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 * $DragonFly: src/sys/kern/vfs_lock.c,v 1.30 2008/06/30 03:57:41 dillon Exp $ 35 */ 36 37 /* 38 * External virtual filesystem routines 39 */ 40 #include "opt_ddb.h" 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/kernel.h> 45 #include <sys/malloc.h> 46 #include <sys/mount.h> 47 #include <sys/proc.h> 48 #include <sys/vnode.h> 49 #include <sys/buf.h> 50 #include <sys/sysctl.h> 51 52 #include <machine/limits.h> 53 54 #include <vm/vm.h> 55 #include <vm/vm_object.h> 56 57 #include <sys/buf2.h> 58 #include <sys/thread2.h> 59 #include <sys/sysref2.h> 60 61 static void vnode_terminate(struct vnode *vp); 62 static boolean_t vnode_ctor(void *obj, void *private, int ocflags); 63 static void vnode_dtor(void *obj, void *private); 64 65 static MALLOC_DEFINE(M_VNODE, "vnodes", "vnode structures"); 66 static struct sysref_class vnode_sysref_class = { 67 .name = "vnode", 68 .mtype = M_VNODE, 69 .proto = SYSREF_PROTO_VNODE, 70 .offset = offsetof(struct vnode, v_sysref), 71 .objsize = sizeof(struct vnode), 72 .mag_capacity = 256, 73 .flags = SRC_MANAGEDINIT, 74 .ctor = vnode_ctor, 75 .dtor = vnode_dtor, 76 .ops = { 77 .terminate = (sysref_terminate_func_t)vnode_terminate 78 } 79 }; 80 81 static TAILQ_HEAD(freelst, vnode) vnode_free_list; /* vnode free list */ 82 83 int freevnodes = 0; 84 SYSCTL_INT(_debug, OID_AUTO, freevnodes, CTLFLAG_RD, 85 &freevnodes, 0, ""); 86 static int wantfreevnodes = 25; 87 SYSCTL_INT(_debug, OID_AUTO, wantfreevnodes, CTLFLAG_RW, 88 &wantfreevnodes, 0, ""); 89 90 /* 91 * Called from vfsinit() 92 */ 93 void 94 vfs_lock_init(void) 95 { 96 TAILQ_INIT(&vnode_free_list); 97 } 98 99 /* 100 * Inline helper functions. vbusy() and vfree() must be called while in a 101 * critical section. 102 * 103 * Warning: must be callable if the caller holds a read spinlock to something 104 * else, meaning we can't use read spinlocks here. 105 */ 106 static __inline 107 void 108 __vbusy(struct vnode *vp) 109 { 110 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 111 freevnodes--; 112 vp->v_flag &= ~(VFREE|VAGE); 113 } 114 115 static __inline 116 void 117 __vfree(struct vnode *vp) 118 { 119 if (vp->v_flag & (VAGE|VRECLAIMED)) 120 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 121 else 122 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 123 freevnodes++; 124 vp->v_flag &= ~VAGE; 125 vp->v_flag |= VFREE; 126 } 127 128 static __inline 129 void 130 __vfreetail(struct vnode *vp) 131 { 132 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 133 freevnodes++; 134 vp->v_flag |= VFREE; 135 } 136 137 /* 138 * Return a C boolean if we should put the vnode on the freelist (VFREE), 139 * or leave it / mark it as VCACHED. 140 * 141 * This routine is only valid if the vnode is already either VFREE or 142 * VCACHED, or if it can become VFREE or VCACHED via vnode_terminate(). 143 */ 144 static __inline boolean_t 145 vshouldfree(struct vnode *vp) 146 { 147 return (vp->v_auxrefs == 0 && 148 (vp->v_object == NULL || vp->v_object->resident_page_count == 0)); 149 } 150 151 /* 152 * Add a ref to an active vnode. This function should never be called 153 * with an inactive vnode (use vget() instead). 154 */ 155 void 156 vref(struct vnode *vp) 157 { 158 KKASSERT(vp->v_sysref.refcnt > 0 && 159 (vp->v_flag & (VFREE|VINACTIVE)) == 0); 160 sysref_get(&vp->v_sysref); 161 } 162 163 /* 164 * Release a ref on an active or inactive vnode. The sysref termination 165 * function will be called when the active last active reference is released, 166 * and the vnode is returned to the objcache when the last inactive 167 * reference is released. 168 */ 169 void 170 vrele(struct vnode *vp) 171 { 172 sysref_put(&vp->v_sysref); 173 } 174 175 /* 176 * Add an auxiliary data structure reference to the vnode. Auxiliary 177 * references do not change the state of the vnode or prevent them 178 * from being deactivated, reclaimed, or placed on the free list. 179 * 180 * An auxiliary reference DOES prevent the vnode from being destroyed, 181 * allowing you to vx_lock() it, test state, etc. 182 * 183 * An auxiliary reference DOES NOT move a vnode out of the VFREE state 184 * once it has entered it. 185 * 186 * MPSAFE 187 */ 188 void 189 vhold(struct vnode *vp) 190 { 191 KKASSERT(vp->v_sysref.refcnt != 0); 192 atomic_add_int(&vp->v_auxrefs, 1); 193 } 194 195 /* 196 * Remove an auxiliary reference from the vnode. 197 * 198 * vdrop needs to check for a VCACHE->VFREE transition to catch cases 199 * where a vnode is held past its reclamation. 200 */ 201 void 202 vdrop(struct vnode *vp) 203 { 204 KKASSERT(vp->v_sysref.refcnt != 0 && vp->v_auxrefs > 0); 205 atomic_subtract_int(&vp->v_auxrefs, 1); 206 if ((vp->v_flag & VCACHED) && vshouldfree(vp)) { 207 /*vp->v_flag |= VAGE;*/ 208 vp->v_flag &= ~VCACHED; 209 __vfree(vp); 210 } 211 } 212 213 /* 214 * This function is called when the last active reference on the vnode 215 * is released, typically via vrele(). SYSREF will give the vnode a 216 * negative ref count, indicating that it is undergoing termination or 217 * is being set aside for the cache, and one final sysref_put() is 218 * required to actually return it to the memory subsystem. 219 * 220 * However, because vnodes may have auxiliary structural references via 221 * v_auxrefs, we must interlock auxiliary references against termination 222 * via the VX lock mechanism. It is possible for a vnode to be reactivated 223 * while we were blocked on the lock. 224 */ 225 void 226 vnode_terminate(struct vnode *vp) 227 { 228 vx_lock(vp); 229 if (sysref_isinactive(&vp->v_sysref)) { 230 /* 231 * Deactivate the vnode by marking it VFREE or VCACHED. 232 * The vnode can be reactivated from either state until 233 * reclaimed. These states inherit the 'last' sysref on the 234 * vnode. 235 * 236 * NOTE: There may be additional inactive references from 237 * other entities blocking on the VX lock while we hold it, 238 * but this does not prevent us from changing the vnode's 239 * state. 240 * 241 * NOTE: The vnode could already be marked inactive. XXX 242 * how? 243 * 244 * NOTE: The vnode may be marked inactive with dirty buffers 245 * or dirty pages in its cached VM object still present. 246 */ 247 if ((vp->v_flag & VINACTIVE) == 0) { 248 vp->v_flag |= VINACTIVE; 249 VOP_INACTIVE(vp); 250 } 251 KKASSERT((vp->v_flag & (VFREE|VCACHED)) == 0); 252 if (vshouldfree(vp)) 253 __vfree(vp); 254 else 255 vp->v_flag |= VCACHED; /* inactive but not yet free */ 256 vx_unlock(vp); 257 } else { 258 /* 259 * Someone reactivated the vnode while were blocked on the 260 * VX lock. Release the VX lock and release the (now active) 261 * last reference which is no longer last. 262 */ 263 vx_unlock(vp); 264 vrele(vp); 265 } 266 } 267 268 /* 269 * Physical vnode constructor / destructor. These are only executed on 270 * the backend of the objcache. They are NOT executed on every vnode 271 * allocation or deallocation. 272 */ 273 boolean_t 274 vnode_ctor(void *obj, void *private, int ocflags) 275 { 276 struct vnode *vp = obj; 277 278 lwkt_token_init(&vp->v_token); 279 lockinit(&vp->v_lock, "vnode", 0, 0); 280 ccms_dataspace_init(&vp->v_ccms); 281 TAILQ_INIT(&vp->v_namecache); 282 RB_INIT(&vp->v_rbclean_tree); 283 RB_INIT(&vp->v_rbdirty_tree); 284 RB_INIT(&vp->v_rbhash_tree); 285 return(TRUE); 286 } 287 288 void 289 vnode_dtor(void *obj, void *private) 290 { 291 struct vnode *vp = obj; 292 293 ccms_dataspace_destroy(&vp->v_ccms); 294 } 295 296 /**************************************************************** 297 * VX LOCKING FUNCTIONS * 298 **************************************************************** 299 * 300 * These functions lock vnodes for reclamation and deactivation related 301 * activities. The caller must already be holding some sort of reference 302 * on the vnode. 303 */ 304 305 void 306 vx_lock(struct vnode *vp) 307 { 308 lockmgr(&vp->v_lock, LK_EXCLUSIVE); 309 } 310 311 static int 312 vx_lock_nonblock(struct vnode *vp) 313 { 314 return(lockmgr(&vp->v_lock, LK_EXCLUSIVE | LK_NOWAIT)); 315 } 316 317 void 318 vx_unlock(struct vnode *vp) 319 { 320 lockmgr(&vp->v_lock, LK_RELEASE); 321 } 322 323 /**************************************************************** 324 * VNODE ACQUISITION FUNCTIONS * 325 **************************************************************** 326 * 327 * These functions must be used when accessing a vnode via an auxiliary 328 * reference such as the namecache or free list, or when you wish to 329 * do a combo ref+lock sequence. 330 * 331 * These functions are MANDATORY for any code chain accessing a vnode 332 * whos activation state is not known. 333 * 334 * vget()/vput() are used when reactivation is desired. 335 * 336 * vx_get() and vx_put() are used when reactivation is not desired. 337 */ 338 int 339 vget(struct vnode *vp, int flags) 340 { 341 int error; 342 343 /* 344 * A lock type must be passed 345 */ 346 if ((flags & LK_TYPE_MASK) == 0) { 347 panic("vget() called with no lock specified!"); 348 /* NOT REACHED */ 349 } 350 351 /* 352 * Reference the structure and then acquire the lock. 0->1 353 * transitions and refs during termination are allowed here so 354 * call sysref directly. 355 */ 356 357 sysref_get(&vp->v_sysref); 358 if ((error = vn_lock(vp, flags)) != 0) { 359 /* 360 * The lock failed, undo and return an error. 361 */ 362 sysref_put(&vp->v_sysref); 363 } else if (vp->v_flag & VRECLAIMED) { 364 /* 365 * The node is being reclaimed and cannot be reactivated 366 * any more, undo and return ENOENT. 367 */ 368 vn_unlock(vp); 369 vrele(vp); 370 error = ENOENT; 371 } else { 372 /* 373 * If the vnode is marked VFREE or VCACHED it needs to be 374 * reactivated, otherwise it had better already be active. 375 * VINACTIVE must also be cleared. 376 * 377 * In the VFREE/VCACHED case we have to throw away the 378 * sysref that was earmarking those cases and preventing 379 * the vnode from being destroyed. Our sysref is still held. 380 */ 381 if (vp->v_flag & VFREE) { 382 __vbusy(vp); 383 sysref_put(&vp->v_sysref); 384 sysref_activate(&vp->v_sysref); 385 } else if (vp->v_flag & VCACHED) { 386 vp->v_flag &= ~VCACHED; 387 sysref_put(&vp->v_sysref); 388 sysref_activate(&vp->v_sysref); 389 } else { 390 KKASSERT(sysref_isactive(&vp->v_sysref)); 391 } 392 vp->v_flag &= ~VINACTIVE; 393 error = 0; 394 } 395 return(error); 396 } 397 398 void 399 vput(struct vnode *vp) 400 { 401 vn_unlock(vp); 402 vrele(vp); 403 } 404 405 /* 406 * XXX The vx_*() locks should use auxrefs, not the main reference counter. 407 */ 408 void 409 vx_get(struct vnode *vp) 410 { 411 sysref_get(&vp->v_sysref); 412 lockmgr(&vp->v_lock, LK_EXCLUSIVE); 413 } 414 415 int 416 vx_get_nonblock(struct vnode *vp) 417 { 418 int error; 419 420 sysref_get(&vp->v_sysref); 421 error = lockmgr(&vp->v_lock, LK_EXCLUSIVE | LK_NOWAIT); 422 if (error) 423 sysref_put(&vp->v_sysref); 424 return(error); 425 } 426 427 /* 428 * Relase a VX lock that also held a ref on the vnode. 429 * 430 * vx_put needs to check for a VCACHE->VFREE transition to catch the 431 * case where e.g. vnlru issues a vgone*(). 432 */ 433 void 434 vx_put(struct vnode *vp) 435 { 436 if ((vp->v_flag & VCACHED) && vshouldfree(vp)) { 437 /*vp->v_flag |= VAGE;*/ 438 vp->v_flag &= ~VCACHED; 439 __vfree(vp); 440 } 441 lockmgr(&vp->v_lock, LK_RELEASE); 442 sysref_put(&vp->v_sysref); 443 } 444 445 /* 446 * Misc functions 447 */ 448 449 void 450 vsetflags(struct vnode *vp, int flags) 451 { 452 crit_enter(); 453 vp->v_flag |= flags; 454 crit_exit(); 455 } 456 457 void 458 vclrflags(struct vnode *vp, int flags) 459 { 460 crit_enter(); 461 vp->v_flag &= ~flags; 462 crit_exit(); 463 } 464 465 /* 466 * Try to reuse a vnode from the free list. NOTE: The returned vnode 467 * is not completely initialized. 468 */ 469 static 470 struct vnode * 471 allocfreevnode(void) 472 { 473 struct vnode *vp; 474 int count; 475 476 for (count = 0; count < freevnodes; count++) { 477 /* 478 * Note that regardless of how we block in this loop, 479 * we only get here if freevnodes != 0 so there 480 * had better be something on the list. 481 * 482 * Try to lock the first vnode on the free list. 483 * Cycle if we can't. 484 * 485 * XXX NOT MP SAFE 486 */ 487 vp = TAILQ_FIRST(&vnode_free_list); 488 if (vx_lock_nonblock(vp)) { 489 KKASSERT(vp->v_flag & VFREE); 490 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 491 TAILQ_INSERT_TAIL(&vnode_free_list, 492 vp, v_freelist); 493 continue; 494 } 495 496 /* 497 * With the vnode locked we can safely remove it 498 * from the free list. We inherit the reference 499 * that was previously associated with the vnode 500 * being on the free list. 501 */ 502 KKASSERT((vp->v_flag & (VFREE|VINACTIVE)) == 503 (VFREE|VINACTIVE)); 504 KKASSERT(sysref_isinactive(&vp->v_sysref)); 505 __vbusy(vp); 506 507 /* 508 * Holding the VX lock on an inactive vnode prevents it 509 * from being reactivated or reused. New namecache 510 * associations can only be made using active vnodes. 511 * 512 * Another thread may be blocked on our vnode lock while 513 * holding a namecache lock. We can only reuse this vnode 514 * if we can clear all namecache associations without 515 * blocking. 516 */ 517 if ((vp->v_flag & VRECLAIMED) == 0) { 518 if (cache_inval_vp_nonblock(vp)) { 519 __vfreetail(vp); 520 vx_unlock(vp); 521 continue; 522 } 523 vgone_vxlocked(vp); 524 /* vnode is still VX locked */ 525 } 526 527 /* 528 * We can reuse the vnode if no primary or auxiliary 529 * references remain other then ours, else put it 530 * back on the free list and keep looking. 531 * 532 * Either the free list inherits the last reference 533 * or we fall through and sysref_activate() the last 534 * reference. 535 * 536 * Since the vnode is in a VRECLAIMED state, no new 537 * namecache associations could have been made. 538 */ 539 KKASSERT(TAILQ_EMPTY(&vp->v_namecache)); 540 if (vp->v_auxrefs || 541 !sysref_islastdeactivation(&vp->v_sysref)) { 542 __vfreetail(vp); 543 vx_unlock(vp); 544 continue; 545 } 546 547 /* 548 * Return a VX locked vnode suitable for reuse. The caller 549 * inherits the sysref. 550 */ 551 return(vp); 552 } 553 return(NULL); 554 } 555 556 /* 557 * Obtain a new vnode from the freelist, allocating more if necessary. 558 * The returned vnode is VX locked & refd. 559 */ 560 struct vnode * 561 allocvnode(int lktimeout, int lkflags) 562 { 563 struct vnode *vp; 564 565 /* 566 * Try to reuse vnodes if we hit the max. This situation only 567 * occurs in certain large-memory (2G+) situations. We cannot 568 * attempt to directly reclaim vnodes due to nasty recursion 569 * problems. 570 */ 571 while (numvnodes - freevnodes > desiredvnodes) 572 vnlru_proc_wait(); 573 574 /* 575 * Try to build up as many vnodes as we can before reallocating 576 * from the free list. A vnode on the free list simply means 577 * that it is inactive with no resident pages. It may or may not 578 * have been reclaimed and could have valuable information associated 579 * with it that we shouldn't throw away unless we really need to. 580 * 581 * HAMMER NOTE: Re-establishing a vnode is a fairly expensive 582 * operation for HAMMER but this should benefit UFS as well. 583 */ 584 if (freevnodes >= wantfreevnodes && numvnodes >= desiredvnodes) 585 vp = allocfreevnode(); 586 else 587 vp = NULL; 588 if (vp == NULL) { 589 vp = sysref_alloc(&vnode_sysref_class); 590 lockmgr(&vp->v_lock, LK_EXCLUSIVE); 591 numvnodes++; 592 } 593 594 /* 595 * We are using a managed sysref class, vnode fields are only 596 * zerod on initial allocation from the backing store, not 597 * on reallocation. Thus we have to clear these fields for both 598 * reallocation and reuse. 599 */ 600 #ifdef INVARIANTS 601 if (vp->v_data) 602 panic("cleaned vnode isn't"); 603 if (bio_track_active(&vp->v_track_read) || 604 bio_track_active(&vp->v_track_write)) { 605 panic("Clean vnode has pending I/O's"); 606 } 607 if (vp->v_flag & VONWORKLST) 608 panic("Clean vnode still pending on syncer worklist!"); 609 if (!RB_EMPTY(&vp->v_rbdirty_tree)) 610 panic("Clean vnode still has dirty buffers!"); 611 if (!RB_EMPTY(&vp->v_rbclean_tree)) 612 panic("Clean vnode still has clean buffers!"); 613 if (!RB_EMPTY(&vp->v_rbhash_tree)) 614 panic("Clean vnode still on hash tree!"); 615 KKASSERT(vp->v_mount == NULL); 616 #endif 617 vp->v_flag = 0; 618 vp->v_lastw = 0; 619 vp->v_lasta = 0; 620 vp->v_cstart = 0; 621 vp->v_clen = 0; 622 vp->v_socket = 0; 623 vp->v_opencount = 0; 624 vp->v_writecount = 0; /* XXX */ 625 626 /* 627 * lktimeout only applies when LK_TIMELOCK is used, and only 628 * the pageout daemon uses it. The timeout may not be zero 629 * or the pageout daemon can deadlock in low-VM situations. 630 */ 631 if (lktimeout == 0) 632 lktimeout = hz / 10; 633 lockreinit(&vp->v_lock, "vnode", lktimeout, lkflags); 634 KKASSERT(TAILQ_FIRST(&vp->v_namecache) == NULL); 635 /* exclusive lock still held */ 636 637 /* 638 * Note: sysref needs to be activated to convert -0x40000000 to +1. 639 * The -0x40000000 comes from the last ref on reuse, and from 640 * sysref_init() on allocate. 641 */ 642 sysref_activate(&vp->v_sysref); 643 vp->v_filesize = NOOFFSET; 644 vp->v_type = VNON; 645 vp->v_tag = 0; 646 vp->v_ops = NULL; 647 vp->v_data = NULL; 648 KKASSERT(vp->v_mount == NULL); 649 650 return (vp); 651 } 652 653 int 654 freesomevnodes(int n) 655 { 656 struct vnode *vp; 657 int count = 0; 658 659 while (n) { 660 --n; 661 if ((vp = allocfreevnode()) == NULL) 662 break; 663 vx_put(vp); 664 --numvnodes; 665 } 666 return(count); 667 } 668 669