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 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 struct vnlru_info { 111 int pass; 112 }; 113 114 static int vnlru_nowhere = 0; 115 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RD, 116 &vnlru_nowhere, 0, 117 "Number of times the vnlru process ran without success"); 118 119 120 static struct lwkt_token mntid_token; 121 122 /* note: mountlist exported to pstat */ 123 struct mntlist mountlist = TAILQ_HEAD_INITIALIZER(mountlist); 124 static TAILQ_HEAD(,mountscan_info) mountscan_list; 125 static struct lwkt_token mountlist_token; 126 static TAILQ_HEAD(,vmntvnodescan_info) mntvnodescan_list; 127 struct lwkt_token mntvnode_token; 128 129 static TAILQ_HEAD(,bio_ops) bio_ops_list = TAILQ_HEAD_INITIALIZER(bio_ops_list); 130 131 /* 132 * Called from vfsinit() 133 */ 134 void 135 vfs_mount_init(void) 136 { 137 lwkt_token_init(&mountlist_token); 138 lwkt_token_init(&mntvnode_token); 139 lwkt_token_init(&mntid_token); 140 TAILQ_INIT(&mountscan_list); 141 TAILQ_INIT(&mntvnodescan_list); 142 } 143 144 /* 145 * Support function called with mntvnode_token held to remove a vnode 146 * from the mountlist. We must update any list scans which are in progress. 147 */ 148 static void 149 vremovevnodemnt(struct vnode *vp) 150 { 151 struct vmntvnodescan_info *info; 152 153 TAILQ_FOREACH(info, &mntvnodescan_list, entry) { 154 if (info->vp == vp) 155 info->vp = TAILQ_NEXT(vp, v_nmntvnodes); 156 } 157 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes); 158 } 159 160 /* 161 * Allocate a new vnode and associate it with a tag, mount point, and 162 * operations vector. 163 * 164 * A VX locked and refd vnode is returned. The caller should setup the 165 * remaining fields and vx_put() or, if he wishes to leave a vref, 166 * vx_unlock() the vnode. 167 */ 168 int 169 getnewvnode(enum vtagtype tag, struct mount *mp, 170 struct vnode **vpp, int lktimeout, int lkflags) 171 { 172 struct vnode *vp; 173 174 KKASSERT(mp != NULL); 175 176 vp = allocvnode(lktimeout, lkflags); 177 vp->v_tag = tag; 178 vp->v_data = NULL; 179 180 /* 181 * By default the vnode is assigned the mount point's normal 182 * operations vector. 183 */ 184 vp->v_ops = &mp->mnt_vn_use_ops; 185 186 /* 187 * Placing the vnode on the mount point's queue makes it visible. 188 * VNON prevents it from being messed with, however. 189 */ 190 insmntque(vp, mp); 191 192 /* 193 * A VX locked & refd vnode is returned. 194 */ 195 *vpp = vp; 196 return (0); 197 } 198 199 /* 200 * This function creates vnodes with special operations vectors. The 201 * mount point is optional. 202 * 203 * This routine is being phased out. 204 */ 205 int 206 getspecialvnode(enum vtagtype tag, struct mount *mp, 207 struct vop_ops **ops, 208 struct vnode **vpp, int lktimeout, int lkflags) 209 { 210 struct vnode *vp; 211 212 vp = allocvnode(lktimeout, lkflags); 213 vp->v_tag = tag; 214 vp->v_data = NULL; 215 vp->v_ops = ops; 216 217 /* 218 * Placing the vnode on the mount point's queue makes it visible. 219 * VNON prevents it from being messed with, however. 220 */ 221 insmntque(vp, mp); 222 223 /* 224 * A VX locked & refd vnode is returned. 225 */ 226 *vpp = vp; 227 return (0); 228 } 229 230 /* 231 * Interlock against an unmount, return 0 on success, non-zero on failure. 232 * 233 * The passed flag may be 0 or LK_NOWAIT and is only used if an unmount 234 * is in-progress. 235 * 236 * If no unmount is in-progress LK_NOWAIT is ignored. No other flag bits 237 * are used. A shared locked will be obtained and the filesystem will not 238 * be unmountable until the lock is released. 239 */ 240 int 241 vfs_busy(struct mount *mp, int flags) 242 { 243 int lkflags; 244 245 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 246 if (flags & LK_NOWAIT) 247 return (ENOENT); 248 /* XXX not MP safe */ 249 mp->mnt_kern_flag |= MNTK_MWAIT; 250 /* 251 * Since all busy locks are shared except the exclusive 252 * lock granted when unmounting, the only place that a 253 * wakeup needs to be done is at the release of the 254 * exclusive lock at the end of dounmount. 255 */ 256 tsleep((caddr_t)mp, 0, "vfs_busy", 0); 257 return (ENOENT); 258 } 259 lkflags = LK_SHARED; 260 if (lockmgr(&mp->mnt_lock, lkflags)) 261 panic("vfs_busy: unexpected lock failure"); 262 return (0); 263 } 264 265 /* 266 * Free a busy filesystem. 267 */ 268 void 269 vfs_unbusy(struct mount *mp) 270 { 271 lockmgr(&mp->mnt_lock, LK_RELEASE); 272 } 273 274 /* 275 * Lookup a filesystem type, and if found allocate and initialize 276 * a mount structure for it. 277 * 278 * Devname is usually updated by mount(8) after booting. 279 */ 280 int 281 vfs_rootmountalloc(char *fstypename, char *devname, struct mount **mpp) 282 { 283 struct vfsconf *vfsp; 284 struct mount *mp; 285 286 if (fstypename == NULL) 287 return (ENODEV); 288 289 vfsp = vfsconf_find_by_name(fstypename); 290 if (vfsp == NULL) 291 return (ENODEV); 292 mp = kmalloc(sizeof(struct mount), M_MOUNT, M_WAITOK | M_ZERO); 293 lockinit(&mp->mnt_lock, "vfslock", VLKTIMEOUT, 0); 294 vfs_busy(mp, LK_NOWAIT); 295 TAILQ_INIT(&mp->mnt_nvnodelist); 296 TAILQ_INIT(&mp->mnt_reservedvnlist); 297 TAILQ_INIT(&mp->mnt_jlist); 298 mp->mnt_nvnodelistsize = 0; 299 mp->mnt_vfc = vfsp; 300 mp->mnt_op = vfsp->vfc_vfsops; 301 mp->mnt_flag = MNT_RDONLY; 302 vfsp->vfc_refcount++; 303 mp->mnt_iosize_max = DFLTPHYS; 304 mp->mnt_stat.f_type = vfsp->vfc_typenum; 305 mp->mnt_flag |= vfsp->vfc_flags & MNT_VISFLAGMASK; 306 strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN); 307 copystr(devname, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, 0); 308 *mpp = mp; 309 return (0); 310 } 311 312 /* 313 * Lookup a mount point by filesystem identifier. 314 */ 315 struct mount * 316 vfs_getvfs(fsid_t *fsid) 317 { 318 struct mount *mp; 319 lwkt_tokref ilock; 320 321 lwkt_gettoken(&ilock, &mountlist_token); 322 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 323 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] && 324 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) { 325 break; 326 } 327 } 328 lwkt_reltoken(&ilock); 329 return (mp); 330 } 331 332 /* 333 * Get a new unique fsid. Try to make its val[0] unique, since this value 334 * will be used to create fake device numbers for stat(). Also try (but 335 * not so hard) make its val[0] unique mod 2^16, since some emulators only 336 * support 16-bit device numbers. We end up with unique val[0]'s for the 337 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. 338 * 339 * Keep in mind that several mounts may be running in parallel. Starting 340 * the search one past where the previous search terminated is both a 341 * micro-optimization and a defense against returning the same fsid to 342 * different mounts. 343 */ 344 void 345 vfs_getnewfsid(struct mount *mp) 346 { 347 static u_int16_t mntid_base; 348 lwkt_tokref ilock; 349 fsid_t tfsid; 350 int mtype; 351 352 lwkt_gettoken(&ilock, &mntid_token); 353 mtype = mp->mnt_vfc->vfc_typenum; 354 tfsid.val[1] = mtype; 355 mtype = (mtype & 0xFF) << 24; 356 for (;;) { 357 tfsid.val[0] = makeudev(255, 358 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); 359 mntid_base++; 360 if (vfs_getvfs(&tfsid) == NULL) 361 break; 362 } 363 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; 364 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; 365 lwkt_reltoken(&ilock); 366 } 367 368 /* 369 * Set the FSID for a new mount point to the template. Adjust 370 * the FSID to avoid collisions. 371 */ 372 int 373 vfs_setfsid(struct mount *mp, fsid_t *template) 374 { 375 int didmunge = 0; 376 377 bzero(&mp->mnt_stat.f_fsid, sizeof(mp->mnt_stat.f_fsid)); 378 for (;;) { 379 if (vfs_getvfs(template) == NULL) 380 break; 381 didmunge = 1; 382 ++template->val[1]; 383 } 384 mp->mnt_stat.f_fsid = *template; 385 return(didmunge); 386 } 387 388 /* 389 * This routine is called when we have too many vnodes. It attempts 390 * to free <count> vnodes and will potentially free vnodes that still 391 * have VM backing store (VM backing store is typically the cause 392 * of a vnode blowout so we want to do this). Therefore, this operation 393 * is not considered cheap. 394 * 395 * A number of conditions may prevent a vnode from being reclaimed. 396 * the buffer cache may have references on the vnode, a directory 397 * vnode may still have references due to the namei cache representing 398 * underlying files, or the vnode may be in active use. It is not 399 * desireable to reuse such vnodes. These conditions may cause the 400 * number of vnodes to reach some minimum value regardless of what 401 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. 402 */ 403 404 /* 405 * This is a quick non-blocking check to determine if the vnode is a good 406 * candidate for being (eventually) vgone()'d. Returns 0 if the vnode is 407 * not a good candidate, 1 if it is. 408 */ 409 static __inline int 410 vmightfree(struct vnode *vp, int page_count, int pass) 411 { 412 if (vp->v_flag & VRECLAIMED) 413 return (0); 414 #if 0 415 if ((vp->v_flag & VFREE) && TAILQ_EMPTY(&vp->v_namecache)) 416 return (0); 417 #endif 418 if (sysref_isactive(&vp->v_sysref)) 419 return (0); 420 if (vp->v_object && vp->v_object->resident_page_count >= page_count) 421 return (0); 422 423 /* 424 * XXX horrible hack. Up to four passes will be taken. Each pass 425 * makes a larger set of vnodes eligible. For now what this really 426 * means is that we try to recycle files opened only once before 427 * recycling files opened multiple times. 428 */ 429 switch(vp->v_flag & (VAGE0 | VAGE1)) { 430 case 0: 431 if (pass < 3) 432 return(0); 433 break; 434 case VAGE0: 435 if (pass < 2) 436 return(0); 437 break; 438 case VAGE1: 439 if (pass < 1) 440 return(0); 441 break; 442 case VAGE0 | VAGE1: 443 break; 444 } 445 return (1); 446 } 447 448 /* 449 * The vnode was found to be possibly vgone()able and the caller has locked it 450 * (thus the usecount should be 1 now). Determine if the vnode is actually 451 * vgone()able, doing some cleanups in the process. Returns 1 if the vnode 452 * can be vgone()'d, 0 otherwise. 453 * 454 * Note that v_auxrefs may be non-zero because (A) this vnode is not a leaf 455 * in the namecache topology and (B) this vnode has buffer cache bufs. 456 * We cannot remove vnodes with non-leaf namecache associations. We do a 457 * tentitive leaf check prior to attempting to flush out any buffers but the 458 * 'real' test when all is said in done is that v_auxrefs must become 0 for 459 * the vnode to be freeable. 460 * 461 * We could theoretically just unconditionally flush when v_auxrefs != 0, 462 * but flushing data associated with non-leaf nodes (which are always 463 * directories), just throws it away for no benefit. It is the buffer 464 * cache's responsibility to choose buffers to recycle from the cached 465 * data point of view. 466 */ 467 static int 468 visleaf(struct vnode *vp) 469 { 470 struct namecache *ncp; 471 472 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 473 if (!TAILQ_EMPTY(&ncp->nc_list)) 474 return(0); 475 } 476 return(1); 477 } 478 479 /* 480 * Try to clean up the vnode to the point where it can be vgone()'d, returning 481 * 0 if it cannot be vgone()'d (or already has been), 1 if it can. Unlike 482 * vmightfree() this routine may flush the vnode and block. Vnodes marked 483 * VFREE are still candidates for vgone()ing because they may hold namecache 484 * resources and could be blocking the namecache directory hierarchy (and 485 * related vnodes) from being freed. 486 */ 487 static int 488 vtrytomakegoneable(struct vnode *vp, int page_count) 489 { 490 if (vp->v_flag & VRECLAIMED) 491 return (0); 492 if (vp->v_sysref.refcnt > 1) 493 return (0); 494 if (vp->v_object && vp->v_object->resident_page_count >= page_count) 495 return (0); 496 if (vp->v_auxrefs && visleaf(vp)) { 497 vinvalbuf(vp, V_SAVE, 0, 0); 498 #if 0 /* DEBUG */ 499 kprintf((vp->v_auxrefs ? "vrecycle: vp %p failed: %s\n" : 500 "vrecycle: vp %p succeeded: %s\n"), vp, 501 (TAILQ_FIRST(&vp->v_namecache) ? 502 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?")); 503 #endif 504 } 505 506 /* 507 * This sequence may seem a little strange, but we need to optimize 508 * the critical path a bit. We can't recycle vnodes with other 509 * references and because we are trying to recycle an otherwise 510 * perfectly fine vnode we have to invalidate the namecache in a 511 * way that avoids possible deadlocks (since the vnode lock is being 512 * held here). Finally, we have to check for other references one 513 * last time in case something snuck in during the inval. 514 */ 515 if (vp->v_sysref.refcnt > 1 || vp->v_auxrefs != 0) 516 return (0); 517 if (cache_inval_vp_nonblock(vp)) 518 return (0); 519 return (vp->v_sysref.refcnt <= 1 && vp->v_auxrefs == 0); 520 } 521 522 /* 523 * Reclaim up to 1/10 of the vnodes associated with a mount point. Try 524 * to avoid vnodes which have lots of resident pages (we are trying to free 525 * vnodes, not memory). 526 * 527 * This routine is a callback from the mountlist scan. The mount point 528 * in question will be busied. 529 * 530 * NOTE: The 1/10 reclamation also ensures that the inactive data set 531 * (the vnodes being recycled by the one-time use) does not degenerate 532 * into too-small a set. This is important because once a vnode is 533 * marked as not being one-time-use (VAGE0/VAGE1 both 0) that vnode 534 * will not be destroyed EXCEPT by this mechanism. VM pages can still 535 * be cleaned/freed by the pageout daemon. 536 */ 537 static int 538 vlrureclaim(struct mount *mp, void *data) 539 { 540 struct vnlru_info *info = data; 541 struct vnode *vp; 542 lwkt_tokref ilock; 543 int done; 544 int trigger; 545 int usevnodes; 546 int count; 547 int trigger_mult = vnlru_nowhere; 548 549 /* 550 * Calculate the trigger point for the resident pages check. The 551 * minimum trigger value is approximately the number of pages in 552 * the system divded by the number of vnodes. However, due to 553 * various other system memory overheads unrelated to data caching 554 * it is a good idea to double the trigger (at least). 555 * 556 * trigger_mult starts at 0. If the recycler is having problems 557 * finding enough freeable vnodes it will increase trigger_mult. 558 * This should not happen in normal operation, even on machines with 559 * low amounts of memory, but extraordinary memory use by the system 560 * verses the amount of cached data can trigger it. 561 */ 562 usevnodes = desiredvnodes; 563 if (usevnodes <= 0) 564 usevnodes = 1; 565 trigger = vmstats.v_page_count * (trigger_mult + 2) / usevnodes; 566 567 done = 0; 568 lwkt_gettoken(&ilock, &mntvnode_token); 569 count = mp->mnt_nvnodelistsize / 10 + 1; 570 571 while (count && mp->mnt_syncer) { 572 /* 573 * Next vnode. Use the special syncer vnode to placemark 574 * the LRU. This way the LRU code does not interfere with 575 * vmntvnodescan(). 576 */ 577 vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes); 578 TAILQ_REMOVE(&mp->mnt_nvnodelist, mp->mnt_syncer, v_nmntvnodes); 579 if (vp) { 580 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, 581 mp->mnt_syncer, v_nmntvnodes); 582 } else { 583 TAILQ_INSERT_HEAD(&mp->mnt_nvnodelist, mp->mnt_syncer, 584 v_nmntvnodes); 585 vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes); 586 if (vp == NULL) 587 break; 588 } 589 590 /* 591 * __VNODESCAN__ 592 * 593 * The VP will stick around while we hold mntvnode_token, 594 * at least until we block, so we can safely do an initial 595 * check, and then must check again after we lock the vnode. 596 */ 597 if (vp->v_type == VNON || /* syncer or indeterminant */ 598 !vmightfree(vp, trigger, info->pass) /* critical path opt */ 599 ) { 600 --count; 601 continue; 602 } 603 604 /* 605 * VX get the candidate vnode. If the VX get fails the 606 * vnode might still be on the mountlist. Our loop depends 607 * on us at least cycling the vnode to the end of the 608 * mountlist. 609 */ 610 if (vx_get_nonblock(vp) != 0) { 611 --count; 612 continue; 613 } 614 615 /* 616 * Since we blocked locking the vp, make sure it is still 617 * a candidate for reclamation. That is, it has not already 618 * been reclaimed and only has our VX reference associated 619 * with it. 620 */ 621 if (vp->v_type == VNON || /* syncer or indeterminant */ 622 (vp->v_flag & VRECLAIMED) || 623 vp->v_mount != mp || 624 !vtrytomakegoneable(vp, trigger) /* critical path opt */ 625 ) { 626 --count; 627 vx_put(vp); 628 continue; 629 } 630 631 /* 632 * All right, we are good, move the vp to the end of the 633 * mountlist and clean it out. The vget will have returned 634 * an error if the vnode was destroyed (VRECLAIMED set), so we 635 * do not have to check again. The vput() will move the 636 * vnode to the free list if the vgone() was successful. 637 */ 638 KKASSERT(vp->v_mount == mp); 639 vgone_vxlocked(vp); 640 vx_put(vp); 641 ++done; 642 --count; 643 } 644 lwkt_reltoken(&ilock); 645 return (done); 646 } 647 648 /* 649 * Attempt to recycle vnodes in a context that is always safe to block. 650 * Calling vlrurecycle() from the bowels of file system code has some 651 * interesting deadlock problems. 652 */ 653 static struct thread *vnlruthread; 654 static int vnlruproc_sig; 655 656 void 657 vnlru_proc_wait(void) 658 { 659 if (vnlruproc_sig == 0) { 660 vnlruproc_sig = 1; /* avoid unnecessary wakeups */ 661 wakeup(vnlruthread); 662 } 663 tsleep(&vnlruproc_sig, 0, "vlruwk", hz); 664 } 665 666 static void 667 vnlru_proc(void) 668 { 669 struct thread *td = curthread; 670 struct vnlru_info info; 671 int done; 672 673 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td, 674 SHUTDOWN_PRI_FIRST); 675 676 crit_enter(); 677 for (;;) { 678 kproc_suspend_loop(); 679 680 /* 681 * Try to free some vnodes if we have too many 682 */ 683 if (numvnodes > desiredvnodes && 684 freevnodes > desiredvnodes * 2 / 10) { 685 int count = numvnodes - desiredvnodes; 686 687 if (count > freevnodes / 100) 688 count = freevnodes / 100; 689 if (count < 5) 690 count = 5; 691 freesomevnodes(count); 692 } 693 694 /* 695 * Nothing to do if most of our vnodes are already on 696 * the free list. 697 */ 698 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) { 699 vnlruproc_sig = 0; 700 wakeup(&vnlruproc_sig); 701 tsleep(td, 0, "vlruwt", hz); 702 continue; 703 } 704 cache_cleanneg(0); 705 706 /* 707 * The pass iterates through the four combinations of 708 * VAGE0/VAGE1. We want to get rid of aged small files 709 * first. 710 */ 711 info.pass = 0; 712 done = 0; 713 while (done == 0 && info.pass < 4) { 714 done = mountlist_scan(vlrureclaim, &info, 715 MNTSCAN_FORWARD); 716 ++info.pass; 717 } 718 719 /* 720 * The vlrureclaim() call only processes 1/10 of the vnodes 721 * on each mount. If we couldn't find any repeat the loop 722 * at least enough times to cover all available vnodes before 723 * we start sleeping. Complain if the failure extends past 724 * 30 second, every 30 seconds. 725 */ 726 if (done == 0) { 727 ++vnlru_nowhere; 728 if (vnlru_nowhere % 10 == 0) 729 tsleep(td, 0, "vlrup", hz * 3); 730 if (vnlru_nowhere % 100 == 0) 731 kprintf("vnlru_proc: vnode recycler stopped working!\n"); 732 if (vnlru_nowhere == 1000) 733 vnlru_nowhere = 900; 734 } else { 735 vnlru_nowhere = 0; 736 } 737 } 738 crit_exit(); 739 } 740 741 /* 742 * MOUNTLIST FUNCTIONS 743 */ 744 745 /* 746 * mountlist_insert (MP SAFE) 747 * 748 * Add a new mount point to the mount list. 749 */ 750 void 751 mountlist_insert(struct mount *mp, int how) 752 { 753 lwkt_tokref ilock; 754 755 lwkt_gettoken(&ilock, &mountlist_token); 756 if (how == MNTINS_FIRST) 757 TAILQ_INSERT_HEAD(&mountlist, mp, mnt_list); 758 else 759 TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list); 760 lwkt_reltoken(&ilock); 761 } 762 763 /* 764 * mountlist_interlock (MP SAFE) 765 * 766 * Execute the specified interlock function with the mountlist token 767 * held. The function will be called in a serialized fashion verses 768 * other functions called through this mechanism. 769 */ 770 int 771 mountlist_interlock(int (*callback)(struct mount *), struct mount *mp) 772 { 773 lwkt_tokref ilock; 774 int error; 775 776 lwkt_gettoken(&ilock, &mountlist_token); 777 error = callback(mp); 778 lwkt_reltoken(&ilock); 779 return (error); 780 } 781 782 /* 783 * mountlist_boot_getfirst (DURING BOOT ONLY) 784 * 785 * This function returns the first mount on the mountlist, which is 786 * expected to be the root mount. Since no interlocks are obtained 787 * this function is only safe to use during booting. 788 */ 789 790 struct mount * 791 mountlist_boot_getfirst(void) 792 { 793 return(TAILQ_FIRST(&mountlist)); 794 } 795 796 /* 797 * mountlist_remove (MP SAFE) 798 * 799 * Remove a node from the mountlist. If this node is the next scan node 800 * for any active mountlist scans, the active mountlist scan will be 801 * adjusted to skip the node, thus allowing removals during mountlist 802 * scans. 803 */ 804 void 805 mountlist_remove(struct mount *mp) 806 { 807 struct mountscan_info *msi; 808 lwkt_tokref ilock; 809 810 lwkt_gettoken(&ilock, &mountlist_token); 811 TAILQ_FOREACH(msi, &mountscan_list, msi_entry) { 812 if (msi->msi_node == mp) { 813 if (msi->msi_how & MNTSCAN_FORWARD) 814 msi->msi_node = TAILQ_NEXT(mp, mnt_list); 815 else 816 msi->msi_node = TAILQ_PREV(mp, mntlist, mnt_list); 817 } 818 } 819 TAILQ_REMOVE(&mountlist, mp, mnt_list); 820 lwkt_reltoken(&ilock); 821 } 822 823 /* 824 * mountlist_scan (MP SAFE) 825 * 826 * Safely scan the mount points on the mount list. Unless otherwise 827 * specified each mount point will be busied prior to the callback and 828 * unbusied afterwords. The callback may safely remove any mount point 829 * without interfering with the scan. If the current callback 830 * mount is removed the scanner will not attempt to unbusy it. 831 * 832 * If a mount node cannot be busied it is silently skipped. 833 * 834 * The callback return value is aggregated and a total is returned. A return 835 * value of < 0 is not aggregated and will terminate the scan. 836 * 837 * MNTSCAN_FORWARD - the mountlist is scanned in the forward direction 838 * MNTSCAN_REVERSE - the mountlist is scanned in reverse 839 * MNTSCAN_NOBUSY - the scanner will make the callback without busying 840 * the mount node. 841 */ 842 int 843 mountlist_scan(int (*callback)(struct mount *, void *), void *data, int how) 844 { 845 struct mountscan_info info; 846 lwkt_tokref ilock; 847 struct mount *mp; 848 thread_t td; 849 int count; 850 int res; 851 852 lwkt_gettoken(&ilock, &mountlist_token); 853 854 info.msi_how = how; 855 info.msi_node = NULL; /* paranoia */ 856 TAILQ_INSERT_TAIL(&mountscan_list, &info, msi_entry); 857 858 res = 0; 859 td = curthread; 860 861 if (how & MNTSCAN_FORWARD) { 862 info.msi_node = TAILQ_FIRST(&mountlist); 863 while ((mp = info.msi_node) != NULL) { 864 if (how & MNTSCAN_NOBUSY) { 865 count = callback(mp, data); 866 } else if (vfs_busy(mp, LK_NOWAIT) == 0) { 867 count = callback(mp, data); 868 if (mp == info.msi_node) 869 vfs_unbusy(mp); 870 } else { 871 count = 0; 872 } 873 if (count < 0) 874 break; 875 res += count; 876 if (mp == info.msi_node) 877 info.msi_node = TAILQ_NEXT(mp, mnt_list); 878 } 879 } else if (how & MNTSCAN_REVERSE) { 880 info.msi_node = TAILQ_LAST(&mountlist, mntlist); 881 while ((mp = info.msi_node) != NULL) { 882 if (how & MNTSCAN_NOBUSY) { 883 count = callback(mp, data); 884 } else if (vfs_busy(mp, LK_NOWAIT) == 0) { 885 count = callback(mp, data); 886 if (mp == info.msi_node) 887 vfs_unbusy(mp); 888 } else { 889 count = 0; 890 } 891 if (count < 0) 892 break; 893 res += count; 894 if (mp == info.msi_node) 895 info.msi_node = TAILQ_PREV(mp, mntlist, mnt_list); 896 } 897 } 898 TAILQ_REMOVE(&mountscan_list, &info, msi_entry); 899 lwkt_reltoken(&ilock); 900 return(res); 901 } 902 903 /* 904 * MOUNT RELATED VNODE FUNCTIONS 905 */ 906 907 static struct kproc_desc vnlru_kp = { 908 "vnlru", 909 vnlru_proc, 910 &vnlruthread 911 }; 912 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp) 913 914 /* 915 * Move a vnode from one mount queue to another. 916 */ 917 void 918 insmntque(struct vnode *vp, struct mount *mp) 919 { 920 lwkt_tokref ilock; 921 922 lwkt_gettoken(&ilock, &mntvnode_token); 923 /* 924 * Delete from old mount point vnode list, if on one. 925 */ 926 if (vp->v_mount != NULL) { 927 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0, 928 ("bad mount point vnode list size")); 929 vremovevnodemnt(vp); 930 vp->v_mount->mnt_nvnodelistsize--; 931 } 932 /* 933 * Insert into list of vnodes for the new mount point, if available. 934 * The 'end' of the LRU list is the vnode prior to mp->mnt_syncer. 935 */ 936 if ((vp->v_mount = mp) == NULL) { 937 lwkt_reltoken(&ilock); 938 return; 939 } 940 if (mp->mnt_syncer) { 941 TAILQ_INSERT_BEFORE(mp->mnt_syncer, vp, v_nmntvnodes); 942 } else { 943 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); 944 } 945 mp->mnt_nvnodelistsize++; 946 lwkt_reltoken(&ilock); 947 } 948 949 950 /* 951 * Scan the vnodes under a mount point and issue appropriate callbacks. 952 * 953 * The fastfunc() callback is called with just the mountlist token held 954 * (no vnode lock). It may not block and the vnode may be undergoing 955 * modifications while the caller is processing it. The vnode will 956 * not be entirely destroyed, however, due to the fact that the mountlist 957 * token is held. A return value < 0 skips to the next vnode without calling 958 * the slowfunc(), a return value > 0 terminates the loop. 959 * 960 * The slowfunc() callback is called after the vnode has been successfully 961 * locked based on passed flags. The vnode is skipped if it gets rearranged 962 * or destroyed while blocking on the lock. A non-zero return value from 963 * the slow function terminates the loop. The slow function is allowed to 964 * arbitrarily block. The scanning code guarentees consistency of operation 965 * even if the slow function deletes or moves the node, or blocks and some 966 * other thread deletes or moves the node. 967 */ 968 int 969 vmntvnodescan( 970 struct mount *mp, 971 int flags, 972 int (*fastfunc)(struct mount *mp, struct vnode *vp, void *data), 973 int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data), 974 void *data 975 ) { 976 struct vmntvnodescan_info info; 977 lwkt_tokref ilock; 978 struct vnode *vp; 979 int r = 0; 980 int maxcount = 1000000; 981 int stopcount = 0; 982 int count = 0; 983 984 lwkt_gettoken(&ilock, &mntvnode_token); 985 986 /* 987 * If asked to do one pass stop after iterating available vnodes. 988 * Under heavy loads new vnodes can be added while we are scanning, 989 * so this isn't perfect. Create a slop factor of 2x. 990 */ 991 if (flags & VMSC_ONEPASS) 992 stopcount = mp->mnt_nvnodelistsize * 2; 993 994 info.vp = TAILQ_FIRST(&mp->mnt_nvnodelist); 995 TAILQ_INSERT_TAIL(&mntvnodescan_list, &info, entry); 996 while ((vp = info.vp) != NULL) { 997 if (--maxcount == 0) 998 panic("maxcount reached during vmntvnodescan"); 999 1000 /* 1001 * Skip if visible but not ready, or special (e.g. 1002 * mp->mnt_syncer) 1003 */ 1004 if (vp->v_type == VNON) 1005 goto next; 1006 KKASSERT(vp->v_mount == mp); 1007 1008 /* 1009 * Quick test. A negative return continues the loop without 1010 * calling the slow test. 0 continues onto the slow test. 1011 * A positive number aborts the loop. 1012 */ 1013 if (fastfunc) { 1014 if ((r = fastfunc(mp, vp, data)) < 0) { 1015 r = 0; 1016 goto next; 1017 } 1018 if (r) 1019 break; 1020 } 1021 1022 /* 1023 * Get a vxlock on the vnode, retry if it has moved or isn't 1024 * in the mountlist where we expect it. 1025 */ 1026 if (slowfunc) { 1027 int error; 1028 1029 switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) { 1030 case VMSC_GETVP: 1031 error = vget(vp, LK_EXCLUSIVE); 1032 break; 1033 case VMSC_GETVP|VMSC_NOWAIT: 1034 error = vget(vp, LK_EXCLUSIVE|LK_NOWAIT); 1035 break; 1036 case VMSC_GETVX: 1037 vx_get(vp); 1038 error = 0; 1039 break; 1040 default: 1041 error = 0; 1042 break; 1043 } 1044 if (error) 1045 goto next; 1046 /* 1047 * Do not call the slow function if the vnode is 1048 * invalid or if it was ripped out from under us 1049 * while we (potentially) blocked. 1050 */ 1051 if (info.vp == vp && vp->v_type != VNON) 1052 r = slowfunc(mp, vp, data); 1053 1054 /* 1055 * Cleanup 1056 */ 1057 switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) { 1058 case VMSC_GETVP: 1059 case VMSC_GETVP|VMSC_NOWAIT: 1060 vput(vp); 1061 break; 1062 case VMSC_GETVX: 1063 vx_put(vp); 1064 break; 1065 default: 1066 break; 1067 } 1068 if (r != 0) 1069 break; 1070 } 1071 1072 next: 1073 /* 1074 * Yield after some processing. Depending on the number 1075 * of vnodes, we might wind up running for a long time. 1076 * Because threads are not preemptable, time critical 1077 * userland processes might starve. Give them a chance 1078 * now and then. 1079 */ 1080 if (++count == 10000) { 1081 /* We really want to yield a bit, so we simply sleep a tick */ 1082 tsleep(mp, 0, "vnodescn", 1); 1083 count = 0; 1084 } 1085 1086 /* 1087 * If doing one pass this decrements to zero. If it starts 1088 * at zero it is effectively unlimited for the purposes of 1089 * this loop. 1090 */ 1091 if (--stopcount == 0) 1092 break; 1093 1094 /* 1095 * Iterate. If the vnode was ripped out from under us 1096 * info.vp will already point to the next vnode, otherwise 1097 * we have to obtain the next valid vnode ourselves. 1098 */ 1099 if (info.vp == vp) 1100 info.vp = TAILQ_NEXT(vp, v_nmntvnodes); 1101 } 1102 TAILQ_REMOVE(&mntvnodescan_list, &info, entry); 1103 lwkt_reltoken(&ilock); 1104 return(r); 1105 } 1106 1107 /* 1108 * Remove any vnodes in the vnode table belonging to mount point mp. 1109 * 1110 * If FORCECLOSE is not specified, there should not be any active ones, 1111 * return error if any are found (nb: this is a user error, not a 1112 * system error). If FORCECLOSE is specified, detach any active vnodes 1113 * that are found. 1114 * 1115 * If WRITECLOSE is set, only flush out regular file vnodes open for 1116 * writing. 1117 * 1118 * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped. 1119 * 1120 * `rootrefs' specifies the base reference count for the root vnode 1121 * of this filesystem. The root vnode is considered busy if its 1122 * v_sysref.refcnt exceeds this value. On a successful return, vflush() 1123 * will call vrele() on the root vnode exactly rootrefs times. 1124 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must 1125 * be zero. 1126 */ 1127 #ifdef DIAGNOSTIC 1128 static int busyprt = 0; /* print out busy vnodes */ 1129 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, ""); 1130 #endif 1131 1132 static int vflush_scan(struct mount *mp, struct vnode *vp, void *data); 1133 1134 struct vflush_info { 1135 int flags; 1136 int busy; 1137 thread_t td; 1138 }; 1139 1140 int 1141 vflush(struct mount *mp, int rootrefs, int flags) 1142 { 1143 struct thread *td = curthread; /* XXX */ 1144 struct vnode *rootvp = NULL; 1145 int error; 1146 struct vflush_info vflush_info; 1147 1148 if (rootrefs > 0) { 1149 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, 1150 ("vflush: bad args")); 1151 /* 1152 * Get the filesystem root vnode. We can vput() it 1153 * immediately, since with rootrefs > 0, it won't go away. 1154 */ 1155 if ((error = VFS_ROOT(mp, &rootvp)) != 0) { 1156 if ((flags & FORCECLOSE) == 0) 1157 return (error); 1158 rootrefs = 0; 1159 /* continue anyway */ 1160 } 1161 if (rootrefs) 1162 vput(rootvp); 1163 } 1164 1165 vflush_info.busy = 0; 1166 vflush_info.flags = flags; 1167 vflush_info.td = td; 1168 vmntvnodescan(mp, VMSC_GETVX, NULL, vflush_scan, &vflush_info); 1169 1170 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { 1171 /* 1172 * If just the root vnode is busy, and if its refcount 1173 * is equal to `rootrefs', then go ahead and kill it. 1174 */ 1175 KASSERT(vflush_info.busy > 0, ("vflush: not busy")); 1176 KASSERT(rootvp->v_sysref.refcnt >= rootrefs, ("vflush: rootrefs")); 1177 if (vflush_info.busy == 1 && rootvp->v_sysref.refcnt == rootrefs) { 1178 vx_lock(rootvp); 1179 vgone_vxlocked(rootvp); 1180 vx_unlock(rootvp); 1181 vflush_info.busy = 0; 1182 } 1183 } 1184 if (vflush_info.busy) 1185 return (EBUSY); 1186 for (; rootrefs > 0; rootrefs--) 1187 vrele(rootvp); 1188 return (0); 1189 } 1190 1191 /* 1192 * The scan callback is made with an VX locked vnode. 1193 */ 1194 static int 1195 vflush_scan(struct mount *mp, struct vnode *vp, void *data) 1196 { 1197 struct vflush_info *info = data; 1198 struct vattr vattr; 1199 1200 /* 1201 * Skip over a vnodes marked VSYSTEM. 1202 */ 1203 if ((info->flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) { 1204 return(0); 1205 } 1206 1207 /* 1208 * If WRITECLOSE is set, flush out unlinked but still open 1209 * files (even if open only for reading) and regular file 1210 * vnodes open for writing. 1211 */ 1212 if ((info->flags & WRITECLOSE) && 1213 (vp->v_type == VNON || 1214 (VOP_GETATTR(vp, &vattr) == 0 && 1215 vattr.va_nlink > 0)) && 1216 (vp->v_writecount == 0 || vp->v_type != VREG)) { 1217 return(0); 1218 } 1219 1220 /* 1221 * If we are the only holder (refcnt of 1) or the vnode is in 1222 * termination (refcnt < 0), we can vgone the vnode. 1223 */ 1224 if (vp->v_sysref.refcnt <= 1) { 1225 vgone_vxlocked(vp); 1226 return(0); 1227 } 1228 1229 /* 1230 * If FORCECLOSE is set, forcibly close the vnode. For block 1231 * or character devices we just clean and leave the vp 1232 * associated with devfs. For all other files, just kill them. 1233 * 1234 * XXX we need to do something about devfs here, I'd rather not 1235 * blow away device associations. 1236 */ 1237 if (info->flags & FORCECLOSE) { 1238 vgone_vxlocked(vp); 1239 #if 0 1240 if (vp->v_type != VBLK && vp->v_type != VCHR) { 1241 vgone_vxlocked(vp); 1242 } else { 1243 vclean_vxlocked(vp, 0); 1244 /*vp->v_ops = &devfs_vnode_dev_vops_p;*/ 1245 insmntque(vp, NULL); 1246 } 1247 #endif 1248 return(0); 1249 } 1250 #ifdef DIAGNOSTIC 1251 if (busyprt) 1252 vprint("vflush: busy vnode", vp); 1253 #endif 1254 ++info->busy; 1255 return(0); 1256 } 1257 1258 void 1259 add_bio_ops(struct bio_ops *ops) 1260 { 1261 TAILQ_INSERT_TAIL(&bio_ops_list, ops, entry); 1262 } 1263 1264 void 1265 rem_bio_ops(struct bio_ops *ops) 1266 { 1267 TAILQ_REMOVE(&bio_ops_list, ops, entry); 1268 } 1269 1270 /* 1271 * This calls the bio_ops io_sync function either for a mount point 1272 * or generally. 1273 * 1274 * WARNING: softdeps is weirdly coded and just isn't happy unless 1275 * io_sync is called with a NULL mount from the general syncing code. 1276 */ 1277 void 1278 bio_ops_sync(struct mount *mp) 1279 { 1280 struct bio_ops *ops; 1281 1282 if (mp) { 1283 if ((ops = mp->mnt_bioops) != NULL) 1284 ops->io_sync(mp); 1285 } else { 1286 TAILQ_FOREACH(ops, &bio_ops_list, entry) { 1287 ops->io_sync(NULL); 1288 } 1289 } 1290 } 1291 1292