237 * If WRITECLOSE is set, flush out unlinked but still open 1238 * files (even if open only for reading) and regular file 1239 * vnodes open for writing. 1240 */ 1241 if ((info->flags & WRITECLOSE) && 1242 (vp->v_type == VNON || 1243 (VOP_GETATTR(vp, &vattr) == 0 && 1244 vattr.va_nlink > 0)) && 1245 (vp->v_writecount == 0 || vp->v_type != VREG)) { 1246 return(0); 1247 } 1248 1249 /* 1250 * If we are the only holder (refcnt of 1) or the vnode is in 1251 * termination (refcnt < 0), we can vgone the vnode. 1252 */ 1253 if (vp->v_sysref.refcnt <= 1) { 1254 vgone_vxlocked(vp); 1255 return(0); 1256 } 1257 1258 /* 1259 * If FORCECLOSE is set, forcibly destroy the vnode and then move 1260 * it to a dummymount structure so vop_*() functions don't deref 1261 * a NULL pointer. 1262 */ 1263 if (info->flags & FORCECLOSE) { 1264 vhold(vp); 1265 vgone_vxlocked(vp); 1266 if (vp->v_mount == NULL) 1267 insmntque(vp, &dummymount); 1268 vdrop(vp); 1269 return(0); 1270 } 1271 #ifdef DIAGNOSTIC 1272 if (busyprt) 1273 vprint("vflush: busy vnode", vp); 1274 #endif 1275 ++info->busy; 1276 return(0); 1277 } 1278 1279 void 1280 add_bio_ops(struct bio_ops *ops) 1281 { 1282 TAILQ_INSERT_TAIL(&bio_ops_list, ops, entry); 1283 } 1284 1285 void 1286 rem_bio_ops(struct bio_ops *ops) 1287 { 1288 TAILQ_REMOVE(&bio_ops_list, ops, entry); 1289 } 1290 1291 /* 1292 * This calls the bio_ops io_sync function either for a mount point 1293 * or generally. 1294 * 1295 * WARNING: softdeps is weirdly coded and just isn't happy unless 1296 * io_sync is called with a NULL mount from the general syncing code. 1297 */ 1298 void 1299 bio_ops_sync(struct mount *mp) 1300 { 1301 struct bio_ops *ops; 1302 1303 if (mp) { 1304 if ((ops = mp->mnt_bioops) != NULL) 1305 ops->io_sync(mp); 1306 } else { 1307 TAILQ_FOREACH(ops, &bio_ops_list, entry) { 1308 ops->io_sync(NULL); 1309 } 1310 } 1311 } 1312 1313 /* 1314 * Lookup a mount point by nch 1315 */ 1316 struct mount * 1317 mount_get_by_nc(struct namecache *ncp) 1318 { 1319 struct mount *mp = NULL; 1320 1321 lwkt_gettoken(&mountlist_token); 1322 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 1323 if (ncp == mp->mnt_ncmountpt.ncp) 1324 break; 1325 } 1326 lwkt_reltoken(&mountlist_token); 1327 return (mp); 1328 } 1329 1330