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 .nom_cache = 256, 73 .flags = SRC_MANAGEDINIT, 74 .ctor = vnode_ctor, 75 .dtor = vnode_dtor, 76 .ops = { 77 .terminate = (sysref_terminate_func_t)vnode_terminate, 78 .lock = (sysref_terminate_func_t)vx_lock, 79 .unlock = (sysref_terminate_func_t)vx_unlock 80 } 81 }; 82 83 /* 84 * The vnode free list hold inactive vnodes. Aged inactive vnodes 85 * are inserted prior to the mid point, and otherwise inserted 86 * at the tail. 87 */ 88 static TAILQ_HEAD(freelst, vnode) vnode_free_list; 89 static struct vnode vnode_free_mid1; 90 static struct vnode vnode_free_mid2; 91 static struct vnode vnode_free_rover; 92 static struct spinlock vfs_spin = SPINLOCK_INITIALIZER(vfs_spin); 93 static enum { ROVER_MID1, ROVER_MID2 } rover_state = ROVER_MID2; 94 95 int freevnodes = 0; 96 SYSCTL_INT(_debug, OID_AUTO, freevnodes, CTLFLAG_RD, 97 &freevnodes, 0, "Number of free nodes"); 98 static int wantfreevnodes = 25; 99 SYSCTL_INT(_debug, OID_AUTO, wantfreevnodes, CTLFLAG_RW, 100 &wantfreevnodes, 0, "Desired number of free vnodes"); 101 #ifdef TRACKVNODE 102 static ulong trackvnode; 103 SYSCTL_ULONG(_debug, OID_AUTO, trackvnode, CTLFLAG_RW, 104 &trackvnode, 0, ""); 105 #endif 106 107 /* 108 * Called from vfsinit() 109 */ 110 void 111 vfs_lock_init(void) 112 { 113 TAILQ_INIT(&vnode_free_list); 114 TAILQ_INSERT_TAIL(&vnode_free_list, &vnode_free_mid1, v_freelist); 115 TAILQ_INSERT_TAIL(&vnode_free_list, &vnode_free_mid2, v_freelist); 116 TAILQ_INSERT_TAIL(&vnode_free_list, &vnode_free_rover, v_freelist); 117 spin_init(&vfs_spin); 118 kmalloc_raise_limit(M_VNODE, 0); /* unlimited */ 119 } 120 121 /* 122 * Misc functions 123 */ 124 static __inline 125 void 126 _vsetflags(struct vnode *vp, int flags) 127 { 128 atomic_set_int(&vp->v_flag, flags); 129 } 130 131 static __inline 132 void 133 _vclrflags(struct vnode *vp, int flags) 134 { 135 atomic_clear_int(&vp->v_flag, flags); 136 } 137 138 void 139 vsetflags(struct vnode *vp, int flags) 140 { 141 _vsetflags(vp, flags); 142 } 143 144 void 145 vclrflags(struct vnode *vp, int flags) 146 { 147 _vclrflags(vp, flags); 148 } 149 150 /* 151 * Inline helper functions. 152 * 153 * WARNING: vbusy() may only be called while the vnode lock or VX lock 154 * is held. The vnode spinlock need not be held. 155 * 156 * MPSAFE 157 */ 158 static __inline 159 void 160 __vbusy_interlocked(struct vnode *vp) 161 { 162 KKASSERT(vp->v_flag & VFREE); 163 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 164 freevnodes--; 165 _vclrflags(vp, VFREE); 166 } 167 168 static __inline 169 void 170 __vbusy(struct vnode *vp) 171 { 172 #ifdef TRACKVNODE 173 if ((ulong)vp == trackvnode) 174 kprintf("__vbusy %p %08x\n", vp, vp->v_flag); 175 #endif 176 spin_lock(&vfs_spin); 177 __vbusy_interlocked(vp); 178 spin_unlock(&vfs_spin); 179 } 180 181 /* 182 * Put a vnode on the free list. The caller has cleared VCACHED or owns the 183 * implied sysref related to having removed the vnode from the freelist 184 * (and VCACHED is already clear in that case). 185 * 186 * MPSAFE 187 */ 188 static __inline 189 void 190 __vfree(struct vnode *vp) 191 { 192 #ifdef TRACKVNODE 193 if ((ulong)vp == trackvnode) { 194 kprintf("__vfree %p %08x\n", vp, vp->v_flag); 195 print_backtrace(-1); 196 } 197 #endif 198 spin_lock(&vfs_spin); 199 KKASSERT((vp->v_flag & VFREE) == 0); 200 201 /* 202 * Distinguish between basically dead vnodes, vnodes with cached 203 * data, and vnodes without cached data. A rover will shift the 204 * vnodes around as their cache status is lost. 205 */ 206 if (vp->v_flag & VRECLAIMED) { 207 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 208 } else if (vp->v_object && vp->v_object->resident_page_count) { 209 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 210 } else if (vp->v_object && vp->v_object->swblock_count) { 211 TAILQ_INSERT_BEFORE(&vnode_free_mid2, vp, v_freelist); 212 } else { 213 TAILQ_INSERT_BEFORE(&vnode_free_mid1, vp, v_freelist); 214 } 215 freevnodes++; 216 _vsetflags(vp, VFREE); 217 spin_unlock(&vfs_spin); 218 } 219 220 /* 221 * Put a vnode on the free list. The caller has cleared VCACHED or owns the 222 * implied sysref related to having removed the vnode from the freelist 223 * (and VCACHED is already clear in that case). 224 * 225 * MPSAFE 226 */ 227 static __inline 228 void 229 __vfreetail(struct vnode *vp) 230 { 231 #ifdef TRACKVNODE 232 if ((ulong)vp == trackvnode) 233 kprintf("__vfreetail %p %08x\n", vp, vp->v_flag); 234 #endif 235 spin_lock(&vfs_spin); 236 KKASSERT((vp->v_flag & VFREE) == 0); 237 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 238 freevnodes++; 239 _vsetflags(vp, VFREE); 240 spin_unlock(&vfs_spin); 241 } 242 243 /* 244 * Return a C boolean if we should put the vnode on the freelist (VFREE), 245 * or leave it / mark it as VCACHED. 246 * 247 * This routine is only valid if the vnode is already either VFREE or 248 * VCACHED, or if it can become VFREE or VCACHED via vnode_terminate(). 249 * 250 * WARNING! This functions is typically called with v_spinlock held. 251 * 252 * MPSAFE 253 */ 254 static __inline boolean_t 255 vshouldfree(struct vnode *vp) 256 { 257 return (vp->v_auxrefs == 0 && 258 (vp->v_object == NULL || vp->v_object->resident_page_count == 0)); 259 } 260 261 /* 262 * Add a ref to an active vnode. This function should never be called 263 * with an inactive vnode (use vget() instead). 264 * 265 * MPSAFE 266 */ 267 void 268 vref(struct vnode *vp) 269 { 270 KKASSERT(vp->v_sysref.refcnt > 0 && 271 (vp->v_flag & (VFREE|VINACTIVE)) == 0); 272 sysref_get(&vp->v_sysref); 273 } 274 275 /* 276 * Release a ref on an active or inactive vnode. The sysref termination 277 * function will be called when the active last active reference is released, 278 * and the vnode is returned to the objcache when the last inactive 279 * reference is released. 280 */ 281 void 282 vrele(struct vnode *vp) 283 { 284 sysref_put(&vp->v_sysref); 285 } 286 287 /* 288 * Add an auxiliary data structure reference to the vnode. Auxiliary 289 * references do not change the state of the vnode or prevent them 290 * from being deactivated, reclaimed, or placed on or removed from 291 * the free list. 292 * 293 * An auxiliary reference DOES prevent the vnode from being destroyed, 294 * allowing you to vx_lock() it, test state, etc. 295 * 296 * An auxiliary reference DOES NOT move a vnode out of the VFREE state 297 * once it has entered it. 298 * 299 * WARNING! vhold() and vhold_interlocked() must not acquire v_spinlock. 300 * The spinlock may or may not already be held by the caller. 301 * vdrop() will clean up the free list state. 302 * 303 * MPSAFE 304 */ 305 void 306 vhold(struct vnode *vp) 307 { 308 KKASSERT(vp->v_sysref.refcnt != 0); 309 atomic_add_int(&vp->v_auxrefs, 1); 310 } 311 312 void 313 vhold_interlocked(struct vnode *vp) 314 { 315 atomic_add_int(&vp->v_auxrefs, 1); 316 } 317 318 /* 319 * Remove an auxiliary reference from the vnode. 320 * 321 * vdrop needs to check for a VCACHE->VFREE transition to catch cases 322 * where a vnode is held past its reclamation. We use v_spinlock to 323 * interlock VCACHED -> !VCACHED transitions. 324 * 325 * MPSAFE 326 */ 327 void 328 vdrop(struct vnode *vp) 329 { 330 KKASSERT(vp->v_sysref.refcnt != 0 && vp->v_auxrefs > 0); 331 spin_lock(&vp->v_spinlock); 332 atomic_subtract_int(&vp->v_auxrefs, 1); 333 if ((vp->v_flag & VCACHED) && vshouldfree(vp)) { 334 _vclrflags(vp, VCACHED); 335 __vfree(vp); 336 } 337 spin_unlock(&vp->v_spinlock); 338 } 339 340 /* 341 * This function is called when the last active reference on the vnode 342 * is released, typically via vrele(). SYSREF will VX lock the vnode 343 * and then give the vnode a negative ref count, indicating that it is 344 * undergoing termination or is being set aside for the cache, and one 345 * final sysref_put() is required to actually return it to the memory 346 * subsystem. 347 * 348 * Additional inactive sysrefs may race us but that's ok. Reactivations 349 * cannot race us because the sysref code interlocked with the VX lock 350 * (which is held on call). 351 * 352 * MPSAFE 353 */ 354 void 355 vnode_terminate(struct vnode *vp) 356 { 357 /* 358 * We own the VX lock, it should not be possible for someone else 359 * to have reactivated the vp. 360 */ 361 KKASSERT(sysref_isinactive(&vp->v_sysref)); 362 363 /* 364 * Deactivate the vnode by marking it VFREE or VCACHED. 365 * The vnode can be reactivated from either state until 366 * reclaimed. These states inherit the 'last' sysref on the 367 * vnode. 368 * 369 * NOTE: There may be additional inactive references from 370 * other entities blocking on the VX lock while we hold it, 371 * but this does not prevent us from changing the vnode's 372 * state. 373 * 374 * NOTE: The vnode could already be marked inactive. XXX 375 * how? 376 * 377 * NOTE: v_mount may be NULL due to assignment to 378 * dead_vnode_vops 379 * 380 * NOTE: The vnode may be marked inactive with dirty buffers 381 * or dirty pages in its cached VM object still present. 382 * 383 * NOTE: VCACHED should not be set on entry. We lose control 384 * of the sysref the instant the vnode is placed on the 385 * free list or when VCACHED is set. 386 * 387 * The VX lock is required when transitioning to 388 * +VCACHED but is not sufficient for the vshouldfree() 389 * interlocked test or when transitioning to -VCACHED. 390 */ 391 if ((vp->v_flag & VINACTIVE) == 0) { 392 _vsetflags(vp, VINACTIVE); 393 if (vp->v_mount) 394 VOP_INACTIVE(vp); 395 } 396 spin_lock(&vp->v_spinlock); 397 KKASSERT((vp->v_flag & (VFREE|VCACHED)) == 0); 398 if (vshouldfree(vp)) 399 __vfree(vp); 400 else 401 _vsetflags(vp, VCACHED); /* inactive but not yet free*/ 402 spin_unlock(&vp->v_spinlock); 403 vx_unlock(vp); 404 } 405 406 /* 407 * Physical vnode constructor / destructor. These are only executed on 408 * the backend of the objcache. They are NOT executed on every vnode 409 * allocation or deallocation. 410 * 411 * MPSAFE 412 */ 413 boolean_t 414 vnode_ctor(void *obj, void *private, int ocflags) 415 { 416 struct vnode *vp = obj; 417 418 lwkt_token_init(&vp->v_token, "vnode"); 419 lockinit(&vp->v_lock, "vnode", 0, 0); 420 ccms_dataspace_init(&vp->v_ccms); 421 TAILQ_INIT(&vp->v_namecache); 422 RB_INIT(&vp->v_rbclean_tree); 423 RB_INIT(&vp->v_rbdirty_tree); 424 RB_INIT(&vp->v_rbhash_tree); 425 return(TRUE); 426 } 427 428 /* 429 * MPSAFE 430 */ 431 void 432 vnode_dtor(void *obj, void *private) 433 { 434 struct vnode *vp = obj; 435 436 KKASSERT((vp->v_flag & (VCACHED|VFREE)) == 0); 437 ccms_dataspace_destroy(&vp->v_ccms); 438 } 439 440 /**************************************************************** 441 * VX LOCKING FUNCTIONS * 442 **************************************************************** 443 * 444 * These functions lock vnodes for reclamation and deactivation related 445 * activities. The caller must already be holding some sort of reference 446 * on the vnode. 447 * 448 * MPSAFE 449 */ 450 void 451 vx_lock(struct vnode *vp) 452 { 453 lockmgr(&vp->v_lock, LK_EXCLUSIVE); 454 } 455 456 /* 457 * The non-blocking version also uses a slightly different mechanic. 458 * This function will explicitly fail not only if it cannot acquire 459 * the lock normally, but also if the caller already holds a lock. 460 * 461 * The adjusted mechanic is used to close a loophole where complex 462 * VOP_RECLAIM code can circle around recursively and allocate the 463 * same vnode it is trying to destroy from the freelist. 464 * 465 * Any filesystem (aka UFS) which puts LK_CANRECURSE in lk_flags can 466 * cause the incorrect behavior to occur. If not for that lockmgr() 467 * would do the right thing. 468 */ 469 static int 470 vx_lock_nonblock(struct vnode *vp) 471 { 472 if (lockcountnb(&vp->v_lock)) 473 return(EBUSY); 474 return(lockmgr(&vp->v_lock, LK_EXCLUSIVE | LK_NOWAIT | LK_NOSPINWAIT)); 475 } 476 477 void 478 vx_unlock(struct vnode *vp) 479 { 480 lockmgr(&vp->v_lock, LK_RELEASE); 481 } 482 483 /**************************************************************** 484 * VNODE ACQUISITION FUNCTIONS * 485 **************************************************************** 486 * 487 * These functions must be used when accessing a vnode via an auxiliary 488 * reference such as the namecache or free list, or when you wish to 489 * do a combo ref+lock sequence. 490 * 491 * These functions are MANDATORY for any code chain accessing a vnode 492 * whos activation state is not known. 493 * 494 * vget() can be called with LK_NOWAIT and will return EBUSY if the 495 * lock cannot be immediately acquired. 496 * 497 * vget()/vput() are used when reactivation is desired. 498 * 499 * vx_get() and vx_put() are used when reactivation is not desired. 500 */ 501 int 502 vget(struct vnode *vp, int flags) 503 { 504 int error; 505 506 /* 507 * A lock type must be passed 508 */ 509 if ((flags & LK_TYPE_MASK) == 0) { 510 panic("vget() called with no lock specified!"); 511 /* NOT REACHED */ 512 } 513 514 /* 515 * Reference the structure and then acquire the lock. 0->1 516 * transitions and refs during termination are allowed here so 517 * call sysref directly. 518 * 519 * NOTE: The requested lock might be a shared lock and does 520 * not protect our access to the refcnt or other fields. 521 */ 522 sysref_get(&vp->v_sysref); 523 if ((error = vn_lock(vp, flags)) != 0) { 524 /* 525 * The lock failed, undo and return an error. 526 */ 527 sysref_put(&vp->v_sysref); 528 } else if (vp->v_flag & VRECLAIMED) { 529 /* 530 * The node is being reclaimed and cannot be reactivated 531 * any more, undo and return ENOENT. 532 */ 533 vn_unlock(vp); 534 vrele(vp); 535 error = ENOENT; 536 } else { 537 /* 538 * If the vnode is marked VFREE or VCACHED it needs to be 539 * reactivated, otherwise it had better already be active. 540 * VINACTIVE must also be cleared. 541 * 542 * In the VFREE/VCACHED case we have to throw away the 543 * sysref that was earmarking those cases and preventing 544 * the vnode from being destroyed. Our sysref is still held. 545 * 546 * We are allowed to reactivate the vnode while we hold 547 * the VX lock, assuming it can be reactivated. 548 */ 549 spin_lock(&vp->v_spinlock); 550 if (vp->v_flag & VFREE) { 551 __vbusy(vp); 552 sysref_activate(&vp->v_sysref); 553 spin_unlock(&vp->v_spinlock); 554 sysref_put(&vp->v_sysref); 555 } else if (vp->v_flag & VCACHED) { 556 _vclrflags(vp, VCACHED); 557 sysref_activate(&vp->v_sysref); 558 spin_unlock(&vp->v_spinlock); 559 sysref_put(&vp->v_sysref); 560 } else { 561 if (sysref_isinactive(&vp->v_sysref)) { 562 sysref_activate(&vp->v_sysref); 563 kprintf("Warning vp %p reactivation race\n", 564 vp); 565 } 566 spin_unlock(&vp->v_spinlock); 567 } 568 _vclrflags(vp, VINACTIVE); 569 error = 0; 570 } 571 return(error); 572 } 573 574 /* 575 * MPSAFE 576 */ 577 void 578 vput(struct vnode *vp) 579 { 580 vn_unlock(vp); 581 vrele(vp); 582 } 583 584 /* 585 * XXX The vx_*() locks should use auxrefs, not the main reference counter. 586 * 587 * MPSAFE 588 */ 589 void 590 vx_get(struct vnode *vp) 591 { 592 sysref_get(&vp->v_sysref); 593 lockmgr(&vp->v_lock, LK_EXCLUSIVE); 594 } 595 596 /* 597 * MPSAFE 598 */ 599 int 600 vx_get_nonblock(struct vnode *vp) 601 { 602 int error; 603 604 sysref_get(&vp->v_sysref); 605 error = lockmgr(&vp->v_lock, LK_EXCLUSIVE | LK_NOWAIT); 606 if (error) 607 sysref_put(&vp->v_sysref); 608 return(error); 609 } 610 611 /* 612 * Relase a VX lock that also held a ref on the vnode. 613 * 614 * vx_put needs to check for a VCACHED->VFREE transition to catch the 615 * case where e.g. vnlru issues a vgone*(). 616 * 617 * MPSAFE 618 */ 619 void 620 vx_put(struct vnode *vp) 621 { 622 spin_lock(&vp->v_spinlock); 623 if ((vp->v_flag & VCACHED) && vshouldfree(vp)) { 624 _vclrflags(vp, VCACHED); 625 __vfree(vp); 626 } 627 spin_unlock(&vp->v_spinlock); 628 lockmgr(&vp->v_lock, LK_RELEASE); 629 sysref_put(&vp->v_sysref); 630 } 631 632 /* 633 * The rover looks for vnodes past the midline with no cached data and 634 * moves them to before the midline. If we do not do this the midline 635 * can wind up in a degenerate state. 636 */ 637 static 638 void 639 vnode_rover_locked(void) 640 { 641 struct vnode *vp; 642 643 /* 644 * Get the vnode after the rover. The rover roves between mid1 and 645 * the end so the only special vnode it can encounter is mid2. 646 */ 647 vp = TAILQ_NEXT(&vnode_free_rover, v_freelist); 648 if (vp == &vnode_free_mid2) { 649 vp = TAILQ_NEXT(vp, v_freelist); 650 rover_state = ROVER_MID2; 651 } 652 KKASSERT(vp != &vnode_free_mid1); 653 654 /* 655 * Start over if we finished the scan. 656 */ 657 TAILQ_REMOVE(&vnode_free_list, &vnode_free_rover, v_freelist); 658 if (vp == NULL) { 659 TAILQ_INSERT_AFTER(&vnode_free_list, &vnode_free_mid1, 660 &vnode_free_rover, v_freelist); 661 rover_state = ROVER_MID1; 662 return; 663 } 664 TAILQ_INSERT_AFTER(&vnode_free_list, vp, &vnode_free_rover, v_freelist); 665 666 /* 667 * Shift vp if appropriate. 668 */ 669 if (vp->v_object && vp->v_object->resident_page_count) { 670 /* 671 * Promote vnode with resident pages to section 3. 672 * (This case shouldn't happen). 673 */ 674 if (rover_state == ROVER_MID1) { 675 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 676 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 677 } 678 } else if (vp->v_object && vp->v_object->swblock_count) { 679 /* 680 * Demote vnode with only swap pages to section 2 681 */ 682 if (rover_state == ROVER_MID2) { 683 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 684 TAILQ_INSERT_BEFORE(&vnode_free_mid2, vp, v_freelist); 685 } 686 } else { 687 /* 688 * Demote vnode with no cached data to section 1 689 */ 690 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 691 TAILQ_INSERT_BEFORE(&vnode_free_mid1, vp, v_freelist); 692 } 693 } 694 695 /* 696 * Try to reuse a vnode from the free list. 697 * 698 * NOTE: The returned vnode is not completely initialized. 699 * 700 * WARNING: The freevnodes count can race, NULL can be returned even if 701 * freevnodes != 0. 702 * 703 * MPSAFE 704 */ 705 static 706 struct vnode * 707 allocfreevnode(void) 708 { 709 struct vnode *vp; 710 int count; 711 712 for (count = 0; count < freevnodes; count++) { 713 /* 714 * Try to lock the first vnode on the free list. 715 * Cycle if we can't. 716 * 717 * We use a bad hack in vx_lock_nonblock() which avoids 718 * the lock order reversal between vfs_spin and v_spinlock. 719 * This is very fragile code and I don't want to use 720 * vhold here. 721 */ 722 spin_lock(&vfs_spin); 723 vnode_rover_locked(); 724 vnode_rover_locked(); 725 vp = TAILQ_FIRST(&vnode_free_list); 726 while (vp == &vnode_free_mid1 || vp == &vnode_free_mid2 || 727 vp == &vnode_free_rover) { 728 vp = TAILQ_NEXT(vp, v_freelist); 729 } 730 if (vp == NULL) 731 break; 732 if (vx_lock_nonblock(vp)) { 733 KKASSERT(vp->v_flag & VFREE); 734 TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); 735 TAILQ_INSERT_TAIL(&vnode_free_list, 736 vp, v_freelist); 737 spin_unlock(&vfs_spin); 738 continue; 739 } 740 741 /* 742 * We inherit the sysref associated the vnode on the free 743 * list. Because VCACHED is clear the vnode will not 744 * be placed back on the free list. We own the sysref 745 * free and clear and thus control the disposition of 746 * the vnode. 747 */ 748 __vbusy_interlocked(vp); 749 spin_unlock(&vfs_spin); 750 #ifdef TRACKVNODE 751 if ((ulong)vp == trackvnode) 752 kprintf("allocfreevnode %p %08x\n", vp, vp->v_flag); 753 #endif 754 /* 755 * Do not reclaim/reuse a vnode while auxillary refs exists. 756 * This includes namecache refs due to a related ncp being 757 * locked or having children. 758 * 759 * We will make this test several times as auxrefs can 760 * get incremented on us without any spinlocks being held 761 * until we have removed all namecache and inode references 762 * to the vnode. 763 * 764 * Because VCACHED is already in the correct state (cleared) 765 * we cannot race other vdrop()s occuring at the same time 766 * and can safely place vp on the free list. 767 * 768 * The free list association reinherits the sysref. 769 */ 770 if (vp->v_auxrefs) { 771 __vfreetail(vp); 772 vx_unlock(vp); 773 continue; 774 } 775 776 /* 777 * We inherit the reference that was previously associated 778 * with the vnode being on the free list. VCACHED had better 779 * not be set because the reference and VX lock prevents 780 * the sysref from transitioning to an active state. 781 */ 782 KKASSERT((vp->v_flag & (VINACTIVE|VCACHED)) == VINACTIVE); 783 KKASSERT(sysref_isinactive(&vp->v_sysref)); 784 785 /* 786 * Holding the VX lock on an inactive vnode prevents it 787 * from being reactivated or reused. New namecache 788 * associations can only be made using active vnodes. 789 * 790 * Another thread may be blocked on our vnode lock while 791 * holding a namecache lock. We can only reuse this vnode 792 * if we can clear all namecache associations without 793 * blocking. 794 * 795 * Because VCACHED is already in the correct state (cleared) 796 * we cannot race other vdrop()s occuring at the same time 797 * and can safely place vp on the free list. 798 */ 799 if ((vp->v_flag & VRECLAIMED) == 0) { 800 if (cache_inval_vp_nonblock(vp)) { 801 __vfreetail(vp); 802 vx_unlock(vp); 803 continue; 804 } 805 vgone_vxlocked(vp); 806 /* vnode is still VX locked */ 807 } 808 809 /* 810 * We can reuse the vnode if no primary or auxiliary 811 * references remain other then ours, else put it 812 * back on the free list and keep looking. 813 * 814 * Either the free list inherits the last reference 815 * or we fall through and sysref_activate() the last 816 * reference. 817 * 818 * Since the vnode is in a VRECLAIMED state, no new 819 * namecache associations could have been made. 820 */ 821 KKASSERT(TAILQ_EMPTY(&vp->v_namecache)); 822 if (vp->v_auxrefs || 823 !sysref_islastdeactivation(&vp->v_sysref)) { 824 __vfreetail(vp); 825 vx_unlock(vp); 826 continue; 827 } 828 829 /* 830 * Return a VX locked vnode suitable for reuse. The caller 831 * inherits the sysref. 832 */ 833 return(vp); 834 } 835 return(NULL); 836 } 837 838 /* 839 * Obtain a new vnode from the freelist, allocating more if necessary. 840 * The returned vnode is VX locked & vrefd. 841 * 842 * All new vnodes set the VAGE flags. An open() of the vnode will 843 * decrement the (2-bit) flags. Vnodes which are opened several times 844 * are thus retained in the cache over vnodes which are merely stat()d. 845 * 846 * MPSAFE 847 */ 848 struct vnode * 849 allocvnode(int lktimeout, int lkflags) 850 { 851 struct vnode *vp; 852 853 /* 854 * Try to reuse vnodes if we hit the max. This situation only 855 * occurs in certain large-memory (2G+) situations. We cannot 856 * attempt to directly reclaim vnodes due to nasty recursion 857 * problems. 858 */ 859 while (numvnodes - freevnodes > desiredvnodes) 860 vnlru_proc_wait(); 861 862 /* 863 * Try to build up as many vnodes as we can before reallocating 864 * from the free list. A vnode on the free list simply means 865 * that it is inactive with no resident pages. It may or may not 866 * have been reclaimed and could have valuable information associated 867 * with it that we shouldn't throw away unless we really need to. 868 * 869 * HAMMER NOTE: Re-establishing a vnode is a fairly expensive 870 * operation for HAMMER but this should benefit UFS as well. 871 */ 872 if (freevnodes >= wantfreevnodes && numvnodes >= desiredvnodes) 873 vp = allocfreevnode(); 874 else 875 vp = NULL; 876 if (vp == NULL) { 877 vp = sysref_alloc(&vnode_sysref_class); 878 KKASSERT((vp->v_flag & (VCACHED|VFREE)) == 0); 879 lockmgr(&vp->v_lock, LK_EXCLUSIVE); 880 numvnodes++; 881 } 882 883 /* 884 * We are using a managed sysref class, vnode fields are only 885 * zerod on initial allocation from the backing store, not 886 * on reallocation. Thus we have to clear these fields for both 887 * reallocation and reuse. 888 */ 889 #ifdef INVARIANTS 890 if (vp->v_data) 891 panic("cleaned vnode isn't"); 892 if (bio_track_active(&vp->v_track_read) || 893 bio_track_active(&vp->v_track_write)) { 894 panic("Clean vnode has pending I/O's"); 895 } 896 if (vp->v_flag & VONWORKLST) 897 panic("Clean vnode still pending on syncer worklist!"); 898 if (!RB_EMPTY(&vp->v_rbdirty_tree)) 899 panic("Clean vnode still has dirty buffers!"); 900 if (!RB_EMPTY(&vp->v_rbclean_tree)) 901 panic("Clean vnode still has clean buffers!"); 902 if (!RB_EMPTY(&vp->v_rbhash_tree)) 903 panic("Clean vnode still on hash tree!"); 904 KKASSERT(vp->v_mount == NULL); 905 #endif 906 vp->v_flag = VAGE0 | VAGE1; 907 vp->v_lastw = 0; 908 vp->v_lasta = 0; 909 vp->v_cstart = 0; 910 vp->v_clen = 0; 911 vp->v_socket = 0; 912 vp->v_opencount = 0; 913 vp->v_writecount = 0; /* XXX */ 914 915 /* 916 * lktimeout only applies when LK_TIMELOCK is used, and only 917 * the pageout daemon uses it. The timeout may not be zero 918 * or the pageout daemon can deadlock in low-VM situations. 919 */ 920 if (lktimeout == 0) 921 lktimeout = hz / 10; 922 lockreinit(&vp->v_lock, "vnode", lktimeout, lkflags); 923 KKASSERT(TAILQ_EMPTY(&vp->v_namecache)); 924 /* exclusive lock still held */ 925 926 /* 927 * Note: sysref needs to be activated to convert -0x40000000 to +1. 928 * The -0x40000000 comes from the last ref on reuse, and from 929 * sysref_init() on allocate. 930 */ 931 sysref_activate(&vp->v_sysref); 932 vp->v_filesize = NOOFFSET; 933 vp->v_type = VNON; 934 vp->v_tag = 0; 935 vp->v_ops = NULL; 936 vp->v_data = NULL; 937 KKASSERT(vp->v_mount == NULL); 938 939 return (vp); 940 } 941 942 /* 943 * MPSAFE 944 */ 945 int 946 freesomevnodes(int n) 947 { 948 struct vnode *vp; 949 int count = 0; 950 951 while (n) { 952 --n; 953 if ((vp = allocfreevnode()) == NULL) 954 break; 955 vx_put(vp); 956 --numvnodes; 957 } 958 return(count); 959 } 960